arxiv_id
stringlengths 11
13
| markdown
stringlengths 2.09k
423k
| paper_doi
stringlengths 13
47
⌀ | paper_authors
listlengths 1
1.37k
| paper_published_date
stringdate 2014-06-03 15:22:49
2024-08-02 17:59:51
| paper_updated_date
stringdate 2014-06-03 15:22:49
2025-04-29 20:34:17
| categories
listlengths 1
7
| title
stringlengths 15
236
| summary
stringlengths 57
2.54k
|
|---|---|---|---|---|---|---|---|---|
1408.0985v1
|
## Speech earthquakes: scaling and universality in human voice
Jordi Luque , 1 Bartolo Luque , 2 and Lucas Lacasa 3 , ∗ 1 Telefonica Research, Edificio Telefonica-Diagonal 00, Barcelona, Spain 2 Departamento de Matem´ atica Aplicada y Estad´ ıstica, EIAE, Universidad Polit´cnica e de Madrid, Spain 3 School of Mathematical Sciences, Queen Mary University of London, ∗
Mile End Road E14NS, London, United Kingdom
(Dated: August 6, 2014)
Speech is a distinctive complex feature of human capabilities. In order to understand the physics underlying speech production, in this work we empirically analyse the statistics of large human speech datasets ranging several languages. We first show that during speech the energy is unevenly released and power-law distributed, reporting a universal robust Gutenberg-Richter-like law in speech. We further show that such earthquakes in speech show temporal correlations, as the interevent statistics are again power-law distributed. Since this feature takes place in the intraphoneme range, we conjecture that the responsible for this complex phenomenon is not cognitive, but it resides on the physiological speech production mechanism. Moreover, we show that these waiting time distributions are scale invariant under a renormalisation group transformation, suggesting that the process of speech generation is indeed operating close to a critical point. These results are put in contrast with current paradigms in speech processing, which point towards low dimensional deterministic chaos as the origin of nonlinear traits in speech fluctuations. As these latter fluctuations are indeed the aspects that humanize synthetic speech, these findings may have an impact in future speech synthesis technologies. Results are robust and independent of the communication language or the number of speakers, pointing towards an universal pattern and yet another hint of complexity in human speech.
PACS numbers: 89.75.Da, 89.75.Fb, 05.70.Jk
The description, understanding and modelling of speech is an interdisciplinary topic of current interest for physics [1], social and cognitive sciences [2, 3], data mining as well as engineering [4-7]. Classical speech synthesis technologies and algorithms were firstly based on linear stochastic models and linear prediction [6-8], and their underlying theory, so called source-filter theory of speech production, was initially relying on several key assumptions including uncoupled vocal tract and speech source, laminar airflow propagating linearly, periodic fold vibration, and homogeneous tract conditions [6, 8, 9]. Despite the successes of this benchmark theory, the synthetic speech generated by linear models fails to be 'natural'. As a matter of fact, current speech synthesizers usually require to incorporate pieces of real speech, e.g. concatenation of smaller speech units (phoneme and diphone-based synthesis [6, 7]) to improve their synthetic output.
With the popularisation of nonlinear dynamics, fractals, and chaos theory, the modelling paradigm slightly shifted and nonlinear speech processing emerged [10]. Accordingly, a number of authors pointed towards low dimensional, chaotic phenomena as the underlying mechanism governing the fluctuations in speech. This paradigm shift has fostered a number of inspiring theoretical vocal-fold and tract models [11] of increasing complexity, displaying a range of nonlinear phenom-
∗ Electronic address: [email protected]
ena such as bifurcations and chaos [12] or irregular oscillations [13] to cite some. Glottal airflow induced by the fluid-tissue interaction has also been modelled using Navier-Stokes equations [14]. Nevertheless, the empirical justification for low dimensional chaos is, up to a certain extent, based on preliminary evidence [15, 16], and this modelling approach only theoretically justified through an analogy between turbulent states, chaos, and fractals [10]. As a matter of fact, it has been recently acknowledged in the nonlinear dynamics community that a word of caution should be taken when empirically analysing the low dimensional nature of short, noisy and nonstationary empirical signals, as not only the computation of dynamical invariants such as fractal dimensions or Lyapunov exponents is difficult in those cases [17], but furthermore correlated stochastic noise can be misleadingly described as to have a low dimensional attractor [18]. It is therefore reasonably unclear what modelling paradigm should we follow to pinpoint the nonlinear nature of the fine grained details of speech.
In this work we make no a priori assumptions about the underlying adequate dynamical model, and we follow a data driven approach. We thoroughly analyse the statistics of speech waveforms in extensive real datasets [19] that extend all the way into the intraphoneme range ( t < 10 -2 s) [20], enabling us to dissect the purely physiological aspects that are playing a role in the production of speech, from other aspects such as cognitive effects [3, 21]. Our main results are the following: (i) energy releases in speech are power law distributed
t
FIG. 1: (Top Panel) Sample speech waveform A t ( ) of a singlespeaker recording (Spanish language). (Bottom panel) Instantaneous energy per unit time ε t ( ) = | A t ( ) | 2 in an excerpt from the top panel. The energy threshold Θ, defined as the instantaneous energy level for which a fixed percentage of the entire data remains above that level, helps us to unambiguously distinguish, for a given threshold, speech events (events for which ε t ( ) > Θ) from silence or speech interevents of duration τ . The energy released E in a speech event is computed from the integration of the instantaneous energy over the duration of that event E ∼ ∑ event ε t ( ) (dark area in the figure denotes the energy released in a given speech event).
with a language-independent universal exponent, and accordingly a Gutenberg-Richter-like law [31] is proposed within speech. (ii) In the intraphoneme range ( t < 10 -2 s), the interevent times (silences of duration τ ) between energy releases are also power-law distributed, suggesting long-range correlations in the time fluctuations of the amplitude signal. (iii) Furthermore, these distributions are invariant under a time Renormalisation Group (RG) transformation, [23-26]). On the basis of these results, we should conclude that the physiological mechanism of speech production self-organises close to a critical point.
## RESULTS
A TV broadcast speech database named KALAKA-2 [19] is employed for analysing language-dependent speech. It was originally designed for language recognition evaluation purposes and consists of wide-band TV broadcast speech recordings (roughly 4 hours per language) featuring 6 different languages: Basque, Catalan, Galician, Spanish, Portuguese and English. TV broadcast shows were recorded and sampled using 2 bytes at 16000 samples/second rate, taking care of including as much diversity as possible regarding speakers and speech modalities. It includes both planned and spontaneous speech throughout diverse environment conditions, such as studio or outside journalist reports but excluding telephonic channel. Therefore audio excerpts may contain voices from several speakers but only a single language. For illustrative purposes in figure 1 we depict a sample speech waveform amplitude A t ( ) and its squared, semi-definite positive instantaneous energy ε t ( ) = | A t ( ) | 2 , respectively. Without loss of generality, dropping the irrelevant constants, ε t ( ) has units of energy per time. Then, a threshold Θ is defined as the instantaneous energy level for which a fixed percentage of data is larger than the threshold. For instance, Θ = 70% is the threshold for which 30% of the data fall under this energy level (this allows to compare data across different empirical signals [22]). Θ not only works as a threshold of 'zero energy' that filters out background (environmental) noise, but help us to unambiguously distinguish a speech event, defined as a sequence of consecutive measurements ε t ( ) > Θ, from a silence event, whose duration is τ (see figure 1). Accordingly, speech can now be seen as a dynamical process of energy releases or 'speech earthquakes' separated by silence events, with in principle different statistical properties given different thresholds Θ. In what follows we address all these properties.
## Energy release: a Gutenberg-Richter-like scaling law in speech
The energy of a speech event is computed from the integration of the instantaneous energy over the duration of that event E = ∫ event ε t dt ( ) ≈ ∑ event ε t ( )∆ t where ∆ = 1 16000 t / s is the inverse of the sampling frequency (and therefore E has arbitrary units of energy). In order to get rid of environmental noise, we set a fixed threshold Θ = 80% and for each language, we compute its histogram P E ( ). In figure 2 we draw, in loglog scales, this histogram for all languages considered (note that a logarithmic binning was used to smooth out the data). We find out a robust power-law scaling P E ( ) ∼ E -α over five decades saturated by the standard finite-size cutoff, where the fitted exponents are all consistent with a language-independent universal behaviour: α = 1 13 . ± 0 06 for Spanish, . α = 1 16 . ± 0 05 for Basque, . α = 1 16 . ± 0 07 for Portuguese, . α = 1 13 . ± 0 05 for Gali-. cian, α = 1 10 . ± 0 05 for Catalan, and . α = 1 15 . ± 0 02 . for English, all having a correlation coefficient r 2 > 0 999 . (for completeness, in the inset of figure 2 we also depict the binned histogram of instantaneous energy P ε ( ) for all languages).
As long as the magnitude in seismicity [31] is related to the logarithm of the energy release, under this definition P E ( ) can be identified as a new 'Gutenberg-Richter-like law' in speech. This may be seen as related to other scaling laws in cognitive sciences [21], although at this stage it is still unclear what particular contributions come from both the mechanical (vocal folds and resonating cavity) and the cognitive systems.
FIG. 2: Log-log plot of the integrated energy distribution P E ( ) released in the statistics of all languages considered in this study, after a logarithmic binning. The figure shows a power law P E ( ) ∼ E -α , that holds for five decades, truncated for large (rare) events by a finite-size cut-off. (Inset panel) Log-log plot of the associated instantaneous energy histogram P ε ( ) of each language, which we include for completeness. We note that this non scaling shape of instantaneous energy is similar to the one found for precipitation rates [29].
## Scaling and universality of waiting time distributions.
In a second part, we study on the temporal orchestration of fluctuations, that is, the arrangement of silences or speech interevents of duration τ . We will pay a special attention to the intraphoneme range (timescales t < 3 10 · -2 s [6, 20]), where we assume no cognitive effects are present, in order to focus on the physiological aspects of speech. At this point we introduce a renormalisation group (RG) transformation to explore the origin of temporal correlations. This technique originates in the statistical physics community and has been previously used in the context of earthquakes [22-24] and tropical-cyclone statistics [25]. The first part of the transformation consists of a decimation: we raise the threshold Θ. This in general leads to different interevent distributions P Θ ( τ ). The second part of the transformation consists in a scale transformation in time, such that renormalized systems become comparable: τ → τ/ τ 〈 〉 Θ , P τ ( ) → P Θ ( τ ) 〈 τ 〉 Θ , where 〈 τ 〉 Θ is the mean interevent time of the system for a particular Θ. Invariant distributions under this RG transformation collapse into a threshold-independent universal curve: an adimensional waiting time distribution. While the complete fixed point structure of this RG is not well understood yet, recent advances [24, 26] rigorously found that stable (attractive) fixed points include the exponential distribution and a somewhat exotic double power-law distribution, which are attractors for both memoryless and short-range correlated stochastic point processes under the RG flow. Invariant distribu- tions other than the previous fixed points are likely to be unstable solutions of the RG flow, therefore encompassing criticality in the underlying dynamics [26].
In the left panel of figure 3 we plot in log-log, for different thresholds, the interevent histogram P Θ ( τ ) associated to the English language. In the intraphoneme range, interevents are power-law distributed in every case P Θ ( τ ) ∼ τ -β . In the right panel of the same figure we plot the rescaled histograms P τ ( ) (Θ = 60% 65% 70% , , for every language, yielding a total of 18 curves), collapsing under a single curve. Note that the collapse is quite good for those timescales that belong to the intraphoneme range, where only physiological mechanisms are in place, and such collapse is lost for larger timescales. This suggests that for every language, the statistics are invariant under this RG transformation, and the pattern is robust across languages. A more careful statistical analysis is required here as the range of the power law, restricted to the intraphoneme regime, is smaller. Accordingly, exponent β is estimated now following Clauset et al.'s method [27] which employs Maximum Likelihood Estimation (MLE) of a power law model, where goodness-of-fit test and confidence interval are based on Kolmogorov-Smirnov (KS) tests (comparing the actual distribution with 100 synthetic power law models whose exponent is the one found in the MLE to obtain p-values). This method yields a fairly universal exponent β = 2 06 . within the 95% confidence interval [1 84 . , 2 20], . KS pvalue of 0 99 and the statistical support for the power-law . hypothesis given by a p-value of 0 95. .
## DISCUSSION
It is well known that the dynamics of speech generation are complex, nonlinear, and certainly poorly explained by the benchmark source-filter theory. The fact that a Gutenberg-Richter law for the energy release probability distribution during speech emerges opens the possibility of understanding speech production in terms of crackling noise, a highly nonlinear behaviour that was first described in condensed matter physics [28] and latter found in a variety of natural hazards systems including earthquakes [22], rain [29] or solar flares [30] to cite some [26]. The underlying theory might then describe energy releases as the resonating response function of the system under airflow perturbation (the so called susceptibility in the statistical physics jargon), and the fact that it is a power law distributed quantity is the first evidence of criticality in these systems.
The fact that the waiting time distributions are different from the well known stable exponential laws while being invariant under the RG transformation confirms that the system is indeed operating close to a critical point. Note that the currently established low dimensional chaotic hypothesis suffers from being a process with short-range correlations and therefore chaotic speech interevents should typically renormalise into the exponential law.
τ <τ>
FIG. 3: (a) Log-log plot of the speech interevent distribution P Θ ( τ ) computed for the English database, for different energy thresholds Θ (there is a vertical translation performed for illustrative purposes, and data has been logarithmically binned). In the intraphoneme range (timescales t < 3 · 10 -2 s ), we find power law distributed interevent times fairly independently from the threshold, what suggests the onset of long-range correlations in this regime. (b) Rescaled interevent distributions P τ ( ) according to the decimation method (see the text), for all languages considered in this study and different thresholds. In the intraphoneme range, interevents collapse in a universal function and are invariant under the RG transformation, suggesting that the underlying physiological mechanism of speech synthesis is critical.
The most plausible conclusion of this work is that the physiological process of speech production evidences long-range correlations and criticality. This argument can be put in the same grounds as equilibrium critical phenomena, where only long-range correlations allow the system to escape from the basin of attraction of the trivial RG (high or low temperature) fixed points. As long as the critical solution is an unstable fixed point of the RG flow, the fact that the underlying dynamics seems to be poised near this point is somewhat remarkable. If we assume that the dynamics of speech generation at the level of the vocal folds are influenced by the properties of the glottal airflow, then our results would suggest that a similar critical behaviour might take place in other physiological processes involving such airflow, something that has been found empirically in the mechanism of lung inflation [32].
tems that self-organise towards a critical state- may be a more adequate modelling scenario than standard low dimensional chaos. If this was to be the case, threshold dynamics [31] would appear as the essential ingredient that encompasses the nonlinear properties in speech. These new approaches could further contribute to the development of both (i) microscopic (generative) models of speech production, and (ii) alternative methods of speech synthesis that would profit from the self-similar properties of speech at the intraphoneme range to make refined speech interpolation without needs to incorporate in the synthesis model pieces of real speech.
Note also that whereas a simple stochastic processes such as a Brownian motion cannot explain these results, under the more general paradigm of fractional Brownian motion with first return distribution P T ( ) ∼ T H -2 , our findings would be consistent with so called 1 /f noise (with Hurst exponent H = 0), on agreement with previous evidence [3], this latter being a trait of long range correlations. This gives further credit to our results, as 1 /f noise is usually found along with criticality [31].
Although we acknowledge that weakly chaotic systems such as intermittent ones can operate close to criticality, our results suggest that the paradigm of self-organised criticality (SOC) [31] - out of equilibrium dissipative sys-
Furthermore, the combined fact that (i) the results are robust for different human languages and that (ii) the timescales involved in this analysis require that the process is purely physiological, lead us to conclude that this mechanism is a universal trait of human beings. It is an open question if the onset of SOC in this system is the result of an evolutionary process [31], where human speech waveforms would have evolved to be independent of speaker and receiver distance and perception thresholds. In this context, further work should be done to investigate whether if similar patterns originate in the communication of other species, and up to which extent other variables, such as ageing, may play a role.
Acknowledgments. The authors would like to thank Luis Javier Rodr´ ıguez-Fuentes and Mikel Pe˜agarikano n for recording and hand-labeling the speech corpus.
- [1] R. Hegger, H. Kantz, and L. Matassini, Denoising human speech signals using chaoslike features, Phys. Rev. Lett. 84 , 14 (2000).
- [2] G. Hickok and D. Poepper, The cortical organization of speech processing, Nature Reviews Neuroscience 8 , 393402 (2007).
- [3] C.T. Kello, G.G. Anderson, J.G. Holden, and G. C. Van Orden, The Pervasiveness of 1/f Scaling in Speech Reflects the Metastable Basis of Cognition , Cognitive Science 32 , 7 (2008).
- [4] Blake S. Wilson, Charles C. Finley, Dewey T. Lawson, Robert D. Wolford, Donald K. Eddington and William M. Rabinowitz, Better speech recognition with cochlear implants, Nature 352 , (1991).
- [5] Fan-Gang Zeng et al, Speech recognition with amplitude and frequency modulations, Proc. Nat. Acad. Sci. USA 102 , 7 (2005).
- [6] James L. Flanagan, Jont B. Allen and Mark A. Hasegawa-Johnson, Speech Analysis Synthesis and Perception, (Springer-Verlag, 3rd ed, 2008).
- [7] Paul Taylor, Text-to-Speech Synthesis, (Cambridge University Press, 2009).
- [8] L.R. Rabiner and R. W. Schafer, Digital Processing of Speech Signals (Prentice-Hall, Englewood Cliffs, NT, 1978).
- [9] R.D. Kent and C. Read. The Acoustic Analysis of Speech (Whurr Publishers - Singular Publishing Group, 1992)
- [10] S. Marshall and G. L. Sicuranza (eds), Advances in Nonlinear Signal and Image Processing , Chapter 5 (Hindawi Publishing Corporation 2006).
- [11] I. Steinecke and H. Herzel, Bifurcations in an asymmetric vocal-fold model, J. Acoust. Soc. Am. , 97 1874-1884 (1995).
- [12] J.J. Jiang, Y. Zhang, and J. Stern, Modeling of chaotic vibrations in symmetric vocal folds J. Acoust. Soc. Am. 110 , 2120-2128 (2001).
- [13] J. Neubauer, P. Mergell, U. Eysholdt, Ulrich and H. Herzel, Spatio-temporal analysis of irregular vocal fold oscillations: Biphonation due to desynchronization of spatial modes J. Acoust. Soc. Am. 110 , 3179-3192 (2001).
- [14] C. Tao and J.J. Jiang, Chaotic component obscured by strong periodicity in voice production system, Phys. Rev. E 77 , 6 (2008).
- [15] A. Kumar and S.K. Mullick, Nonlinear dynamical analysis of speech, J. Acoust. Soc. Am. , 100 1 (1996).
- [16] P. Maragos, and A. Potamianos, Fractal dimensions of speech sounds: Computation and application to automatic speech recognition, J. Acoust. Soc. Am. , 105 3
(1999).
- [17] H. Kantz and T. Schreiber, Nonlinear Time Series Analysis (Cambridge University Press, 2nd ed, New York, 2005).
- [18] A.R. Ozborne and A. Provenzale, Finite correlation dimension for stochastic systems with power-law spectra, Physica D 35 , 3 (1989).
- [19] Luis Javier Rodr´ ıguez-Fuentes, Mikel Pe˜agarikano, Amn paro Varona and Mireia D´ ıez and Germ´ an Bordel, KALAKA-2: a TV Broadcast Speech Database for the Recognition of Iberian Languages in Clean and Noisy Environments, LREC , 99-105, (2012).
- [20] Thomas H. Crystal and Arthur S. House, Segmental durations in connected-speech signals: Syllabic stress, J. Acoust. Soc. Am. 83 , 1574-1585 (1988).
- [21] C.T. Kello, G.D.A. Brown, R. Ferrer-i-Cancho, J. Holden, K. Linkenkaer-Hansen, T. Rhoded, and G. C. Van Orden, Scaling laws in cognitive sciences, Trends in Cognitive Sciences 14 , 5 (2010).
- [22] P. Bak, L. Danon, and T. Scanlon, Unified law for earthquakes, Phys. Rev. Lett. 88 , 178501 (2002).
- [23] A. Corral, Long-Term Clustering, Scaling, and Universality in the Temporal Occurrence of Earthquakes, Phys. Rev. Lett. 92 , 10 (2004).
- [24] A. Corral, Renormalization-Group Transformations and Correlations of Seismicity, Phys. Rev. Lett. 95 , 2 (2005).
- [25] A. Corral, A. Osso and J.E. Llebot, Scaling of tropicalcyclone dissipation, Nature Phys. 6 (2010).
- [26] A. Corral,Point-occurrence self-similarity in cracklingnoise systems and in other complex systems, J. Stat. Mech. P01022 , (2009).
- [27] A. Clauset, C.R. Shalizi, and M.E.J. Newman, PowerLaw Distributions in Empirical Data, SIAM Rev 51 (4) (2009).
- [28] JP Sethna, KA Dahmen and CR Myers, Cracking Noise, Nature 418 , 242 (2001).
- [29] O. Peters, A. Deluca, A. Corral, J.D. Neelin and C.E. Holloway, Universality of rain event size distributions, J. Stat. Mech. P11030 , (2010).
- [30] L. de Arcangelis, C. Godano, E. Lippiello, and M. Nicodemi, Universality in solar flare and earthquake occurrence, Phys. Rev. Lett. 96 , 051102 (2006).
- [31] K. Christensen, NR Moloney, Complexity and Criticality (Imperial College Press, London, 2005)
- [32] B. Suki, A-L Barabasi, Z. Hantos, F. Petak and H.E. Stanley, Avalanches and power-law behaviour in lung inflation, Nature 368 , 14 (1994).
| null |
[
"Jordi Luque",
"Bartolo Luque",
"Lucas Lacasa"
] |
2014-08-05T14:34:20+00:00
|
2014-08-05T14:34:20+00:00
|
[
"physics.soc-ph",
"cs.CL",
"q-bio.NC"
] |
Speech earthquakes: scaling and universality in human voice
|
Speech is a distinctive complex feature of human capabilities. In order to
understand the physics underlying speech production, in this work we
empirically analyse the statistics of large human speech datasets ranging
several languages. We first show that during speech the energy is unevenly
released and power-law distributed, reporting a universal robust
Gutenberg-Richter-like law in speech. We further show that such earthquakes in
speech show temporal correlations, as the interevent statistics are again
power-law distributed. Since this feature takes place in the intra-phoneme
range, we conjecture that the responsible for this complex phenomenon is not
cognitive, but it resides on the physiological speech production mechanism.
Moreover, we show that these waiting time distributions are scale invariant
under a renormalisation group transformation, suggesting that the process of
speech generation is indeed operating close to a critical point. These results
are put in contrast with current paradigms in speech processing, which point
towards low dimensional deterministic chaos as the origin of nonlinear traits
in speech fluctuations. As these latter fluctuations are indeed the aspects
that humanize synthetic speech, these findings may have an impact in future
speech synthesis technologies. Results are robust and independent of the
communication language or the number of speakers, pointing towards an universal
pattern and yet another hint of complexity in human speech.
|
1408.0986v1
|
## A new correction of stellar oscillation frequencies for near-surface effects
Warrick H. Ball 1 and L. Gizon 2 1 ;
- 1 Institut für Astrophysik, Georg-August-Universität Göttingen, Friedrich-Hund-Platz 1, 37077 Göttingen, Germany e-mail: [email protected]
- 2 Max-Planck-Institut für Sonnensystemforschung, Justus-von-Liebig-Weg 3, 37077 Göttingen, Germany
August 6, 2014
## ABSTRACT
Context. Space-based observations of solar-like oscillations present an opportunity to constrain stellar models using individual mode frequencies. However, current stellar models are inaccurate near the surface, which introduces a systematic di GLYPH<11> erence that must be corrected.
Aims. We introduce and evaluate two parametrizations of the surface corrections based on formulae given by Gough (1990). The first we call a cubic term proportional to GLYPH<23> 3 = I and the second has an additional inverse term proportional to GLYPH<23> GLYPH<0> 1 = I , where GLYPH<23> and I are the frequency and inertia of an oscillation mode.
Methods. We first show that these formulae accurately correct model frequencies of two di GLYPH<11> erent solar models (Model S and a calibrated MESA model) when compared to observed BiSON frequencies. In particular, even the cubic form alone fits significantly better than a power law. We then incorporate the parametrizations into a modelling pipeline that simultaneously fits the surface e GLYPH<11> ects and the underlying stellar model parameters. We apply this pipeline to synthetic observations of a Sun-like stellar model, solar observations degraded to typical asteroseismic uncertainties, and observations of the well-studied CoRoT target HD 52265. For comparison, we also run the pipeline with the scaled power-law correction proposed by Kjeldsen et al. (2008).
Results. The fits to synthetic and degraded solar data show that the method is unbiased and produces best-fit parameters that are consistent with the input models and known parameters of the Sun. Our results for HD 52265 are consistent with previous modelling e GLYPH<11> orts and the magnitude of the surface correction is similar to that of the Sun. The fit using a scaled power-law correction is significantly worse but yields consistent parameters, suggesting that HD 52265 is su GLYPH<14> ciently Sun-like for the same power-law to be applicable.
Conclusions. We find that the cubic term alone is suitable for asteroseismic applications and it is easy to implement in an existing pipeline. It reproduces the frequency dependence of the surface correction better than a power-law fit, both when comparing calibrated solar models to BiSON observations and when fitting stellar models using the individual frequencies. This parametrization is thus a useful new way to correct model frequencies so that observations of individual mode frequencies can be exploited.
Key words. asteroseismology - stars: oscillations - stars: individual: HD 52265
## 1. Introduction
The Kepler (Borucki et al. 2010) and CoRoT (Auvergne et al. 2009) missions have ushered in a new era for asteroseismology. Hundreds of main-sequence and subgiant stars have now been observed at su GLYPH<14> ciently short cadence to resolve their solar-like oscillations. Of these, individual oscillation frequencies have been identified in dozens of stars (e.g. Appourchaux et al. 2012) for which many have complementary spectroscopic observations (e.g Bruntt et al. 2012; Molenda- ˙ akowicz et al. 2013) and a handful Z additionally have interferometric constraints on their radii (Huber et al. 2012; White et al. 2013). This wealth of observational data presents the opportunity to better constrain stellar model parameters, as well as the physics of the models themselves.
approximation when calculating oscillation frequencies, and the absence of a description of interactions between convection and the oscillations.
The main obstruction to constraining stellar models using the individually-identified frequencies is the known systematic di GLYPH<11> erence between models and observations caused by improper modelling of the near-surface layers. The frequency shifts are produced by several neglected or poorly-modelled physical processes and are collectively known as the surface e GLYPH<11> ects or surface term . Specific examples include poor modelling of temperature gradients in the superadiabatic layer, the use of the adiabatic
Several methods have already been employed to reduce the bias induced by these surface e GLYPH<11> ects on the underlying parameters of stellar models. First, Kjeldsen et al. (2008) proposed modelling the surface e GLYPH<11> ects as a power-law in frequency, calibrated to radial oscillations in the Sun around the frequency of maximum oscillation power GLYPH<23> max, and fit to the di GLYPH<11> erences between observed and modelled frequencies, the latter rescaled by the mean stellar density relative to the Sun. This scaled power-law approach has been adopted widely (e.g. Tang & Gai 2011; Mathur et al. 2012; Deheuvels et al. 2012). However, the fit overpredicts the magnitude of the surface e GLYPH<11> ect at lower frequencies (see Figs 1 and 2), and the correction assumes that the frequency shifts follow the same power law in other stars.
Second, Roxburgh & Vorontsov (2003) suggested fitting models to observations using particular ratios of frequency di GLYPH<11> erences, which have been used by several groups (e.g. Silva Aguirre et al. 2013) to fit stellar models. By approximating the eigenfunctions with partial wave solutions, one can express the surface e GLYPH<11> ect as a phase shift that can be suppressed by choosing appropriate ratios
Fig. 1. Frequency di GLYPH<11> erences between a standard solar model (Model S, Christensen-Dalsgaard et al. 1996) and observations of low-degree modes ( ' GLYPH<20> 3) by BiSON. The lines show fits made using an inverse term (dashed), cubic term (dotted) or both terms (solid). The fit with the inverse term is quite poor, with the cubic term much better and with both terms somewhat better still. The dot-dashed lines show a power law, fit to nine radial orders about GLYPH<23> max = 3090 GLYPH<22> Hz, as is used in the frequency correction proposed by Kjeldsen et al. (2008).
of frequency di GLYPH<11> erences. Otí Floranes et al. (2005) computed the structural sensitivity kernels for these ratios and demonstrated that they are indeed more sensitive to the stellar core. Because this method uses di GLYPH<11> erences and ratios, there is some loss of information, but the advantage is that stellar model parameters should not be biased by uncertainty at the surface.
Third, Gruberbauer et al. (2012) proposed a Bayesian method in which an additional frequency o GLYPH<11> set parameter is introduced for each oscillation mode and then marginalized against some prior. They used their method to compare the statistical evidence for solar models with di GLYPH<11> erent input physics (Gruberbauer & Guenther 2013), and to model a number of Kepler targets (Gruberbauer et al. 2013). The method appears successful in solar modelling, but, as noted by Gruberbauer et al. (2012) in their closing sections, the results appear biased towards larger masses when too few low-frequency modes are available.
Motivated by a lack of a leading method for modelling surface e GLYPH<11> ects, we propose here a new method in which the surface e GLYPH<11> ects are modelled by one or both of terms proportional to GLYPH<23> GLYPH<0> 1 = I and GLYPH<23> 3 = I , where GLYPH<23> is the frequency of an oscillation mode and I its corresponding inertia, normalized by the total displacement at the photosphere. In Section 2, we motivate these parametrizations and demonstrate the quality of fits to the di GLYPH<11> erences between modelled and observed low-degree oscillations in the Sun. In Section 3, we test the robustness of the parametrizations against both synthetic and real solar data. In Section 4, we apply the method to the planethosting CoRoT target HD 52265, with promising results, given the quality of the fit and consistency with previous results. Finally, we discuss the performance of the method and its potential flaws.
Fig. 2. Frequency di GLYPH<11> erences between a solar model calibrated with MESA and observations of low-degree modes ( ' GLYPH<20> 3) by BiSON. The lines show fits made using an inverse term (dashed), cubic term (dotted) or both terms (solid). The relative performances of the fits is the same as Fig. 1. The dot-dashed lines show a power law, fit to nine radial orders about GLYPH<23> max = 3090 GLYPH<22> Hz, as is used in the frequency correction proposed by Kjeldsen et al. (2008).
## 2. Modelling and fitting of surface effects
## 2.1. Parametrizations
Gough (1990) discussed potential asymptotic forms for the frequency shifts observed by Libbrecht & Woodard (1990) over the solar activity cycle. By considering perturbations near the surface under a variational principle and taking the asymptotic form of the displacement eigenfunctions in the evanescent layer, Gough (1990) concluded that the frequency shifts should generally take the form (equation 9.3 of Gough 1990)
$$\underset { \text{video} } { \text{test some} } \quad \delta \nu \, \infty \, \frac { Q ( \nu ^ { 2 } ) } { \nu \mathcal { I } },$$
where Q x ( ) is a quadratic function in x ; GLYPH<23> is the cyclic frequency of an oscillation mode; and GLYPH<14>GLYPH<23> is the frequency shift induced by a perturbation near the stellar surface. The normalized mode inertia I is defined by (e.g. Aerts et al. 2010, equation 3.140)
$$\underset { \ e e \, \text{effects} } { g \, \text{surface} } \quad \mathcal { I } = \frac { 4 \pi \int _ { 0 } ^ { R } \left [ | \xi _ { r } ( r ) | ^ { 2 } + \ell ( \ell + 1 ) | \xi _ { h } ( r ) | ^ { 2 } \right ] \rho r ^ { 2 } d r } { M \left [ | \xi _ { r } ( R ) | ^ { 2 } + \ell ( \ell + 1 ) | \xi _ { h } ( R ) | ^ { 2 } \right ] },$$
where GLYPH<24> r and GLYPH<24> h are the radial and horizontal components of the displacement eigenvector, R the photospheric radius, GLYPH<26> the unperturbed stellar density, M the total stellar mass, and ' the degree of the mode. In the solar models, the mode inertia decreases rapidly with frequency below about 2000 GLYPH<22> Hz before levelling out and reaching a minimum around 4000 GLYPH<22> Hz. This behaviour suppresses the magnitude of the frequency shifts at low frequency.
Gough (1990) further argued that a perturbation caused by a magnetic field concentrated into filaments, which would mostly modify the sound speed without much a GLYPH<11> ecting the gas density, would cause a shift proportional to GLYPH<23> 3 = I . This is the same form
suggested by Goldreich et al. (1991) for photospheric perturbations. A perturbation caused by a change in the description of convection, which would presumably modify the pressure scale height, would cause a shift proportional to GLYPH<23> GLYPH<0> 1 = I . Following from their distinct dependences on frequency, we refer to these two terms as cubic and inverse surface e GLYPH<11> ects, and to their combination as a combined surface e GLYPH<11> ect.
Fig. 1 shows the di GLYPH<11> erence between low-degree ( ' GLYPH<20> 3) solar oscillation frequencies 1 observed over 8640 days by the Birmingham Solar-Oscillations Network (BiSON, Broomhall et al. 2009) and adiabatic oscillation frequencies computed for a standard solar (Model S, Christensen-Dalsgaard et al. 1996). The curves represent uncertainty-weighted least-squares fits to the observed frequency di GLYPH<11> erences using either the cubic term, the inverse term or both. The inverse term alone does not produce a good fit to the di GLYPH<11> erences ( GLYPH<31> 2 = 424092), whereas the cubic term does ( GLYPH<31> 2 = 16150). The combination of both terms provides a significantly better fit ( GLYPH<31> 2 = 11 250). Similarly, Fig. 2 shows the di GLYPH<11> erence between the BiSON frequencies and a solar model computed using the solar calibration test case distributed with the Modules for Experiments in Stellar Astrophysics (MESA 2 , Paxton et al. 2011, 2013). Again, the inverse term alone fits badly ( GLYPH<31> 2 = 517678), the cubic term better ( GLYPH<31> 2 = 27287), and their combination slightly better still ( GLYPH<31> 2 = 25728). The larger GLYPH<31> 2 of the MESA solar model is chiefly contributed by a handful of modes below 1500 GLYPH<22> Hz, where the MESA model shows systematically higher frequencies and the observed uncertainties are smallest. In fact, the two lowest-frequency modes alone contribute about 17 000 to GLYPH<31> 2 . Despite this di GLYPH<11> erence, both solar models fit best with similar coe GLYPH<14> cients for the surface terms.
Figs 1 and 2 also show power-law fits as suggested by Kjeldsen et al. (2008). For each of the solar models, we fit a power law to the frequency di GLYPH<11> erences for nine radial orders about 3090 GLYPH<22> Hz. Though the fits are reasonably good in the fitted range, they are distinctly worse at higher and lower frequencies, as can be seen by the overprediction of the size of the correction around 2200 GLYPH<22> Hz. When one considers all the modes, the power-law fits give GLYPH<31> 2 = 49981 and 90 657 for Model S and the MESA solar model, or 46 626 and 75 172 when also fit to all the observed frequencies. The deviation is largely a result of the change of the frequency dependence of the mode inertiae around the 2500 GLYPH<22> Hz. Thus, when restricted to frequencies around 3090 GLYPH<22> Hz, the fit is comparably good, but its accuracy declines as the frequency range increases.
We thus propose two potential parametrizations for surface e GLYPH<11> ects. Firstly, we suggest a purely cubic term,
$$\delta \nu & = a _ { 3 } ( \nu / \nu _ { \text{ac} } ) ^ { 3 } / \mathcal { I }, & ( 3 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
or, secondly, a combined expression,
$$\delta \nu & = \left ( a _ { - 1 } ( \nu / \nu _ { \text{ac} } ) ^ { - 1 } + a _ { 3 } ( \nu / \nu _ { \text{ac} } ) ^ { 3 } \right ) / \mathcal { I }, & ( 4 ) \quad \left ( \begin{array} { c } a _ { - 1 } \\ a _ { 3 } \end{array} \right ),$$
where a GLYPH<0> 1 and a 3 are coe GLYPH<14> cients that are to be fit for the stellar model under consideration, and GLYPH<23> ac is the acoustic cuto GLYPH<11> frequency, used here purely to non-dimensionalize the frequencies. For convenience, we compute GLYPH<23> ac using the scaling relation
$$\nu _ { \text{ac} } / \nu _ { \text{ac}, \odot } & = \frac { g } { g _ { \odot } } \left ( \frac { T _ { \text{eff} } } { T _ { \text{eff}, \odot } } \right ) ^ { - 1 / 2 } & & ( 5 ) & \text{ pseudoin} \\ & & \text{model, tl}$$
where we take GLYPH<23> ac ; GLYPH<12> = 5000 GLYPH<22> Hz.
1 We have used the 'quiet-Sun' frequencies given in Table 3 of Broomhall et al. (2009).
2 http://mesa.sourceforge.net/
## 2.2. Fitting the surface effect
This parametrization uses only global parameters of the oscillations (i.e. frequency GLYPH<23> and mode inertia I ), both of which are typically computed by oscillation packages (e.g. ADIPLS, Christensen-Dalsgaard 2008). In addition, the parametrization only introduces linear terms, which can be fit analytically. It is not necessary to construct any additional stellar models or use iterative methods to compute the coe GLYPH<14> cients a GLYPH<0> 1 and a 3 , so the fit can easily be incorporated into existing model-fitting pipelines. We now describe in detail how this can be performed.
Suppose that we are comparing a stellar model, whose mode frequencies GLYPH<23> i ; mod and inertiae I i have been computed, to a set of observed mode frequencies GLYPH<23> i ; obs with uncertainties GLYPH<27> i . Here, i 2 f 1 ; : : : ; N modes g , where N modes is the number of observed modes. The best-fitting coe GLYPH<14> cients for the surface e GLYPH<11> ect are determined by minimizing
$$\text{how the} \\ \text{ir model} \\ \text{ted with} \\ \text{MFSA} ^ { 2 }.$$
with respect to the coe GLYPH<14> cients a 3 and a GLYPH<0> 1 , with GLYPH<14>GLYPH<23> i given by either of equations 3 or 4, above. Note that the coe GLYPH<14> cients are optimized with respect to all the observed frequencies rather than just the radial modes, as in the formula of Kjeldsen et al. (2008).
Because the surface e GLYPH<11> ect is linear in the coe GLYPH<14> cients, the bestfitting values can be determined analytically by linear regression. Let us first define the vector y by
$$\underset { \text{by } K \text{jeld} ^ { - } } { \text{odd} } y _ { i } = \frac { \nu _ { i, \text{obs} } - \nu _ { i, \text{mod} } } { \sigma _ { i } }.$$
For the cubic surface e GLYPH<11> ect (equation 3), we further define the vector X by
$$\begin{smallmatrix} \i \arg, \\ s, \, \text{as can} \\ n \, \text{around} \\ r \text{-law fits} \end{smallmatrix} \ X _ { i } = \frac { v _ { i, \text{mod} } ^ { 3 } } { \mathcal { I } _ { i } \sigma _ { i } },$$
in which case the best-fitting coe GLYPH<14> cient a 3 is
$$\underset { \substack { \text{bserved} \\ \text{ge of the} \\ 5 0 0 \mu \text{Hz.} \\ \text{$\text{the fact} } } } a _ { 3 } = \frac { \sum _ { i } X _ { i } y _ { i } } { \sum _ { i } X _ { i } ^ { 2 } }.$$
For the combined surface e GLYPH<11> ect (equation 4), we instead define a matrix X by
$$\Gamma \text{ surface} \quad X _ { i, 1 } = \frac { \nu _ { i, \text{mod} } ^ { - 1 } } { \mathcal { I } _ { i } \sigma _ { i } }, \ \ X _ { i, 2 } = \frac { \nu _ { i, \text{mod} } ^ { 3 } } { \mathcal { I } _ { i } \sigma _ { i } },$$
in which case the best-fitting coe GLYPH<14> cients a GLYPH<0> 1 and a 3 are given by
$$( 4 ) \ \left ( \begin{array} { c } a _ { - 1 } \\ a _ { 3 } \end{array} \right ) = \left ( X ^ { T } X \right ) ^ { - 1 } X ^ { T } \mathbf y.$$
This matrix equation does not reduce to as simple a form as the cubic case but the equations are still easily solved analytically. The matrix X X T is only 2 GLYPH<2> 2-dimensional, so its inverse is trivially computed. Note that ( X X T ) GLYPH<0> 1 X T is simply the Moore-Penrose pseudoinverse of X .
Thus, given the mode frequencies and inertiae of a stellar model, the analytic calculation described above returns the optimal values of the coe GLYPH<14> cients in a least-squares sense. Unlike the method of Kjeldsen et al. (2008), the frequencies are not rescaled by the ratio of the modelled and observed large separation before being corrected. In their formulation, this constrains the overall scale of the frequency correction to the same as those models that
have a similar large separation. Without such a rescaling, it is possible that an incorrect model could be found to fit well with unrealistically large values of the coe GLYPH<14> cients. In fact, in the new parametrization presented here, even the sign of the surface correction is currently not constrained, so there is nothing to prevent a best-fitting model that has a surface e GLYPH<11> ect of opposite sign to that of the Sun. We have not found this to be a problem but we cannot demonstrate that it never will be.
A possible solution of this issue would be to penalize large values of the coe GLYPH<14> cients by choosing an appropriate prior. For example, one could assume that the logarithms of the coe GLYPH<14> cients are uniformly distributed, which is often done for parameters that represent a scale rather than a position in parameter space. This would also prevent values of opposite sign but would require modification of the analytic calculation above.
## 2.3. Stellar models and fitting procedure
We tested the parametrizations, using the method outline above, with the downhill simplex method (Nelder & Mead 1965) implemented in MESA (revision 6022) optimizes the values of a stellar model's age t , mass M , initial metallicity [Fe H] , initial / i helium abundance Yi and mixing-length parameter GLYPH<11> . The combined spectroscopic constraints and corrected frequencies were used to evaluate the objective function given by
$$\chi ^ { 2 } & = \sum _ { i = 1 } ^ { N } \left ( \frac { x _ { i, \text{mod} } - x _ { i, \text{obs} } } { \sigma _ { i } } \right ) ^ { 2 }, & ( 1 2 ) \underset { \text{used as t} } { \text{observat} }$$
where xi ; obs represent all of the N observations usedspectroscopic or seismic-, GLYPH<27> i their uncertainties, and xi ; mod the corresponding model values. That is, we do not distinguish between any of the parameters in computing GLYPH<31> 2 , as in e.g. Metcalfe et al. (2014).
For comparison, we also ran the pipeline using the surface correction proposed by Kjeldsen et al. (2008), in which the surface correction is modelled as a power-law in frequency, calibrated to the Sun. By fitting the frequency di GLYPH<11> erences between the BiSON data and the MESA solar model above for nine radial orders about GLYPH<23> = 3090 GLYPH<22> Hz, we determined the appropriate index to be b = 4 81. This is similar to the value : b = 4 9 reported by Kjeldsen : et al. (2008), which is often used without recalibration.
To compute an initial model for each evolutionary track, we first initialized a pre-main-sequence model with a central temperature 300 000 K and allowed it to evolve until log L = L GLYPH<12> = 0 4, : where L and L GLYPH<12> are the total stellar and solar luminosities. At the beginning of each track, this model was then rescaled by mass to the value being evaluated, from its existing mass of 1 M GLYPH<12> for the synthetic and degraded solar fits, and 1 2 M : GLYPH<12> for HD 52265. In both cases, these models had ages less than 2 Myr. Stellar models were evolved starting from the pre-main-sequence model described above until a minimum of GLYPH<31> 2 had been found and one of the spectroscopic constraints was beyond its observed value by 5 GLYPH<27> . The parameters of the stellar model (including the age of the optimal fit along that track) are then used to update the simplex.
For the constitutive physics, we use opacities from the OPAL tables (Iglesias & Rogers 1996) and Ferguson et al. (2005) at high and low temperatures, respectively, blended linearly in the range 4 00 : GLYPH<20> log T GLYPH<20> 4 10. A standard Eddington grey atmosphere : is employed and convective processes are described by mixinglength theory according to Henyey et al. (1965). In the synthetic and Sun-as-a-star fits, convective overshooting was fixed at its solar-calibrated value. For HD 52265, no overshooting was included. The composition was scaled from the solar values given by Grevesse & Sauval (1998). Di GLYPH<11> usion and gravitational settling are included through the method of Thoul et al. (1994). Nuclear reaction rates are drawn from Caughlan & Fowler (1988) or the NACRE collaboration (Angulo et al. 1999), with preference given to the latter when available. We used newer specific reaction rates for the reactions 14 N( p ; GLYPH<13> ) 15 O(Imbriani et al. 2005) and 12 C( GLYPH<11>; GLYPH<13> ) 16 O (Kunz et al. 2002). All other options took default values, which are described in the MESA instrument papers (Paxton et al. 2011, 2013) and available in the source code.
The oscillation frequencies and mode inertiae were computed using the Aarhus adiabatic oscillation package (ADIPLS, Christensen-Dalsgaard 2008), without remeshing the stellar model. The models were computed with a mesh-spacing coefficient C = 1 2, which corresponds to about 1500 meshpoints. = This is su GLYPH<14> cient to correctly compute oscillation frequencies and mode inertiae along the main sequence. Calls to ADIPLS are made from MESA during the stellar evolution calculation and by default return the frequencies and inertiae.
We set the parameter tolerance in the downhill simplex method to 10 GLYPH<0> 5 . For the synthetic data, the initial parameters used the same parameters as for the input model. For the Sun, the initial parameters were the calibrated values given by Paxton et al. (2013). Finally, for HD 52265, we began with parameters computed by a fit to the spectroscopic parameters and averaged large and small separations, as measured by Ballot et al. (2011), and then found best-fit model parameters for the unperturbed observations with a cubic surface e GLYPH<11> ect. These parameters were used as the initial guess for all the fits to HD 52265.
We then used the observed uncertainties to generate 100 random realizations of the observations and fit a stellar model to each realization, using the parameters of the primary model as the initial guess. The reported values are the 15 85, 50 and 84 15 : : percentiles for the relevant distributions. i.e. the median and a 68 3 per cent confidence interval for the sample of 100 fits. :
Because di GLYPH<11> erent choices of timestep can lead to slightly di GLYPH<11> erent models for a given set of parameters, the timesteps were fixed at 10 Myr when fitting the synthetic data, the same value used to compute the input model. For the fits to the degraded BiSON observations and HD 52265, the timestep was decreased adaptively from 30 Myr to 300 000 yr based on partial evaluations of GLYPH<31> 2 dropping below certain limits.
The results of all of the fits are presented in Table 1.
## 3. Tests on synthetic and solar data
## 3.1. Synthetic data
We first tested whether our method correctly recovers synthetic input data. We constructed a stellar model with input parameters taken from the solar calibration presented by Paxton et al. (2013) and computed its oscillation frequencies and mode inertiae. We then used its observable properties as inputs for our modelling pipeline, using the same selection of observables as is available for HD 52265, along with the corresponding uncertainties. The spectroscopic parameters are the e GLYPH<11> ective temperature T e GLYPH<11> , surface gravity log g , metallicity [Fe H] / s and luminosity log L = L . GLYPH<12> We fit the frequencies of 28 oscillation modes: 10 radial, 10 dipole, and 8 quadrupole modes, centred on GLYPH<23> max GLYPH<25> 3000. These observable constraints are fairly typical for a star with good spectroscopic and intermediate seismic observations, with a timeseries of 117 days.
We first tested the cubic-only surface e GLYPH<11> ect, first by fitting the synthetic data without any surface e GLYPH<11> ect (case A) and second by fitting with a surface e GLYPH<11> ect with a 3 = GLYPH<0> 1 87 : GLYPH<2> 10 GLYPH<0> 7 GLYPH<22> Hz (case B).
## Table 1. Results of all fits presented here, as well as fits for HD 52265 presented by Escobar et al. (2012) and spectroscopic inputs for all fits (labelled Input ). The central values are the medians and the upper and lower limits are the locations of the 15 85 and 85 15 percentiles. In the case of the synthetic data, we also include the parameter values used to generate the observations. : :
Sun-as-a-star
| R = R GLYPH<12> | 0 : 9918 0 : 9911 + 0 : 0006 GLYPH<0> 0 : 0016 0 : 9911 + 0 : 0006 GLYPH<0> 0 : 0014 0 : 9911 + 0 : 0041 GLYPH<0> 0 : 0032 0 : 9913 + 0 : 0053 GLYPH<0> 0 : 0035 |
|-------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [Fe / H] s | 0 : 04 GLYPH<6> 0 : 05 0 : 040 + 0 : 012 GLYPH<0> 0 : 014 0 : 038 + 0 : 014 GLYPH<0> 0 : 020 0 : 035 + 0 : 039 GLYPH<0> 0 : 019 0 : 034 + 0 : 040 GLYPH<0> 0 : 035 |
| log L = L GLYPH<12> | GLYPH<0> 0 : 01 GLYPH<6> 0 : 05 GLYPH<0> 0 : 010 + 0 : 012 GLYPH<0> 0 : 014 GLYPH<0> 0 : 010 + 0 : 014 GLYPH<0> 0 : 017 GLYPH<0> 0 : 009 + 0 : 023 GLYPH<0> 0 : 026 GLYPH<0> 0 : 011 + 0 : 029 GLYPH<0> 0 : 018 |
| log g | 4 : 45 GLYPH<6> 0 : 09 4 : 4451 + 0 : 0006 GLYPH<0> 0 : 0008 4 : 4452 + 0 : 0004 GLYPH<0> 0 : 0009 4 : 4451 + 0 : 0016 GLYPH<0> 0 : 0015 4 : 4453 + 0 : 0017 GLYPH<0> 0 : 0019 |
| T e GLYPH<11> = K | 5773 GLYPH<6> 60 5766 + 43 GLYPH<0> 41 5769 + 46 GLYPH<0> 54 5772 + 70 GLYPH<0> 78 5768 + 78 GLYPH<0> 52 |
| a GLYPH<0> 1 = 10 GLYPH<0> 9 GLYPH<22> Hz | 0 : 71 + 6 : 90 GLYPH<0> 7 : 88 4 : 22 + 6 : 77 GLYPH<0> 6 : 37 |
| a 3 = 10 GLYPH<0> 7 GLYPH<22> Hz | 0 : 00 + 0 : 06 GLYPH<0> 0 : 10 GLYPH<0> 1 : 90 + 0 : 10 GLYPH<0> 0 : 08 GLYPH<0> 0 : 01 + 0 : 23 GLYPH<0> 0 : 20 GLYPH<0> 2 : 05 + 0 : 32 GLYPH<0> 0 : 24 |
| GLYPH<11> | 1 : 908 : 90 + 0 : 03 GLYPH<0> 0 : 05 : 91 + 0 : 03 GLYPH<0> 0 : 06 : 91 + 0 : 08 GLYPH<0> 0 : 09 : 91 + 0 : 11 GLYPH<0> 0 : 07 |
| Y i | 0 : 2784 0 : 2780 + 0 : 0006 GLYPH<0> 0 : 0004 0 : 2781 + 0 : 0005 GLYPH<0> 0 : 0004 0 : 2778 + 0 : 0009 GLYPH<0> 0 : 0011 0 : 2778 + 0 : 0011 GLYPH<0> 0 : 0015 |
| [Fe / H] i | 0 : 097 0 : 098 + 0 : 011 GLYPH<0> 0 : 014 0 : 096 + 0 : 014 GLYPH<0> 0 : 020 0 : 093 + 0 : 037 GLYPH<0> 0 : 019 0 : 093 + 0 : 037 GLYPH<0> 0 : 033 |
| M = M GLYPH<12> | 1 : 000 0 : 999 + 0 : 002 GLYPH<0> 0 : 005 0 : 999 + 0 : 002 GLYPH<0> 0 : 005 0 : 999 + 0 : 012 GLYPH<0> 0 : 009 1 : 000 + 0 : 015 GLYPH<0> 0 : 011 |
| t = Gyr | 4 : 61 4 : 64 + 0 : 23 GLYPH<0> 0 : 27 4 : 58 + 0 : 26 GLYPH<0> 0 : 21 4 : 64 + 0 : 28 GLYPH<0> 0 : 25 4 : 62 + 0 : 31 GLYPH<0> 0 : 23 |
| Synthetic | Input A B C D |
HD 52265
| : 000 + 0 : 005 GLYPH<0> 0 : 003 : 000 + 0 : 004 GLYPH<0> 0 : 003 : 001 + 0 : 003 GLYPH<0> 0 : 002 |
|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| 0 : 00 GLYPH<6> 0 : 05 0 : 023 + 0 : 022 GLYPH<0> 0 : 047 1 0 : 026 + 0 : 017 GLYPH<0> 0 : 047 1 0 : 033 + 0 : 013 GLYPH<0> 0 : 012 1 |
| 4 : 44 GLYPH<6> 0 : 09 4 : 4382 + 0 : 0020 GLYPH<0> 0 : 0013 4 : 4383 + 0 : 0020 GLYPH<0> 0 : 0013 4 : 4381 + 0 : 0015 GLYPH<0> 0 : 0009 |
| 5778 GLYPH<6> 60 5768 + 48 GLYPH<0> 46 5779 + 38 GLYPH<0> 43 5786 + 58 GLYPH<0> 50 |
| 1 : 83 + 0 : 05 GLYPH<0> 0 : 07 GLYPH<0> 2 : 1 : 84 + 0 : 04 GLYPH<0> 0 : 05 GLYPH<0> 2 : 1 : 87 + 0 : 04 GLYPH<0> 0 : 05 |
| 0 : 276 + 0 : 006 GLYPH<0> 0 : 018 0 : 278 + 0 : 006 GLYPH<0> 0 : 017 0 : 277 + 0 : 005 GLYPH<0> 0 : 011 |
| + 0 : : GLYPH<0> 0 : 19 1 : 000 : GLYPH<0> 0 : 007 0 : 085 : GLYPH<0> 0 : 043 + 0 : 25 GLYPH<0> 0 : 32 1 : 003 + 0 : 010 GLYPH<0> 0 : 007 0 : 091 + 0 : 012 GLYPH<0> 0 : 012 |
| GLYPH<0> 0 : 19 1 : 000 015 GLYPH<0> 0 : 010 0 : 083 + 0 : 020 GLYPH<0> 0 : 045 + 0 30 + 0 014 + 0 017 |
| : 68 : 75 |
| 4 4 |
| G |
| F |
| Input E |
| : : |
| 4 74 |
| + 0 18 |
| 314 + 0 : 008 GLYPH<0> 0 : + 0 : GLYPH<0> 0 | 002 004 : 003 + 0 : 008 GLYPH<0> 0 : 003 GLYPH<6> 0 : 02 | 1.321 1.310 |
|-----------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------|
| 0 : 19 GLYPH<6> 0 : 05 0 : 161 + 0 : 034 GLYPH<0> 0 : 042 0 : 164 + 0 : 033 GLYPH<0> 0 : 029 0 : 218 + 0 : 030 GLYPH<0> 0 : 039 0 : 20 GLYPH<6> 0 : 04 | 0 : 19 GLYPH<6> 0 : 05 0 : 161 + 0 : 034 GLYPH<0> 0 : 042 0 : 164 + 0 : 033 GLYPH<0> 0 : 029 0 : 218 + 0 : 030 GLYPH<0> 0 : 039 0 : 20 GLYPH<6> 0 : 04 | 0 : 19 GLYPH<6> 0 : 05 0 : 161 + 0 : 034 GLYPH<0> 0 : 042 0 : 164 + 0 : 033 GLYPH<0> 0 : 029 0 : 218 + 0 : 030 GLYPH<0> 0 : 039 0 : 20 GLYPH<6> 0 : 04 |
| 0 : 328 GLYPH<6> 0 : 024 0 : 322 + 0 : 015 GLYPH<0> 0 : 013 0 : 320 + 0 : 014 GLYPH<0> 0 : 010 0 : 329 + 0 : 018 GLYPH<0> 0 : 012 0 : 348 GLYPH<6> 0 : 006 | 0.313 | 0.328 |
| 4 : 35 GLYPH<6> 0 : 09 4 : 2853 + 0 : 0052 GLYPH<0> 0 : 0012 4 : 2842 + 0 : 0009 GLYPH<0> 0 : 0013 4 : 2854 + 0 : 0016 GLYPH<0> 0 : 0010 : 284 GLYPH<6> 0 : 002 | 4.282 | 4.282 |
| 6100 GLYPH<6> 60 6056 + 40 GLYPH<0> 40 7 : 39 5 : 77 6057 + 55 GLYPH<0> 38 6073 + 56 GLYPH<0> 58 6120 GLYPH<6> 20 | 6019 | 6097 |
| 11 | 11 | 11 |
| 0 : 292 + 0 : 006 GLYPH<0> 0 : 029 0 : 294 + 0 : 005 GLYPH<0> 0 : 004 0 : 301 + 0 : 006 GLYPH<0> 0 : 009 0 : 28 GLYPH<6> 0 : 02 | 0 : 280 | 0 : 298 |
| 0 : 259 + 0 : 030 GLYPH<0> 0 : 040 0 : 264 + 0 : 029 GLYPH<0> 0 : 026 0 : 312 + 0 : 024 GLYPH<0> 0 : 035 0 : 27 GLYPH<6> 0 : 04 | 0 : 23 | 0 : 215 |
| 1 : 215 + 0 : 030 GLYPH<0> 0 : 008 1 : 211 + 0 : 007 GLYPH<0> 0 : 007 1 : 222 + 0 : 020 GLYPH<0> 0 : 005 1 : 24 GLYPH<6> 0 : 02 | 1 : 22 | 1 : 20 |
| 2 : 44 + 0 : 16 GLYPH<0> 0 : 15 2 : 44 + 0 : 13 GLYPH<0> 0 : 09 2 : 26 + 0 : 18 GLYPH<0> 0 : 11 2 : 6 GLYPH<6> 0 : 2 | 3 : 00 | 2 : 38 |
| Input H I J TGEC | AMP(a) | AMP(b) |
Fig. 3. Uncorrected and corrected frequency di GLYPH<11> erences between observed frequencies for the degraded solar data and MESA models fit using only the cubic term (equation 3). The solid and empty points show the uncorrected and corrected di GLYPH<11> erences with error bars that correspond to the observed uncertainties. The shaded regions show the spread of the modelled frequencies. i.e. taken from the 100 fits to random realizations of the observations. There is no obvious remaining trend in the corrected frequencies.
We then tested the combined surface e GLYPH<11> ect, again first without a surface e GLYPH<11> ect (case C) and then with an artificial surface e GLYPH<11> ect (case D) with a GLYPH<0> 1 = 2 07 : GLYPH<2> 10 GLYPH<0> 7 GLYPH<22> Hz and a 3 = GLYPH<0> 2 78 : GLYPH<2> 10 GLYPH<0> 9 GLYPH<22> Hz. The results are labelled correspondingly in Table 1.
In all four cases, the fitted parameters are consistent with the input parameters, be it with or without a surface e GLYPH<11> ect, cubic or combined. We also confirmed that the parameters of the input model give a perfect fit. i.e. GLYPH<31> 2 = 0, as expected.
In cases C and D, a GLYPH<0> 1 is consistent with the input value but comes with large uncertainties. In fact, in case D, the result is consistent with a GLYPH<0> 1 = 0. This is unsurprising, given that the range of observed frequencies does not extend to the bend in the surface e GLYPH<11> ect caused by the sharp frequency dependence of the mode inertiae at low frequencies. In addition, Figs 1 and 2 show that the surface e GLYPH<11> ect is dominated by the cubic term, especially at frequencies about about 2500 GLYPH<22> Hz.
## 3.2. Solar data
As a second test, we perform a Sun-as-a-star experiment, where we degrade observations of the Sun to the same selection and uncertainties as observations of HD 52265. We fit the observations with the cubic term (case E), the combined terms (case F), or the scaled power law (case G), the results of which are presented in Table 1.
The fitted parameters are all consistent with our knowledge of the Sun, although the mixing length parameter GLYPH<11> is consistently somewhat lower than the calibrated value of 1 908. The oscillation : frequencies tightly constrain the mechanical structure of the star and thus the mass, radius and combinations thereof, like mean density and surface gravity. As noted when fitting the synthetic data, the inverse term is not well-constrained, but its inclusion
Fig. 4. As in Fig. 3 but with a fit made using the combined surface term (equation 4).
Fig. 5. As in Fig. 3 but with a fit made using the rescaled power law proposed by Kjeldsen et al. (2008). Note that, as seen for the fits to the solar data in Figs 1 and 2, the correction is slightly overestimated at low frequencies.
does not appear to greatly worsen the quality of the fit. The surface terms themselves are consistent with the coe GLYPH<14> cients determined by fitting the BiSON frequencies to the calibrated solar model (see Fig. 2): a 3 = GLYPH<0> 2 13 : GLYPH<2> 10 GLYPH<0> 7 GLYPH<22> Hz for the cubic fit, and a 3 = GLYPH<0> 2 25 : GLYPH<2> 10 GLYPH<0> 7 GLYPH<22> Hz and a GLYPH<0> 1 = 1 73 : GLYPH<2> 10 GLYPH<0> 9 GLYPH<22> Hz for the combined fit.
The frequency di GLYPH<11> erences for cases E, F, and G are shown in Figs 3, 4, and 5. The error bars on the data points represent the observed uncertainties; the shaded regions around the corrected and uncorrected di GLYPH<11> erences represent the spread in the 100 fits
to independent realizations. The quality of the new fits to the corrected frequencies is encouraging, given that there are no obvious remaining trends in either case. The power-law fit shows a marginal trend in the residual di GLYPH<11> erences, with the corrected model frequencies being too high at low frequencies. This is consistent with the fits to solar data in Figs 1 and 2.
It is clear from Figs 3 and 4 that simultaneously fitting the surface e GLYPH<11> ect with the cubic or combined terms allows a greater range of models than the observed uncertainties. This can be seen in the fact that the envelope of model uncertainties is broader for the uncorrected than the corrected frequencies. In essence, the uncertainty about the surface e GLYPH<11> ects is reflected in the uncertainties of the underlying stellar parameters but every model is optimally corrected to match the observed frequencies. For example, consider case E, where only the cubic term was used. The uncertainty in the coe GLYPH<14> cient a 3 is about 10 per cent, which naively permits an uncertainty of about 1 GLYPH<22> Hz for a 10 GLYPH<22> Hz frequency shift, all else being equal. In reality, the coe GLYPH<14> cients are correlated with the underlying model parameters, so this estimate is only approximate. When using both terms, an even greater range of models is allowed and the uncertainties are correspondingly larger.
In addition, the uncertainties of the corrected model frequencies are generally smaller than the observed uncertainties. This follows from the random realization process. In the observations, the uncertainties are uncorrelated, so the observed frequencies are perturbed independently. In the stellar models, the variations of the frequencies with respect to the stellar parameters are instead tightly correlated. It is highly unlikely that a random realization of the observations reproduces the correlated variations of the model frequencies, so the underlying parameters vary relatively little from their best-fitting values.
## 4. Application to HD 52265
HD 52265 is a metal-rich G0V planet-hosting star that was observed by CoRoT for 117 days between November 2008 and March 2009. Ballot et al. (2011) reported spectroscopic and seismic observations that Escobar et al. (2012) used to perform a detailed analysis of the star. This involved the direct comparison of manually-selected models (TGEC) and two automated fits using the Asteroseismic Modelling Portal (AMP, 3 Metcalfe et al. 2009): one using all the frequencies given by Ballot et al. (2011) (fit (a)) and one omitting the lowest three frequencies, as is done here (fit (b)). Lebreton & Goupil (2012) estimated the extent of inward convective overshooting using stellar models with t GLYPH<25> 2Gyr, M GLYPH<25> 1 25 M : GLYPH<12> , and R GLYPH<25> 1 3 R : GLYPH<12> , in reasonable agreement with the results presented here, although they do not report quantitative uncertainties on the model parameters. They do estimate the depth of the convective zone, with penetrative overshooting, to be R cz = 0 800 : GLYPH<6> 0 004 : R GLYPH<3> , where R GLYPH<3> is the radius of the star. This is compatible with our result of R cz = R GLYPH<3> = 0 803 : for the best-fitting stellar model. Finally, HD 52265 was modelled using the SEEK pipeline (Quirion et al. 2010), as reported in Escobar et al. (2012) and Gizon et al. (2013), which returned best-fit parameters t = 2 37 : GLYPH<6> 0 29 Gyr, : M = 1 27 : GLYPH<6> 0 03 M : GLYPH<12> , and R = 1 34 : GLYPH<6> 0 02 R : GLYPH<12> . The higher mass is somewhat discrepant, and Escobar et al. (2012) argued that it represents a local, secondary, optimum of the model parameters. The spectroscopic observations are listed in Table 1, along with model parameters for the TGEC and AMP fits by Escobar et al. (2012), and the results presented in this paper. For the frequency selection, we used the reported frequencies for 10 ' = 0, 10 ' = 1 and 8 ' = 2 oscillation
3 https://amp.phys.au.dk/
Fig. 6. As in Fig. 3 for the fit to HD 52265 using the cubic term (equation 3).
modes between 1600 and 2550 GLYPH<22> Hz. We omitted the lowest three frequencies given by Ballot et al. (2011), which are reportedly less reliable.
The fits with the cubic, combined, and power-law surface terms (cases H, I, and J) are included in Table 1. We have also included the best-fit parameters reported by Escobar et al. (2012) for comparison, and our results with the cubic and combined surface terms are entirely consistent. Like previous models, we find that the surface gravity is lower (though not significantly) than the spectroscopic value. In addition, the parameter fits made with the combined surface term are consistent with those made with the cubic term only.
Figs 6, 7, and 8 show the frequency di GLYPH<11> erences for the cubic, combined, and power-law surface terms. For the cubic and combined terms, there is again no obvious residual, though the highest-frequency ' = 2 mode shows a significant remaining di GLYPH<11> erence. In both cases, the overall scale of the surface e GLYPH<11> ect is similar to that of the Sun, taking a value of roughly 7 GLYPH<22> Hz at GLYPH<23> max GLYPH<25> 2100 GLYPH<22> Hz. The coe GLYPH<14> cients of the cubic term also of similar size, when the oscillation frequencies are normalized against the acoustic cuto GLYPH<11> frequency. The coe GLYPH<14> cient of the inverse term is non-zero at the 2 GLYPH<27> level, although still consistent with the Sunas-a-star value because of the large uncertainties. As noted before, the addition of the inverse term does not appear to improve the fit significantly, although it similarly does no harm. This is also presumably due to the limited number of radial orders covered by our frequency selection.
The power-law fit again shows a slight residual trend but the overall correction is similar. However, the quality of the fit is significantly worse than the cubic term. The di GLYPH<11> erences between the corrected and observed frequencies for the best-fitting models give GLYPH<31> 2 = 124 for the scaled power-law compared with GLYPH<31> 2 = 31 for the cubic term. The underlying model parameters deviate only within uncertainties but we expect that they would deviate further as the number of observed modes increases and the shape of the surface e GLYPH<11> ect deviates from a power law. Thus, given the smaller values of GLYPH<31> 2 in both HD 52265 and the full set of BiSON
Fig. 7. As in Fig. 3 for the fit to HD 52265 using the combined term (equation 4).
Fig. 8. As in Fig. 3 but for a fit to HD 52265 made using the rescaled power law proposed by Kjeldsen et al. (2008). Again, the correction is slightly overestimated at low frequencies.
frequencies, we believe that the cubic term produces better best-fit parameters than the scaled power law.
## 5. Discussion
## 5.1. Quality of parametrization fits
Based on the results presented here, it appears that the proposed parametrizations of the surface e GLYPH<11> ects can adequately describe the true surface e GLYPH<11> ects, certainly enough to correctly recover the underlying stellar model parameters. In all cases, the price to pay is increased uncertainties in the uncorrected frequencies, even though the corrected frequencies have small uncertainties. Despite this loss of accuracy, the exploitation of the full set of frequencies, even given a relatively meagre selection of mode frequencies, allows precise determination of the underlying stellar model parameters.
For the degraded BiSON data, the scaled power-law correction gives consistent results. In this case the index of the power law has been fit directly to the star in question, and the fit is about as good as the new parameterizations over the observed range. There is, however, an uncorrected trend, which would induce greater discrepancies in the stellar model parameters given a larger frequency range. In the case of HD 52265, the fit is significantly worse, although the stellar parameters continue to be consistent. The poor fit is not because a power law describes the frequency di GLYPH<11> erences badly, but rather because the solarcalibrated value of the index is incorrect. For comparison, if we fit a power law to the frequency di GLYPH<11> erences of the best fit model with the cubic term, we find b = 4 11, which is smaller than the : solar-calibrated value of b = 4 81. The use of the mode inertia : appears to capture the variation of the power law between the two stars studied here.
In the Sun, the surface e GLYPH<11> ect appears to be dominated by the cubic term and the frequency range used here does not strongly constrain the inverse term. We recommend that stellar models are first fit using only the cubic term and that the inverse term be introduced as an additional parameter after the initial fit has been performed. A greater number of lower-order frequencies would better constrain the inverse term to determine whether its contribution remains small. In fact, generally speaking, lower order modes constrain any surface term better, because those modes are a GLYPH<11> ected least. Such observations are most easily made when frequencies are determined from radial velocity data rather than photometric data. The proposed Stellar Observations Network Group (SONG, Grundahl et al. 2014), of which one node is now essentially operational, would be invaluable in making such observations.
## 5.2. Uncertainties
Our uncertainties for the parameters of HD 52265 are generally smaller than those reported by Escobar et al. (2012) for the TGEC models. Their model parameters were derived from the mean and standard deviation of three manually chosen best-fitting models. The results from the AMP are reported without uncertainties. Thus, our smaller uncertainties are not necessarily at odds with the TGEC fit. The uncertainty on the initial helium abundance Yi may seem particularly small, but it is consistent with the uncertainty on the initial metallicity. If we suppose that the initial helium and metal abundances are related by some enrichment law Y = Y 0 + GLYPH<1> ( Y = GLYPH<1> Z Z ) and that
$$[ F e / H ] & = \log ( Z / X ) - \log ( Z / X ) _ { \odot }, & ( 1 3 )$$
with some primordial helium abundance Y 0 , then it is possible to solve for the Y given [Fe H]. By taking, for example, / Y 0 = 0 245 : and GLYPH<1> Y = GLYPH<1> Z = 1 4, then a change in [Fe H] : / i from 0 27 to 0 30 : : induces a change of just 0 002 in : Yi . Thus, the uncertainties, though small, are compatible with our knowledge of the initial metallicity, even though the helium and metal abundance were fit independently.
The uncertainties of the fits to the synthetic data are smaller still because the central values, about which the random realizations are made, are perfectly consistent. Similarly, central values of the degraded solar data are based on the BiSON data, for which
the uncertainties are also very small. Thus, even though we have used greater uncertainties to randomly realize observations, the parameter space is dominated by a deeper optimal solution.
We also note that the uncertainties reported here are purely statistical. They do not reflect uncertainties in the input model physics that would be induced by changing, for example, the opacity tables, equation of state, or atmospheric model. This is an interesting problem that has not yet been extensively studied for fits using individual frequencies. Chaplin et al. (2014) analyzed several hundred Kepler short-cadence targets using six stellar evolution codes across 11 pipelines and concluded that the uncertainties induced by di GLYPH<11> erent codes and pipelines are about 3.7, 1.7, and 16 per cent in mass, radius, and age, when spectroscopic constraints on T e GLYPH<11> , log g and [Fe H] are available. These fits were / not based on individually-resolved frequencies, which presumably constrain models better than overall seismic properties, but they still inform us on the di GLYPH<11> erences caused by using di GLYPH<11> erent stellar models. In the future, we intend to study the systematic errors in detail by exploiting the modular design of MESA to isolate the e GLYPH<11> ects of individual choices of input physics.
## 5.3. Potential shortcomings
Despite performing well on the targets studied here, the method does contain several potential flaws. First, as is always the case with optimization problems, we cannot guarantee that we have found the global best-fit solution. To some extent, the repeated random realizations mitigate this problem because if a much better minimum were available, we might expect at least a few of the random realizations to converge there. Indeed, this is the case in HD 52265 where, in cases H and I, about 15 and 3 models settled on a secondary minimum with a mass of roughly 1 30 M : GLYPH<12> , close to the value identified by Escobar et al. (2012). The asymmetry of the uncertainties in Table 1 captures the existence of this alternative local optimum. An initial grid-based search could also rule out the existence of better minima in distinctly di GLYPH<11> erent regions of parameter space. The greater problem is that the introduction of one or two new free parameters, which are simultaneously optimized during the fit and currently not limited in scale or sign, might introduce new local minima that confuse the optimization process. We propose to mitigate this by choosing initial guesses that weight non-seismic observations more heavily, though the final sample of fits should continue to be performed without artificially weighting the observables.
Second, there is an additional implicit free parameter: the depth of the normalization of the mode inertiae, which controls the depth of the perturbation that produces the surface e GLYPH<11> ect. Here, we have used the radius of the photosphere. This is a reasonable choice and fits the Sun well but there is no guarantee that this will be the case for stars significantly di GLYPH<11> erent from the Sun. Rosenthal et al. (1999) computed the frequency di GLYPH<11> erences between a solar model and a detailed simulation of the Sun's surface convection zone and found that the surface e GLYPH<11> ect is largely (but not entirely) produced by the suspension of the envelope by turbulent pressure, which peaks in a narrow region that is essentially at the photosphere. Recent simulations (e.g. Beeck et al. 2013; Magic et al. 2013; Trampedach et al. 2013) should allow similar studies to be carried out for a range of stellar surface properties and allow some forward computation of the expected location and scale of the surface correction. Presently, it seems reasonable to continue to normalize the mode inertiae at the photosphere.
Third, our assessment of the parametrizations is specific to the traditional mixing-length theory and adiabatic pulsations.
Though not widely implemented, other one-dimensional descriptions of convection exist (e.g. full spectrum turbulence , Canuto &Mazzitelli 1991) and these predict di GLYPH<11> erent properties for the superadiabatic layers near the surfaces of Sun-like stars. Since this region introduces some component of the surface e GLYPH<11> ect, the parametrizations presented here may no longer fit so well. Similarly, non-adiabatic pulsation frequencies exhibit a kind of surface e GLYPH<11> ect, which also may or may not be described precisely by the parametrizations.
## 6. Conclusion
We have presented two parametrizations of the known systematic di GLYPH<11> erence between modelled and observed stellar oscillation frequencies induced by poor-modelling of the near-surface layers. The first contains a term proportional to GLYPH<23> 3 = I (the cubic term, equation 3) and the second an additional term proportional to GLYPH<23> GLYPH<0> 1 = I (the inverse term, equation 4). The second parametrization corrects most of the frequency di GLYPH<11> erences between two calibrated solar models and observed solar oscillation frequencies, in the sense that no obvious trends remain. Both parametrizations fit the frequency di GLYPH<11> erences significantly better than power laws.
We implemented a simultaneous fit to the surface e GLYPH<11> ects in MESA and constructed a model-fitting pipeline that fits stellar models to 100 random realizations of the observable data. By testing the new pipeline on data for a synthetic target, the Sun, and the CoRoT target HD 52265, we have shown that these parametrizations give unbiased results that are compatible with independent measurements and model fits. We compared our results with calculations made with the same pipeline, but using the scaled power-law proposed by Kjeldsen et al. (2008). In the case of the Sun, both methods perform similarly well, but for HD 52265, the power-law fit produces a significantly worse fit to the frequencies but consistent stellar parameters. Thus, the cubic term alone is arguably an improvement over a scaled power law.
Generally, the cubic term dominates the surface correction. The contribution of the inverse term is clear when comparing observations to calibrated solar models, but poorly-constrained by the small range of radial orders that we have used in the fitting pipeline. In all cases, the stellar model parameters are consistent, irrelevant of which surface term is used, and the overall scale of the surface term is similar in the Sun and HD 52265.
The simultaneous fit potentially introduces new local minima to the models' parameter space and we recommend that initial guesses be found by weighting complementary non-seismic observations more strongly. Given a reasonable initial guess, however, we have found that the pipeline converges on a reasonable solution even though the magnitudes and signs of the coe GLYPH<14> cients are not constrained. In upcoming work, we will apply our pipeline to more stars, covering a greater range of parameters, to ensure that the parametrizations are generally useful in allowing us to exploit present and future observations of resolved oscillation modes.
Acknowledgements. WHB and LG acknowledge research funding by Deutsche Forschungsgemeinschaft (DFG) under grant SFB 963 1 'Astrophysical flow in-/ stabilities and turbulence' (Project A18). LG also acknowledges support from EU FP7 Collaborative Project 'Exploitation of Space Data for Innovative Helioand Asteroseismology" (SPACEINN). Part of this work was completed during a visit to the Aarhus Stellar Astrophysics Centre.
## References
Aerts, C., Christensen-Dalsgaard, J., & Kurtz, D. W. 2010, Asteroseismology,
Astronomy and Astrophysics Library (Springer) Angulo, C., Arnould, M., Rayet, M., et al. 1999, Nuclear Physics A, 656, 3
Appourchaux, T., Chaplin, W. J., García, R. A., et al. 2012, A&A, 543, A54 Auvergne, M., Bodin, P., Boisnard, L., et al. 2009, A&A, 506, 411 Ballot, J., Gizon, L., Samadi, R., et al. 2011, A&A, 530, A97 Beeck, B., Cameron, R. H., Reiners, A., & Schüssler, M. 2013, A&A, 558, A48 Borucki, W. J., Koch, D., Basri, G., et al. 2010, Science, 327, 977 Broomhall, A.-M., Chaplin, W. J., Davies, G. R., et al. 2009, MNRAS, 396, L100 Bruntt, H., Basu, S., Smalley, B., et al. 2012, MNRAS, 423, 122 Canuto, V. M. & Mazzitelli, I. 1991, ApJ, 370, 295 Caughlan, G. R. & Fowler, W. A. 1988, Atomic Data and Nuclear Data Tables, 40, 283
Chaplin, W. J., Basu, S., Huber, D., et al. 2014, ApJS, 210, 1
Christensen-Dalsgaard, J. 2008, Ap&SS, 316, 113
Christensen-Dalsgaard, J., Dappen, W., Ajukov, S. V., et al. 1996, Science, 272, 1286
Deheuvels, S., García, R. A., Chaplin, W. J., et al. 2012, ApJ, 756, 19 Escobar, M. E., Théado, S., Vauclair, S., et al. 2012, A&A, 543, A96 Ferguson, J. W., Alexander, D. R., Allard, F., et al. 2005, ApJ, 623, 585 Gizon, L., Ballot, J., Michel, E., et al. 2013, Proceedings of the National Academy of Science, 110, 13267
Goldreich, P., Murray, N., Willette, G., & Kumar, P. 1991, ApJ, 370, 752 Gough, D. O. 1990, in Lecture Notes in Physics, Berlin Springer Verlag, Vol. 367, Progress of Seismology of the Sun and Stars, ed. Y. Osaki & H. Shibahashi, 283
Grevesse, N. & Sauval, A. J. 1998, Space Sci. Rev., 85, 161
Gruberbauer, M. & Guenther, D. B. 2013, MNRAS, 432, 417
Gruberbauer, M., Guenther, D. B., & Kallinger, T. 2012, ApJ, 749, 109
Gruberbauer, M., Guenther, D. B., MacLeod, K., & Kallinger, T. 2013, MNRAS,
435, 242
Grundahl, F., Christensen-Dalsgaard, J., Pallé, P. L., et al. 2014, in IAU Symposium, Vol. 301, IAU Symposium, ed. J. A. Guzik, W. J. Chaplin, G. Handler, &A. Pigulski, 69-75
Henyey, L., Vardya, M. S., & Bodenheimer, P. 1965, ApJ, 142, 841
Huber, D., Ireland, M. J., Bedding, T. R., et al. 2012, ApJ, 760, 32
Iglesias, C. A. & Rogers, F. J. 1996, ApJ, 464, 943
Imbriani, G., Costantini, H., Formicola, A., et al. 2005, European Physical Journal A, 25, 455
Kjeldsen, H., Bedding, T. R., & Christensen-Dalsgaard, J. 2008, ApJ, 683, L175
Kunz, R., Fey, M., Jaeger, M., et al. 2002, ApJ, 567, 643
Lebreton, Y. & Goupil, M. J. 2012, A&A, 544, L13
Libbrecht, K. G. & Woodard, M. F. 1990, Nature, 345, 779
Magic, Z., Collet, R., Asplund, M., et al. 2013, A&A, 557, A26
Mathur, S., Metcalfe, T. S., Woitaszek, M., et al. 2012, ApJ, 749, 152
Metcalfe, T. S., Creevey, O. L., & Christensen-Dalsgaard, J. 2009, ApJ, 699, 373
Metcalfe, T. S., Creevey, O. L., Dogan, G., et al. 2014, ArXiv e-prints
Molenda- ˙ akowicz, J., Sousa, S. G., Frasca, A., et al. 2013, MNRAS, 434, 1422
Z
Nelder, J. A. & Mead, R. 1965, The Computer Journal, 7, 308
Otí Floranes, H., Christensen-Dalsgaard, J., & Thompson, M. J. 2005, MNRAS, 356, 671
Paxton, B., Bildsten, L., Dotter, A., et al. 2011, ApJS, 192, 3
Paxton, B., Cantiello, M., Arras, P., et al. 2013, ApJS, 208, 4
Quirion, P.-O., Christensen-Dalsgaard, J., & Arentoft, T. 2010, ApJ, 725, 2176
Rosenthal,
C.
S.,
Christensen-Dalsgaard,
J.,
Nordlund,
Å.,
Stein,
R.
F.,
&
Trampedach, R. 1999, A&A, 351, 689
Roxburgh, I. W. & Vorontsov, S. V. 2003, A&A, 411, 215
Silva Aguirre, V., Basu, S., Brandão, I. M., et al. 2013, ApJ, 769, 141
Tang, Y. K. & Gai, N. 2011, A&A, 526, A35
Thoul, A. A., Bahcall, J. N., & Loeb, A. 1994, ApJ, 421, 828
- Trampedach, R., Asplund, M., Collet, R., Nordlund, Å., & Stein, R. F. 2013, ApJ, 769, 18
White, T. R., Huber, D., Maestro, V., et al. 2013, MNRAS, 433, 1262
|
10.1051/0004-6361/201424325
|
[
"Warrick H. Ball",
"L. Gizon"
] |
2014-08-05T14:44:53+00:00
|
2014-08-05T14:44:53+00:00
|
[
"astro-ph.SR"
] |
A new correction of stellar oscillation frequencies for near-surface effects
|
Space-based observations of solar-like oscillations present an opportunity to
constrain stellar models using individual mode frequencies. However, current
stellar models are inaccurate near the surface, which introduces a systematic
difference that must be corrected. We introduce and evaluate two
parametrizations of the surface corrections based on formulae given by Gough
(1990). The first we call a cubic term proportional to $\nu^3/\mathcal{I}$ and
the second has an additional inverse term proportional to
$\nu^{-1}/\mathcal{I}$, where $\nu$ and $\mathcal{I}$ are the frequency and
inertia of an oscillation mode. We first show that these formulae accurately
correct model frequencies of two different solar models ... We then incorporate
the parametrizations into a modelling pipeline that simultaneously fits the
surface effects and the underlying stellar model parameters. We apply this
pipeline to synthetic observations of a Sun-like stellar model, solar
observations degraded to typical asteroseismic uncertainties, and observations
of the well-studied CoRoT target HD52265. For comparison, we also run the
pipeline with the scaled power-law correction proposed by Kjeldsen et al.
(2008). The fits to synthetic and degraded solar data show that the method is
unbiased and produces best-fit parameters that are consistent with the input
models and known parameters of the Sun. Our results for HD52265 are consistent
with previous modelling efforts and the magnitude of the surface correction is
similar to that of the Sun. The fit using a scaled power-law correction is
significantly worse but yields consistent parameters ... We find that the cubic
term alone is suitable for asteroseismic applications and it is easy to
implement in an existing pipeline. ... This parametrization is thus a useful
new way to correct model frequencies so that observations of individual mode
frequencies can be exploited.
|
1408.0987v3
|
ISSN 2310-2144
www.aaatec.org
## Reconstruction of Ancient Egyptian Sundials
## Larisa N. Vodolazhskaya
Department of Space Physics, Southern Federal University (SFU), Rostov-on-Don, Russian Federation; E-mails: [email protected], [email protected]
## Abstract
The article presents the results of a study of the design features of vertical and L-shaped ancient Egyptian sundials, developed their models, on the basis of which the reconstruction of the sundial was carried out. An original scheme of a simple method for a fairly accurate time measurement with their help is also proposed. The model proposed by us, which describes a vertical sundial, is a vertical sundial with an inclined gnomon that takes into account the latitude of the area. The model proposed by us is characterized by marking the hour lines from 6 to 12 o'clock every hour. From 12:00 a shift is made in the marking of the hour lines for half an hour, and then the marking is repeated every hour. As a consequence of the reconstruction of the vertical sundial, we have developed and proposed a model that describes the design features and operation of two types of L-shaped sundial. They had to work in conjunction with an inclined gnomon, like a vertical sundial, or directly with a vertical sundial. In this case, the L-shaped sundial could complement the vertical sundial by providing the ability to read hour markers and interpret the readings of the vertical sundial, as there were no inscriptions on the vertical sundial. The article also provides a decoding of the inscription from the tomb of Seti I, which has long intrigued researchers. It is proved that the inscription contains the lengths of the intervals between adjacent marks of the L-shaped sundial of the second type, where the first interval corresponds not to one hour, but to half an hour.
Keywords: sundial, model, astronomical methods, archaeoastronomy, ancient Egypt.
## Introduction
Technologies of ancient Egypt, including measurement of time, reached a high level. In ancient Egypt, there were clepsydras and sundials, which had a different design: L-shaped, stepped, vertical sundials. Examples of such tools are stored in the Cairo and Berlin museums.
The earliest written evidence of the existence of a sundial in Ancient Egypt dates from the reign of Pharaoh Thutmose III (1521-1473 BC). This pharaoh made several dozen campaigns both in Asia and in Nubia. In the description of one battle in the Manedo Gorge, which he fought during his first campaign in Asia, there is a mention that the army set out at noon, when the shadow of the sun "turns" [1], [2], [3] . It is believed that this moment could only be determined by the sundial.
However, if the very first mention of the ancient Egyptian sundial is still associated with the name of Thutmose III, then, as the Egyptologist D. G. Breasted rightly points out [4], this does not mean that there were no sundials in ancient Egypt before him.
Figure 1. Egyptian sundial (XIII century BC).
Also in the tomb of Seti I (1300 BC) there is an image of a sundial [7]. Currently, there are several examples of Egyptian sundial stored in various museums around the world. The Egyptian Museum in Berlin has a green slate sundial dating from the reign of Thutmose III and a Faiyum sundial dating from 1000-600 BC [8]. Another sundial is known from the reign of Pharaoh Merneft (1258-1239 BC), which were found in Palestine (Fig. 1) [9]. On a sundial from Palestine, hour lines emanating from the center are plotted on a flat surface. In ancient times, this sundial was used to determine time, as well as modern sundial, with the help of the direction of the shadow.
It is believed that the division into hours first appeared in Egypt. Already from 2100 BC. The Egyptian priests divided the day into 24 hours [5]. However, the length of the hour constantly changed throughout the year and was equal to 1/12 of the time from sunrise to sunset or from sunset to sunrise. Thus, the duration of an hour fluctuated depending on latitude and time of year [6]. According to the vertical sundial in ancient Egypt, they could more or less accurately determine the hours of the day only during the spring and autumn equinoxes; at other times they showed the hours of the day approximately.
In 2013, during archaeological excavations in the Valley of the Kings, an expedition from the University of Basel (Switzerland) led by S. Bickel and E. Paulin-Grothe discovered a limestone tile with a hole and fan-shaped lines diverging from it ( Fig. 2). The authors of the find suggested that it was a sundial.
Figure 2. Limestone tiles discovered during archaeological excavations in the Valley of the Kings 1
The sundial was a semicircular piece of local limestone. The front is relatively flat, while the back is very uneven. Dimensions are 15,5 cm high, 17,5 cm wide and a maximum depth of 3,6 cm. The upper right edge of the image is broken off. The dial is painted black on the front. It consists of a horizontal line, in the middle of which a hole about 10 mm deep and about 6 mm in diameter is drilled. A gnomon was attached here, which was not found. It could be a wooden, bronze or lead rod.
In the case of a vertical sundial, the gnomon could be installed either horizontally or obliquely along the axis of rotation of the Earth. The horizontal line on the tile marks the hour lines for 6 am (left) and 6 pm (right). The midday line (12 o'clock) runs almost at right angles (91° or 89°) to the horizontal line. Five more hour lines are marked between them, from the hole to the edge of the tile. A dot is drawn in each of the sectors. Behind the left half of the bounding line, you can see the first erased line. The beginning of the incomplete line is also visible on the first oblique on the right (17:00) [10].
1 http://aegyptologie.unibas.ch/forschung/projekte/university-of-basel-kings-valley-project/report-2013/ "Preliminary
Report on the Work carried out during the season 2013", fig. 5 (accessed on 01.08.2014)
Figure 3. Limestone tile from the Valley of the Kings, drawing with applied hour lines calculated for a vertical sundial with a horizontal gnomon. The noon line was chosen as the zero point. The lines on the limestone tile are shown in the drawing as black solid lines, the hour lines calculated for the equinox are shown as gray solid lines, for the summer solstice gray dotted lines, for the winter solstice - gray lines of dots [11].
In 2014, a study by S. Bickel and R. Gautschy was published, in which they note that the greatest coincidence with the lines on the tile is given by the calculated hour lines of the vertical sundial, oriented to the south, and an inclined gnomon oriented along the Earth's axis of rotation. At the same time, the researchers noted that a good match between the calculated lines and those drawn on the tile is observed only in the range from 7 to 12 noon.
The authors of the study were unable to explain the poor match between 12 and 17 o'clock and concluded that the lines on the tile are approximate and possibly belong to another type of clock - a vertical sundial with a horizontal gnomon. However, the hour lines calculated by them for this type of watch also poorly coincided with the lines on the tile (Fig. 3) [12].
The aim of the study presented in this article was to reconstruct and develop a sundial model that describes the found sundial from the Valley of the Kings with greater accuracy than the previously proposed model, as well as the reconstruction of methods for measuring time using vertical and L-shaped sundials in ancient Egypt.
## Vertical sundial from the Valley of the Kings
In the process of research, we analyzed the markings of the sundial from the Valley of the Kings. Having carried out calculations for horizontal and vertical sundials, we came to the conclusion that the markings of the found clocks are most similar to those of vertical sundial with an inclined gnomon, especially in the range from 6 to 12 hours. When studying the lines depicted on the limestone tile, one can notice the asymmetry of the sectors of the semicircle relative to the central line (the assumed noon line). If these lines on the tile are hourly, then the central line will be the 12 o'clock line, and the sectors symmetrical with respect to it, formed by the drawn lines, must have equal angular dimensions. As part of this study, in order to explain the asymmetry of the sectors, a hypothesis was put forward that hour lines could be plotted not only for whole hours, but also for time intervals greater or less than an hour by half an hour. The fact that half an hour could be measured and taken into account is indicated by dots drawn approximately in the middle of each sector between the lines on the tile (Fig. 2, 3). To test the hypothesis, hourly angles were calculated every half hour using the standard formula 1 for a vertical sundial with an inclined gnomon [13]:
$$\cdot t g H ),$$
Where - the hour angle of the Sun (for noon t= 12, H= 0º), t - the time (hours), H - the angle formed by the hour line with the noon line, / φ - the latitude (Fig. 4).
The results of calculating the values of hour angles in the range from 6 to 18 hours for the geographical latitude of the Valley of the Kings (Egypt) φ = 25°44' N are presented in tables 1 and 2. Swiss researchers measured and calculated hour angles relative to the six o'clock hour line. In the table, the hour angles are recalculated relative to the noon line.
Figure 4. Scheme illustrating the principle of operation of a vertical sundial with an inclined gnomon [14].
The average deviation of the calculated hourly angles from the measured ones in the range from 7 to 12 hours is ≈0,8º (Table 1). In the range from 12:00 to 17:00, the average deviation of the calculated hour angles from the measured ones is ≈5,7º (for whole hours), and ≈1º for half an hour. The value of the average deviation of the calculated hour angles from the measured ones, taking into account the shift by half an hour after noon, is close to the value of the average deviation before noon without a shift.
On the image of a tile from the Valley of the Kings, in accordance with the calculations, hour lines were applied corresponding to whole hours in the range from 6 to 12 hours and hour lines corresponding to half hours in the range from 12,5 to 16,5 hours (Fig. 5). When drawing lines on a photograph, the vertices of the calculated hour lines were aligned with the vertices of the corresponding lines on the photograph of the tile.
Table 1. Hour angles of a vertical sundial with an inclined gnomon before noon for latitude 25 44 N (relative to the noon line): 0 / H - the hour angle of the Sun, H / - the hour angle of the sundial calculated by formula 1, Hf - the measured hour angle on the tile from the Valley of the Kings [15], H // - the hour angle of the sundial, calculated for the equinoxes (recalculated relative to the noon line) [16], t - time.
| | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) |
|---------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|
| | 6,0 | 6,5 | 7,0 | 7,5 | 8,0 | 8,5 | 9,0 | 9,5 | 10,0 | 10,5 | 11,0 | 11,5 | 12,0 |
| H, (º) | -90,0 | -82,5 | -75,0 | -67,5 | -60,0 | -52,5 | -45,0 | -37,5 | -30,0 | -22,5 | -15,0 | -7,5 | 0,0 |
| H f, (º) | - | - | -73,0 | - | -57,0 | - | -43,0 | - | -30,0 | - | -14,0 | - | 0,0 |
| H, / (º) | -90,0 | -81,7 | -73,4 | -65,3 | -57,3 | -49,6 | -42,0 | -34,7 | -27,5 | -20,5 | -13,6 | -6,8 | 0,0 |
| H // , (º) | - | - | -73,5 | - | -57,5 | - | -42,0 | - | -27,5 | - | -13,5 | - | 0,0 |
| |H f - H| / | - | - | 0,4 | - | 0,3 | - | 1,0 | - | 2,5 | - | 0,4 | - | 0,0 |
| |H f - H // | | - | - | 0,5 | - | 0,5 | - | 1,0 | - | 2,5 | - | 0,5 | - | 0,0 |
Table 2. Hour angles of a vertical sundial with an inclined gnomon after noon for latitude 25 44 N (relative to the noon line): 0 / H - the hour angle of the Sun, H / - the hour angle of the sundial calculated by formula 1, Hf - the measured hour angle on the tile from the Valley of the Kings, H // - the hour angle of the sundial, calculated for the equinoxes (recalculated relative to the noon line), t - time.
| | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) | t, (hour) |
|---------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|-------------|
| | 12,0 | 12,5 | 13,0 | 13,5 | 14,0 | 14,5 | 15,0 | 15,5 | 16,0 | 16,5 | 17,0 | 17,5 | 18,0 |
| H, (º) | 0,0 | 7,5 | 15,0 | 22,5 | 30,0 | 37,5 | 45,0 | 52,5 | 60,0 | 67,5 | 75,0 | 82,5 | 90,0 |
| H f, (º) | 0,0 | 10,0 | 10,0 | 21,0 | 21,0 | 34,0 | 34,0 | 49,0 | 49,0 | 66,0 | 66,0 | - | - |
| H, / (º) | 0,0 | 6,8 | 13,6 | 20,5 | 27,5 | 34,7 | 42,0 | 49,6 | 57,3 | 65,3 | 73,4 | 81,7 | 90,0 |
| H // , (º) | 0,0 | - | 13,5 | - | 27,5 | - | 42,0 | - | 57,5 | - | 73,5 | - | - |
| |H f - H| / | 0,0 | 3,2 | - | 0,5 | - | 0,7 | - | 0,6 | - | 0,7 | - | - | - |
| |H f - H // | | 0,0 | - | 3,5 | - | 6,5 | - | 8 | - | 8,5 | - | 7,5 | - | - |
It should also be noted that the hour line at 12,5 o'clock is rather carelessly applied to the tile. It is possible that the time from 12 to 13,5 hours corresponded to the midday rest of the workers, next to whose stone dwelling this sundial was discovered. It seems that after the afternoon rest, the working time, just as before noon, was divided into whole hours. At the same time, the first working hour began at 13,5 hours, and each subsequent one - exactly one hour later. At 16,5 o'clock the working day most likely ended, since there are no hour lines or dot marks on the tile after 16,5 to 18 o'clock.
Figure 5. Limestone slab from the Valley of the Kings, photograph with hour lines applied to it, calculated hour lines before noon (black lines) and calculated hour lines with a shift of half an hour after noon (blue lines).
The time of noon has been singled out and associated with rest and eating in many traditions. So, for example, it is known that in Ancient Greece the working day began at dawn and lasted until noon, which marked the end of working hours (Anth. Pa1., X, 43 - Anthologia Palatina). Lunch time in Ancient Rome (Mart., IV, 8 - Marcus Valerius Martialis), fell on a period close to noon [17]. In many countries with a hot climate, a midday rest - a siesta - is still widespread. Perhaps the markings of the sundial found in the Valley of the Kings is one of the oldest evidence of the existence of this tradition as far back as the era of Ancient Egypt.
## L-shaped sundial
The image of a sundial in the form of an L-shaped bar from the tomb of Seti I is well known (Fig. 6).
Figure 6. Image of a L-shaped sundial from the tomb of Seti I at Abydos [18]
The L-shaped sundial made of green slate dating from the reign of Thutmose III is kept in the Berlin Egyptian Museum under accession number 19744, and a Faiyum sundial dating from 1000-600 BC, are stored under inventory number 19743 (Fig. 7) [19].
Figure 7. Sundial with a linear scale: a - L-shaped sundial with inventory number 19743 , 2 b - L-shaped sundial with inventory number 19744 . 3
Also quite well known in the scientific world is a fragment of a papyrus papyrus from Tanis of the Roman time, on which in the upper part of the drawing there are lines resembling the hour lines of a vertical clock, and under them, presumably, a fragment of an L-shaped sundial is depicted (Fig. 8a) [20]. The papyrus fragment with the upper part of the L-shaped sundial and the center of the fan-shaped divergent lines is unfortunately missing.
As part of this study, a reconstruction of a more complete image was proposed, which is illustrated in Figure 8b. At the bottom of the figure, in gray, an L-shaped sundial is depicted, as in Figure 7, without any additional crossbar or bar. In the upper part of the figure, at the point
2 http://members.aon.at/sundials/berlin-egypte.htm (accessed on 01.08.2014)
3 http://www.aegyptisches-museum-berlin-verein.de/c31.php (accessed on 01.08.2014)
where the fan-shaped lines converge, the place of attachment of the gnomon is assumed, and the fan-shaped lines are the hour lines of the vertical sundial, similar to the lines on the tiles from the Valley of the Kings.
a
b
Figure 8. Papyrus fragment, which supposedly depicts a vertical sundial (partial reconstruction by Flinders Petrie [21]): a - drawing of a papyrus fragment, b - reconstruction of the image of an L-shaped sundial (highlighted in gray) on the same papyrus fragment.
Based on this interpretation of the images, it can be assumed that a vertical sundial was used in conjunction with an L-shaped sundial. To test this assumption, we calculated the relative lengths of segments ai (ratios ai /a1 ) formed by the intersection of adjacent hour lines of a vertical sundial with an inclined gnomon in the range from 8 to 11 o'clock with a line perpendicular to the noon line, according to formula 2 (Fig. 9 ):
$$\Phi _ { \Psi \Psi } / \Phi _ { \Psi \Psi }$$
where Hi / - hour angle between the noon line (H0 =0) and the -th hour line, calculated by / i formula 1; - the number of the hour line, where =1 for the hour line at 11 o'clock, and =4 for i i i 8 o'clock; ai - the length of the segment bounded by ( -1)-th and -th hour lines, corresponding i i to the ith hour line.
The results of calculations according to formula 2 are presented in table 3.
Figure 9. Scheme illustrating the location of segments ai formed by the intersection of adjacent hour lines from 8 to 12 o'clock with a line perpendicular to the noon line, c - the distance from the place of attachment of the gnomon to the line perpendicular to the noon line.
Table 3. Calculated length of segments ai relative to a1 for a vertical sundial with an inclined gnomon for latitude φ =25°44' N Hi / : - angle between the noon line and the i -th hour line, calculated by formula 1, t i - time corresponding to i -th hour line, ai is the length of the segment bounded by ( -1)-th and -th hour lines, and corresponding to the -th hour line. i i i
| i | 1 | 2 | 3 | 4 |
|----------------|------|------|------|------|
| t i , ( hour ) | 11 | 10 | 9 | 8 |
| H i / , (º) | 13,6 | 27,5 | 42 | 57,3 |
| tg(H i / ) | 0,24 | 0,52 | 0,90 | 1,56 |
| a i /a 1 | 1,00 | 1,15 | 1,57 | 2,72 |
Similar calculations were carried out for the hour markings on the tiles found in the Valley of the Kings (Tab. 4).
Table 4. Calculated relative length of segments ai for sundials from the Valley of the Kings: Hi / = Hf i - measured angle between the noon line and the i -th hour line, t i - time corresponding to the -th hour line, i ai - segment length bounded by ( -1)-th and -th hour lines, i i and corresponding to the -th hour line. i
| i | 1 | 2 | 3 | 4 |
|----------------|------|------|------|------|
| t i , ( hour ) | 11 | 10 | 9 | 8 |
| H,(º) / | 14 | 30 | 43 | 57 |
| tg(H i / ) | 0,25 | 0,58 | 0,93 | 1,54 |
| a i /a 1 | 1,00 | 1,36 | 1,47 | 2,51 |
Taking into account the linear parameters of the L-shaped sundial from the Berlin Museum, the distances between hour marks were determined from photographs (Table 5), and the relative lengths of segments ai between adjacent hour marks were calculated (Table 6).
Table 5. Linear parameters of the L-shaped sundial: ai - distance between " i-1 " and " " i hour markers.
,
| Museum number of sundials | Distance from the beginning of the bar to the first mark, (cm) | a 1 , ( cm ) | a 2 , ( cm ) | a 3 , ( cm ) | a 4 , ( cm ) | Distance from the last mark to the edge of the bar, (cm) | Total length of the bar 4,5 , ( cm ) |
|-----------------------------|------------------------------------------------------------------|----------------|----------------|----------------|----------------|------------------------------------------------------------|----------------------------------------|
| №19743 | 3,2 | 2,5 | 3,8 | 5,1 | 6,3 | 9,4 | 30,3 |
| №19744 | 2,9 | 2,4 | 3,4 | 4,3 | 5,7 | 4,5 | 23,2 |
Table 6. Lengths of segments of a sundial: ai /a1 -relative length of the segment corresponding to the i -th hour line for a vertical sundial or relative length of the segment between adjacent marks on L-shaped sundial. - vertical sundial with an inclined gnomon for I latitude 25 44 0 / N (Valley of the Kings), II - vertical sundial found in the Valley of the Kings, III - L-shaped sundial № 19743, IV - L-shaped sundial № 19744.
| | a 1/ a 1 | a 2/ a 1 | a 3/ a 1 | a 4/ a 1 |
|-----|------------|------------|------------|------------|
| I | 1,00 | 1,15 | 1,57 | 2,72 |
| II | 1,00 | 1,36 | 1,47 | 2,51 |
| III | 1,00 | 1,52 | 2,04 | 2,52 |
| IV | 1,00 | 1,42 | 1,79 | 2,38 |
For comparison, we also calculated the absolute length of the first segment for a vertical sundial at the level of its lower edge using formula 3:
$$l _ { 1 } = c \cdot t g \left ( H _ { 1 } ^ { \prime } \right )$$
4 http://members.aon.at/sundials/berlin-egypte.htm#1.2 (accessed on 01.08.2014)
5 http://www.aegyptisches-museum-berlin-verein.de/c31.php (accessed on 01.08.2014)
where c=11,8 (cm) - the height of the gnomon attachment relative to the lower edge of the tile from the Valley of the Kings (the length of the noon line), H1 / is the angle of 11 o'clock (between the hour lines of 12 and 11 o'clock).
As a result of calculations according to formula 4, for a latitude of 25°44 / N , we obtain l 1 =2,86 (cm) for a vertical sundial with an inclined gnomon and l 1 =2,94 (cm) for marking a sundial from the Valley of the Kings. When taking into account the height of the L-shaped sundial bar, which, in the case of a clock from the Berlin Museum, is approximately equal to one inch - the ancient Egyptian unit of length one "djeba", equal to 1,875 (сm) [22], for vertical sundial we get l 1 =2,4 (cm) (as for sundial № 19744), and for sundial from the Valley of the Kings l 1 =2,5 (cm) (as for sundial № 19743).
All these results support the hypothesis of the combined use of the Egyptian L-shaped sundial and vertical sundial with an oblique gnomon. Figure 10 illustrates this combination with a sundial found in the Valley of the Kings and an L-shaped sundial from the Berlin Egyptian Museum.
Figure 10. Complex of vertical sundial with inclined gnomon, found in the Valley of the Kings, and L-shaped sundial № 19744.
The L-shaped sundial complements the vertical sundial by making it possible to read the inscriptions near the hour markers and interpret the readings of the vertical sundial accordingly, since there are no inscriptions on the vertical sundial. Perhaps such a complex use of two types of sundials was conceived in order to limit the circle of people able to determine the time during the day, including the start and end times of work during the day, by vertical sundials.
In the tomb of Seti I in Abydos, above the image of an L-shaped sundial, next to the marks, there are signatures in the form of Egyptian numbers: 3, 6, 9, 12 (Fig. 6). Until now, there is no satisfactory hypothesis to explain how a sundial with such markings measured time, and what these numbers actually meant.
In the course of this study, it was suggested that these numbers could be related to the relative lengths of the segments between adjacent marks bi in the case when the countdown starts from the noon line, and the first hour line corresponds to 12,5 o'clock, the second to 13,5 o'clock, etc. (Fig. 11). The calculation was carried out according to formula 4, similar to formula 2:
$$\tilde { \ } t n / \Delta$$
where Hi / - hour angle between the noon line and the i -th hour line, H1 / -hour angle between the noon line and the first hour line, i - the number of the hour line, where =1 for the i hour line at 12,5 o'clock, and =4 for 15,5 o'clock; i bi - the length of the segment bounded by ( i 1)-th and -th hour lines, corresponding to the i ith hour line.
Figure 11. Scheme illustrating the location of the segments bi formed by the intersection of adjacent hour lines from 12 to 15,5 o'clock with a line perpendicular to the noon line, c - the distance from the place of attachment of the gnomon to the line perpendicular to the noon line.
The results of calculations by formula 5 for a vertical sundial with an oblique gnomon at latitude 25 44 0 / N (Valley of the Kings, Egypt), in the time range from 12 to 15,5 hours are presented in Table 7.
The calculated relative lengths of the segments bi, when rounded to integers and increased by a factor of three, give the same series of numbers as in the fresco in the tomb of Seti I: 3, 6, 9, 12. Since these numbers have a common factor, it is logical to assume that the fresco depicted not the relative lengths of the segments, but their absolute values, for example, the lengths in ancient Egyptian inches (djeba). Thus, a series of numbers on a fresco in the tomb of Seti I also confirms the existence of an Egyptian sundial with hour markings shifted by half an hour after 12 noon.
Table 7. Calculated length of segments ai relative to a1 for a vertical sundial with an inclined gnomon for latitude φ =25°44' N Hi / : - angle between the noon line and the i -th hour line, calculated by formula 1, t i - time corresponding to i -th hour line, ai - the length of the segment bounded by ( -1)-th and -th hour lines, and corresponding to the -th hour line. i i i
| i | 1 | 2 | 3 | 4 |
|----------------|------|------|------|------|
| t i , ( hour ) | 12,5 | 13,5 | 14,5 | 15,5 |
| H,( / 0 ) | 6,8 | 20,5 | 34,7 | 49,6 |
| tg(H i / ) | 0,12 | 0,37 | 0,69 | 1,17 |
| b i | 1,0 | 2,1 | 2,7 | 4,0 |
| ≈b i | 1 | 2 | 3 | 4 |
| ≈ b i ×3 | 3 | 6 | 9 | 12 |
The proposed method of measuring time with the help of an L-shaped sundial does not contradict the instructions for their use, placed on a fresco in the tomb of Seti I under the drawing of a sundial, the translation of which was completed and published in part [23].
Figure 12. Complex of a vertical sundial with an inclined gnomon found in the Valley of the Kings and an L-shaped sundial with hour markings shifted by half hour after 12 noon. The dotted lines illustrate the directions of the gnomon's shadow corresponding to the hour markers on the L-shaped sundial.
Figure 12 illustrates the combined use of a vertical sundial (using the Valley of the Kings sundial as an example) and an L-shaped sundial with afternoon hour markers shifted by half an hour.
It seems that L-shaped sundials without hour offsets were used to measure time and determine the start and end of work in the morning, and sundials with a half-hour offset were used to determine lunch and working hours after 12 noon.
## Conclusion
Thus, as a result of this study, a new interpretation of the hourly markings of the ancient Egyptian sundial with an inclined gnomon, found in the Valley of the Kings, was proposed. It was concluded that the markings on the tile represent the hourly markings of a sundial with an oblique gnomon for the latitude of the Valley of the Kings, offset by half an hour after 12 noon. The half-hour shift could be associated with an hour and a half rest for workers after 12 noon a traditional siesta, typical for countries with a hot climate.
A similar marking with the allocation of one and a half hours in the afternoon has already been found on the sundial of the XIII - XII centuries BC [24]. Within the framework of our proposed model, the average error of marking a sundial from the Valley of the Kings is ≈1º or ≈4 minutes. Accuracy offered by S. Bickel and R. Gautschy models: for a vertical sundial with an inclined gnomon 3.9º or 15.6 minutes, and with a horizontal gnomon 6.6º or 42.8 minutes. Thus, the accuracy of the reconstructed model of the ancient Egyptian sundial proposed in this article, found in the Valley of the Kings, is noticeably higher than the accuracy of the models proposed in the article by S. Bickel and R. Gautschy, which testifies in favor of the truth of the model described in this article.
The article also presents a reconstruction of the technology of joint time measurement using an L-shaped sundial and a vertical sundial with an inclined gnomon. In the process of research, it was found that there were two types of L-shaped sundial: with hourly markings without offset before noon and with a shift of half an hour after noon. The L-shaped sundial of these two types supplemented the vertical sundial by allowing the inscriptions next to the hour lines to be read.
## References
- 1. Antoniadi E.M. L'astronomie eguptienne: Depuis les temps les plus recuiles jusqu'a la fin de 1epoque Alexandrine. P., 1934.
- 2. Kann G. Kratkaya istoriya chasovogo iskusstva [Brief history of watchmaking]. L., 1926.
- 3. Pipunyrov V.N. Istoriya chasov s drevnejshix vremen do nashix dnej [The history of watches from ancient times to the present day]. Moscow: Nauka, 1982, p. 21.
- 4. Breamsted J.H. The beginnings of time measurement and the origins of our calendar.In: Time and its mysteries . Ser. 1.N.Y.; L., 1936.
- 5. Bikerman E. Xronologiya drevnego mira. Blizhnij Vostok i antichnost'. [Chronology of the ancient world. Near East and antiquity]. - M., 1976 - P. 11.
- 6. Borchardt L. Agyptische Zeitmessung, 1920
- 7. Frankfort H. The Cenotaph of Seti I at Abydos (Egypt Exploration Society. Excavation Memoirs 39), London, 1933, vol.2, plate 33.
- 8. Borchardt L. Altägyptische Sonnenuhren, ZÄS, 1911, vol. 48, pp. 9-17.
- 9. Pipunyrov V.N. Istoriya chasov s drevnejshix vremen do nashix dnej [The history of watches from ancient times to the present day]. Moscow: Nauka, 1982, fig. 5.
- 10. Bickel S., Gautschy R. Eine ramessidische Sonnenuhr im Tal der Könige. Zeitschrift für Ägyptische Sprache und Altertumskunde 2014; 141(1): 3-14.
- 11. Bickel S., Gautschy R. Eine ramessidische Sonnenuhr im Tal der Könige. Zeitschrift für Ägyptische Sprache und Altertumskunde 2014; 141(1): 3-14, Abb. 6.
- 12. Bickel S., Gautschy R. Eine ramessidische Sonnenuhr im Tal der Könige. Zeitschrift für Ägyptische Sprache und Altertumskunde 2014; 141(1): 3-14.
- 13. Rohr RRJ. Sundials: History, Theory, and Practice. New York: Dover, 1996, p. 65.
- 14. Savoie D. Sundials design construction and use. Springer, 2009, fig. 7.2.
- 15. Bickel S., Gautschy R. Eine ramessidische Sonnenuhr im Tal der Könige. Zeitschrift für Ägyptische Sprache und Altertumskunde 2014; 141(1): 3-14, Abb. 1.
- 16. Bickel S., Gautschy R. Eine ramessidische Sonnenuhr im Tal der Könige. Zeitschrift für Ägyptische Sprache und Altertumskunde 2014; 141(1): 3-14, Abb. 1.
- 17. Ideler L. Lehrbuch der Chronologie. Publisher: A. Rücker, 1831, p. 260.
- 18. Frankfort H. The Cenotaph of Seti I at Abydos (Egypt Exploration Society. Excavation Memoirs 39), London, 1933, vol.2, plate 33.
- 19. Borchardt L. Altägyptische Sonnenuhren, ZÄS, 1911, vol. 48, pp. 9-17.
- 20. Symons S. Ancient Egyptian astronomy: timekeeping and cosmography in the new kingdom. University of Leicester, 1999, p. 133.
- 21. Griffith F. Ll. Flinders P. W. M., Brugsch H. Two Hieroglyphic Papyri from Tanis. London, Trübner & Co., 1889, pl. XV.
- 22. Clagett М. Ancient Egyptian Science: Ancient Egyptian mathematics. American Philosophical Society, 1999, pp. 7-8.
- 23. Symons S. Ancient Egyptian astronomy: timekeeping and cosmography in the new kingdom. University of Leicester, 1999, p. 130.
- 24. Vodolazhskaya L. Analemmatic and horizontal sundials of the Bronze Age (Northern Black Sea Coast). Archaeoastronomy and Ancient Technologies 2013, 1(1), pp. 68-88.
\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_
© This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).
| null |
[
"Larisa Vodolazhskaya"
] |
2014-08-04T10:43:08+00:00
|
2023-04-27T13:58:46+00:00
|
[
"physics.hist-ph"
] |
Reconstruction of Ancient Egyptian Sundials
|
The article presents the results of a study of the design features of
vertical and L-shaped ancient Egyptian sundials, developed their models, on the
basis of which the reconstruction of the sundial was carried out. An original
scheme of a simple method for a fairly accurate time measurement with their
help is also proposed. The model proposed by us, which describes a vertical
sundial, is a vertical sundial with an inclined gnomon that takes into account
the latitude of the area. The model proposed by us is characterized by marking
the hour lines from 6 to 12 o'clock every hour. From 12:00 a shift is made in
the marking of the hour lines for half an hour, and then the marking is
repeated every hour. As a consequence of the reconstruction of the vertical
sundial, we have developed and proposed a model that describes the design
features and operation of two types of L-shaped sundial. They had to work in
conjunction with an inclined gnomon, like a vertical sundial, or directly with
a vertical sundial. In this case, the L-shaped sundial could complement the
vertical sundial by providing the ability to read hour markers and interpret
the readings of the vertical sundial, as there were no inscriptions on the
vertical sundial. The article also provides a decoding of the inscription from
the tomb of Seti I, which has long intrigued researchers. It is proved that the
inscription contains the lengths of the intervals between adjacent marks of the
L-shaped sundial of the second type, where the first interval corresponds not
to one hour, but to half an hour.
|
1408.0988v1
|
## Distance partitions of extremal and largest known circulant graphs of degree 2 to 9
## R. R. Lewis
Department of Mathematics and Statistics The Open University Milton Keynes, UK [email protected]
5th August 2014
## Abstract
This paper considers the degree-diameter problem for extremal and largest known undirected circulant graphs of degree 2 to 9 of arbitrary diameter. As these graphs are vertex transitive it is possible to define their distance partition. The number of vertices in each level of the distance partition is shown to be related to an established upper bound for the order of Abelian Cayley graphs. Furthermore these graphs are all found to have odd girth which is maximal for their diameter. Therefore the type of each vertex in a level may be well-defined by the number of adjacent vertices in the preceding level. With this definition the number of vertices of each type in each level is also shown to be related to the same Abelian Cayley graph upper bound. Finally some implications are discussed for circulant graphs of higher degree.
## 1 Introduction
The degree-diameter problem is to identify extremal graphs, having the largest possible number of vertices for a given maximum degree and diameter. For undirected circulant graphs of given degree and arbitrary diameter, extremal graphs have been identified only for degree 2 to 5. For degree 2 and 3, the solutions are straightforward. For degree 4 Chen and Jia included a proof in their 1993 paper [1], and for degree 5 Dougherty and Faber presented a proof in 2004 [2]. For degree 6 to 9 families of largest known circulant graphs have been discovered which are conjectured to be extremal but have been proven so only for a limited range of diameters: degree 6 and 7 by Dougherty and Faber [2] and degree 8 and 9 by the author [4]. For degree 10 and above a lower bound has been established for all even degrees by Chen and Jia [1].
Adopting the terminology of Macbeth, Siagiova and Siran [5] we will use d for degree, k for diameter, and CC d,k ( ) for the order of an extremal undirected circulant graph of degree d and diameter k . We also use CJ d,k ( ) for Chen and Jia's lower bound, DF d,k ( ) for the order of Dougherty and Faber's graphs of degree 6 and 7, and L d, k ( ) for the order of the author's graphs of degree 8 and 9. Formulae, depending on the diameter, for the order of the extremal and largest known circulant graphs of degree 2 to 9 are shown in table 1.
Table 1: The order and number of isomorphism classes of the extremal and largest known circulant graphs of degree 2 to 9 of arbitrary diameter k .
| Degree, d | Status | Validity | Order | Isomorphism classes |
|-------------|---------------|----------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------|-----------------------|
| 2 | Extremal | k ≥ 1 | CC (2 , k ) = 2 k +1 | 1 |
| 3 | Extremal | k ≥ 1 | CC (3 , k ) = 4 k | 1 |
| 4 | Extremal | k ≥ 1 | CC (4 , k ) = 2 k 2 +2 k +1 | 1 |
| 5 | Extremal | k ≥ 2 | CC (5 , k ) = 4 k 2 | 1 |
| 6 | Largest known | k ≡ 0 (mod 3), k ≥ 3 k ≡ 1 (mod 3), k ≥ 1 k ≡ 2 (mod 3), k ≥ 2 | DF (6 , k ) = (32 k 3 +48 k 2 +54 k +27) / 27 DF (6 , k ) = (32 k 3 +48 k 2 +78 k +31) / 27 DF (6 , k ) = (32 k 3 +48 k 2 +54 k +11) / 27 | 2 1 2* |
| 7 | Largest known | k ≡ 0 (mod 3), k ≥ 3 k ≡ 1 (mod 3), k ≥ 4 k ≡ 2 (mod 3), k ≥ 5 | DF (7 , k ) = (64 k 3 +108 k ) / 27 DF (7 , k ) = (64 k 3 +60 k - 16) / 27 DF (7 , k ) = (64 k 3 +60 k +16) / 27 | 1 2 2 |
| 8 | Largest known | k ≡ 0 (mod 2), k ≥ 4 k ≡ 1 (mod 2), k ≥ 3 | L (8 , k ) = ( k 4 +2 k 3 +6 k 2 +4 k ) / 2 L (8 , k ) = ( k 4 +2 k 3 +6 k 2 +6 k +1) / 2 | 1 1 |
| 9 | Largest known | k ≡ 0 (mod 2), k ≥ 6 k ≡ 1 (mod 2), k ≥ 5 | L (9 , k ) = k 4 +3 k 2 +2 k L (9 , k ) = k 4 +3 k 2 | 1 2 |
* For k = 2 there are three isomorphism classes
An undirected circulant graph X ( Z n , C ) of order n may be defined as a Cayley graph whose vertices are the elements of the cyclic group Z n where two vertices i, j are connected by an arc ( i, j ) if and only if j -i is an element of C , an inverse-closed subset of Z n \ 0, called the connection set. In common with all Cayley graphs, circulant graphs are vertex transitive and regular, with the degree d of each vertex equal to the size of C . If n is odd then Z n \ 0 has no elements of order 2. Therefore C has even size, say d = 2 f , and is comprised of f complementary pairs of elements with one of each pair strictly between 0 and n/ 2. The set of f elements of C between 0 and n/ 2 is defined to be the generator set G for X . If n is even then Z n \ 0 has just one element of order 2, namely n/ 2. In this case C is comprised of f complementary pairs of elements, as for odd n , with or without the addition of the self-inverse element n/ 2. If C has odd size, so that d = 2 f +1, then the value of the self-inverse element n/ 2 is fixed by n . Therefore for a circulant graph of given order and degree, its connection set C is completely defined by specifying its generator set G . The size of the connection set is equal to the degree d of the graph, and the size of the generator set f is defined to be the dimension of the graph. In summary, undirected circulant graphs of odd degree d must have even order. They have dimension f = ( d -1) / 2. Graphs of even degree d may have odd or even order. They have dimension f = d/ 2.
## 2 Distance partitions of extremal and best circulant graphs
As Cayley graphs are vertex transitive it is possible to define their distance partition. If n l denotes the number of vertices in the l th level of the partition with respect to any reference vertex, then the distance partition profile of any vertex transitive graph is defined to be
the ( k + 1)-vector (1 , n 1 , n 2 , . . . , n k ) where k is the diameter of the graph. For extremal circulant graphs of one dimension we see easily that for degree 2 the distance partition profiles for increasing diameter k ≥ 1 are (1 2) (1 2 2) (1 , , , , , , 2 2 2) (1 2 2 2 2) , , , , , , , , . . . ; and for degree 3 are (1 , 3) (1 3 4) (1 , , , , , 3 4 4) (1 3 4 4 4) , , , , , , , , . . . . In both cases the size of each successive level is a constant (2 or 4 respectively). For extremal circulant graphs of two dimensions we find for degree 4 the sequence (1 , 4) (1 4 8) , , , , (1 4 8 12) (1 4 8 12 16) , , , , , , , , , etc, so that each successive level is 4 more than the previous. For degree 5, each new level does not immediately take its final size. From Level 2 onwards the size of each new level is 2 below its maximum value, which it reaches when the next level is added. See Table 8.
Table 2: Distance partition profiles for extremal circulant graphs of degree 5.
| Diameter | Order | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level |
|------------|-------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|
| k | CC (5 , k ) | 0 | 1 | 2 | 3 | 4 | 5 | 6 |
| 1 | 6 | 1 | 5 | | | | | |
| 2 | 16 | 1 | 5 | 10 | | | | |
| 3 | 36 | 1 | 5 | 12 | 18 | | | |
| 4 | 64 | 1 | 5 | 12 | 20 | 26 | | |
| 5 | 100 | 1 | 5 | 12 | 20 | 28 | 34 | |
| 6 | 144 | 1 | 5 | 12 | 20 | 28 | 36 | 42 |
From level 3 onwards the maximum size of each successive level is 8 more than the previous. So for both degree 3 and 4 the size of the levels increase at a constant rate.
For three dimensions the evolution of the distance partition profiles for increasing diameter k becomes more complicated. The profiles for the largest known graphs of degree 6 for diameter k ≤ 15 are shown in Table 3. For diameters with two isomorphism classes both have the same distance partition profile.
Table 3: Distance partition profiles for largest known circulant graphs of degree 6.
| Diameter | Order | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | | | | | | | | | |
|------------|-------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|-----|-----|-----|-----|-----|-----|-----|-----|-----|
| k | DF (6 , k ) | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 |
| 1 | 7 | 1 | 6 | | | | | | | | | | | | | | |
| 2 | 21 | 1 | 6 | 14 | | | | | | | | | | | | | |
| 3 | 55 | 1 | 6 | 18 | 30 | | | | | | | | | | | | |
| 4 | 117 | 1 | 6 | 18 | 38 | 54 | | | | | | | | | | | |
| 5 | 203 | 1 | 6 | 18 | 38 | 62 | 78 | | | | | | | | | | |
| 6 | 333 | 1 | 6 | 18 | 38 | 66 | 94 | 110 | | | | | | | | | |
| 7 | 515 | 1 | 6 | 18 | 38 | 66 | 102 | 134 | 150 | | | | | | | | |
| 8 | 737 | 1 | 6 | 18 | 38 | 66 | 102 | 142 | 174 | 190 | | | | | | | |
| 9 | 1027 | 1 | 6 | 18 | 38 | 66 | 102 | 146 | 190 | 222 | 238 | | | | | | |
| 10 | 1393 | 1 | 6 | 18 | 38 | 66 | 102 | 146 | 198 | 246 | 278 | 294 | | | | | |
| 11 | 1815 | 1 | 6 | 18 | 38 | 66 | 102 | 146 | 198 | 254 | 302 | 334 | 350 | | | | |
| 12 | 2329 | 1 | 6 | 18 | 38 | 66 | 102 | 146 | 198 | 258 | 318 | 366 | 398 | 414 | | | |
| 13 | 2943 | 1 | 6 | 18 | 38 | 66 | 102 | 146 | 198 | 258 | 326 | 390 | 438 | 470 | 486 | | |
| 14 | 3629 | 1 | 6 | 18 | 38 | 66 | 102 | 146 | 198 | 258 | 326 | 398 | 462 | 510 | 542 | 558 | |
| 15 | 4431 | 1 | 6 | 18 | 38 | 66 | 102 | 146 | 198 | 258 | 326 | 402 | 478 | 542 | 590 | 622 | 638 |
As for degree 5, the size of each new level is initially below its maximum value, but now the number of increments to reach its maximum is not fixed at 1 but increases with increasing diameter, so that the maximal zone, where the levels have reached their maximum value, covers about the first two thirds of the levels. Also the difference between the maximum
size of successive levels is no longer constant but increases linearly. Degree 7 is similar, with distance partition profile independent of isomorphism class and maximal zone covering two thirds of the levels.
For the largest known circulant graphs of dimension 4 the evolution of the distance partition profiles follows a similar structure but with certain differences. For both degree 8 and 9 the maximal zone covers about the first half of the levels, and the difference between the maximum size of successive levels increases as a quadratic. For degree 9 the two isomorphism classes for odd diameter share the same profile.
The use of distance partitions in the analysis of extremal graphs can be taken a stage further by considering an extension of the intersection array. We define the total intersection array of a vertex transitive graph to have the same format as a standard intersection array but where each element is the total number of adjacent vertices summed across all vertices within each level of the distance partition. So taking Godsil and Royle's example of the dodecahedron [3] which has distance partition profile (1 3 6 6 3 1) and standard , , , , , intersection array
$$\begin{pmatrix} - & 1 & 1 & 1 & 2 & 3 \\ 0 & 0 & 1 & 1 & 0 & 0 \\ 3 & 2 & 1 & 1 & 1 & - \end{pmatrix}, \\ \text{becomes}$$
its total intersection array becomes
$$\left ( \begin{array} { c c c c c } - & 3 & 6 & 6 & 6 & 3 \\ 0 & 0 & 6 & 6 & 0 & 0 \\ 3 & 6 & 6 & 6 & 3 & - \\ \end{array} \right ). \\ \right ) \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \ tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde{ \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde{ \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \delta \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \ tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e \, \dots \, \tilde { \ } e$$
With this definition the sum of the elements in each column of the total intersection array is equal to the corresponding element of the distance partition profile multiplied by the degree. The advantage for our analysis of circulant graphs is that the total intersection array is defined for all vertex transitive graphs whereas standard intersection arrays are only defined for distance regular graphs. Total intersection arrays provide a useful view on the structure of the graphs. The first non-zero element in the middle row determines the odd girth of the graph. They can also distinguish non-isomorphic graphs of common degree, diameter and order that might have the same distance partition profile. An example is given by the four isomorphism classses of extremal circulant graphs of degree 9 and diameter 3. These are easily proven to be distinct by determining their total intersection arrays, which are all different; see Table 4.
Reverting to the discussion on distance partition profiles, some obvious questions arise. What is the structure behind the maximum size of each level? What is the logic behind the evolution of the size of each level until it reaches its maximum? What determines the number of levels in the maximal zone?
Table 4: Extremal graphs of each of the four isomorphism classes for d = 9 and k = 3.
| Diameter, k | Order, n | Generator set (excluding n/ 2) | Distance partition profile | Total intersection array | Total intersection array | Total intersection array | Total intersection array | Total intersection array | Total intersection array |
|---------------|------------|----------------------------------|------------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|
| 3 | 130 | { 1 , 8 , 14 , 47 } | (1 , 9 , 40 , 80) | 0 | - 9 | 9 0 72 | 72 44 244 | 244 476 - | |
| 3 | 130 | { 1 , 8 , 20 , 35 } | (1 , 9 , 40 , 80) | 9 | - 0 | 9 0 72 | 72 46 242 | 242 478 - | |
| 3 | 130 | { 1 , 26 , 49 , 61 } | (1 , 9 , 40 , 80) | 9 | - 0 | 9 0 72 | 72 2 286 | 286 434 - | |
| 3 | 130 | { 2 , 8 , 13 , 32 } | (1 , 9 , 40 , 80) | 9 | - 0 | 9 0 72 | 72 54 234 | 234 486 - | |
## 3 Distance partitions and maximal levels
First we consider the maximum size of each level. A useful upper bound for the order of any Abelian Cayley graph is given by Doughterty and Faber [2]. For a graph of degree d and diameter k the upper bound M AC ( d, k ) is defined as follows:
$$M _ { A C } ( d, k ) = \begin{cases} \begin{array} { c c } S ( f, k ) & \text{for even $d$, where $f=d/2$} \\ S ( f, k ) + S ( f, k - 1 ) & \text{for odd $d$, where $f=(d-1)/2$} \end{array} \end{cases} \end{cases}$$
where f is the dimension of the graph and S f, k ( ) = ∑ f i =0 2 i ( )( ) f i k i .
If we consider the implied upper bound for the size of level l determined by the first order difference of the upper bound M AC ( d, l ) for increasing l , so that LM AC ( d, l ) = M AC ( d, l ) -M AC ( d, l -1) for degree d ≥ 2 and level l ≥ 2, then the result is shown in Table 5.
Now these values are precisely the maximum size of each corresponding level of the distance partitions. So over the range of degrees and diameters considered, we can see that for each extremal and largest known circulant graph each distance partition level l is filled to the maximum determined by the upper bound M AC ( d, l ). This suggests that not only the extremal graphs but also the graphs of degree 6 to 9 and the upper bound M AC ( d, l ) are optimal in some sense. Formulae for LM AC ( d, l ) as a function of l are presented for degree 2 to 9 in Table 6. For the graphs of dimension 1 and 2 all the levels are maximal, with the exception of the last level for degree 5. For dimension 3 we see that the proportion of maximal levels is about 2 / 3 and for dimension 4 about 1 / 2. For f ≥ 2 we note that these proportions are represented by the expression 2 /f . The exact number of maximal levels in each case is also shown in Table 6. These values are proved for degree 2, 4, 6 and 8 in the following six theorems. The proofs for odd degrees are similar.
Table 5: LM AC ( d, l ), first order difference of the upper bound, M AC ( d, l ).
| Dimension | Degree | Distance, l , from the reference vertex at level 0 | Distance, l , from the reference vertex at level 0 | Distance, l , from the reference vertex at level 0 | Distance, l , from the reference vertex at level 0 | Distance, l , from the reference vertex at level 0 | Distance, l , from the reference vertex at level 0 | Distance, l , from the reference vertex at level 0 | Distance, l , from the reference vertex at level 0 | Distance, l , from the reference vertex at level 0 |
|-------------|----------|------------------------------------------------------|------------------------------------------------------|------------------------------------------------------|------------------------------------------------------|------------------------------------------------------|------------------------------------------------------|------------------------------------------------------|------------------------------------------------------|------------------------------------------------------|
| f | d | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
| 1 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 |
| | 3 | 4 | 4 | 4 | 4 | 4 | 4 | 4 | 4 | 4 |
| 2 | 4 | 8 | 12 | 16 | 20 | 24 | 28 | 32 | 36 | 40 |
| | 5 | 12 | 20 | 28 | 36 | 44 | 52 | 60 | 68 | 76 |
| 3 | 6 | 18 | 38 | 66 | 102 | 146 | 198 | 258 | 326 | 402 |
| | 7 | 24 | 56 | 104 | 168 | 248 | 344 | 456 | 584 | 728 |
| 4 | 8 | 32 | 88 | 192 | 360 | 608 | 952 | 1408 | 1992 | 2720 |
| | 9 | 40 | 120 | 280 | 552 | 968 | 1560 | 2360 | 3400 | 4712 |
Table 6: Maximal distance partition levels of extremal and largest known circulant graphs of degree 2 to 9 and diameter k : size of each maximal level, LM AC ( d, l ), and the position of the last maximal level.
| Dimension, f | Degree, d | LM AC ( d, 1) | LM AC ( d, l ) , l ≥ 2 | Last maximal level |
|----------------|-------------|-----------------|---------------------------------------------------|-------------------------------------------------------------------|
| 1 | 2 3 | 2 3 | 2 4 | k k |
| 2 | 4 5 | 4 5 | 4 l 8 l - 4 | k k - 1 |
| 3 | 6 7 | 6 7 | 4 l 2 +2 8 l 2 - 8 l +8 | /floorleft (2 k +1) / 3 /floorright /floorleft 2 k/ 3 /floorright |
| 4 | 8 9 | 8 9 | (8 l 3 +16 l ) / 3 (16 l 3 - 24 l 2 +56 l - 24) / | /floorleft ( k +1) / 2 /floorright /floorleft k/ 2 /floorright |
Theorem 1. For extremal degree 2 circulant graphs of arbitrary diameter k , every distance partition level l is maximal in the sense determined by the upper bound M AC (2 , l ) .
Proof. For an extremal circulant graph of degree 2 and arbitrary diameter k , the order is CC (2 , k ) = 2 k +1 and a generator set is { g 1 } where g 1 = 1. The graph is a cycle graph, and it is immediately evident that each level contains exactly 2 vertices and is therefore maximal.
Theorem 2. For extremal degree 4 circulant graphs of arbitrary diameter k , every distance partition level l is maximal in the sense determined by the upper bound M AC (4 , l ) .
Proof. For an extremal circulant graph of degree 4 and arbitrary diameter k , the order is CC (4 , k ) = n = 2 k 2 +2 k +1 and a generator set is { g , g 1 2 } where g 1 = 1 and g 2 = 2 k +1. Let v a and v b be vertices in level l ≤ k , relative to an arbitrary root vertex. As cyclic groups are Abelian the generating elements for each vertex can be taken in any order. Therefore v a ≡ a g 1 1 + a g 2 2 (mod n ) and v b ≡ b g 1 1 + b 2 g 2 (mod n ) where | a 1 | + | a 2 | = | b 1 | + | b 2 | = l for some a , a 1 2 , b 1 , b 2 ∈ Z . The level is maximal if and only if a 1 = b 1 and a 2 = b 2 whenever v a = v b .
Suppose that v a = v b . Then ( a 1 -b 1 ) g 1 + ( a 2 -b 2 ) g 2 ≡ 0 (mod n ). Now | ( a 1 -b 1 ) g 1 + ( a 2 -b 2 ) g 2 | ≤ 2 kg 2 = 4 k 2 + 2 k < 2 n . Thus ( a 1 -b 1 ) g 1 + ( a 2 -b 2 ) g 2 = 0 or ± n . If
( a 1 -b 1 ) g 1 +( a 2 -b 2 ) g 2 = 0 then a 1 -b 1 ≡ 0 (mod 2 k +1). But | a 1 | , | b 1 | ≤ k . Hence a 1 = b 1 and so a 2 = b 2 . Therefore in this case level l is maximal. If ( a 1 -b 1 ) g 1 +( a 2 -b 2 ) g 2 = ± n , then without loss of generality assume ( a 1 -b 1 ) g 1 +( a 2 -b 2 ) g 2 = n . Noting that n ≡ k +1 (mod 2 k + 1) we have a 1 -b 1 ≡ k + 1 (mod 2 k + 1), so that a 1 -b 1 = -k or k + 1. If a 1 -b 1 = -k then ( a 2 -b 2 ) g 2 = n + k = 2 k 2 + 3 k + 1 so that a 2 -b 2 = k + 1. If a 1 -b 1 = k +1 then ( a 2 -b 2 ) g 2 = n -k -1 = 2 k 2 + k so that a 2 -b 2 = k . In either case | a 1 -b 1 | + | a 2 -b 2 | = 2 k +1 contrary to the premise. Hence we must have a 1 = b 1 and a 2 = b 2 , and the level is maximal.
Theorem 3. For largest known degree 6 circulant graphs of arbitrary diameter k , every distance partition level l is maximal for l ≤ /floorleft (2 k +1) 3 / /floorright in the sense determined by the upper bound M AC (6 , l ) .
Proof. We give the proof just for the case k ≡ 0 (mod 3), k ≥ 3, for a graph of isomorphism class 1. The proofs for isomorphism class 2 and for cases k ≡ 1 and k ≡ 2 (mod 3) are similar. So let k = 3 m for m ≥ 1. For a best circulant graph of degree 6 and arbitrary diameter 3 m , the order is DF (6 , k ) = n = 32 m 3 + 16 m 2 + 6 m + 1 and a generator set for isomorphism class 1 is { g , g 1 2 , g 3 } where g 1 = 1, g 2 = 4 m + 1 and g 3 = 16 m 2 +4 m + 1. Let v a and v b be vertices in level l ≤ 2 m , relative to an arbitrary root vertex. Then v a ≡ a g 1 1 + a g 2 2 + a g 3 3 (mod n ) and v b ≡ b 1 g 1 + b 2 g 2 + b 3 g 3 (mod n ) where | a 1 | + | a 2 | + | a 3 | = | b 1 | + | b 2 | + | a 3 | = l ≤ 2 m for some a , a 1 2 , a 3 , b 1 , b 2 , b 3 ∈ Z . Also let w a = a g 1 1 + a g 2 2 + a g 3 3 and w b = b 1 g 1 + b 2 g 2 + b 3 g 3 . The level is maximal if and only if a 1 = b 1 , a 2 = b 2 and a 3 = b 3 whenever v a = v b .
Suppose that v a = v b . Then ( a 1 -b 1 ) g 1 + ( a 2 -b 2 ) g 2 + ( a 3 -b 3 ) g 3 ≡ 0 (mod n ). Now lg 3 ≤ 2 mg 3 = 32 m 3 +8 m 2 +2 m<n . So | w a -w b | < n 2 , and w a -w b = 0 or ± n .
First we consider the case w a -w b = ± n . Without loss of generality we may assume w a -w b = n , so that ( a 1 -b 1 ) g 1 +( a 2 -b 2 ) g 2 +( a 3 -b 3 ) g 2 = n . Now n = 32 m 3 +16 m 2 + 6 m +1 = 2 mg 3 +2 mg 2 +(2 m +1) g 1 . This is equivalent to 6 m +1 edges and there is no construction with fewer edges. However w a and w b are each the sum of at most 2 m edges, or 4 m in total. Hence there exists no pair w a and w b such that w a -w b = n . Therefore this case admits no solution.
/negationslash
/negationslash
Now we consider the case w a -w b = 0. Here ( a 1 -b 1 ) g 1 +( a 2 -b 2 ) g 2 +( a 3 -b 3 ) g 2 = 0. Suppose for a contradiction that a 3 = b 3 . Then | a 3 -b 3 | ≥ 1, so that | ( a 1 -b 1 ) g 1 +( a 2 -b 2 ) g 2 | ≥ g 3 = 16 m 2 + 4 m + 1. But | ( a 1 -b 1 ) + ( a 2 -b 2 ) | ≤ | a 1 -b 1 | + | a 2 -b 2 | ≤ 4 m , so that | ( a 1 -b 1 ) g 1 +( a 2 -b 2 ) g 2 | ≤ 4 mg 2 = 16 m 2 +4 m . This contradiction proves that a 3 = b 3 . Now suppose that a 2 = b 2 . Then | a 2 -b 2 | ≥ 1, so that | a 1 -b 1 | g 1 ≥ g 2 = 4 m +1. But | a 1 -b 1 | g 1 ≤ 4 m . Hence a 2 = b 2 and thus also a 1 = b 1 . This proves that if v a = v b then we must have a 1 = b 1 , a 2 = b 2 and a 3 = b 3 . Hence the level is maximal.
Theorem 4. For largest known degree 6 circulant graphs of arbitrary diameter k , every distance partition level l is submaximal for l > /floorleft (2 k + 1) / 3 /floorright in the sense determined by the upper bound M AC (6 , l ) .
Proof. We give the proof just for the case k ≡ 0 (mod 3), k ≥ 3, for a graph of isomorphism class 1. The proofs for isomorphism class 2 and for cases k ≡ 1 and k ≡ 2 (mod 3) are similar. So let k = 3 m for m ≥ 1. For a best circulant graph of degree 6 and
arbitrary diameter 3 m , the order is DF (6 , k ) = n = 32 m 3 + 16 m 2 + 6 m + 1 and a generator set for isomorphism class 1 is { g , g 1 2 , g 3 } where g 1 = 1, g 2 = 4 m + 1 and g 3 = 16 m 2 + 4 m +1. Let v a and v b be vertices in level 2 m +1, relative to an arbitrary root vertex. Then v a ≡ a g 1 1 + a g 2 2 + a g 3 3 (mod n ) and v b ≡ b 1 g 1 + b 2 g 2 + b 3 g 3 (mod n ) where | a 1 | + | a 2 | + | a 3 | = | b 1 | + | b 2 | + | a 3 | = 2 m + 1 for some a , a 1 2 , a 3 , b 1 , b 2 , b 3 ∈ Z . The level is maximal if and only if a 1 = b 1 , a 2 = b 2 and a 3 = b 3 whenever v a = v b . Now consider the case where a 1 = 0 , a 2 = 2 m +1 , a 3 = 0 , b 1 = -1 , b 2 = -(2 m -1) , g 3 = 1. Then v a = (2 m +1)(4 m +1) = 8 m 2 +6 m +1 and v b = - -1 (2 m -1)(4 m +1)+(16 m 2 +4 m +1) = 8 m 2 +6 m +1. Thus v a = v b and so level 2 m +1 is submaximal. If a level is submaximal then the next level must necessarily also be submaximal. Therefore every level l ≥ 2 m +1 is submaximal.
Theorem 5. For largest known degree 8 circulant graphs of arbitrary diameter k , every distance partition level l is maximal for l ≤ /floorleft ( k + 1) / 2 /floorright in the sense determined by the upper bound M AC (8 , l ) .
Proof. We give the proof just for the case k ≡ 0 (mod 2), k ≥ 4. The proof for the case k ≡ 1 (mod 2) is similar. So let k = 2 m for m ≥ 2. For a largestknown circulant graph of degree 8 and arbitrary diameter 2 m , the order is L (8 , k ) = n = 8 m 4 +8 m 3 +12 m 2 +4 m and a generator set is { g , g 1 2 , g 3 , g 4 } where g 1 = 1, g 2 = 4 m 3 + 4 m 2 + 6 m + 1, g 3 = 4 m 4 + 4 m 2 -4 m and g 4 = 4 m 4 + 4 m 2 -2 m . Let v a and v b be vertices in level l ≤ m , relative to an arbitrary root vertex. Then v a ≡ ∑ 4 i =1 a g i i (mod n ) and v b ≡ ∑ 4 i =1 b g i i (mod n ) where ∑ 4 i =1 | a i | = ∑ 4 i =1 | b i | = l ≤ m for some a , b i i ∈ Z . The level is maximal if and only if a i = b i for i = 1 2 3 4 whenever , , , v a = v b .
Suppose that v a = v b . Then ∑ 4 i =1 ( a i -b i ) g i = cn for some c ∈ Z . We note that n ≡ g 3 ≡ g 4 ≡ 0 (mod 2 m ) and g 1 ≡ g 2 ≡ 1 (mod 2 m ). Hence ( a 1 -b 1 )+( a 2 -b 2 ) ≡ 0 (mod 2 m ). First suppose | ( a 1 -b 1 ) + ( a 2 -b 2 ) | = 2 m . Then l = m, a | 1 + a 2 | = m, b | 1 + b 2 | = m , and a 3 = a 4 = b 3 = b 4 = 0. Without loss of generality suppose that a 1 + a 2 = -m and b 1 + b 2 = m . Then v a -v b ≡ ( a 1 -b 1 ) +(( m -a 1 ) -( m -b 1 ))(4 m 3 +4 m 2 +6 m +1) ≡ ( b 1 -a 1 )(4 m 3 +4 m 2 +6 m ) (mod n ). But ( b 1 -a 1 )(4 m 3 +4 m 2 +6 m ) ≤ 2 m (4 m 3 +4 m 2 +6 m < n ) . Thus v a -v b = ( b 1 -a 1 )(4 m 3 +4 m 2 +6 m ) = 0. Hence a 1 = b 1 . So a 2 = b 2 and the level is maximal.
Now suppose that ( a 1 -b 1 )+( a 2 -b 2 ) = 0, so that ( a 1 -b 1 )( g 1 -g 2 )+( a 3 -b 3 ) g 3 +( a 4 -b 4 ) g 4 = cn . Note that ∑ 4 i =1 | a i -b i | ≤ 2 m , and let a i -b i = x m i + y i for x , y i i ∈ Z for i = 1 3 4. , , Then we have -( x m 1 + y 1 )(4 m 3 +4 m 2 +6 m ) + ( x m 3 + y 3 )(4 m 4 +4 m 2 -4 m ) + ( x m 4 + y 4 )(4 m 4 +4 m 2 -2 m ) = c (8 m 4 +8 m 3 +12 m 2 +4 m ). Equating coefficients of powers of m m ; 5 : 4 x 3 +4 x 4 = 0. m 4 : -4 x 1 +4 y 3 +4 y 4 = 8 . c m 3 : -4 y 1 -4 x 1 +4 x 3 +4 x 4 = 8 . c m 2 : -4 y 1 -6 x 1 +4 y 3 -4 x 3 +4 y 4 -2 x 4 = 12 . c m : -6 y 1 -4 y 3 -2 y 4 = 4 . c
As x 4 = -x 3 and ∑ 4 i =1 | a i -b i | ≤ 2 m we must have x 3 = 0 or x 3 = 1. If x 3 = 0 then x 4 = 0 , x 1 = -2 c, y 1 = 0 , y 3 = -2 c and y 4 = 2 . c Thus ∑ 4 i =1 | a i -b i | = 2 | cm | + | 2 cm | + 2 | c | + 2 | c | = 4 | c | ( m + 1) ≤ 2 m . So c = 0 and a i = b i for i = 1 2 3 4. , , , If x 3 = 1 then x 4 = -1 , x 1 = 1 -2 c, y 1 = -1 , y 3 = 2 -2 c and y 4 = 2 c -1. Thus ∑ 4 i =1 | a i -b i | = (1 | -2 ) c m - | 1 + (1 | -2 ) c m - | 1 + | m +2 -2 c | + | -m +2 c - | 1 . If c = 0, then ∑ 4 i =1 | a i -b i | = 4 m + 1. If c = 1, then ∑ 4 i =1 4 | a i -b i | = 4 m . For any other value of c , ∑ 4 i =1 | a i -b i | ≥ 8 m . However we have that ∑ i =1 | a i -b i | ≤ 2 m . Hence there is no solution with x 3 = 1. This completes the proof that level l is maximal for any l ≤ m .
Theorem 6. For largest known degree 8 circulant graphs of arbitrary diameter k , every distance partition level l is submaximal for l > /floorleft ( k +1) 2 / /floorright in the sense determined by the upper bound M AC (8 , l ) .
Proof. We give the proof just for the case k ≡ 0 (mod 2), k ≥ 4. The proof for the case k ≡ 1 (mod 2) is similar. So let k = 2 m for m ≥ 2. For a best circulant graph of degree 8 and arbitrary diameter 2 m , the order is L (8 , k ) = n = 8 m 4 +8 m 3 +12 m 2 +4 m and a generator set is { g , g 1 2 , g 3 , g 4 } where g 1 = 1, g 2 = 4 m 3 +4 m 2 +6 m +1, g 3 = 4 m 4 +4 m 2 -4 m and g 4 = 4 m 4 +4 m 2 -2 m . Let v a and v b be vertices in level m +1, relative to an arbitrary root vertex. Then v a ≡ ∑ 4 i =1 a g i i (mod n ) and v b ≡ ∑ 4 i =1 b g i i (mod n ) where ∑ 4 i =1 | a i | = ∑ 4 i =1 | b i | = m + 1 for some a , b i i ∈ Z . The level is maximal if and only if a i = b i for i = 1 2 3 4 whenever , , , v a = v b . Now consider the case where a 1 = m,a 2 = 0 , a 3 = 1 , a 4 = 0 , b 1 = -m,b 2 = 0 , g 3 = 0 , b 4 = 1. Then v a = m +(4 m 3 +4 m 2 -4 m ) = 4 m 3 +4 m 2 -3 m and v b = -m + (4 m 3 + 4 m 2 -2 m ) = 4 m 3 + 4 m 2 -3 m . Thus v a = v b and so level m +1 is submaximal. If a level is submaximal then the next level must necessarily also be submaximal. Therefore every level l ≥ m +1 is submaximal.
## 4 Submaximal levels
Now we consider the evolution of the size of the submaximal levels in a largest known graph's distance partition. Clearly the difference between the order of the graph, DF d,k ( ) or L d, k ( ), and the corresponding upper bound M AC ( d, k ) will be equal to the sum for all levels of the shortfall in each level relative to its maximum value.
So for a graph of degree d and diameter k we define the level defect D d,k,l ( ) for each level l in its distance partition profile to be the difference between its actual size s l and the maximal size LM AC ( d, l ): D d,k,l ( ) = LM AC ( d, l ) -s l . Then for largest known graphs:
$$\sum _ { l = 0 } ^ { k } D ( d, k, l ) = \begin{cases} M _ { A C } ( d, k ) - D F ( d, k ) & \text{for $d=6,7$} \\ M _ { A C } ( d, k ) - L ( d, k ) & \text{for $d=8,9$.} \end{cases}$$
For example for degree d = 6
$$\sum _ { l = 0 } ^ { k } D ( 6, k, l ) = \begin{cases} ( 4 k ^ { 3 } + 6 k ^ { 2 } + 1 8 k ) / 2 7 & \text{ for } k \equiv 0 \pmod { 3 } \\ ( 4 k ^ { 3 } + 6 k ^ { 2 } - 6 k - 4 ) / 2 7 & \text{ for } k \equiv 1 \pmod { 3 } \\ ( 4 k ^ { 3 } + 6 k ^ { 2 } + 1 8 k + 1 6 ) / 2 7 & \text{ for } k \equiv 2 \pmod { 3 }. \end{cases} \\ \text{Considering degree } 6, \text{ for diameter } k \leq 15, \text{ the submaximal level defects are shown in}$$
Considering degree 6, for diameter k ≤ 15, the submaximal level defects are shown in Table 7.
Reading the columns of the table upward from the zero elements (where the level is maximal), there is one sequence of submaximal level defects for even levels: 4 12 36 68 108 164 , , , , , , 228 , . . . , and one for odd levels: 8 24 48 88 136 192 264 , , , , , , , . . . . The reason there are two different alternating sequences depends on the fact that for every increase of 3 in the diameter an extra two levels become maximal. The odd levels are related to an increase of 2 in the diameter and the even levels to an increase of 1. The second order differences of both theses sequences display the same repeating cycle of length three: 8 8 16 , , , . . . .
Table 7: Submaximal level defects, D (6 , k, l ), of the distance partition profiles for largest known graphs of degree 6
| DiameterOrder | DiameterOrder | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l | Distance partition level, l |
|-----------------|-----------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|-------------------------------|
| k | n | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 |
| 3 | 55 | 0 | 0 | 0 | 8 | | | | | | | | | | | | |
| 4 | 117 | 0 | 0 | 0 | 0 | 12 | | | | | | | | | | | |
| 5 | 203 | 0 | 0 | 0 | 0 | 4 | 24 | | | | | | | | | | |
| 6 | 333 | 0 | 0 | 0 | 0 | 0 | 8 | 36 | | | | | | | | | |
| 7 | 515 | 0 | 0 | 0 | 0 | 0 | 0 | 12 | 48 | | | | | | | | |
| 8 | 737 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 24 | 68 | | | | | | | |
| 9 | 1027 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 8 | 36 | 88 | | | | | | |
| 10 | 1393 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 12 | 48 | 108 | | | | | |
| 11 | 1815 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 24 | 68 | 136 | | | | |
| 12 | 2329 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 8 | 36 | 88 | 164 | | | |
| 13 | 2943 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 12 | 48 | 108 | 192 | | |
| 14 | 3629 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 24 | 68 | 136 | 228 | |
| 15 | 4431 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 8 | 36 | 88 | 164 | 264 |
Taken in reverse order these sequences define the evolution of a new level as the diameter is increased. For each level l the level defect decreases in a fixed pattern depending on the parity of the level until the level reaches its maximal value LM AC (6 , l ) as defined by the upper bound M AC (6 , l ). Reading along each row of Table 7, these level evolutions translate into the submaximal level defects of the distance partition profile of a graph of the corresponding diameter. For the same reason that every increase of 3 in the diameter is paired with an extra two levels becoming maximal, we find three different submaximal level defect sequences, one for each value of the diameter k (mod 3):
$$\begin{cases} 8, 3 6, 8 8, 1 6 4, 2 6 4, \dots & \text{ for } k \equiv 0 \pmod { 3 } \\ 1 2, 4 8, 1 0 8, 1 9 2, 3 0 0, \dots & \text{ for } k \equiv 1 \pmod { 3 } \\ 4, 2 4, 6 8, 1 3 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 }. \end{cases} \\ \text{order differences of all three sequences have the same constant} \, \begin{cases} 8, 3 6, 8 8, 1 6 4, 2 6 4, \dots & \text{ for } k \equiv 0 \pmod { 3 } \\ 1 2, 4 8, 1 0 8, 1 9 2, 3 0 0, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 3, 4 8, 1 0 8, 1 9 2, 3 0 0, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 4, 2 4, 6 8, 1 3 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 5, 4 8, 1 0 8, 1 9 2, 3 0 0, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 6, 2 4, 6 8, 1 3 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 7, 2 4, 6 8, 1 3 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 8, 2 4, 6 8, 1 3 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, 0 5, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 1 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 3 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 3 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \ dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text { for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \equiv 2 \pmod { 3 } \\ 1 9, 2 4, 6 8, 1 3 2 6, 2 2 8, \dots & \text{ for } k \end{cases}$$
The second order differences of all three sequences have the same constant value of 24. The level defects for degree 7 follow a similar form in having two alternating sequences along levels, both with second order differences displaying a repeating cycle of length three, 16 16 32 , , , . . . ; and three along the profiles for each diameter, all with a constant second order difference of 48.
For degree 8 there is only one sequence of submaximal level defects valid for all levels: 8 24 64 , , , 128 232 376 576 832 1160 1560 , , , , , , , . . . , with the exception that the final value in any of the sequences is always augmented by 1. The reason they are all the same, as distinct from the two versions for dimension 3, stems from the fact that for every increase of 2 in the diameter a single extra level becomes maximal. Therefore the context is similar for each level. The third order differences of this sequence form a repeated cycle of length two, 0 16 , , . . . , except that the final level in any sequence is 1 higher. For the same reason, reading along the profile for each diameter we find two different submaximal level defect
sequences (instead of three as for dimension 3):
$$\begin{cases} 8, 6 4, 2 3 2, 5 7 6, 1 1 6 0, 2 0 4 8 \dots & \text{ for } k \equiv 0 \pmod { 2 } \\ 2 4, 1 2 8, 3 7 6, 8 3 2, 1 5 6 0, 2 6 2 4 \dots & \text{ for } k \equiv 1 \pmod { 2 }. \end{cases}$$
both with a constant third order difference of 64. Again the value of the final level is increased by 1. The level defects for degree 9 follow a similar form in having a single sequence along levels with third order difference displaying a repeating cycle of length two, 0 32 , , . . . ; and two along the profiles for each diameter, both with a constant third order difference of 128.
The sequential second or third order differences between the size of maximal levels and submaximal level defects for degree 6 to 9 is summarised in Table 8, along with a parametrisation in terms of the dimension f . For each degree the relevant order difference for submaximal level defects is greater than the size of the maximal levels by a multiple of the dimension, indicating a relation between the sequence of submaximal level defects and the Abelian Cayley upper bound function M AC ( d, k ).
Table 8: Sequential differences of size of maximal levels and submaximal level defects for largest known graphs of degree 6 to 9, also parametrised in terms of the dimension f .
| Dimension, f | Degree, d | Order of difference | Difference for maximal levels | Difference for submaximal level defects |
|----------------|-------------|-----------------------|---------------------------------|-------------------------------------------|
| 3 | 6 | 2 | 8 | 24 |
| 4 | 8 9 | 3 3 | 16 32 | 64 128 |
| f | 2 f 2 f +1 | f - 1 f - 1 | 2 f 2 f +1 | f 2 f f 2 f +1 |
## 5 Odd girth of extremal and largest known circulant graphs of degree 2 to 9
Any circulant graph of degree d ≥ 3 has at least two distinct generators which, taken as pairs in either order, generate two distinct paths of length 2 between a single pair of vertices. Hence these graphs have girth of at most 4. However it is noteworthy that the extremal and largest known circulant graphs of degree 2 to 9 all have odd girth that is maximal for their diameter. The proofs that such graphs of diameter k have odd girth 2 k +1 are presented for even degree only, but the proofs for odd degree are similar. On the other hand, the circulant graphs corresponding to Chen and Jia's lower bound have lower odd girth: for example, odd girth k for degree 8 where k ≡ 1 (mod 4), and (4 k +1) 5 for / degree 10 where k ≡ 1 (mod 5).
Theorem 7. For extremal degree 2 circulant graphs of arbitrary diameter k their odd girth is maximal, 2 k +1 .
Proof. For an extremal circulant graph of degree 2 and arbitrary diameter k , the order is CC (2 , k ) = n = 2 k + 1 and a generator set is { g 1 } where g 1 = 1. The graph is a cycle
graph, and it is immediately clear that it has girth 2 k + 1 and therefore also odd girth 2 k +1.
Theorem 8. For extremal degree 4 circulant graphs of arbitrary diameter k their odd girth is maximal, 2 k +1 .
Proof. For an extremal circulant graph of degree 4 and arbitrary diameter k , the order is CC (4 , k ) = n = 2 k 2 + 2 k + 1 and a generator set is { g , g 1 2 } where g 1 = 1 and g 2 = 2 k +1. If the graph contains a cycle of length 2 l +1 then d g 1 1 + d g 2 2 ≡ 0 (mod n ) with | d 1 | + | d 2 | ≤ 2 l +1 and | d 1 | + | d 2 | ≡ 1 (mod 2) for some d , d 1 2 ∈ Z .
When l = k we have two solutions: d 1 = -k, d 2 = k +1, and d 1 = k +1 , d 2 = k , both giving d g 1 1 + d g 2 2 = n . Therefore the graph contains cycles of length 2 k +1. Now suppose l < k , so that | d 1 | + | d 2 | ≤ 2 k -1. Then | d g 1 1 + d g 2 2 | ≤ (2 k -1) g 2 = (2 k -1)(2 k +1) = 4 k 2 -1 < n 2 , so that d g 1 1 + d g 2 2 = 0 or ± n . First consider the case d g 1 1 + d g 2 2 = 0. We note that g 1 ≡ 1 (mod 2 k +1) and g 2 ≡ 0 (mod 2 k +1). So d 1 ≡ 0 (mod 2 k +1), and hence d 1 = 0. Therefore d 2 = 0 and there is no odd cycle of length 2 +1. Now consider l d g 1 1 + d g 2 2 = ± n . Without loss of generality we assume d g 1 1 + d g 2 2 = n . Noting that n ≡ k +1 (mod 2 k +1) we have d 1 ≡ k + 1 (mod 2 k + 1). Hence d 1 = -k or d 1 = k + 1. If d 1 = -k then d g 2 2 = 2 k 2 +3 +1 and k d 2 = k +1. If d 1 = k +1 then d g 2 2 = 2 k 2 + k and d 2 = k . In both cases | d 1 | + | d 2 | = 2 k +1. Therefore there are no odd cycles of length 2 l +1 for l < k .
Theorem 9. For largest known degree 6 circulant graphs of arbitrary diameter k their odd girth is maximal, 2 k +1 .
Proof. We give the proof just for the case k ≡ 0 (mod 3), k ≥ 3 for isomorphism class 1. The proofs for isomorphism class 2 and for cases k ≡ 1 and k ≡ 2 (mod 3) are similar. So let k = 3 m for m ≥ 1. For a largest known circulant graph of degree 6 and arbitrary diameter 3 m , the order is DF (6 , k ) = n = 32 m 3 +16 m 2 +6 m +1 and a generator set is { g , g 1 2 , g 3 } where g 1 = 1, g 2 = 4 m +1 and g 3 = 16 m 2 +4 m +1.
If the graph contains a cycle of length 2 l +1 then d g 1 1 + d g 2 2 + d g 3 3 ≡ 0 (mod n ) with | d 1 | + | d 2 | + | d 3 | ≤ 2 l +1 and | d 1 | + | d 2 | + | d 3 | ≡ 1 (mod 2) for some d , d 1 2 , d 3 ∈ Z . When l = 3 m we have a solution: d 1 = 2 m +1 , d 2 = 2 m and d 3 = 2 m , giving d g 1 1 + d g 2 2 + d g 3 3 = n . Therefore the graph contains a cycle of length 6 m +1. Now suppose l < 3 m , so that | d 1 | + | d 2 | + | d 3 | ≤ 6 m -1. Then | d g 1 1 + d g 2 2 + d g 3 3 | ≤ (6 m -1) g 3 = (6 m -1)(16 m 2 + 4 m + 1) = 96 m 3 + 8 m 2 + 2 m -1 < 3 n . So d g 1 1 + d g 2 2 + d g 3 3 = 0 , ± n or ± 2 n . But if d g 1 1 + d g 2 2 + d g 3 3 = 0 or ± 2 n then d g 1 1 + d g 2 2 + d g 3 3 ≡ 0 (mod 2), whereas g 1 ≡ g 2 ≡ g 3 ≡ 1 (mod 2) and | d 1 | + | d 2 | + | d 3 | = 1 (mod 2), which gives a contradiction and thus admits no solution. So consider d g 1 1 + d g 2 2 + d g 3 3 = ± n . Then without loss of generality we may assume d g 1 1 + d g 2 2 + d g 3 3 = n . Now n = 32 m 3 +16 m 2 +6 m +1 = 2 mg 3 +2 mg 2 +(2 m +1) g 1 . This is equivalent to 6 m + 1 edges and there is no construction with fewer edges. Thus the graph has no odd cycle with length shorter than 6 m +1.
Theorem 10. For largest known degree 8 circulant graphs of arbitrary diameter k their odd girth is maximal, 2 k +1 .
Proof. We give the proof just for the case k ≡ 0 (mod 2), k ≥ 4. The proof for the case k ≡ 1 (mod 2) is similar. So let k = 2 m for m ≥ 2. For a largest known circulant graph of
degree 8 and arbitrary diameter 2 m , the order is L d, k ( ) = n = 8 m 4 +8 m 3 +12 m 2 +4 m and a generator set is { g , g 1 2 , g 3 , g 4 } where g 1 = 1, g 2 = 4 m 3 + 4 m 2 + 6 m + 1, g 3 = 4 m 4 +4 m 2 -4 m and g 4 = 4 m 4 +4 m 2 -2 m .
If the graph contains a cycle of length 2 l +1 then ∑ 4 i =1 d g i i = cn with ∑ 4 i =1 | d i | ≤ 2 l +1 and ∑ 4 i =1 | d i | ≡ 1 (mod 2) for some d , c i ∈ Z . Let d i = x m i + y i for some x , y i i ∈ Z for i = 1 2 3 4, and let , , , c = x 5 + y 5 for some x , y 5 5 ∈ Z . So ∑ 4 i =1 | x i | ≤ 4. Then we have ( x m 1 + y 1 ) + ( x m 2 + y 2 )(4 m 3 +4 m 2 +6 m +1) + ( x m 3 + y 3 )(4 m 4 +4 m 2 -4 m ) + ( x m 4 + y 4 )(4 m 4 +4 m 2 -2 m ) = ( x m 5 + y 5 )(8 m 4 + 8 m 3 +12 m 2 +4 m ). Equating coefficients of powers of m m ; 5 : 4 x 3 + 4 x 4 = 8 x 5 . So x 3 + x 4 is even. Hence x 3 + x 4 = -4 2 0 2 or , , , 4. m 4 : 4 x 2 + 4 y 3 + 4 y 4 = 8 x 5 + 8 y 5 . m 3 : 4 x 2 + 4 y 2 + 4 x 3 + 4 x 4 = 12 x 5 + 8 y 5 . m 2 : 6 x 2 +4 y 2 -4 x 3 +4 y 3 -2 x 4 +4 y 4 = 4 x 5 +12 y 5 . m 1 : x 1 + x 2 +6 y 2 -4 y 3 -2 y 4 = 4 y 5 . m 0 : y 1 + y 2 = 0. Noting that n ≡ g 3 ≡ g 4 ≡ 0 (mod 2 m ) and g 1 ≡ g 2 ≡ 1 (mod 2 m ), we have d 1 + d 2 ≡ 0 (mod 2 m ). Thus x 1 + x 2 = -4 , -2 0 2 or 4. , , By straighforward manipulation we find that there are no solutions for odd cycle length less than 4 m + 1. For length 4 m +1 there are exactly 16 distinct combinations of generator elements corresponding to each combination of x i = ± 1 for i = 1 2 3 4. , , , For example with x 1 = x 2 = x 3 = x 4 = 1 we have y 1 = y 2 = y 3 = 0 and y 4 = 1, giving ∑ 4 i =1 d g i i = 8 m 5 +8 m 4 +12 m 3 +4 m 2 = mn . This proves the graph has odd girth 2 k +1.
## 6 Maximal distance partition levels by vertex type
We have seen that all the extremal and largest known circulant graphs of degree 2 to 9 of arbitrary diameter k , above some threshold, have maximal odd girth, 2 k +1. This means that only in the final distance partition level, k , relative to an arbitrary root vertex, is any vertex adjacent to another in the same level. Thus any vertex in level l for 1 ≤ l ≤ k -1 is adjacent only to vertices in level l -1 or l +1 and to none in level l . In this sense such a vertex may be defined as thin . As the degree, d , of each vertex is fixed, if it is adjacent to s vertices in level l -1 then it must be adjacent to d -s in level l +1, and such a vertex is defined to be of type T s . Vertices in level k may be adjacent to others in the same level but not of course to any in a further level. Therefore the type of these vertices is also well-defined by the number of adjacent vertices in the preceding level.
Analysis of the extremal and largest known circulant graphs of degree 2 to 9 reveals a regular structure in the number of vertices of each type in each distance partition level. Examples for graphs of degree 4, 6 and 8, all with diameter 12, are shown in Tables 9, 10 and 11.
Table 9: Distance partition profile by vertex type: extremal graph of degree 4 and diameter 12
| Vertex | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | | | | | Differences Maximal |
|----------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----|----|----|----|-----------------------|
| type | 0 | 1 | 2 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | |
| T 0 T 1 | 1 | 4 4 | 4 | 4 | 4 | 4 | 4 4 | 4 | 4 | | 4 | 4 | ∆ 0 = 4 |
| T 2 | | 4 | 8 | 12 | 16 | 20 24 | 28 | 32 | | 36 | 40 | 44 | ∆ 1 = 4 |
| Total | 1 | 4 | 8 12 | 16 | 20 24 | 28 | 32 | 36 | | 40 | 44 | 48 | ∆ 1 = 4 |
In these examples for graphs of even degree we can see that the number of different vertex types increases with degree. Graphs of every degree contain type T 1 vertices, with
Table 10: Distance partition profile by vertex type: largest known graph of degree 6, diameter 12
| Vertex | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Distance partition level | Differences |
|----------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|----------------------------|--------------------|
| type | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 9 | | 10 | 11 | 12 | Maximal Submaximal |
| T 0 T 1 | 1 | 6 6 | 6 | 6 | | 6 | 6 | 6 | 6 | | | | ∆ 0 | = 6 ∆ 0 = 0 1 |
| T 2 | | 12 | 24 | 36 | | 48 | 60 | 72 84 | 88 | 76 | 64 | | 52 ∆ 1 | = 12 ∆ = - 12 |
| T 3 | | | 8 | 24 | 48 | 80 | 120 | 168 | 226 | 274 | 306 | 322 | ∆ 2 | = 8 ∆ 2 = - 16 1 |
| T 4 | | | | | | | | | 4 | 16 | 28 | 40 | | ∆ = 12 |
| Total | 1 | 6 18 | 38 | 66 | 102 | 146 | 198 | 258 | 318 | 366 | 398 | 414 | ∆ 2 = 8 | ∆ 2 = - 16 |
Table 11: Distance partition profile by vertex type: largest known graph of degree 8, diameter 12
| Vertex | Distance partition level Differences | Distance partition level Differences | Distance partition level Differences | Distance partition level Differences | Distance partition level Differences | Distance partition level Differences | Distance partition level Differences | Distance partition level Differences | Distance partition level Differences | Distance partition level Differences | Distance partition level Differences | Distance partition level Differences | Distance partition level Differences | |
|----------|----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|------------|
| type | 0 | 1 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | Maximal | Submaximal |
| T 0 | 1 | | | | | | | | | | | | | |
| T 1 | | 8 8 | 8 | 8 | 8 | 8 | 4 | | | | | | ∆ 0 = 8 | ∆ 0 = 0 |
| T 2 | | 24 | 48 | 72 | 96 | 120 | 136 | 124 | 100 | 76 | 52 | 26 | ∆ 1 = 24 | ∆ 1 = - 24 |
| T 3 | | | 32 | 96 | 192 | 320 | 476 | 624 | 720 | 752 | 720 | 624 | ∆ 2 = 32 | ∆ 2 = - 64 |
| T 4 | | | | 16 | 64 | 160 | 328 | 564 | 844 | 1124 | 1356 | 1495 | ∆ 3 = 16 | ∆ 3 = - 48 |
| T 5 | | | | | | | | 32 | 96 | 192 | 320 | 476 | | ∆ 2 = 32 |
| T 6 | | | | | | | | | | | | 2 | | |
| Total | 1 | 8 32 | 88 | 192 | 360 | 608 | 944 | 1344 | 1760 | 2144 | 2448 | 2623 | ∆ 3 = 16 ∆ | 3 = - 48 |
their number remaining constant and equal to the degree within the maximal zone, while becoming absent from the submaximal zone within the first few levels. The number of each subsequent vertex type grows with a constant higher order difference in the maximal zone before reversing with a constant difference of the same order in the submaximal zone. In each case the values of these constants depend only on the degree and are independent of diameter. The graphs of odd degree display a similar structure. Common parameters of all the distance partition profiles by vertex type are presented in Table 12 for the maximal levels and in Table 13 for the submaximal levels.
Table 12: Number of vertices of each type within maximal levels of extremal and largest known graphs of degree d = 4 to 9 (dimension f = 2 to 4).
| Vertex type T | Even degree d = 2 f | Even degree d = 2 f | Odd degree d = 2 f +1 | Odd degree d = 2 f +1 |
|-----------------|-----------------------|-----------------------|-------------------------|-------------------------|
| s | Difference order | Value | Difference order | Value |
| 1 ≤ s ≤ f | ∆ s - 1 | 2 s ( f s ) | ∆ s - 1 | 2 s f s ) |
| s = f +1 | - | - | ∆ f - 1 | ( 2 f |
We will now prove for circulant graphs of any degree d and arbitrary diameter k that the number of vertices of each type in each maximal level is determined by the same upper bound M AC ( d, k ) that has been proven to determine the total number of vertices in each maximal level.
Theorem 11. For circulant graphs of any degree d , the number of vertices, V T ( d, s, l ) , of type T s in distance partition level l ≥ 1 , where the level is maximal, so that the size of the level is LM AC ( d, l ) = M AC ( d, l ) -M AC ( d, l -1) , is given by the following formulae.
Table 13: Number of vertices of each type within submaximal levels of extremal and largest known graphs of degree d = 6 to 9 (dimension f = 3 and 4).
| Vertex | Even degree d = 2 f | Even degree d = 2 f | Odd degree d = 2 f +1 | Odd degree d = 2 f +1 |
|-----------|-----------------------|-----------------------|-------------------------|-------------------------|
| type T s | Difference order | Value | Difference order | Value |
| 2 ≤ s ≤ f | ∆ s - 1 | - ( s - 1)2 s ( f s ) | ∆ s - 1 | - ( s - 1)2 s f s ) |
| s = f +1 | ∆ f - 2 | f 2 f - 1 | ∆ f - 1 | ( - ( f - 1)2 f |
| s = f +2 | - | - | ∆ f - 2 | f 2 f - 1 |
For even degree d = 2 f where f is the dimension, we have:
$$\ V T ( d, s, l ) = \binom { f } { s } \sum _ { i = 1 } ^ { s } { ( - 1 ) ^ { s - i } \binom { s } { i } L M _ { A C } ( 2 i, l ) }$$
and hence
$$\ V T ( d, s, l ) = \begin{cases} \ d & \text{for $s=1$} \\ 2 f s \prod _ { i = 1 } ^ { s - 1 } 2 ( f - i ) ( l - i ) / ( i + 1 ) ^ { 2 } & \text{for $s\geq 2$.} \end{cases} \\ \L o d d \ e g r e e \ d = 2 f + 1 \text{ where $f$ is the dimension, we have the recurrence relation:}$$
For odd degree d = 2 f +1 where f is the dimension, we have the recurrence relation:
$$V T ( d, s, l ) = V T ( 2 f, s, l ) + V T ( 2 f, s - 1, l - 1 ) \text{ for } s, l \geq 2$$
and hence
$$\text{and hence} \\ \ V T ( d, s, l ) = \begin{cases} \ d & \text{for $s=1$} \\ 2 f ( f - 1 ) ( l - 1 ) + 2 f & \text{for $s=2$} \\ 2 f s \prod _ { i = 1 } ^ { s - 1 } 2 ( f - i ) ( l - i ) / ( i + 1 ) ^ { 2 } \\ + 2 f ( s - 1 ) \prod _ { i = 1 } ^ { s - 2 } 2 ( f - i ) ( l - 1 - i ) / ( i + 1 ) ^ { 2 } & \text{for $s\geq 3$.} \end{cases} \\ \text{Proof} \text{. First consider a circulant graph of even degree $d$ and diameter $k$, being the Cayley}$$
Proof. First consider a circulant graph of even degree d and diameter k , being the Cayley graph of a cyclic group with generator set G = { g , ..., g 1 f } where f = d/ 2 is the dimension of the graph. Then the connection set is C = {± g , ..., 1 ± g f } .
Let v be a vertex at distance l < k from an arbitrary root vertex u . Suppose for a contradiction that a path of length l from u to v contains an edge generated by g i and another edge generated by -g i for some g i ∈ G . Then as the group is Abelian the path from u generated from the same set of edges after removing this pair would lead to v after a distance of only l -2, contradicting the premise that v is distant l from u . Hence if v is distant l from u then for any g i ∈ G no path of length l from u to v contains edges generated by both g i and -g i .
Suppose there exists a path p of length l ≥ 2 from u to v with two of its edges generated by different generators, say c 1 , c 2 where | c 1 | = g i and | c 2 | = g j for some i, j with 1 ≤ i < j ≤ f . Then as the group is Abelian we may reorder the edges of p to generate two distinct paths p 1 = ( x , ..., x 1 l -2 , c 1 , c 2 ) and p 2 = ( x , ..., x 1 l -2 , c 2 , c 1 ) from u to v . Now consider the two
vertices v , v 1 2 both distant l -1 from u , reached by following paths p ′ 1 = ( x , ..., x 1 l -2 , c 1 ) and p ′ 2 = ( x , ..., x 1 l -2 , c 2 ) from u . These are distinct vertices within distance partition level l -1 that are adjacent to v in level l . Thus v is connected to more than one level l -1 vertex and so is not a type T 1 vertex. Therefore for any type T 1 vertex in level l ≥ 1 there is only one path from u of length l and each edge of the path is generated by the same element of the connection set. Also every vertex on this path is also a type T 1 vertex generated by the same element. Conversely every element of the connection set generates a unique path from u passing through vertices which are all distinct type T 1 vertices while the distance partition level remains in the maximal zone by the definition of the upper bound M AC ( d, l ). Therefore within the maximal zone the number of type T 1 vertices in each level, V T ( d, 1 , l ), will be equal to the degree d of the graph. This can also be expressed as the product of the number of such vertices for each generator, LM AC (2 , l ) = 2 where LM AC ( d, l ) = M AC ( d, l ) -M AC ( d, l -1) as defined in an earlier section, and the number of generators, f , giving V T ( d, 1 , l ) = 2 f = d .
Now consider any two generators g , g i j ∈ G and all vertices in level l that can be reached from root vertex u with a path of length l comprised only of edges ± g i and ± g j . As level l is maximal, by definition of the upper bound M AC ( d, k ) there are LM AC (4 , l ) such vertices. We now restrict the vertex set to only those vertices where the path includes at least one edge ± g i and one edge ± g j , so that l ≥ 2. As the group is Abelian, each of these vertices will have at least one path from u with final edge ± g i and at least one path with final edge ± g j , and clearly no paths with any other final edge. Therefore all these vertices are of type T 2 . The vertices with paths only of edges ± g i or only of edges ± g j are excluded. Thus the number of excluded vertices is 2 LM AC (2 , l ), and so the number of T 2 vertices in level l reached by paths generated by the pair g , g i j is given by LM AC (4 , l ) -2 LM AC (2 , l ) = 4( l -1). As there are f ( f -1) / 2 distinct pairs of generators, the total number of T 2 vertices in level l is given by V T ( d, 2 , l ) = f ( f -1) / 2 × 4( l -1) = d d ( -2)( l -1) / 2.
Similarly the number of vertices of type T 3 in level l ≥ 3 from any given triad of generators is LM AC (6 , l ) -3 LM AC (4 , l )+3 LM AC (2 , l ) = 4( l -1)( l -2). As there are f ( f -1)( f -2) / 6 distinct triads of generators, the total number of type T 3 vertices in level is V T ( d, 3 , l ) = d d ( -2)( d -4)( l -1)( l -2) / 12. Also the number of vertices of type T 4 in level l ≥ 4 from any given set of four generators is LM AC (8 , l ) -4 LM AC (6 , l ) + 6 LM AC (4 , l ) -4 LM AC (2 , l ) = 8( l -1)( l -2)( l -3) / 3, and so the total number of type T 4 vertices in level l is V T ( d, 4 , l ) = d d ( -2)( d -4)( d -6)( l -1)( l -2)( l -3) / 144. More generally, for even degree d = 2 f and any s ≥ 1,
$$\ V T ( d, s, l ) = \binom { f } { s } \sum _ { i = 1 } ^ { s } ( - 1 ) ^ { s - i } \binom { s } { i } L M _ { A C } ( 2 i, l ). \\.$$
This can be reformulated as: V T ( d, s, l ) = ds ∏ s -1 i =1 ( d -2 )( i l -i ) / i ( +1) 2 for s ≥ 2. Now consider a circulant graph of odd degree d = 2 f + 1 where f is the dimension, and order n . If the generator set is G = { g , ..., g 1 f } then the connection set will be C = {± g , ..., 1 ± g f , n/ 2 . } As the element n/ 2 has order 2, it can only generate a path of length 1 to create one additional type T 1 vertex in level 1 but no extra vertices of type T 1 in any higher levels. Consider any level l ≥ 2 within the maximal zone, and any vertex v in this level. It is possible to reach v by a path comprised either of edges generated by the non order 2 elements {± g , . . . , 1 ± g f } alone or else also by including a single edge
generated by the order 2 element n/ 2. It is not possible to reach any given vertex v via paths of both cases as the level is within the maximal zone. So for any s ≥ 2 the total number of vertices of type T s in level l is the sum of the vertices reached by paths of either case. The number in the first case is simply the result just determined for a graph of even degree d = 2 f : V T (2 f, s, l ). For the vertices in the second case, where the path includes an edge n/ 2, as the group is Abelian we may consider only those paths where the final edge is n/ 2. As vertex v is of type T s in level l , then the preceding vertex v ′ on each path must be of type T s -1 in level l -1, where the path to v ′ is comprised of edges from the connection set C = {± g , ..., 1 ± g f } . Therefore, invoking the result for even degree again, the number of vertices in this case is V T (2 f, s -1 , l -1). Hence for a circulant graph of odd degree d = 2 f + 1 we have V T ( d, s, l ) = V T (2 f, s, l ) + V T (2 f, s -1 , l -1) for s, l ≥ 2.
We now consider the number of type T 1 vertices in extremal, largest known and lower bound circulant graphs of degree 2 to 10. For these graphs the total number of type T 1 vertices is a linear function of the diameter in each case. For degree 2 there are two vertices in each distance partition level after the root and, being a cycle graph, they are all type T 1 . For degree 3 there are four vertices in each level from level 2; two type T 1 and two type T 2 . For degree d ≥ 4 the total number of T 1 vertices increases by 4 for each increase of 1 in the diameter. This surprising result is true not only for the extremal and largest known graphs but also for the Chen and Jia lower bound constructions. The functions are summarised in Table 14. The proofs for certain degree 6 and 8 cases are given in Theorems 12 and 13. The other cases are proved similarly.
Table 14: Total number of type T 1 vertices in extremal, largest known and lower bound graphs up to degree 10 for arbitrary diameter k ≥ 2.
| Degree, d | Status | Order | Isomorphism class | Number of T 1 vertices |
|-------------|---------------|--------------|---------------------|--------------------------|
| 2 | Extremal | CC (2 , k ) | - | 2 k |
| 3 | Extremal | CC (3 , k ) | - | 2 k +1 |
| 4 | Extremal | CC (4 , k ) | - | 4 k |
| 5 | Extremal | CC (5 , k ) | - | 4 k - 1 |
| 6 | Largest known | DF (6 , k ) | 1 | 4 k |
| 6 | Largest known | DF (6 , k ) | 2 | 4 k +2 |
| 7 | Largest known | DF (7 , k ) | 1 | 4 k - 1 |
| 7 | Largest known | DF (7 , k ) | 2 | 4 k +1 |
| 8 | Largest known | L (8 , k ) | - | 4 k +4 |
| 9 | Largest known | L (9 , k ) | 1 | 4 k +1 |
| 9 | Largest known | L (9 , k ) | 2 | 4 k +1 |
| 8 | Lower bound | CJ (8 , k ) | - | 4 k - 10 |
| 10 | Lower bound | CJ (10 , k ) | - | 4 k - 10 |
Theorem 12. For largest known degree 6 circulant graphs of diameter k , the total number of type T 1 vertices is equal to 4 k for isomorphism class 1 and equal to 4 k + 2 for isomorphism class 2 .
Proof. We give the proof just for the case k ≡ 0 (mod 3), k ≥ 3, both for a graph of isomorphism class 1 and of isomorphism class 2. The proofs for cases k ≡ 1 and k ≡ 2 (mod 3) are similar. So let k = 3 m for m ≥ 1. By Theorem 3 every distance partition level l of a largest known degree 6 graph of diameter k is maximal for l ≤ /floorleft (2 k +1) 3 / /floorright = 2 m . Then by Theorem 11 the number of type T 1 vertices in level l is equal to the degree d = 6 for 1 ≤ l ≤ 2 m .
For a largest known circulant graph of degree 6 of isomorphism class 1 and arbitrary diameter 3 m , the order is DF (6 , k ) = n = 32 m 3 + 16 m 2 + 6 m + 1 and a generator set is { g , g 1 2 , g 3 } where g 1 = 1, g 2 = 4 m + 1 and g 3 = 16 m 2 + 4 m + 1. Now we consider the six potential type T 1 vertices within level 2 m +1: (2 m +1)( ± g 1 ), (2 m +1)( ± g 2 ) and (2 m +1)( ± g 3 ), relative to root vertex 0. First consider the vertex g 2 -2 mg 1 . It can be reached by a path of length 2 m +1 with edges from two distinct generators. Therefore if it is in level 2 m + 1 then it is a type T 2 vertex or possibly T 3 but not T 1 . But we have g 2 -2 mg 1 = 4 m +1 -2 m = 2 m +1 = (2 m +1) g 1 . Thus (2 m +1) g 1 is not a type T 1 vertex in level 2 m +1, and so neither is (2 m +1)( -g 1 ). Similarly the vertex g 3 -(2 m -1) g 2 -g 1 can be reached by a path of length 2 m + 1 with edges from three distinct generators. Therefore if it is in level 2 m +1 then it is a type T 3 vertex. Now g 3 -(2 m -1) g 2 -g 1 = 16 m 2 +4 m +1 -(2 m -1)(4 m +1) -1 = 8 m 2 +6 m +1 = (2 m +1)(4 m +1) = (2 m +1) g 2 . Thus (2 m +1) g 2 is not a type T 1 vertex in level 2 m +1, and so neither is (2 m +1)( -g 2 ). Finally the vertex -2 mg 3 -g 2 (mod n ) can be reached by a path of length 2 m + 1 with edges from two distinct generators. Therefore if it is in level 2 m +1 then it is a type T 2 vertex or possibly T 3 . Here 2 n -2 mg 3 -g 2 = 64 m 3 +32 m 2 +12 m +2 -2 m (16 m 2 +4 m +1) -(4 m +1) = 32 m 3 +24 m 2 +6 m + 1 = (2 m + 1)(16 m 2 + 4 m + 1) = (2 m + 1) g 3 . Thus (2 m + 1) g 3 is not a type T 1 vertex in level 2 m + 1, and so neither is (2 m + 1)( -g 3 ). Hence there are no type T 1 vertices in the first submaximal level, 2 m +1, and consequently none in any later levels. Therefore for isomorphism class 1 the total number of type T 1 vertices is 6 × 2 m = 12 m = 4 k .
For a largest known circulant graph of degree 6 of isomorphism class 2 and arbitrary diameter 3 m , the order is DF (6 , k ) = n = 32 m 3 + 16 m 2 + 6 m + 1 and a generator set is { g , g 1 2 , g 3 } where g 1 = 1, g 2 = 8 m 2 +2 m and g 3 = 8 m 2 +6 m +2. We again consider the six potential type T 1 vertices within level 2 m + 1: (2 m + 1)( ± g 1 ), (2 m + 1)( ± g 2 ) and (2 m +1)( ± g 3 ). First consider the vertex -g 3 -(2 m -1) g 2 + g 1 (mod n ). It can be reached by a path of length 2 m +1 with edges from three distinct generators. Therefore if it is in level 2 m + 1 then it is a type T 3 vertex. But n -g 3 -(2 m -1) g 2 + g 1 = 32 m 3 +16 m 2 +6 m +1 -(8 m 2 +6 m +2) -(2 m -1)(8 m 2 +2 m ) +1 = 16 m 3 +12 m 2 +2 m = (2 m + 1)(8 m 2 + 2 m ) = (2 m + 1) g 2 . Thus (2 m + 1) g 2 is not a type T 1 vertex in level 2 m + 1, and so neither is (2 m + 1)( -g 2 ). The vertex -(2 m -1) g 3 + g 2 -g 1 (mod n ) can also be reached by a path of length 2 m +1 with edges from three distinct generators. Therefore if it is in level 2 m +1 then it is a type T 3 vertex. Now n -(2 m -1) g 3 + g 2 -g 1 = 32 m 3 +16 m 2 +6 m +1 -(2 m -1)(8 m 2 +6 m +2)+(8 m 2 +2 m ) -1 = 16 m 3 +20 m 2 +10 m +2 = (2 m +1)(8 m 2 +6 m +2) = (2 m +1) g 3 . Thus (2 m +1) g 3 is not a type T 1 vertex in level 2 m +1, and so neither is (2 m +1)( -g 3 ). Suppose for a contradiction that the vertex (2 m +1) g 1 is not a type T 1 vertex in level 2 m + 1. Then we have a g 1 1 + a g 2 2 + a g 3 3 ≡ 2 m + 1 (mod n ) with | a 1 | + | a 2 | + | a 3 | ≤ 2 m + 1 and | a 2 | + | a 3 | ≥ 1 for some a , a 1 2 , a 3 ∈ Z . Now | a g 1 1 + a g 2 2 + a g 3 3 | ≤ (2 m + 1) g 3 = 16 m 3 + 20 m 2 + 10 m + 2 < n -(2 m + 1). Thus a g 1 1 + a g 2 2 + a g 3 3 = 2 m + 1. We have g 1 ≡ g 3 ≡ 1 (mod 4 m + 1) and g 2 ≡ 0
(mod 4 m +1). So a 1 + a 3 ≡ 2 m +1 (mod 4 m +1). But | a 1 | + | a 2 | + | a 3 | ≤ 2 m +1. So a 1 + a 3 = 2 m +1 , a 1 ≥ 0 and a 3 ≥ 0. Hence a 2 = 0. So we have (2 m +1) -a 3 + a g 3 3 = 2 m +1. Hence a 3 ( g 3 -1) = 0 and a 3 = 0 contradicting the premise. Therefore vertex (2 m +1) g 1 is a type T 1 vertex in level 2 m +1 and consequently so is (2 m +1)( -g 1 ). Finally we need to prove that the vertex (2 m +2) g 1 is not a type T 1 vertex. The vertex g 3 -g 2 -2 mg 1 can be reached by a path of length 2 m + 2 with edges from three distinct generators. Therefore if it is in level 2 m + 2 then it is a type T 3 vertex. Here g 3 -g 2 -2 mg 1 = 8 m 2 + 6 m + 2 -(8 m 2 + 2 m ) -2 m = 2 m + 1 = (2 m + 1) g 1 . Thus (2 m + 2) g 1 is not a type T 1 vertex in level 2 m + 2, and so neither is (2 m + 2)( -g 1 ). Hence there are two type T 1 vertices in the first submaximal level, 2 m + 1, none in the second, 2 m + 2, and consequently none in any later levels. Therefore for isomorphism class 2 the total number of type T 1 vertices is 6 × 2 m +2 = 12 m +2 = 4 k +2.
Theorem 13. For largest known degree 8 circulant graphs of diameter k , the total number of type T 1 vertices is equal to 4 k +4 .
Proof. We give the proof just for the case k ≡ 0 (mod 2), k ≥ 4. The proof for the case k ≡ 1 (mod 2) is similar. So let k = 2 m for m ≥ 2. By Theorem 5 every distance partition level l of a largest known degree 8 graph of diameter k is maximal for l ≤ /floorleft ( k +1) 2 / /floorright = m . Then by Theorem 11 the number of type T 1 vertices in level l is equal to the degree d = 8 for 1 ≤ l ≤ m .
For a largest known circulant graph of degree 8 and arbitrary diameter 2 m , the order is L d, k ( ) = n = 8 m 4 + 8 m 3 + 12 m 2 + 4 m and a generator set is { g , g 1 2 , g 3 , g 4 } where g 1 = 1, g 2 = 4 m 3 + 4 m 2 + 6 m + 1, g 3 = 4 m 4 + 4 m 2 -4 m and g 4 = 4 m 4 + 4 m 2 -2 m . Now we consider the eight potential type T 1 vertices within level m + 1: ( m + 1)( ± g 1 ), ( m +1)( ± g 2 ), ( m +1)( ± g 3 ) and ( m +1)( ± g 4 ), relative to root vertex 0.
First consider the vertex g 4 -g 3 -( m -1) g 1 . It can be reached by a path of length m + 1 with edges from three distinct generators. Therefore if it is in level m + 1 then it is a type T 3 vertex or possibly T 4 but not T 1 . But we have g 4 -g 3 -( m -1) g 1 = 4 m 4 +4 m 2 -2 m -(4 m 4 +4 m 2 -4 m ) -( m -1) = m +1 = ( m +1) g 1 . Thus ( m +1) g 1 is not a type T 1 vertex in level m +1, and so neither is ( m +1)( -g 1 ). Similarly the vertex -g 4 + g 3 -( m -1) g 2 (mod n ) can be reached by a path of length m +1 with edges from three distinct generators. Therefore if it is in level m + 1 then it is a type T 3 vertex or possibly T 4 . Now n -g 4 + g 3 -( m -1) g 2 = 8 m 4 + 8 m 3 + 12 m 2 + 4 m -(4 m 4 + 4 m 2 -2 m ) + (4 m 4 +4 m 2 -4 m ) -( m -1)(4 m 3 +4 m 2 +6 m +1) = 4 m 4 +8 m 3 +10 m 2 +7 m +1 = ( m + 1)(4 m 3 + 4 m 2 + 6 m +1) = ( m + 1) g 2 . Thus ( m + 1) g 2 is not a type T 1 vertex in level m +1, and so neither is ( m +1)( -g 2 ).
Suppose for a contradiction that the vertex ( m +1) g 3 is not a type T 1 vertex in level m +1. Then we have ∑ 4 i =1 a g i i ≡ ( m +1) g 3 (mod n ) with ∑ 4 i =1 | a i | ≤ m +1 and | a 1 | + | a 2 | + | a 4 | ≥ 1 for some a i ∈ Z . Now g 1 ≡ g 2 ≡ 1 (mod 2 m ) and g 3 ≡ g 4 ≡ n ≡ 0 (mod 2 m ). Thus a 1 + a 2 ≡ 0 (mod 2 m ), and so a 2 = -a 1 . Also g 1 ≡ 1 , g 2 ≡ 2 m +1 , g 3 ≡ 0 , g 4 ≡ 2 m,n ≡ 0 (all mod 4 m ). So a 1 +(2 m +1) a 2 +2 ma 4 ≡ 0 (mod 4 m ), a 1 -a 1 (2 m +1)+2 ma 4 ≡ 0 (mod 4 m ), and 2 m a ( 4 -a 1 ) ≡ 0 (mod 4 m ). Hence a 4 ≡ a 1 (mod 2). Let a i = x m i + y i where x , y i i ∈ Z for i = 1 2 3 4. , , , Then x 2 = -x 1 and y 2 = -y 1 . Also x 4 ≡ x 1 and y 4 ≡ y 1 (both mod 2). As ∑ 4 i =1 | a i | ≤ m +1, we have 2 | x 1 | + | x 3 | + | x 4 | ≤ 1. Thus x 1 = 0. Then x 4 is even, and so x 4 = 0. Therefore we have a 1 = y , a 1 2 = -y , a 1 3 = x m 3 + y 3 and a 4 = y 4 .
Hence ∑ 4 i =1 a g i i -( m +1) g 3 = 4( x 3 -1) m 5 +4( y 3 + y 4 -1) m 4 +4( -y 1 + x 3 -1) m 3 + 4( -y 1 + y 3 -x 3 + y 4 ) m 2 +2( -3 y 1 -2 y 3 -y 4 +2) m ≡ 0 (mod n ). But n ≡ 0 (mod 4 m ). So 2 m y ( 1 +2 y 4 ) ≡ 0 (mod 4 m ), and y 4 = -y 1 (mod 2). As this polynomial is a quintic in m we may set it equal to ( x m 5 + y 5 ) n where x , y 5 5 ∈ Z . Now we equate coefficients of powers of m m . 5 : x 3 -1 = 2 x 5 . m 4 : y 3 + y 4 -1 = 2 x 5 +2 y 5 . m 3 : -y 1 + x 3 -1 = 3 x 5 +2 y 5 . m 2 : -y 1 + y 3 -x 3 + y 4 = x 5 +3 y 5 . m : -3 y 1 -2 y 3 -y 4 +2 = 2 y 5 . If x 3 = 0 then this would imply 2 x 5 = -1 and so x 3 = ± 1. First consider x 3 = 1. Then by simple manipulation we find x 5 = 0 , y 5 = 0 , y 1 = 0 , y 3 = 1 and y 4 = 0. This gives ∑ 4 i =1 a g i i = ( m +1) g 3 , which does not satisfy the premise that we must have | a 1 | + | a 2 | + | a 4 | ≥ 1. Now consider x 3 = -1. Then we find x 5 = -1 , y 5 = 0 , y 1 = 1 , y 3 = 0 and y 4 = -1. So ∑ 4 i =1 a g i i = g 1 -g 2 -mg 3 -g 4 , which does not satisfy the premise that ∑ 4 i =1 | a i | ≤ m +1. Therefore vertex ( m +1) g 3 is a type T 1 vertex in level m +1 and consequently so is ( m +1)( -g 3 ).
Similarly, we suppose for a contradiction that the vertex ( m +1) g 4 is not a type T 1 vertex in level m +1. Then we have ∑ 4 i =1 a g i i ≡ ( m +1) g 4 (mod n ) with ∑ 4 i =1 | a i | ≤ m +1 and | a 1 | + | a 2 | + | a 3 | ≥ 1 for some a i ∈ Z . Following the same argument as for vertex ( m +1) g 3 we again find a 1 = y , a 1 2 = -y , a 1 3 = x m 3 + y 3 and a 4 = y 4 . Hence ∑ 4 i =1 a g i i -( m +1) g 4 = 4( x 3 -1) m 5 +4( y 3 + y 4 -1) m 4 +4( -y 1 + x 3 -1) m 3 +2( -2 y 1 +2 y 3 -2 x 3 +2 y 4 -1) m 2 + 2( -3 y 1 -2 y 3 -y 4 + 1) m ≡ 0 (mod n ). As this polynomial is a quintic in m we may set it equal to ( x m 5 + y 5 ) n where x , y 5 5 ∈ Z . Now we equate coefficients of powers of m . m 5 : x 3 -1 = 2 x 5 . m 4 : y 3 + y 4 -1 = 2 x 5 + 2 y 5 . m 3 : -y 1 + x 3 -1 = 3 x 5 + 2 y 5 . m 2 : -2 y 1 +2 y 3 -2 x 3 +2 y 4 -1 = 2 x 5 +6 y 5 . m : -3 y 1 -2 y 3 -y 4 +1 = 2 y 5 . If x 3 = 0 then this would imply 2 x 5 = -1 and so x 3 = ± 1. First consider x 3 = 1. Then by simple manipulation we find 2 y 5 = 1 which is not possible. Now consider x 3 = -1. This implies 2 y 5 = -1 which is also not possible. Therefore vertex ( m + 1) g 4 is a type T 1 vertex in level m +1 and consequently so is ( m +1)( -g 4 ).
Now we need to prove that the vertex ( m +2) g 3 is not a type T 1 vertex in level m +2. The vertex -g 4 -( m -1) g 3 -g 2 + g 1 can be reached by a path of length m +2 with edges from four distinct generators. Therefore if it is in level m +2 then it is a type T 4 vertex. We have mn -g 4 -( m -1) g 3 -g 2 + g 1 = 8 m 5 +8 m 4 +12 m 3 +4 m 2 -(4 m 4 +4 m 2 -2 m ) -( m -1)(4 m 4 +4 m 2 -4 m ) -(4 m 3 +4 m 2 +6 m +1)+1 = 4 m 5 +8 m 4 +4 m 3 +4 m 2 -8 m = ( m +2)(4 m 4 +4 m 2 -4 m ) = ( m +2) g 3 . Thus ( m +2) g 3 is not a type T 1 vertex in level m +2, and so neither is ( m +2)( -g 3 ). Finally we prove that the vertex ( m +2) g 4 is not a type T 1 vertex in level m + 2. The vertex -( m -1) g 4 + g 3 + g 2 -g 1 (mod n ) can be reached by a path of length m +2 with edges from four distinct generators. Therefore if it is in level m +2 then it is a type T 4 vertex. Here ( m -1) n -( m -1) g 4 + g 3 + g 2 -g 1 = ( m -1)(8 m 4 +8 m 3 +12 m 2 +4 m ) -( m -1)(4 m 4 +4 m 2 -2 m )+(4 m 4 +4 m 2 -4 m )+(4 m 3 + 4 m 2 +6 m +1) -1 = 4 m 5 +8 m 4 +4 m 3 +6 m 2 -4 m = ( m +2)(4 m 4 +4 m 2 -2 m ) = ( m +2) g 4 . Thus ( m +2) g 4 is not a type T 1 vertex in level m +2, and so neither is ( m +2)( -g 4 ).
Hence there are four type T 1 vertices in the first submaximal level, m + 1, none in the second, m +2, and consequently none in any later levels. Therefore the total number of type T 1 vertices is 8 m +4 = 4 k +4.
## 7 Conclusion
We have seen how an analysis of their distance partitions reveals much interesting structure of extremal and largest known circulant graphs up to degree 9. These graphs were all found to have odd girth which is maximal for their diameter. The maximum number of vertices in each level of the distance partition was shown to be related to an established upper bound for the order of Abelian Cayley graphs, M AC ( d, k ). These graphs all have a maximal zone where the levels achieve this upper bound, and for degree d ≥ 5 a submaximal zone where they are smaller. Defining the type of each vertex in a level according to the number of adjacent vertices in the preceding level, the number of vertices of each type in each maximal level was also shown to be related to the same upper bound. Finally the total number of type T 1 vertices in each of these graphs was determined to be a linear function of their diameter.
We have observed for all the extremal and largest known graphs of degree 4 to 9 that the total number of type T 1 vertices increases by 4 for every increase by 1 in the diameter. We have also established that the number of type T 1 vertices in each level l ≥ 2 within the maximal zone is twice the dimension f , giving 2 f . The resultant ratio of 2 /f gives a value of 1 for degree 4, 2/3 for degree 6, and 1/2 for degree 8. This correlates with the proportion of levels that lie within the maximal zone for each even degree. We also note for the largest known graphs of degree 6 to 9, having a submaximal zone, that the number of type T 2 vertices in each level is initially 0 in level 1, increases by 4 ( f 2 ) = 2 f ( f -1) per level in the maximal zone, and then decreases at the same rate in the submaximal zone after a limited transition adjustment between the two zones. As the number of such vertices in a level can never be negative, this progression can only exist as long as the submaximal zone does not contain more levels than the maximal zone. This limit is reached at dimension 4, when the maximal zone covers half of the levels. Indeed for the best graph of degree 8 the number of type T 2 vertices in level k remains constant at 26 for any diameter k ≥ 3. Extrapolating these relationships to the case of degree 10, the proportion of levels within the maximal zone would be 2/5 and the number of type T 2 vertices in the final level would consequently be negative for any diameter above some threshold, which is of course impossible.
A prime objective of this analysis of properties of the extremal and best graphs of degree 2 to 9 is to discover relationships that may be parametrised by degree and enable extrapolation to degree 10 and beyond. An unfortunate consequence of the failure of the type T 2 vertex number calculation for degree 10 is that some key relationships that are valid for circulant graphs up to dimension 4 are found to be invalid for any larger dimension. This means that extremal graphs of degree 10 and above must differ in structure from the solutions for degree 9 and below. Nevertheless it is hoped that elements of the approach presented in this paper will be useful in the search for extremal circulant graphs of higher dimension. It is conjectured that such graphs will continue to have maximal zones where the proportion of levels in the zone depends on the degree but is independent of the diameter. It is also conjectured that the total number of type T 1 vertices will be a linear function of the diameter with coefficient 4, independent of the degree. However it is quite possible that the graphs will not have odd girth that is maximal, 2 k +1.
The question also arises whether such an analysis of distance partitions and vertex types
might provide the basis for a proof that the largest known circulant graphs of degree 6 to 9 are extremal for all diameters above their known thresholds.
## References
- [1] S. Chen and X.-D. Jia. 'Undirected loop networks', Networks, Vol 23 (1993) 257-260
- [2] R. Dougherty and V. Faber. 'The degree-diameter problem for several varieties of Cayley graphs, I: The Abelian case', SIAM Journal on Discrete Mathematics, Vol. 17, No. 3 (2004) 478-519
- [3] C. Godsil and G. Royle. Algebraic Graph Theory, New York, Springer-Verlag (2001)
- [4] R. R. Lewis. 'The degree-diameter problem for circulant graphs of degree 8 and 9', arXiv:1404.3948 (2014)
- [5] H. Macbeth, J. Siagiova and J. Siran. 'Cayley graphs of given degree and diameter for cyclic, Abelian and metacyclic groups', Discrete Mathematics 312 (2012) 94-99
| null |
[
"Robert Lewis"
] |
2014-08-05T14:45:11+00:00
|
2014-08-05T14:45:11+00:00
|
[
"math.CO"
] |
Distance partitions of extremal and largest known circulant graphs of degree 2 to 9
|
This paper considers the degree-diameter problem for extremal and largest
known undirected circulant graphs of degree 2 to 9 of arbitrary diameter. As
these graphs are vertex transitive it is possible to define their distance
partition. The number of vertices in each level of the distance partition is
shown to be related to an established upper bound for the order of Abelian
Cayley graphs. Furthermore these graphs are all found to have odd girth which
is maximal for their diameter. Therefore the type of each vertex in a level may
be well-defined by the number of adjacent vertices in the preceding level. With
this definition the number of vertices of each type in each level is also shown
to be related to the same Abelian Cayley graph upper bound. Finally some
implications are discussed for circulant graphs of higher degree.
|
1408.0989v1
|
## Delivery of Liquid Metal to the Target Vessels as Vascular Embolic Agent to Starve Diseased Tissues or Tumors to Death
Qian Wang 1,3 , Yang Yu 1,3 and Jing Liu 1,2 *
1. Department of Biomedical Engineering, School of Medicine,
Tsinghua University, Beijing 100084, China.
2. Technical Institute of Physics and Chemistry, Chinese Academy of Sciences,
Beijing 100190, China.
3. These authors contributed equally to this work.
Correspondence should be addressed to J.L. ([email protected])
## Abstract
Tumor growth strongly depends on the continuous blood and nutrients supply. Theoretically, it is an ideal therapeutic way of treating tumor by vascular embolization, which also reduces the harmful effects on surrounding normal tissues. However, most of the currently available vascular embolic agents are still rather insufficient in performances to fulfill the real clinical need due to the reasons like: incomplete filling of target vasculature, being easily washed away by blood or body solution, or just producing toxicity to human body. Here from an alternative way, the body temperature liquid metal, a kind of soft and highly compliant material, was proposed for the first time as a new conceptual blood vessel embolization strategy towards tumor physical therapy. With the unique capability of easy phase transition between liquid and solid state and sub-cooling behavior, such agent can be fluently injected into and to fully fill the tiny vessels including the ending capillaries. The in vitro cytotoxicity experiments have been performed which indicated that treating localized diseased tissues through liquid metal embolic agent is acceptable. Endowed with a high density, the liquid metal-filled vessels are highly visible under the CT scan, which offers the potential of the diagnosis-treatment integration with only one single material. To further demonstrate the new liquid metal vascular embolization therapy, several typical experiments in vivo on vasculatures of rabbit ears and mouse tails have been performed to provide the evidences of destroying the targeted tissues due to liquid metal filling. To interpret the liquid metal starvation therapy effects, theoretical model has been established to simulate the tumor growth with zero, partial or complete filling of the metal agent inside the vessels. All the results support that, given appropriate administration, the liquid metal embolization is able to destruct the target regions effectively and might be able to starve the tumors to death through a
relatively easy and compliant way. This study lays the foundation of a promising tumor starvation therapy with both high contrast image guidance capability and overall treatment outcome.
Keywords: Liquid metal; Tumor treatment; Blood vessel; Embolic agent; Starvation therapy; Diagnosis-therapy integration.
## 1. Introduction
Originated from somatic mutations, the tumor cells generally proliferate and differentiate continuously under the regulation of various growth factors as well as the oxygen and nutrients provided throughout the blood vessels [1-4]. At the early stage when the death and the regeneration of tumor cells are nearly balanced, the existing vascular network near the tumor tissues supplies nutrients and carries away metabolites. Afterwards, a new tumor vascular network generates around the tumor tissues due to the vascular endothelial growth factor secreted by the tumor cells [5]. Then, the rapidly increasing tumor cells would diffuse to other parts via the tumor-induced angiogenesis. Clearly, the blood vessels play a significant role on the growth and metastasis of the tumor [6].
Considering the dependency of the tumor growth on the blood vessels, the treatment of vascular targeting therapy for starving the tumors to death has increasingly attracted the attentions of researchers across the world [7, 8]. One of such strategies is to prevent the tumor-induced angiogenesis by the angiogenesis inhibitors, which was originally proposed by Folkman in the early 1970s [9]. Ever since then, various inhibitive drugs have been developed [10-13], and such embolization therapy has also been combined with chemotherapy or radiotherapy for better treatment [14]. However, the angiogenesis inhibitors have to be used for a long time. And some studies found that within this period, the tumor itself would gradually generate resistance to these drugs, which makes it not as effective as expected [15]. As another approach focusing on blocking the existing vessels, colchicine was reported to destroy the vessels, yet with rather high toxicity [16]. Afterwards, some other vascular disrupting agents were studied including small molecular agents [17, 18].
Except for these drugs, there are some other physical embolization ways to offer the possibility of starving the tissues or tumor to death by means of embolic agents,
such as autologous clots, coils, gelfoam, balloons, glue, nanoparticles, microspheres embolic agents and so on [19-23]. However, the failure to embolize the vessels may often occur. One key reason lies in that some solid embolic agents cannot conform well to the vessels, leaving certain space at the interface for nutrients to leak in. Another existing issue is that some embolic agents are easily rushed away by the blood flow, such as the balloons. What's more, the catheters or other more complex equipments are needed in order to place some of embolic agents into the target positions. Different from these approaches, gas embolotherapy uses gas bubbles to occlude the blood flow [24]. The bubbles originate as small diameter liquid droplets of dodecafluoropentane (DDFP) mixed in saline and albumin. After injected into the vascular system, these droplets are small enough to pass through the capillary beds. At the strategic location, the droplets are vaporized into larger bubbles through acoustic wave to occlude flow. However, the entry of gas into the vessel system is a risk in virtually all areas of clinical care [25].
In view of the above tumor treatment situations, this paper is dedicated to proposing and demonstrating an alternative vascular embolization therapy, the body temperature liquid metal, as a kind of unconventional vessel embolic agent for starving the target tissues or tumors to death. Here, to emphasize the method itself, although plenty of liquid metal candidates can serve as the treatment agents, we mainly tested for the first trial the pure gallium and Ga75.5In24.5 alloy as the illustration examples. To characterize the physical behaviors of such metal agent, the DSC experiments on the liquid metal gallium were conducted, which revealed the thermal properties of the gallium and verified its feasibility for the delivery of therapy. The cytotoxicity of the gallium and indium in vitro were also evaluated with the tests of CCK-8 and flow cytometry. As an additional advantage, the liquid metal in vessels showed high image contrast under the X-ray scan, which indicated its great potential of the diagnosis-treatment integration. For further investigation, we experimented on several different tissues or transplanted tumor to demonstrate the liquid metal embolization effect in vivo. Finally, we presented the theoretical modeling and simulation of tumor growth affected by liquid metal enabled blockage of blood vessels near the tumor. This study opened the way of applying the liquid metal vascular embolization for a potential tumor starvation therapy in the coming time.
## 2. Principle and methods
## 2.1 Basic feature of liquid metal vascular embolic therapy
Figure 1 The principle illustration of liquid metal based tumor vascular embolization therapy. (a) The oxygen and nutrition supplies from the vessels to tissues without any embolic agents . (b) The incomplete occlusion. (c) The complete occlusion by the liquid metal . (d) The physical occlusion of the blood supply to tumor. (e) The liquid metal could be injected into or sucked out of the vessel in case of need.
As a soft, compliant and fluidic material, the liquid metal can be easily injected and filled into the vessels and capillaries sufficiently, which stops the blood flow and all the related physiological events thus enabled because of its high density and fluidity. Fig. 1 illustrates the basic principle of the liquid metal embolization therapy and surgical operation of its injection and removal. As is widely known, tissues acquire oxygen and nutrition from the nearby feeding vessels (Fig. 1a). Though the vascular blockage can shut down the blood supply, yet the stream is still able to pass through the vessels interior when the embolic agent fails to completely occlude the vessels. As a result, the reduction of blood flow does not guarantee a strong enough effect on starving the necrotic cells (Fig. 1b) and the embolic agent can even possibly be rushed away within a short period after it was injected. Being able to maintain in
liquid state at body temperature, the liquid metal agent can easily shape itself along the vessels and fully occlude the channels due to its perfect compliance (Fig. 1c). Compared with the normal tissues, the tumor is more dependent on the nutrients supplies from the surrounding vessels. After injected into the vessels, the liquid metal results in a physical occlusion and thus leads to tumor regressions (Fig. 1d). Another merit of the current therapy still lies in that, theoretically, the liquid metal could also be well sucked out in case of need after completing its treatment role (Fig. 1e).
## 2.2 Characterization of the physical properties of the liquid metal agents
We measured the thermal property of gallium (99.999%, purchased from Anhui Rare New Materials Co. LTD, China) with the differential scanning calorimetry device (DSC 200 F3 Maia, Netzsch Scientific Instrument Trading Co. Ltd, Germany). The 14.63mg of gallium was put into the alumina crucible with another empty one as the reference, which were both placed in a heating furnace. The test temperature range included three cycles, which were set as 40° C → 60° C for 1min → -60° C for 1min → 60° C for 1min → -60° C for 1min → 60° C for 1min → -60° C for 1min → 60° C for 1min → 45° C → 70° C, and the heating and cooling rates were both prescribed at 10° C /min. The results were analyzed by the software of NETZSCH Proteus Thermal Analysis.
## 2.3 Evaluation of the metal agents cytotoxicity in vitro
Two tests including the cell counting kit-8 (CCK-8) and the flow cytometry were applied on the mouse embryonic fibroblast cells (NIH3T3) to evaluate the cytotoxicity of the liquid metal.
A drop of the gallium was put on the six-well plates filled with culture solution, soaking for 24 hours and 48 hours in the incubator. Meanwhile, a piece of polished and sterilized indium and copper were soaked under the same condition.
CCK-8 method: Cells at logarithmic growth phase were treated by trypsin and tuned into single cell suspension with a concentration of 50000/ml, then seeded in a 96-well plate with 100μL in each well. After being cultured for 24 hours, the cells were observed under the microscope and the samples in bad state were eliminated. The original culture solution was removed and the wells were washed with PBS solution. Then, the soaked solution of gallium, indium, copper and the original culture
solution (control group) were added into the wells with 5 parallel samples in each group. After being cultured for 24 hours, each well was washed and added with a mixture of 100μL fresh culture solution and 10μL CCK-8. The 96-well plate were put into the incubator for 2 hours, then optical density (OD) at 450nm wavelength were measured with multifunctional microplate reader (EnVision, USA PerkinElmer). The cell viability was calculated according the following equation:
$$\text{Cell viability} = (OD-OD_{0})/ (ODc-OD_{0}))$$
where, ODc was the OD value of the control group and OD0 was the OD value of black wells without any cells.
Flow cytometry method: The cells were seeded into the sterile dish of 60mm diameter with 300000 cells each. After cultured for 24 hours in the incubator, the original culture medium was removed and the wells were washed with PBS solution. Then, the three soaked solutions and the original culture solution were added separately. After cultured for another 24 hours, the cells in each dish were collected into the centrifuge tubes. By centrifuge at 1500r/min for 5min, the supernatant was removed and 1mL cell staining buffer and 5uL PI staining solution were added, then a 30-min ice bath followed. Finally, the cell solution were placed in the flow cytometer, detecting the red fluorescence by flow cytometry (BD Calibur, the USA BD).
## 2.4 Liquid metal angiography of the tumor vessels
It was recently found that, different from other embolic agents, liquid metal has excellent visibility under the X-ray [26]. Thus in a further study, the tumor vessels injected with the liquid metal went through a CT scan to investigate the distribution of the material. In order to assess this special characteristic, several representative objects such as in vitro swine kidney, a healthy 8-week-old female CD1 mice and an 8-week-old female Barb/c nude mice (purchased by the Center of Biomedical Analysis of Tsinghua University) with a tumor were selected to serve as the experimental testing cases.
About 0.8mL liquid metal gallium was infused into the artery branches of the swine kidney. Then the filled kidney was X-ray photographed by a micro-CT scanner (XM-Tracer-130, from Institute of High Energy Physics, Chinese Academy of Sciences).
The mouse has both arteries broad enough for liquid metal Ga75.5In24.5 alloy infusion and thin capillaries to display the alloy's high contrast. First, the mouse was
excessively anesthetized to death, then the heart was exposed by cutting open its chest.
A thin syringe needle was used to penetrate into the left ventricle of the mouse heart and it was clamped to prevent leakage and reverse flow. With a slight cut near the right atrium, the liquid metal alloy could finally be infused into the circulatory tunnels and flow throughout the body. When the silvery alloy flows to the right heart, it was supposed to finish the infusion and that the sample had been stuffed with the contrast agent. Then the sample was taken to the imaging instrument (Super Nova CT, from Technical Institute of Physics and Chemistry, Chinese Academy of Sciences). Images with higher resolution were obtained from the micro-CT (XM-Tracer-130, from Institute of High Energy Physics, Chinese Academy of Sciences).
The structures of the vessels in the tumor tissues and normal tissues may be different. In order to offer more evidences on various tissue vasculatures, we also managed to infuse the liquid metal Ga75.5In24.5 alloy into the tumor vessels. The mouse breast cancer cells (EMT6, 1x10 cells) were injected directly into the subcutaneous at 6 the back of the nude mice. The tumor reached approximately a diameter of 5mm in two weeks and the vessels were visible on the skin. After the mice was anesthetized with intraperitoneal injection of 0.2mL 1% (g/L) sodium pentobarbital (Nembutal), the liquid metal was injected directly into a tumor vessel via a special syringe needle with an inner diameter of 0.3mm. Afterwards, the tumor region went under a CT scan (Super Nova CT, from Technical Institute of Physics and Chemistry, Chinese Academy of Sciences) for visualizing the vessels injected with liquid metal.
## 2.5 Liquid metal as embolic agents of blood vessels in vivo
A two-month-old male New Zealand rabbit (purchased from the Center of Biomedical Analysis of Tsinghua University), weighted about 2.2kg, were used as the experimental subject, because the vessels in the ears are broad and visible. Anesthetized with intraperitoneal injection of 4% (g/L) sodium pentobarbital (Nembutal) by the standard of 1 ml/kg, the rabbit was injected with liquid metal gallium into the vein of one ear. It was observed that the material filled the vessels at the rim and the tip of the ear quickly. The other ear was used for contrast. The infrared camera (Thermovision A40, FLIR, Wilsonville, OR) was adopted to record the temperature changes of the rabbit ear during the operational process. Afterwards, the rabbit was observed for half an hour and then put back to the caring center. The rabbit continued to be observed every day. Blood analysis were carried out before the
experiments and 1 day, 3 days, 7 days afterwards with the automatic hematology analyzer (BM800), each time with 3 repeated trials. After the occurrence of necrosis at the ear tip, the rabbit was sacrificed through standard procedure. Then the tissues of both the necrotic region and the same position of the other ear were surgically cut and immediately restored in formalin for 24 hours. The samples were then sectioned, stained with HE methods (Hematoxylin-eosin) and observed under the microscope.
Afterwards, we injected the liquid metal gallium into the vessels of the mouse tail and the rabbit leg in the same way. All experiments above have been approved by the Ethics Committee of Tsinghua University, Beijing, China under contract [SYXK (Jing) 2009-0022].
## 2.6 Theoretical modeling and simulation of tumor growth with liquid metal embolization
Although the liquid metal embolization of blood vessels in vivo was implemented on the vessels of rabbit ears and legs, a case of tumor might be more illustrative. Here, a theoretical modeling and simulation of tumor growth affected by the metal blockage of blood vessels nearby was (see Fig. 2) carried out to interpret the starvation effect of the current therapy.
Figure 2 Illustration of the theoretical tumor growth model. (a) Without the liquid metal in the vessel. (b) With the liquid metal embolic agent in the vessel.
Without losing any generality, there are generally four dependent variables to dominate the tumor growth model, including cancer cell density, oxygen density, extracellular matrix density and matrix degrading enzymes. The functions of each variable (further details about the functions could be seen in Alexander's work [27]) in non-dimensional are expressed as
$$\frac { \partial n } { \partial t } = d _ { n } \nabla ^ { 2 } n - \rho \nabla \cdot ( n \nabla f )$$
$$\frac { \hat { \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \boldsymbol \Psi } } { \hat { \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm } } = d _ { o } \nabla ^ { 2 } o + \nu f - \omega n - \varphi o$$
$$\frac { \partial f } { \partial t } = - \eta m f$$
$$\frac { \hat { \sigma } m } { \hat { \sigma } t } = d _ { m } \nabla ^ { 2 } m + \kappa n - \sigma m$$
where n , o , f denote the density of cells, oxygen and extracellular matrix respectively, and m denotes the matrix degrading enzymes .
The oxygen is supplied by the vessel and we make the assumption that the oxygen concentration is zero at the metal blockage position. The probabilities of movement of an individual cell are generated by discretizing the Equation (2).
Table 1 The non-dimensional parameter values used in the simulations
| Parameters | Value | Parameters | Value |
|------------------------------|------------|--------------|---------|
| n d | 0.0005 | | 1 |
| 0 d （ out of /in the tumors | 0.05/0.025 | | 0 |
| m d | 0.0005 | | 0.5 |
| | 0.01 | | 0.57 |
| | 50 | | 0.025 |
Besides, it is assumed that the cancer cells migrate through a combination of diffusion and haptotaxis as well as undergoing proliferation. And each cell is allowed to migrate when the number of neighbor cells is more than one. Cell death can occur due to lack of oxygen. In the model, we also make the assumption that each cell produces two daughter cells when the cell has reached maturity. If no empty space around the parent cell exists, the cell becomes quiescent until space appears. We assume that quiescent cells only consume half amount of oxygen of the active tumor cells. The tumor cells would not only maintain one phenotype, but also becomes more
unstable and more aggressive phenotypes due to mutations. In the model, we mainly consider three phenotypes. The irreversible mutations occur at the mitosis stage.
The computation domain is divided by 200× 200 grids with a space step of 0.005and a time step of h 0.0001. Initially, 36 tumor cells are in the center. The k initial value of oxygen density, extracellular matrix density and matrix degrading enzymes are 1, 1 and 0, respectively. The non-dimensional parameter values used in the simulations are listed in Table 1 [27].
## 3. Results
## 3.1 Physical properties of the liquid metal
Fig. 3 depicts the DSC curves of the gallium in reciprocally reverse processes. In either the cooling or the heating process, the measurement goes for three trials and the results are almost identical. There is a very sharp exothermic peak over the cooling process, and the freezing temperature is 4.5 ° C, while the endothermic peak is also steep in the heating process and the melting point is 28.3° C. The curves indicate an easy transition between the solid and liquid phase of the gallium, which is an excellent property for the phase control. What's more, owing to its sub-cooling effect, the material is able to maintain in the liquid state for a long time at the body temperature. Thanks to its liquid phase, compliance and fluidity, the material is able to fluently flow into and fill the tiny vessels including the capillaries whose diameter is at a scale of micrometers.
3.5-
-0.5
20
Figure 3 The DSC curves of the gallium. (a) Cooling process. (b) Heating process.
## 3.2 Cytotoxicity of the liquid metal in vitro
Fig. 4a shows the viability of the cells after cultured in different metal soaked solutions with CCK-8 method. The cell viability in the copper soaked solution in 24 hours was measured as 7%, and almost all died in 48 hours. The cell viability of the gallium soaked solution in 24 hours and 48 hours were 100.6% and 75.7%, and that of indium were 107% and 79.7%, respectively. According to the cytotoxicity evaluation criteria of the national standard, cell viability above 75% is qualified and safe. The viability of the cells can also be visually demonstrated through the microscope, which is shown in Fig. 4b. The cells of the gallium, indium soaked solution and original culture solution were exactly of the right size and shape large, while those in the copper soaked solution distributed sparsely and the shapes were relatively round, which means most of them were dead.
Figure 4 Cytotoxicity in vitro. (a) The cell viability after cultured in different metal soaking solution with CCK-8 method. (b) The cellular morphology in each experimental group observed under the microscope. (c) The cell apoptosis with flow cytometry method. (d) The scatter images with flow cytometry method.
Fig. 4c indicates cell apoptosis of each experimental group. In the PI fluorescence intensity histograms, M1 stands for apoptosis cells while M2 represents living cells. Compared with the control group, the relative apoptosis rates of gallium and indium group was 10.18% and 7.9% respectively while 78.44% for the copper group. In the scatter images in Fig. 4d, it can be further found that the strength of forward scattered light reduced and the one side scattered light increased in the copper soaked solution, which also indicated that there were more apoptosis and necrosis cells. However, the gallium and indium group showed tiny difference with the control group.
These experiments have revealed that the cytotoxicity of the gallium and indium ion are relatively low and their applications on localized severe disease treatment are clinically acceptable, given its unique values as will be revealed in later sections.
## 3.3 Liquid metal angiography of the tumor vessels
Endowed with a high density, the liquid metal is an excellent contrast agent for the vessel-like tissues under the X-ray [26], which is a unique advantage compared with most other embolic agents.
The X-ray images of renal artery of the in vitro swine kidney are presented in Fig. 5a. Clearly, a whole branch of the renal artery network is intact and the texture of the small vessels is rather clear. As shown in Fig. 5b, the vessels are separated with the other tissues obviously in highlight in the CT scan reconstruction of a whole-body infused mouse. In the image, it is easy to identify the liquid metal's flow path throughout the carotid, limb and tail vessels even to the end tips. Besides, the branches in the lung, head and abdominal organs can also be distinguished with ultra-high contrast and clarity. In order to offer more evidence of the image guidance ability of the liquid metal alloy enhancement, Fig. 5c provides the closer X-ray photographs of vascular network in the abdomen of the mouse which has plenty of even tinier vessels. In the figure, the sensor focused in a square field with each side 6.6mm long, and the resolution is thus set as 13μm/pixel. Accordingly, it can be referred to that the thinnest visible vessel is less than 30μm wide, which means a tremendous improvement of the blood vessel visualization. Besides, the tumor vessels filled with liquid metal are highly visible under the CT scan as expected. As can be seen, the tumarized vasculature appeares somewhat irregular (Fig. 5d) as compared with normal tissue vessels (Fig.5a). This clearly reflected the growth behavior of such diseased vessels.
Figure 5 The images of vessels filled with the liquid metal under the X-ray. (a) The X-ray image of renal artery of the in vitro swine kidney. (b) The top view of a liquid metal infused mouse in a whole-body CT scan and 3D reconstruction. (c) The X-ray image of vessels in the mouse abdomen with higher resolution. (d) 3D reconstruction image of tumor vessels.
## 3.4 Liquid metal embolization to starve target living tissues to death
As shown in Fig. 6a, the liquid metal has filled the rabbit ear vein at the rim and the tip. Without abnormal reactions, the rabbit showed obvious symptoms of necrosis only in the ear tip in three weeks, which finally come out like a dry leaf. The necrotic changes appeared not so serious at the rim of the ear, which might be due to the supply of certain surrounding tiny vessels. However, the very side of the ear had been crimped, which was a reflection of the liquid metal obstruction.
As an attendant phenomenon, the temperature of the liquid metal obstructed region would decrease due to stop of the blood flow. As a special evaluation of the material's influence, the temperature distributions in 0h, 24h, 48h and 72h from the thermal infrared images are depicted in Fig. 6b. Clearly, at the moment of the injection, the blood was immediately blocked, causing the temperature rise in the upstream and decrease in the downstream regions. With the time went on, the area of the low temperature region shrunk gradually, which indicated the concentration of the material to the tip.
In order to provide a biological inspection, Fig. 6c displays the sections of both the necrotic and normal tissues. It can be seen that near the liquid metal blocked
vessels, the surrounding area tends to be emptier and the cells are much smaller than the situation in the similar region in a normal tissue. This shows the damage effect of the liquid metal embolization treatment.
Figure 6 The effects of liquid metal embolization at the rabbit ears. (a) The symptoms of necrosis in the ear. (b) The temperature distribution changes after 0h, 24h, 48h and 72h. (c) Comparison between the necrosis tissue sections and control normal tissue section. (d) The results of WBC, LYM%, MON%, GRA%, HGB, RBC and PLT * changed over time. (e) WBC, RBC and PLT dynamic histograms distribution.
* WBC: white blood cell; LYM: lymphocyte; MON: monocyte; GRA: granulocyte; HGB: hemoglobin; RBC: red blood cell; PLT: platelet.
Other than the morphologic analysis, blood tests were also carried out in a week and seven parameters including WBC, LYM%, MON%, GPA%, HGB, RBC and PLT are shown in Fig. 6d. Fig. 6e shows the WBC, RBC and PLT dynamic histograms distribution, respectively. As can be seen from the figure, the second peak of WBC
distribution was lowered after one day but returned to normal after 3 days, which indicated the reduction of the neutrophil granulocytes. These blood analysis results demonstrated that the blood parameters would soon return to normal level after the liquid metal gallium was injected into the blood.
Figure 7 The effects of liquid metal embolization at the mice tail. (a) The liquid metal injected into the tail vein. (b) The regenerated vascular branches.
It should be pointed out that, not all the injections will be able to completely fill the vessels and thus lead to damage of the target tissues. The output in fact depends on the specific surgical injection and delivery of liquid metal. To explain this point, we select the tail vessels of the mice as an additional test object to evaluate the blockage effect of liquid metal embolization on the tissues in vivo. As shown in Fig. 7a, three tail vessels of the mice were filled with liquid metal gallium. However, there was no visible necrosis after three weeks later. Moreover, certain new small blood vessels were even observed to regenerate (Fig.7b). The reason can be interpreted as follows. With a single flow direction and different from former cases, it was more difficult in current situation to fully block the femoral vessels at the tail of the mice. This led some of the liquid metal to flow away and no visible necrosis occurs at the target tissues over the whole experiments. Besides, these regenerated blood vessels near the tail vessels have further contributed for the nutrition and oxygen supply. This reminds us that, without appropriate administration of injecting the liquid metal agent, the embolization effect as anticipated before may not always necessarily be guaranteed. Therefore, for therapeutic purpose, careful treatment planning should be made in
advance to surgically deliver liquid metal into the target vessels along specific directions. Lastly, it was also worth mentioning that these mice filled with liquid metal were kept alive for a rather long time like several weeks. Such findings indicate the safety and non-toxicity of the liquid metal materials staying inside the vessels on condition that they will not cause physically physiological danger.
## 3.5 Theoretical results of tumor growth with liquid metal embolic agents
Theoretical simulations on the two cases with and without liquid metal embolic agents in the vessels near the tumor have been implemented to interpret and evaluate the performance of tumor growth suppression by the present method. Fig. 8 depicts the changes of oxygen concentration near the tumor at six times (5 days, 9days, 11days, 13days, 15days and 17days later). The tumor growth consumes oxygen and the oxygen concentration in tumor center is the lowest. In order to maintain growth, the tumor needs adequate oxygen supply by the vessel. However, as a result of liquid metal blockage in the vessel, the oxygen concentrations of the nearby tissues are decreasing. Fig. 9 shows the predicted results for tumor growth. Different color represents different phenotype cells, and red is the most aggressive cells. The tumor grows towards the vessel since there exists only one single vessel on one side. The tumor cell distribution shows a dead central region, which is due to lack of the oxygen as shown in Fig. 8. For the case without liquid metal embolic agent, the dead central region occures after the 17 days, while for the case with liquid metal embolic agent, the dead central region occurs after the 11 days. As the time goes on, the dead central regions continue to enlarge.
Figure 8 The oxygen distribution with the change of time. (a) Without the liquid metal in the vessel. (b) With the liquid metal embolic agent in the vessel.
Figure 9 Spatial distribution of tumor cells with the time. (a) Without the liquid metal in the vessel. (b) With the liquid metal embolic agent in the vessel.
## 4. Discussion
As is well known, the abundant vessels in the tumor tissues make it difficult for tumor treatment. Curatie surgery can easily lead to a massive hemorrhage, and large blood vessels would produce localized cooling in heated tissues during tumor hyperthermia, for instance. In this study, body temperature liquid metal yields a promising injectable tumor treatment. With its easy phase conversion, such material owns the capability to flow into tiny vessels at liquid state and stay at the target tissues at the solid state. Once the liquid metal was injected into the vessels, the nearby tissues would be possibly starved to death due to cutting off the feeding route of the nutrition and oxygen supplies.
Though the unprecedented application of the liquid metal on the tumor treatment seems to be beyond imagination, there remains a wide range for the further utilization of the material. In the present experiments, only gallium and Ga75.5 In24.5 alloy were tested for brief. In fact, the melting points of the liquid metal could be modified by adding some other metals into alloys and adjusting the compounds ratio.
As a metal, the gallium is endowed with good electrical and thermal conductivity. Hence on the base of the blocking therapy, newer treatment might still be developed through combining the electrical and thermal stimulation together. We could even add certain substances into the liquid metal to obtain or enhance specific properties. Ma and Liu proposed a method to fabricate liquid metal with desired properties by loading with nanoparticles [28]. These nanoparticles include magnetic oxides such as Fe3O4, metallic particles and some semi-conductive particles, which further expands the application of liquid metal as the embolic agents. What's more, it is possible to
combine this approach with chemotherapy or radiotherapy by injecting mixed liquid metal with chemotherapy and radiotherapy substances.
In this study, it is worth emphasizing the role of the liquid metal angiography, which reveals an extra advantage of the vessel blocking therapy. The visibility of the material under the X-ray has provided the possibility for monitoring the flow process along the gallium injection and evaluating vascular changes and damages. The combination of the liquid metal blocking and the X-ray imaging even opens a promising approach to the investigation of the tumor vessel growth and distribution, which also indicates a great potential of the diagnosis-treatment integration.
However, there still exist many issues to be solved in order to further improve the efficiency of the liquid metal embolization. Considering the difficulties to inject the liquid metal into the artery due to fast blood flow there and deep position in the tissues, we almost administrated these experiments on the veins. However, as we all know, the blood in artery flows into the capillary vessel, while that in vein blood flows out of the capillary vessel. Therefore, it is better to inject the liquid metal into the artery to occlude the vessel. Besides, the regeneration of the tumor vessel network under the obstructive condition is also a common challenge in the blocking therapy.
The safety is also a non-negligible issue in this new conceptual blood vessel embolization strategy. As mentioned above, some of the liquid metal could be washed away from the target tissues since injection from the veins. As is shown in Fig. 10, the liquid metal has entered the heart and lungs of the rabbit in the experiments of leg infusion yet the rabbit is still alive. Therefore it is very important to adopt measures to prevent this problem from happening in further studies. One possible solution is to apply proper pressure on the proximal end to slow the blood flow velocity. Another method to decrease the blood flow is to reduce the temperature of the blood in the target tissues, which could also speed up the solidification of the liquid metal. It also indicates that the in vivo application is highly sophisticated due to safety control and thus skilled surgical operation is required as well. In this paper, we only used gallium and Ga75.5In24.5 alloy as the embolic agents, whose melting point are both under the 37° C. Next, we will try to synthetize more other liquid metal alloys with a little higher melting point than 37° C and test their embolization effects. In addition, a comprehensive evaluation of long-term toxicity still requests further research.
Figure 10 Traces of the liquid metal in the heart and lungs of the rabbit.
## 5. Conclusion
In summary, the liquid metal, as a fluidic material at body temperature, has been proposed for the first time as the blood vessel embolic agent to starve the necrotic tissues or target tumors to death. With its merits of easy phase transition and sub-cooling, this kind of material can be conveniently injected into the vessels with relatively easy operations. A series of in vitro experiments on the material cytotoxicity have been implemented to evaluate its safety and the results have indicated that both gallium and indium irons show low toxicity on normal cell growth. On considering its application against tumor or other severe diseases, such a performance is quite acceptable. As an extra advantage, the liquid metal is endowed with excellent contrast under the X-ray over the tissues, which provided a powerful soft tool for tumor vascular research and offered great potential to realize the diagnosis-treatment integration. Furthermore, both in vivo experiments and theoretical model simulations have preliminarily demonstrated the performance of the liquid metal to starve the tissues or tumors to death. Though still facing certain challenges, the unprecedented utilization of the liquid metal agent opens a new way for further practice in future tumor vessel blocking therapy.
## Acknowledgement
This work was partially supported by the NSFC under Grant 51376102.
## References
- [1] Esther W, Michael S, and Yosef Y. Roles for growth factors in cancer progression. Phys, 2010, 25(2): 85-101.
- [2] Zhimin L, Guo J, Peter B, and Tony H. Epidermal Growth factor-Induced tumor cell invasion and metastasis initiated by dephosphorylation and downregulation of focal adhesion kinase. Mol Cell Biol, 2001, 21(12): 4016-4031.
- [3] Wells A. Tumor invasion: Role of growth factor-induced cell motility. Adv Cancer Res, 2000, 78:31-101.
- [4] Wells A, Kassis J and Solava J, Turner T and Lauffenburger DA. Growth factor-induced cell motility in tumor invasion. Acta Oncol, 2002, 41(2):124-30.
- [5] Folkman J. Tumor angiogenesis factor. Cancer Research, 1974, 34(8): 2109-2113.
- [6] Jain R K. Vascular and interstitial barriers to delivery of therapeutic agents in tumors. Cancer Metast Re, 1990, 9(3): 253-266.
- [7] Eichhorn M E, Strieth S and Dellian M. Anti-vascular tumor therapy: Recent advances, pitfalls and clinical perspectives. Drug Resis Update, 2004, 7(2): 125-138.
- [8] Thorpe P E. Vascular targeting agents as cancer therapeutics. Clin Cancer Res, 2004, 10(2): 415-427.
- [9] Folkman J. What is the evidence that tumors are angiogenesis dependent? J Natl Cancer I, 1990, 82(1): 4-7.
- [10]Kerbel R S. Tumor angiogenesis: Past, present and the near future. Carcinogenesis, 2000, 21(3): 505-515.
- [11]John A and Tuszynski G. The role of matrix metalloproteinases in tumor angiogenesis and tumor metastasis. Pathol Oncol Res, 2001, 7(1): 14-23.
- [12]Kim K J, Li B, Winer J, Armanini M, Gillett N, Philips H S and Ferrara N. Inhibition of vascular endothelial growth factor-induced angiogenesis suppresses tumour growth in vivo. Nature, 1993, 362(6423): 841 - 844.
- [13]O'Reilly M S, Boehm T, Shing Y, Fukai N, Vasios G, Lane W S, Flynn E, Birkhead J R, Olsen B R and Folkman J. Endostatin: an endogenous inhibitor of angiogenesis and tumor growth. Cell, 1997, 88(2): 277-285.
- [14]Feldman A L and Libutti S K. Progress in antiangiogenic gene therapy of cancer. Cancer, 2000, 89(6): 1181-1194.
- [15]Quesada A R, Medina M A and Alba E. Playing only one instrument may be not enough: Limitations and future of the antiangiogenic treatment of cancer.
Bioessays, 2007, 29(11): 1159-1168.
[16]Philip E T, David J C, and David C B. The first international conference on vascular targeting: Meeting overview. Cancer Res, 2003, 63(5):1144-1147.
- [17]Denekamp J. The tumour microcirculation as a target in cancer therapy: A clearer perspective. Eur J Clin Invest, 1999, 29(9): 733-736.
- [18]Hori K, Saito S and Kubota K. A novel combretastatin A-4 derivative, AC7700, strongly stanches tumour blood flow and inhibits growth of tumours developing in various tissues and organs. Brit J Cancer, 2002, 86(10): 1604-1614.
- [19]Lovet J M and Bruix J. Systematic review of randomized trials for unresectable hepatocellular carcinoma: Chemoembolization improves survival. Hepatology, 2003, 37(2): 429-442.
- [20]Liu L X, Zhang W H and Jiang H C. Current treatment for liver metastases from colorectal cancer. World J Gastroentero, 2003, 9(2): 193-200.
- [21]Yu H, Zhu G and Xu R et al. Arterial embolization hyperthermia using As2O3 nanoparticles in VX2 carcinoma-induced liver tumors. PLoS One, 2011, 6(3): e17926.
- [22]Chen X, Lv H, Ye M, Wang S, Ni E, Zeng F, Cap C, Luo F and Yan J. Novel superparamagnetic iron oxide nanoparticles for tumor embolization application: Preparation, characterization and double targeting. Int J Pharm, 2012, 426(1-2):248-255.
- [23]Teng M M, Chang C Y, Chiang J H, Lirng J F, Luo C B, Chen S S, Chang F C, Guo W Y. Double-balloon technique for embolization of carotid cavernous fistulas. Am J Neuroradiol, 2000, 21(9):1753-1756.
- [24]Joseph L B. Cardiovascular bubble dynamics. Crit Rev Biomed Eng, 2005, 33(4):299-346.
- [25]Claus M M and Erik S S. Gas embolism. New Engl J Med, 2000, 342(7): 476-482.
[26]Wang Q, Yu Y Pan K and Liu J. Liquid metal angiography for mega contrast X-ray visualization of vascular network in reconstructing in-vitro organ anatomy. IEEE Tran Biomed Eng, 2014, 61(7): 2161 - 2166.
[27]Anderson A R A. A hybrid mathematical model of solid tumour invasion: The importance of cell adhesion. Math Med Biol, 2005, 22(2): 163-186.
- [28]Ma K Q and Liu J. Nano liquid-metal fluid as ultimate coolant. Phys Lett A, 2007, 361(3): 252-256.
| null |
[
"Qian Wang",
"Yang Yu",
"Jing Liu"
] |
2014-08-04T15:09:35+00:00
|
2014-08-04T15:09:35+00:00
|
[
"physics.med-ph"
] |
Delivery of Liquid Metal to the Target Vessels as Vascular Embolic Agent to Starve Diseased Tissues or Tumors to Death
|
Tumor growth relies heavily on the continuous blood and nutrients supply.
Theoretically, it is an ideal therapeutic way of killing tumor by only vascular
embolization. However, most of the existing vascular embolic agents are still
rather insufficient to fulfill the real clinical need due to the reasons like:
incomplete filling of target vasculature, being easily washed away by blood or
body solution, or just producing toxicity to tissues. Here from an alternative
way, the body temperature liquid metal, a kind of soft and highly compliant
material, was proposed for the first time as blood vessel embolization agent
for tumor physical therapy. With its unique capability of easy phase transition
between liquid and solid state and sub-cooling behavior, such material can be
fluently injected into the tiny vessels including ending capillaries and fully
block them. The in vitro cytotoxicity experiments were performed which showed
that treating localized diseased tissues through liquid metal embolic agent is
acceptable. Endowed with a high density, the liquid metal-filled vessels are
highly visible under the CT scan, which offers the potential of
diagnosis-treatment integration. To further demonstrate the new conceptual
liquid metal vascular embolization therapy, several experiments on in vivo
vasculatures of rabbit ears and mouse tails were performed to provide evidences
of destroying the targeted tissues. To interpret the liquid metal starvation
therapy effects, a theoretical model was established to simulate the tumor
growth with zero, partial or complete filling of the metal agent inside the
vessels. All the results support that, given appropriate administration, the
liquid metal embolization is able to destruct the target regions and might
starve the tumors to death through a relatively easy way. This study lays the
foundation of a promising tumor starvation therapy in the coming time.
|
1408.0990v1
|
## Assessment of Response Time for New Multi Level Feedback Queue Scheduler
M.V. Panduranga Rao , K.C. Shet 1 2
1
Research Scholar, NITK, Surathkal , Mangalore, India 2 Professor, NITK, Surathkal, Mangalore, India
Abstract - Response time is one of the characteristics of scheduler, happens to be a prominent attribute of any CPU scheduling algorithm. The proposed New Multi Level Feedback Queue [NMLFQ] Scheduler is compared with dynamic, real time, Dependent Activity Scheduling Algorithm (DASA) and Locke's Best Effort Scheduling Algorithm (LBESA). We abbreviated beneficial result of NMLFQ scheduler in comparison with dynamic best effort schedulers with respect to response time.
Keywords - NMLFQ, scheduler, process, queue, deadline, preemption, priority, Multi Level, DASA, LBESA, Dynamic, Best Effort and algorithm.
## I. INTRODUCTION
In a multiprocessing environment of computer, the operating system's theme is to assign the execution time for various processes. The main component of kernel that creates the 'selection' is called the scheduler [1][4][31][44][51]. The representation of the policy of scheduler employed is called the scheduling algorithm. Numerous different scheduling algorithms have been deduced over the years.
## II. REVIEW OF DYNAMIC BEST EFFORT REAL-TIME SCHEDULING ALGORITHMS
At least two flavors of best-effort real-time scheduling algorithms exists. This comprises of the Dependent Activity Scheduling Algorithm (DASA) and Locke's Best Effort Scheduling Algorithm (LBESA) [2][6][38][42][49].
In fact, the DASA and LBESA scheduling algorithms are comparable with the canonic Earliest Deadline First (EDF) algorithm during under-loaded conditions. In this case, the EDF promises to meet entire deadlines and is always optimal. However, in the case of an over-loaded circumstance, DASA and LBESA attempt to increase the total task benefit [12][17][34].
The author R.K. Clark et al. [43], demonstrates that DASA more often, overtakes LBESA for the duration of over-loaded circumstances. Peng Li and Ravindran [41], confirms that DASA overtakes the Robust Earliest Deadline First (RED) scheduling algorithm [15][19][35].
Burns A et al. [10]. asserts that alternative features of DASA and LBESA are utilized for the development of MK7.3 kernel. R.K. Clark et al. [43], emphasizes that wide-ranging
characteristics of DASA and LBESA are exploited for the growth of Alpha real-time operating system [13][16].
According to R.K. Clark et al. [43], DASA and LBESA are the decent, benefit accrual scheduling algorithms. Both of these algorithms are employed to utmost extent for the growth of mission critical systems [25][26].
The characteristic and performance measurement parameters of each of the scheduler are known from previous work of others. The literature survey provided several outstanding noticeable features [9][14][27][32].
After studying policy mechanisms of different available schedulers, a New Multilevel Feedback Queue (NMLFQ) scheduling algorithm is proposed. NMLFQ includes all necessary modules to compete as a real time scheduler applied in embedded system domain [18][22].
The MLFQ principle of operation is used in NMLFQ scheduling algorithm in such a way that the response time is reduced and the functionality of the scheduling is improved [21][23][33]. The maximum number of queues and the quantum burst time for each queue are found using dynamic method. In NMLFQ scheduling, the operating system can modify the number of queues and the quantum of each queue according to the obtainable processes, deadline and as well as urgency consideration of the process (Garcia P. et al. [16]).
In Earliest-Deadline-First scheduling (EDF) priority level for every task is neglected. It primarily concentrates on completion of each task. It focuses on selection of task with nearest deadline for execution. Respectable CPU utilization will be gained, only when the chosen task is having precise deadline [3][20][36][42]. The canonic algorithm for real-time operating systems is EDF. It practices dynamic scheduling principle. The ready process from pool of priority queue is to be selected. The chosen process from the queue will be searched, for the process nearest to its deadline. The selected process is then scheduled to begin execution. The main drawback of the algorithm is, as and when the system is overloaded, the set of processes will miss deadlines and the system leads to deadlock.
Fig. 1, demonstrates the problem of missing of tasks in response to a deadline. The X axis corresponds to the values of time in milliseconds. Y axis corresponds to light weight
## International Journal of Computer Trends and Technology (IJCTT) - volume 13 number 3 - Jul 2014
processes or small threads. In the example of a load sharing of several time restraint jobs, response is expected at accurate time [5][7][39][47][53]. As the fig. 1, illustrates, request is made at 3 milliseconds. The response is expected within deadline period. Here a dispatch latency of 1 millisecond is suffered during the operation. Henceforth the response is sent after the deadline period. The response is sent at 4 milliseconds missing the deadline, in turn results in invalid response. The data object is passing some vital data to another object involving many processes and threads. We expect to have a response to a request within three milliseconds. The requirement will be perfect if the system reacts within two milliseconds. If the response comes after three milliseconds, the job execution crashes. There exists a sleek time of 1 millisecond from 3 to 4 milliseconds. So here the task missing the deadline results in invalid response.
p=Runtime.getRuntime().exec("ps -ef");
In general a scheduler is written precisely to respect application priorities. It is necessary for real time applications need to be developed with limited dispatch latency [8][10][28][40][45]. The term dispatch latency designates the amount of time a system takes to respond to a request for a process to begin operation. The complete application response time comprises of the interrupt response time, the dispatch latency and the application's response time. The system detects that a process with higher priority than the interrupted process is now ready to dispatch and dispatches the process. The time to switch context from a lower priority process to a higher priority process is comprised in the dispatch latency time.
The shorter version of the pseudo-code of NMLFQ scheduler is described in the following section. This includes dynamic priority of arriving processes. Construction of queue is performed by loading with, certain number of processes to carry out small jobs [11][20][29][37].
Depending on the load of jobs for the scheduler, the numbers of processes are dynamically varied in ready queue. We can designate several levels of queues, depending on waiting condition of processes and urgency conditions of tasks. At any instant of time and with any number of processes and jobs to be completed, scheduler must guarantee meeting the deadlines of the hard real-time tasks.
PUBLIC CLASS BEGIN : DynamicPriority
VARIABLE : Vector<Process> readyQueue =new Vector();
VARIABLE : Hashtable<Integer, Process> pQueue = new
Hashtable<Integer, Process>();
VARIABLE : public Process p=null;
VARIABLE : public Process removedProcess=null;
CONSTRUCTOR BEGIN : DynamicPriority()
//Initially loading Queue with some number of
processes.
MONITOR BEGIN :
FOR BEGIN :(int i=1;i<=pQueue.size();i++) //Creating process
insertIntoPQueue(i*10,p);
FOR END
MONITOR END
CATCH BEGIN (Exception ex)
ex.printStackTrace();
CATCH END
CONSTRUCTOR END
CONSTRUCTOR BEGIN : public DynamicPriority(int i)
MONITOR BEGIN :
//Removing the requested number of processes from
Queue
OUTER IF BEGIN (i<10)
FOR BEGIN (int j=0;j<i;j++)
removedProcess=removeAtEnd(j);
INNER IF BEGIN (removedProcess != null)
System.out.println("process is removed from
queue");
INNER IF END
INNER ELSE BEGIN
System.out.println("process removed from the
queue is null");
INNER ELSE END
FOR END
OUTER IF END
OUTER ELSE BEGIN
addAtFront(i, Runtime.getRuntime().exec("ps -
ef"));
OUTER ELSE END
MONITOR END
CATCH BEGIN :(Exception ex)
ex.printStackTrace();
CATCH END
CONSTRUCTOR END
- /* This method add the process to the readyQueue based on priority. It will check the priority and moves the other processes in readyQueue accordingly. */
METHOD BEGIN : public void addAtFront(int priority,
Process temp)
Enumeration enumitr = pQueue.keys();
WHILE BEGIN :(enumitr.hasMoreElements())
IF BEGIN : (!(priority
<(Integer)enumitr.nextElement()))
pQueue.put(priority, temp);
readyQueue.add(temp);
IF END
WHILE END
METHOD END
/* This method will insert some processes initially into Queue and also readyQueue(implementation details) */ METHOD BEGIN : public void insertIntoPQueue(int priority, Process process) //Adding process to the Queue and also into readyQueue
## International Journal of Computer Trends and Technology (IJCTT) - volume 13 number 3 - Jul 2014
IF BEGIN :(process != null)
pQueue.put(priority, process);
readyQueue.addAll(pQueue.values()); System.out.println("process is added into Ready
Queue");
IF END
ELSE BEGIN
System.out.println("process is null");
ELSE END
METHOD END
/* This method removes the process from the Queue. */
METHOD BEGIN : public Process removeAtEnd(int index)
//Removing the process from the index specified from both readyQueue and pQueue
readyQueue.remove(index);
RETURN : return(pQueue.remove(index));
METHOD END
- /* Algorithms guarantee that if a task is accepted for execution, the task and all previous tasks accepted by the algorithm will meet their time constraints [22][26][30][48]. The planning based algorithms attempt to improve the response and performance of a system to aperiodic and soft real-time tasks while continuing to guarantee meeting the deadlines of the hard real-time tasks. */
METHOD BEGIN : public void planningBased()
IF BEGIN (true)
//Task is accepted for execution
//Resource are allocated.
//Responds in time
IF END
METHOD END
METHOD BEGIN : public void bestEffortBased()
IF BEGIN : (true)
MONITOR BEGIN
Enumeration enumitr = pQueue.keys();
Vector sorted = new Vector();
WHILE BEGIN : (enumitr.hasMoreElements())
int dynArr =(Integer)enumitr.nextElement();
sorted.add(dynArr);
WHILE END
Collections.sort(sorted);
FOR BEGIN (int i=0;i<sorted.size();i++)
//Before calling shortest job first set the priority for
the job.
ExecutingProcess((Process)sorted.get(i++)); FOR END MONITOR END CATCH BEGIN (Exception ex) ex.printStackTrace(); CATCH END IF END
METHOD END
/* Actual execution started. */
METHOD BEGIN : public void ExecutingProcess(Process
process)
//Actual Execution happens
System.out.println("Execution started...");
METHOD END
/* We are creating pQueue as a placeholder for processes which are ready to execute. As it is DynamicPriorityQueue, we have to use HashTable for storing both priority and processes. Processes will be added based on the priority to the Queue and will removed from the end or again, based on priority form [24][46][50]. */
METHOD MAIN BEGIN : public static void main(String[]
args)
// TO-DO Auto-generated method stub
//Creating Queue
CALLING METHOD : new DynamicPriority();
//Removing processes from Queue
CALLING METHOD : new DynamicPriority(2);
CALLING METHOD : new DynamicPriority(110);
MONITOR BEGIN
CALLING METHOD : new
DynamicPriority().addAtFront(2,Runtime.getRuntime().exec(
"ps -ef"));
CALLING METHOD : new
DynamicPriority().planningBased();
CALLING METHOD : new
DynamicPriority().bestEffortBased();
MONITOR END
CATCH BEGIN(IOException e)
// TO-DO Auto-generated catch block
e.printStackTrace();
CATCH END
METHOD MAIN END
CLASS END
A job is a function of many tasks or processes. Division of jobs into task modules is performed from analysis of jobs. All the steps of NMLFQ scheduler are summarized in fig. 2. This model depicts the coordination amongst various modules of scheduler. Jobs are divided into task modules after the analysis of the nature of jobs. The history of jobs is maintained for future use. The processes are allocated to lock the resources with the control of interrupt handler. The task controller is made to govern for time stamping of deadline. It also supervises ready, pending, blocked and sleeping tasks. The evaluation of priority levels of processes, ordering of processes and division into time slices is controlled by scheduler through the task controller. The peripherals are associated to the CPU through input output controller. This occasionally makes use of context switching. Several queues are supervised by a substantial queue controller. As stated earlier, mapping of processes to critical resource is achieved by semaphores. Dynamic prioritization is a vital task of
scheduler, which necessitate raising or scaling of priority of processes.
## III. NMLFQ COMPARISON WITH BEST-EFFORT REAL-TIME SCHEDULING ALGORITHMS - DASA AND LBESA
The NMLFQ real time scheduler is compared with existing best-effort real-time scheduling algorithms. This comprise of the Dependent Activity Scheduling Algorithm (DASA) and Locke's Best Effort Scheduling Algorithm (LBESA) [20][52][54]. The comparison of NMLFQ scheduler with DASA and LBESA is accomplished for twenty test cases. The results of NMLFQ, DASA and LBESA for twenty different set of inputs are reckoned. The results provided ameliorate characteristics for CPU utilization, overall turnaround time, average turnaround time, average waiting time and average response time of schedulers. The comparative results are analyzed for several processes correspondingly, as shown in fig. 3 and generalized graph of hundreds of processes can also be drawn. Fig. 3, exemplifies, Average Response time for NMLFQ scheduler, compared with real time DASA and LBESA scheduers for twenty testcases, proves 10 to 25% reduction of response time. In this research paper, We have depicted the performance with respect to Average Response time for NMLFQ scheduler
Fig. 3, proves as per the achieved results, average response time for NMLFQ scheduler, compared with real time DASA and LBESA scheduers for twenty testcases, illustrates 10 to 25% reduction of response time in each subsidiary stage.
Y axis
Fig. 1 Task missing the deadline results in invalid response.
Fig. 2 Model of detailed conceptual view of NMLFQ including several modules.
Fig. 3 Average Response time for NMLFQ scheduler, compared with real time DASA and LBESA scheduers for twenty testcases, depicts 10 to 25% reduction of response time in each subsidiary stage.
## IV. CONCLUSIONS
In this research paper, we have discussed the themes associated with NMLFQ scheduler. The basic review of dynamic best effort real-time scheduling algorithms is explained. Comparison of proposed scheduler is made, with best-effort real-time scheduling algorithms -DASA and LBESA. Eventually, Model of detailed conceptual view of NMLFQ including several modules is also discussed in short.
## REFERENCES
- [1] Abdelzaher T. F., Sharma V., and Lu C.,'A Utilization Bound for Aperiodic Tasks and Priority Driven Scheduling', in IEEE Transactions on Computer , Volume 53, Number 3, Page(s): 334-350, 2004.
- [2] Albert Haque, 'An Analysis and Comparison of Processor Scheduling Techniques', University of Texas at Austin, [email protected], December 7, 2012.
- [3] Arezou Mohammadi and Selim G. Akl, 'Scheduling Algorithms for Real-Time Systems', Technical Report No. 2005-499, Natural Sciences and Engineering Research Council of Canada , School of Computing, Queen's University, July 15, 2005.
- [4] Ashiq Anjum, Richard McClatchey, Arshad Ali and Ian Willers, 'Bulk Scheduling With Diana Scheduler', in proceedings of IEEE Transactions on Nuclear Science, ISSN 0018-9499 , Volume 53, Issue 6, Page(s):: 3818-3829, 2006.
- [5] Ayan Bhunia, 'Enhancing the Performance of Feedback Scheduling', International Journal of Computer Applications, ISSN 0975 - 8887, Volume 18, Number.4, March 2011.
- [6] Baskiyar S and Meghanathan N, 'A Survey of Contemporary Realtime Operating Systems', Informatica, Volume 29, Auburn University, Auburn, USA, Page(s):: 233-240, 2005.
- [7] Becchetti, L., Leonardi, S. and Marchetti S.A., 'Average-Case and Smoothed Competitive Analysis of the Multilevel Feedback Algorithm', Mathematics of Operation Research, Volume 31, Number 1, February, Page(s):: 85-108, 2006.
- [8] Behera H. S., Rakesh Mohanty, Sabyasachi Sahu and Sourav Kumar Bhoi, 'Comparative Performance Analysis of Multi-Dynamic Time Quantum Round Robin (MDTQRR) Algorithm with Arrival Time', Indian Journal of Computer Science and Engineering, IJCSE, ISSN: 09765166, Volume 2, Number 2, Page(s):: 262-271, Apr-May 2011.
- [9] Brebner G and Diessel O, 'Chip-based Reconfigurable Task Management', in Field-Programmable Logic and Applications (FPL'01), Springer Verlag, Berlin, Germany, Page(s): 182-191, 2001.
- [10] Burns A, Tindell K and Wellings A.J. 'Fixed priority scheduling with deadline prior to completion', in proceedings of Sixth Euromicro Workshop on Real-Time systems, Research Group Department of computer Science University of York , UK. Page(s): 138 - 142, 1994.
- [11] Cai Sufeng and Hugh Anderson, 'Queueing Theory', Operating System, Group Project Report, Group 36, NUS, National University of Singapore, school of computing, 2007 - 2008.
- [12] David B. Stewart, Donald E. Schmitz, and Pradeep K. Khosla, 'The Chimera II Real-Time Operating System for Advanced Sensor-Based Control Applications', IEEE Transactions on Systems, Man, and Cybernetics , Volume 22, Number 6, Page(s):: 1282-1295, Nov / Dec 1992.
- [13] Dharamendra Chouhan, SM Dilip Kumar and Jerry Antony Ajay, 'A MLFQ Scheduling Technique using M/M/c Queues for Grid Computing', International Journal of Computer Science Issues, IJCSI, ISSN: 1694-0784, Volume 10, Issue 2, Number 1, Page(s):: 357-364, March 2013.
- [14] Dodd R.B, 'Coloured Petri Net Modelling of a Generic Avionics Mission Computer' under 'Avionics Enabling Research and Development', Defense Science and Technology Organization, DSTO, Australia, DSTO-TN-0692, April 2006.
- [15] Dror G. Feitelson, 'Notes on Operating Systems', School of Computer Science and Engineering, The Hebrew University of Jerusalem, Israel, 2011.
- [16] Garcia P, Compton K, Schulte M, Blem E and Fu W., 'An overview of reconfigurable hardware in embedded systems', in EURASIP Journal on Embedded Systems , Page(s): 1-19, 2006.
- [17] Gary J. Nutt, 'Implementing Processes, Threads, and Resources' and 'scheduling' along with 'Basic Synchronization Principles', in text book Operating Systems, Third Edition, ISBN 0-201-77344-9, Addison-Wesley, 2004.
- [18] Hoganson, Kenneth, 'Reducing MLFQ Scheduling Starvation with Feedback and Exponential Averaging', Consortium for Computing Sciences in Colleges , Southeastern Conference, Georgia, 2009.
- [19] Jean-François Hermant and Gérard Le Lann, 'Fast asynchronous uniform consensus in real-time distributed systems', in IEEE Transactions on Computers , Volume 51, Number 8, Page(s): 931-944, 2002.
- [20] krithi Ramamritham and John A. Stankovic, 'Scheduling Algorithms and Operating Systems Support for Real-Time Systems', Proceedings of the IEEE, Volume 82, Issue 1, Page(s):: 55-67, Jan 1994.
- [21] Kruk L, Lehoczky J. P, Shreve S. E and Yeung S.N, 'Earliest deadline first service in heavy traffic acyclic networks', in Annals of Applied Probability , Volume 14, Number 3, Page(s): 1306-1352, 2004.
- [22] Lauzac S and Melhem R, 'An Improved Rate-Monotonic Admission Control and Its Applications', in IEEE Transactions on Computers, Volume 52, Number 3, Page(s): 337-350, 2003.
- [23] Li Lo and Liang-Teh Lee, 'An Interactive Oriented Fair Scheduler with Bounded Starvation for Desktop Time-Sharing Systems', Thesis for Master of Science, Department of Computer Science and Engineering, Tatung University, January 2008.
- [24] Maricruz Valiente, Gonzalo Genova and Jesus Carretero, 'UML 2.0 Notation for Modeling Real Time Task Scheduling', proceedings in Journal of Object Technology, Volume 5, Number 4, Page(s):: 91-105, May-June 2006.
- [25] Michael B. Jones et al, 'CPU Reservations and Time Constraints: Efficient, Predictable Scheduling of Independent Activities', in Proceedings of the sixteenth ACM Symposium on Operating Systems Principles, ACM SOSP '97, ISBN:0-89791-916-5, Page(s):: 198-211, 1997.
- [26] Micro digital, 'Simple Multitasking Executive', in proceedings of online document SMX RTOS, smxinfo, Embedded Software Outfitters, SMX , Symmetry Innovations, Australia, December 2004.
- [27] Mohammad Reza Effat Parvar, Karim Faez, Mehdi EffatParvar, Mehdi Zarei, Saeed Safari, 'An Intelligent MLFQ Scheduling Algorithm (IMLFQ) with Fault Tolerant Mechanism', in proceedings of Sixth International Conference on Intelligent Systems Design and Applications, ISDA'06, Volume 3, Page(s):: 80-85, Oct 2006.
- [28] Mohammad Reza Effat Parvar, Mehdi Effat Parvar, Abolfazl Toroghi Haghighat, Reza Mahini, Mehdi Zarei, 'An Intelligent MLFQ Scheduling Algorithm (IMLFQ)', in proceedings of International Conference on Parallel and Distributed Processing Techniques and Applications, PDPTA, Las Vegas, Nevada, USA, Page(s):: 1033-1036, June 2006.
- [29] Mohammad Reza Effat Parvar, Akbar Bemana, Mehdi EffatParvar, 'IMLFQ Scheduling Algorithm with Combinational Fault Tolerant Method', in proceedings of International Conference on Enformatika Systems Sciences and Engineering, Prague, Czech Republic, Jan 2006.
- [30] Mohammad Reza Effat Parvar, Mehdi Effat Parvar, Saeed Safari, 'A Starvation Free IMLFQ Scheduling Algorithm Based on Neural Network', International Journal of Computational Intelligence Research, ISSN 0973-1873, Volume.4, Number.1, Page(s):: 27-36, 2008.
- [31] Panduranga Rao M.V. and Shet K.C, 'Study and Development of a New Multi Level Feedback Queue Scheduler for Embedded Processor', International Journal of Computer Sciences and Engineering Systems. IJCSES, ISSN 0973-4406. Volume 4, Number 3, Page(s): 203213, July 2010.
- [32] Panduranga Rao M.V. and Shet K.C, 'A Simplified Study of Processor Scheduler for Real Time Operating System', International Journal of Computational Cognition . IJCC, ISSN 1542-5908 (online); ISSN 15428060 (print). Volume 8, Number 3, Page(s): 5-16, September 2010.
- [33] Panduranga Rao M.V. and Shet K.C, 'A Research in Real Time Scheduling Policy for Embedded System Domain', CLEI Electronic Journal, ISSN 0717- 5000 . Volume 12, Number 2, Paper 4, August 2009.
## International Journal of Computer Trends and Technology (IJCTT) - volume 13 number 3 - Jul 2014
- [34] Panduranga Rao M.V. and Shet K.C, 'Analysis and Refining of Scheduler for Real Time Operating System', ICFAI University Journal of Computer Sciences ISSN: 0973-9904. . Volume. 3, Number 4, Page(s): 7-22, October 2009.
[49]
Sinnen O and Sousa L, 'On Task Scheduling Accuracy: Evaluation Methodology and Results', in The Journal of Supercomputing , Volume 27, Number 2, Page(s): 177-194, 2004.
- [35] Panduranga Rao M.V. and Shet K.C, 'New Approaches to Improve CPU Process Scheduler in the Embedded System Domain', imanager's Journal on Future Engineering and Technology, JFET, ISSN: 0973- 2632 . Volume 4, Number 1, Page(s): 72-84, July September 2009 .
- [36] Panduranga Rao M.V. and Shet K.C, 'A Simplistic Study of Scheduler for Real Time and Embedded System Domain', International Journal of Computer Science and Applications, IJCSA, ISSN: 0974-1003 . Volume 2, Number 2, Page(s): 99-108, November 2009.
[37] Panduranga Rao M.V. and Shet K.C, R. Balakrishna and K. Roopa, 'Development of Scheduler for Real Time and Embedded System Domain', 22 nd IEEE International Conference on Advanced Information Networking and Applications - Workshops, WAINA '08, 2008. FINA 2008, Fourth International Symposium on Frontiers in Networking with Applications, Gino-wan, Okinawa, JAPAN. IEEE Computer Society 2008. DOI 10.1109/WAINA.2008.33, Page(s):1 - 6, 25 to 28 th March 2008.
- [38] Panduranga Rao M.V., Shet K.C and K. Roopa. 'Efficient and predictable process scheduling', In proceedings of International Conference on Contemporary Computing IC3MACMILLAN JOURNAL, Noida, New Delhi, India, University of FLORIDA (UFL & JIITU) and Jaypee Institute of Information Technology University, page(s): 131-140, August 7-9 2008.
- [39] Panduranga Rao M.V., Shet K.C, K. Roopa and K.J. Sri Prajna, 'Implementation of a simple co-routine based scheduler', In proceedings of Knowledge based computing systems & Frontier Technologies NCKBFT, MIT Manipal, Karnataka, INDIA , Page(s): 161-164, 19 to 20 th Feb 2007.
- [40] Paolo Di Francesco, 'Design and implementation of a MLFQ scheduler for the Bacula backup software', Master thesis in Global Software Engineering, Universita degli Studi dell'Aquila, Italy, 2011 / 2012.
- [41] Peng Li and Binoy Ravindran, 'Fast, Best-Effort Real-Time Scheduling Algorithms', IEEE Transactions On Computers, Volume 53, Issue 9, Page(s):: 1159-1175, Sept 2004.
- [42] Raymond Keith Clark, 'Scheduling Dependent Real-Time Activities', PhD Thesis, School of Computer Science, Carnegie Mellon University, CMU-CS-90-155, August 1990.
- [43] Raymond Keith Clark, Jensen E. D and Rouquette N. F. 'Software organization to facilitate dynamic processor scheduling', in IEEE Workshop on Parallel and Distributed Real-Time Systems (WPDRTS), Page(s): 122b, 2004.
[44] Sami Khuri, Hsiu-Chin Hsu, 'Visualizing the CPU Scheduler and Page Replacement Algorithms', Proceedings of the 30th ACM SIGCSE Technical Symposium on Computer Science Education, SIGCSE '99, New Orleans, Louisiana, USA, ACM Press, New York, Page(s):: 227231, March 1999.
- [45] Shahrooz Feizabadi et al., 'Utility Accrual Scheduling With Real-Time Java', On The Move to Meaningful Internet Systems, OTM Workshops , ISBN 978-3-540-39962-9, Page(s):: 550-563, 2003.
- [46] Shahrooz S. Feizabadi, 'Garbage Collection Scheduling for Utility Accrual Real-Time Systems', PhD Thesis, Virginia Polytechnic Institute & State University, Blacksburg, Virginia, December 7, 2006.
- [47] Sifat Islam et al., 'Concurrency Compliant Embedded System Modeling Methodology', in 2nd Annual IEEE International Systems Conference, SysCon 2008, ISBN: 978-1-4244-2149-7, Montreal, Canada, Page(s):: 1-8, April 2008 .
- [48] Silberschatz, A., P.B. Galvin and G. Gagne, 'Operating Systems Concepts', 7th Edn., John Wiley and Sons, USA., ISBN: 13: 9780471694663 , Page(s):: 944, 2004.
- [50] Sukanya Suranauwarat, 'A CPU Scheduling Algorithm Simulator', Session F2H, 37th ASEE / IEEE Frontiers in Education Conference, Milwaukee, WI, Page(s):: F2H-19 - F2H-24, October 10-13, 2007.
- [51] Tanenbaum A.S., 'Modern Operating Systems', 3rd Edn., Prentice Hall, ISBN: 13: 9780136006633 , Page(s):: 1104, 2008.
- [52] Torrey, L.A., Coleman, J and Miller, B.P., 'A Comparison of the Interactivity in the Linux 2.6 Scheduler and an MLFQ Scheduler', Software Practice and Experience , John Wiley & Sons, Ltd, Volume 37, Number 4, Page(s):: 347-364, 2007.
- [53] Yaashuwanth C and Ramesh R, A New Scheduling Algorithms for Real Time Tasks, International Journal of Computer Science and Information Security, IJCSIS, ISSN 1947-5500, Volume 6, Number 2, Page(s):: 61-66, 2009.
[54] Zhi Quan and Jong-Moon Chang, 'A Statistical Framework for EDF
Scheduling', in
Journal on
IEEE
Communication Letters
, Volume 7, Number 10,
Page(s): 493-495, 2003.
## V. AUTHOR BIOGRAPHIES
Prof. M.V. Panduranga Rao is a research scholar at National Institute of Technology
Karnataka, Mangalore, India. He has completed Master of Technology in computer science and Bachelor of Engineering in electronics and communication.
His research interests are in the field of Real-Time and Embedded Systems on Linux platform. He has published various research papers in journal and conferences across India, Also in the IEEE international conference in Okinawa, Japan. He has authored two reference books on Linux Internals. He is the Life member of Indian Society for Technical Education and IAENG.
His webpage can be found via http://www.pandurangarao.i8.com/ .
Dr. K.C.Shet obtained his PhD degree from Indian Institute of Technology, Bombay, Mumbai, India, in 1989. He has been working as a Professor in the Department of Computer Engineering, National Institute of Technology, Surathkal, Karnataka, India, since 1980.
He has published over 200 papers in the area of Electronics, Communication, & computers. He is a member of Computer Society of India, Mumbai, India, and Indian Society for Technical Education, New Delhi, India.
His webpage can be found via http://www.nitk.ac.in/~kcshet/index.html .
|
10.14445/22312803/IJCTT-V13P124
|
[
"M. V. Panduranga Rao",
"K. C. Shet"
] |
2014-08-02T19:56:24+00:00
|
2014-08-02T19:56:24+00:00
|
[
"cs.OS",
"68U20",
"D.3.2"
] |
Assessment of Response Time for New Multi Level Feedback Queue Scheduler
|
Response time is one of the characteristics of scheduler, happens to be a
prominent attribute of any CPU scheduling algorithm. The proposed New Multi
Level Feedback Queue [NMLFQ] Scheduler is compared with dynamic, real time,
Dependent Activity Scheduling Algorithm (DASA) and Lockes Best Effort
Scheduling Algorithm (LBESA). We abbreviated beneficial result of NMLFQ
scheduler in comparison with dynamic best effort schedulers with respect to
response time.
|
1408.0991v1
|
Varieties of groupoids and quasigroups generated by linear-bivariate polynomials over the ring Z n ∗†
E. Ilojide Department of Mathematics, Obafemi Awolowo University, Ile-Ife 220005, Nigeria. [email protected]
T. G. Jaiy´o . l´ e a ‡ Department of Mathematics, Obafemi Awolowo University, Ile-Ife 220005, Nigeria. [email protected] [email protected]
O. O. Owojori
Department of Mathematics, Obafemi Awolowo University, Ile-Ife 220005, Nigeria.
[email protected],[email protected]
## Abstract
Some varieties of groupoids and quasigroups generated by linear-bivariate polynomials P x,y ( ) = a + bx + cy over the ring Z n are studied. Necessary and sufficient conditions for such groupoids and quasigroups to obey identities which involve one, two, three (e.g. Bol-Moufang type) and four variables w.r.t. a , b and c are established. Necessary and sufficient conditions for such groupoids and quasigroups to obey some inverse properties w.r.t. a , b and c are also established. This class of groupoids and quasigroups are found to belong to some varieties of groupoids and quasigroups such as medial groupoid(quasigroup), F-quasigroup, semi automorphic inverse property groupoid(quasigroup) and automorphic inverse property groupoid(quasigroup).
## 1 Introduction
## 1.1 Groupoids, Quasigroups and Identities
Let G be a non-empty set. Define a binary operation ( ) on · G . ( G, · ) is called a groupoid if G is closed under the binary operation ( ). · A groupoid ( G, · ) is called a quasigroup if the
∗ 2010 Mathematics Subject Classification. Primary 20N02, 20NO5; Secondary 05B15
† Keywords and Phrases : groupoids, quasigroups, linear-bivariate polynomials
‡ All correspondence to be addressed to this author.
equations a · x = b and y · c = d have unique solutions for x and y for all a, b, c, d ∈ G . A quasigroup ( G, · ) is called a loop if there exists a unique element e ∈ G called the identity element such that x · e = e · x = x for all x ∈ G .
A function f : S × S → S on a finite set S of size n > 0 is said to be a Latin square (of order n ) if for any value a ∈ S both functions f ( a, · ) and f ( · , a ) are permutations of S . That is, a Latin square is a square matrix with n 2 entries of n different elements, none of them occurring more than once within any row or column of the matrix.
Definition 1.1 A pair of Latin squares f 1 ( · , · ) and f 2 ( · , · ) is said to be orthogonal if the pairs ( f 1 ( x, y ) , f 2 ( x, y ) ) are all distinct, as x and y vary.
For associative binary systems, the concept of an inverse element is only meaningful if the system has an identity element. For example, in a group ( G, · ) with identity element e ∈ G , if x ∈ G then the inverse element for x is the element x -1 ∈ G such that
$$x \cdot x ^ { - 1 } = x ^ { - 1 } \cdot x = e.$$
In a loop ( G, · ) with identity element e , the left inverse element of x ∈ G is the element x λ ∈ G such that
$$x ^ { \lambda } \cdot x = e$$
while the right inverse element of x ∈ G is the element x ρ ∈ G such that
$$x \cdot x ^ { \rho } = e$$
In case ( G, · ) is a quasigroup, then for each x ∈ G , the elements x ρ ∈ G and x λ ∈ G such that xx ρ = e ρ and x x λ = e λ are called the right and left inverse elements of x respectively. Here, e ρ ∈ G and e λ ∈ G satisfy the relations xe ρ = x and e λ x = x for all x ∈ G and are respectively called the right and left identity elements. Whenever e ρ = e λ , then ( G, · ) becomes a loop.
In case ( G, · ) is a groupoid, then for each x ∈ G , the elements x ρ ∈ G and x λ ∈ G such that xx ρ = e ρ ( x ) and x x λ = e λ ( x ) are called the right and left inverse elements of x respectively. Here, e ρ ( x ) ∈ G and e λ ( x ) ∈ G satisfy the relations xe ρ ( x ) = x and e λ ( x x ) = x for each x ∈ G and are respectively called the local right and local left identity elements of x . Whenever e ρ ( x ) = e λ ( x ), then we simply write e x ( ) = e ρ ( x ) = e λ ( x ) and call it the local identity of x .
The basic text books on quasigroups, loops are Pflugfelder [19], Bruck [1], Chein, Pflugfelder and Smith [2], Dene and Keedwell [3], Goodaire, Jespers and Milies [4], Sabinin [25], Smith [26], Ja´ ıy´ol´ [5] and Vasantha Kandasamy [28]. e . a
Groupoids, quasigroups and loops are usually studied relative to properties or identities. If a groupoid, quasigroup or loop obeys a particular identity, then such types of groupoids, quasigroups or loops are said to form a variety. In this work, our focus will be on groupoids and quasigroups. Some identities that describe groupoids and quasigroups which would be of interest to us here are categorized as follows:
- (A) Those identities which involve one element only on each side of the equality sign:
$$a a = a \text{ \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\text{$a$} a = a \text{ \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\text{$i$} \text{ \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\text{$i$} \text{ \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\text{$i$} \text{\quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\text{\text$i$} \text{ \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\\text$i$} \text{ \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\\text$i$} \text{\quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\\text$i$} \quad \text{\quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\\text$i$} \text$} \text{\quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\\text$i$}$$
$$a a = b b \quad \quad \text{unipotent law}$$
(B) Those identities which involve two elements on one or both sides of the equality sign:
| ab = ba | commutative law | (3) |
|---------------------|----------------------------|-------|
| ( ab ) b = a | Sade right Keys law | (4) |
| b ( ba ) = a | Sade left keys law | (5) |
| ( ab ) b = a ( bb ) | right alternative law | (6) |
| b ( ba ) = ( bb ) a | left alternative law | (7) |
| ( ba ) = ( ab ) a | medial alternative law | (8) |
| ( ba ) = b | law of right semisymmetry | (9) |
| ( ab ) a = b | law of left semisymmetry | (10) |
| a ( ab ) = ba | Stein first law | (11) |
| a ( ba ) = ( ba ) a | Stein second law | (12) |
| a ( ab ) = ( ab ) b | Schroder first law | (13) |
| ( ab )( ba ) = a | Schroder second law | (14) |
| ( ab )( ba ) = b | Stein third law | (15) |
| ab = a | Sade right translation law | (16) |
| ab = b | Sade left translation law | (17) |
- (C) Those identities which involve three distinct elements on one or both sides of the equality sign:
| ( ab ) c = a ( bc ) | associative law | (18) |
|-----------------------|-----------------------------|--------|
| a ( bc ) = c ( ab ) | law of cyclic associativity | (19) |
| ( ab ) c = ( ac ) b | law of right permutability | (20) |
| a ( bc ) = b ( ac ) | law of left permutability | (21) |
| a ( bc ) = c ( ba ) | Abel-Grassman law | (22) |
| ( ab ) c = a ( cb ) | commuting product law | (23) |
| ( ba ) = ( bc ) a | dual of commuting product | (24) |
| ( ab )( bc ) = ac | Stein fourth law | (25) |
| ( ba )( ca ) = bc | law of right transitivity | (26) |
## (D) Those involving four elements:
| ( ab )( ac ) = bc | law of left transitivity | (27) |
|-------------------------------|--------------------------------------|--------|
| ( ab )( ac ) = cb | Schweitzer law | (28) |
| ( ba )( ca ) = cb | dual of Schweitzer law | (29) |
| ( ab ) c = ( ac )( bc ) | law of right self-distributivity law | (30) |
| c ( ba ) = ( cb )( ca ) | law of left self-distributivity law | (31) |
| ( ab ) c = ( ca )( bc ) | law of right abelian distributivity | (32) |
| c ( ba ) = ( cb )( ac ) | law of left abelian distributivity | (33) |
| ( ab )( ca ) = [ a ( bc )] a | Bruck-Moufang identity | (34) |
| ( ab )( ca ) = a [ bc ) a ] | dual of Bruck-Moufang identity | (35) |
| [( ab ) c ] b = a [ b ( cb )] | Moufang identity | (36) |
| [( bc ) b ] a = b [ c ( ba )] | Moufang identity | (37) |
| [( ab ) c ] b = a [( bc ) b ] | right Bol identity | (38) |
| [ b ( cb )] a = b [ c ( ba )] | left Bol identity | (39) |
| [( ab ) c ] a = a [ b ( ca )] | extra law | (40) |
| [( ba ) a ] c = b [( aa ) c | ] RC 4 law | (41) |
| [ b ( aa )] c = b [ a ( ac )] | LC 4 law | (42) |
| ( aa )( bc ) = [ a ( ab )] c | LC 2 law | (43) |
| [( bc ) a ] a = b [( ca ) a ] | RC 1 law | (44) |
| [ a ( ab )] c = a [ a ( bc )] | LC 1 law | (45) |
| ( bc )( aa ) = b [( ca ) a ] | RC 2 law | (46) |
| [( aa ) b ] c = a [ a ( bc )] | LC 3 law | (47) |
| [( bc ) a ] a = b [ c ( aa )] | RC 3 law | (48) |
| [( ba ) a ] c = b [ a ( ac )] | C-law | (49) |
| a [ b ( ca )] = cb | Tarski law | (50) |
| a [( bc )( ba )] = c | Neumann law | (51) |
| ( ab )( ca ) = ( ac )( ba ) | specialized medial law | (52) |
| involving four elements: | | |
| ( ab )( cd ) = ( ad )( cb ) | first rectangle rule | (53) |
| ( ab )( ac ) = ( db )( dc ) | second rectangle rule | (54) |
| ( ab )( cd ) = ( ac )( bd ) | internal mediality or medial law | (55) |
(E) Those involving left or right inverse elements:
$$x ^ { \lambda } \cdot x y = y \quad \text{
} | <? </? |
| | |& + - . ) \end ]& | & !& +
+&$$
yx
·
x
ρ
=
y
$$\text{right inverse property}$$
$$x ( y x ) ^ { \rho } = y ^ { \rho } \text{ or } ( x y ) ^ { \lambda } x = y ^ { \lambda } \text{ \quad \ \ weakel inverse property(WIP)}$$
xy x · ρ = y or x yx · ρ = y or x λ · ( yx ) = y or x y x λ · = y cross inverse property(CIP) (59)
$$( x y ) ^ { \rho } = x ^ { \rho } y ^ { \rho } \text{ or } ( x y ) ^ { \lambda } = x ^ { \lambda } y ^ { \lambda } \text{ automorphic inverse property } ( \text{AIP} ) \text{ \quad } ( 6 0 )$$
$$( x y ) ^ { \rho } = y ^ { \rho } x ^ { \rho } \text{ or } ( x y ) ^ { \lambda } = y ^ { \lambda } x ^ { \lambda } \text{ anti-automorphic inverse property } ( A A I P ) \text{ \quad } ( 6 1 )$$
( xy · x ) ρ = x y ρ ρ · x ρ or ( xy · x ) λ = x y λ λ · x λ semi-automorphic inverse property (SAIP)
(62)
## Definition 1.2 (Trimedial Quasigroup)
A quasigroup is trimedial if every subquasigroup generated by three elements is medial.
Medial quasigroups have also been called abelian, entropic, and other names, while trimedial quasigroups have also been called triabelian, terentropic, etc.
There are two distinct, but related, generalizations of trimedial quasigroups. The variety of semimedial quasigroups(also known as weakly abelian, weakly medial, etc.) is defined by the equations
$$x x \cdot y z = x y \cdot x z$$
$$z y \cdot x x = z x \cdot y z$$
Definition 1.3 (Semimedial Quasigroup)
A quasigroup satisfying (63) (resp. (64)) is said to be left (resp. right) semimedial.
## Definition 1.4 (Medial-Like Identities)
A groupoid or quasigroup is called an external medial groupoid or quasigroup if it obeys the identity
$$a b \cdot c d = d b \cdot c a \quad \ e x t e r n a l \ m e d i a l \ o r \ p a r a m e d i a l i t y \ l a w$$
A groupoid or quasigroup is called a palindromic groupoid or quasigroup if it obeys the identity
$$a b \cdot c d = d c \cdot b a \quad \ \ \ p a l i d r o m i t y \ l a w$$
Other medial like identities of the form ( ab )( cd ) = ( π a π b ( ) ( ))( π c π d ( ) ( )) , where π is a certain permutation on { a, b, c, d } are given as follows:
$$a b \cdot c d = a b \cdot d c \text{ \quad \ \ } C _ { 1 }$$
ab
·
cd
=
ba
·
cd
C
2
(68)
$$a b \cdot c d = b a \cdot d c \text{ \quad \ \ } C _ { 3 }$$
$$a b \cdot c d = c d \cdot a b \text{ \quad \ \ } C _ { 4 }$$
ab
·
cd
=
cd
·
ba
C
5
(71)
$$a b \cdot c d = d c \cdot a b \text{ \quad \ \ } C _ { 6 }$$
ab
·
cd
=
ac
·
db
CM
1
(73)
$$a b \cdot c d = a d \cdot b c \text{ \quad \ \ } C M _ { 2 }$$
ab
·
cd
=
ad
·
cb
CM
3
(75)
$$a b \cdot c d = b c \cdot a d \text{ \quad \ \ } C M _ { 4 }$$
ab
·
cd
=
bc
·
da
CM
5
(77)
$$a b \cdot c d = b d \cdot a c \text{ \quad \ \ } C M _ { 6 }$$
ab
·
cd
=
bd
·
ca
CM
7
(79)
ab
·
cd
=
ca
·
bd
CM
8
(80)
$$a b \cdot c d = c a \cdot d b \text{ \quad \ \ } C M _ { 9 }$$
ab
·
cd
=
cb
·
ad
CM
10
(82)
ab
·
cd
=
cb
·
da
CM
11
(83)
$$a b \cdot c d = d a \cdot b c \text{ \quad \ \ } C M _ { 1 2 }$$
ab
·
cd
=
da
·
cb
CM
13
(85)
$$a b \cdot c d = d b \cdot a c \text{ \quad \ \ } C M _ { 1 4 }$$
The variety of F-quasigroups was introduced by Murdoch [18].
## Definition 1.5 (F-quasigroup)
An F-quasigroup is a quasigroup that obeys the identities
$$x \cdot y z = x y \cdot ( x \nwarrow x ) z \text{ \quad \ \ } l e f t \ F { - l a w }$$
$$z y \cdot x = z ( x / x ) \cdot y x \quad \ \ \ r i g h t \ F { - l a w }$$
A quasigroup satisfying (87) (resp. (88)) is called a left (resp. right) F-quasigroup.
## Definition 1.6 (E-quasigroup)
An E-quasigroup is a quasigroup that obeys the identities
$$x \cdot y z = e _ { \lambda } ( x ) y \cdot x z \quad \quad \ \ E _ { l } \ l a w$$
$$z y \cdot x = z x \cdot y e _ { \rho } ( x ) \text{ \quad \ \ } E _ { r } \, L a w$$
A quasigroup satisfying (89) (resp. (90)) is called a left (resp. right) E-quasigroup.
Some identities will make a quasigroup to be a loop, such are discussed in Keedwell [6, 7].
Definition 1.7 (Linear Quasigroup and T-quasigroup) A quasigroup ( Q, · ) of the form x · y = xα + yβ + c where ( Q, +) is a group, α is its automorphism and β is a permutation of the set Q , is called a left linear quasigroup.
- A quasigroup ( Q, · ) of the form x · y = xα + yβ + c where ( Q, +) is a group, β is its automorphism and α is a permutation of the set Q , is called a right linear quasigroup.
- A T-quasigroup is a quasigroup ( Q, · ) defined over an abelian group ( Q, +) by x · y = c + xα + yβ , where c is a fixed element of Q and α and β are both automorphisms of the group ( Q, +) .
Whenever one considers mathematical objects defined in some abstract manner, it is usually desirable to determine that such objects exist. Although occasionally this is accomplished by means of an abstract existential argument, most frequently, it is carried out through the presentation of a suitable example, often one which has been specifically constructed for the purpose. An example is the solution to the open problem of the axiomization of rectangular quasigroups and loops by Kinyon and Phillips [12] and the axiomization of trimedial quasigroups by Kinyon and Phillips [10, 11].
Chein et. al. [2] presents a survey of various methods of construction which has been used in the literature to generate examples of groupoids and quasigroups. Many of these constructions are ad hoc-designed specifically to produce a particular example; while others are of more general applicability. More can be found on the construction of ( r, s, t )-inverse quasigroups in Keedwell and Shcherbacov [8, 9], idempotent medial quasigroups in Krˇadinac c and Volenec [14] and quasigroups of Bol-Moufang type in Kunen [15, 16].
Remark 1.1 In the survey of methods of construction of varieties and types of quasigroups highlighted in Chein et. al. [2], it will be observed that some other important types of quasigroups that obey identities (1) to (90) are not mentioned. Also, examples of methods of construction of such varieties that are groupoids are also scarce or probably not in existence by our search. In Theorem 1.4 of Kirnasovsky [13], the author characterized T-quasigroups with a score and two identities from among identities (1) to (90). The present work thus proves some results with which such groupoids and quasigroups can be constructed.
## 1.2 Univariate and Bivariate Polynomials
Consider the following definitions.
Definition 1.8 A polynomial P x ( ) = a 0 + a x 1 + · · · + a x n n , n ∈ N is said to be a permutation polynomial over a finite ring R if the mapping defined by P is a bijection on R .
Definition 1.9 A bivariate polynomial is a polynomial in two variables, x and y of the form P x, y ( ) = Σ i,j a ij x y i j .
Definition 1.10 (Bivariate Polynomial Representing a Latin Square)
A bivariate polynomial P x, y ( ) over Z n is said to represent (or generate) a Latin square if ( Z n , ∗ ) is a quasigroup where ∗ : Z n × Z n → Z n is defined by x ∗ y = P x, y ( ) for all x, y ∈ Z n .
Mollin and Small [17] considered the problem of characterizing permutation polynomials. They established conditions on the coefficients of a polynomial which are necessary and sufficient for it to represent a permutation.
Shortly after, Rudolf and Mullen [23] provided a brief survey of the main known classes of permutation polynomials over a finite field and discussed some problems concerning permutation polynomials (PPs). They described several applications of permutations which indicated why the study of permutations is of interest. Permutations of finite fields have become of considerable interest in the construction of cryptographic systems for the secure transmission of data. Thereafter, the same authors in their paper [24], described some results that had appeared after their earlier work including two major breakthroughs.
Rivest [22] studied permutation polynomials over the ring ( Z n , + ) where , · n is a power of 2: n = 2 w . This is based on the fact that modern computers perform computations modulo 2 w efficiently (where w = 2 8 16 32 or 64 is the word size of the machine), and so it was , , , of interest to study PPs modulo a power of 2. Below is an important result from his work which is relevant to the present study.
## Theorem 1.1 (Rivest [22])
A bivariate polynomial P x, y ( ) = Σ i,j a ij x y i j represents a Latin square modulo n = 2 w , where w ≥ 2 , if and only if the four univariate polynomials P x, ( 0) , P x, ( 1) , P (0 , y ) , and P (1 , y ) are all permutation polynomial modulo n .
/negationslash
Vadiraja and Shankar [27] motivated by the work of Rivest continued the study of permutation polynomials over the ring ( Z n , + ) by studying Latin squares represented by linear , · and quadratic bivariate polynomials over Z n when n = 2 w with the characterization of some PPs. Some of the main results they got are stated below.
## Theorem 1.2 (Vadiraja and Shankar [27])
/negationslash
A bivariate linear polynomial a + bx + cy represents a Latin square over Z n , n = 2 w if and only if one of the following equivalent conditions is satisfied:
- (i) both b and c are coprime with n ;
- (ii) a + bx , a + cy , ( a + ) + c bx and ( a + ) + b cy are all permutation polynomials modulo n .
Remark 1.2 It must be noted that P x, y ( ) = a + bx + cy represents a groupoid over Z n . P x, y ( ) represents a quasigroup over Z n if and only if ( Z n , P ) is a T-quasigroup. Hence whenever ( Z n , P ) is a groupoid and not a quasigroup, ( Z n , P ) is neither a T-quasigroup nor left linear quasigroup nor right linear quasigroup. Thus, the present study considers both T-quasigroup and non-T-quasigroup.
## Theorem 1.3 (Vadiraja and Shankar [27])
If P x, y ( ) is a bivariate polynomial having no cross term, then P x, y ( ) gives a Latin square if and only if P x, ( 0) and P (0 , y ) are permutation polynomials.
/negationslash
/negationslash
The authors were able to establish the fact that Rivest's result for a bivariate polynomial over Z n when n = 2 w is true for a linear-bivariate polynomial over Z n when n = 2 w . Although the result of Rivest was found not to be true for quadratic-bivariate polynomials over Z n when n = 2 w with the help of counter examples, nevertheless some of such squares can be forced to be Latin squares by deleting some equal numbers of rows and columns.
/negationslash
Furthermore, Vadiraja and Shankar [27] were able to find examples of pairs of orthogonal Latin squares generated by bivariate polynomials over Z n when n = 2 w which was found impossible by Rivest for bivariate polynomials over Z n when n = 2 w .
## 1.3 Some Important Results on Medial-Like Identities
Some important results which we would find useful in our study are stated below.
## Theorem 1.4 (Polonijo [21])
For any groupoid ( Q, · ) , any two of the three identities (55), (65) and (66) imply the third one.
## Theorem 1.5 (Polonijo [21])
Let ( Q, · ) be a commutative groupoid. Then ( Q, · ) is palindromic. Furthermore, the constraints (55) and (65) are equivalent, i.e a commutative groupoid ( Q, · ) is internally medial if and only if it is externally medial.
## Theorem 1.6 (Polonijo [21])
For any quasigroup ( Q, · ) and i ∈ { 1 2 , , ..., 6 } , C is valid if and only if the quasigroup is i commutative.
## Theorem 1.7 (Polonijo [21])
For any quasigroup ( Q, · ) and i ∈ { 1 2 , , ..., 14 } , CM holds if and only if the quasigroup i is both commutative and internally medial.
## Theorem 1.8 (Polonijo [21])
For any quasigroup ( Q, · ) and i ∈ { 1 2 , , ..., 14 } , CM is valid if and only if the quasigroup i is both commutative and externally medial.
Theorem 1.9 A quasigroup ( Q, · ) is palindromic if and only if there exists an automorphism α such that
$$\alpha ( x \cdot y ) = y \cdot x \ \forall \ x, y \in Q$$
## holds.
It is important to study the characterization of varieties of groupoids and quasigroups represented by linear-bivariate polynomials over the ring Z n even though very few of such have been sighted as examples in the past.
## 2 Main Results
Theorem 2.1 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over { Z n , Z p } such that 'HYPO' is true. P x, y ( ) represents a 'NAME' { groupoid, quasigroup } { ( Z n , P ) , ( Z p , P ) } over { Z n , Z p } if and only if 'N and S' is true. (Table 1)
## Proof
There are 66 identities for which the theorem above is true for in a groupoid or quasigroup. For the sake of space, we shall only demonstrate the proof for one identity for each category.
- (A) Those identities which involve one element only on each side of the equality sign:
Lemma 2.1 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents a unipotent groupoid ( Z n , P ) over Z n if and only if ( b + )( c x -y ) = 0 for all x, y ∈ Z n .
## Proof
P x, y ( ) satisfies the unipotent law ⇔ P x, x ( ) = P y, y ( ) ⇔ a + bx + cx = a + by + cy ⇔ a + bx -cx -a -by -cy = 0 ⇔ ( b + )( c x -y ) = 0 as required.
Lemma 2.2 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents a unipotent quasigroup ( Z n , P ) over Z n if and only if ( b + )( c x -y ) = 0 and ( b, n ) = ( c, n ) = 1 for all x, y ∈ Z n .
## Proof
This is proved by using Lemma 2.1 and Theorem 1.2
Theorem 2.2 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents a unipotent groupoid ( Z n , P ) over Z n if and only if b + c ≡ 0( mod n ) .
## Proof
This is proved by using Lemma 2.1.
Theorem 2.3 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents a unipotent quasigroupp ( Z n , P ) over Z n if and only if b + c ≡ 0(mod n ) and ( b, n ) = ( c, n ) = 1 .
## Proof
This is proved by using Lemma 2.2.
Example 2.1 P x, y ( ) = 5 x + y is a linear bivariate polynomial over Z 6 . ( Z 6 , P ) is a unipotent groupoid over Z 6 .
Example 2.2 P x, y ( ) = 1 + 5 x + y is a linear bivariate polynomial over Z 6 . ( Z 6 , P ) is a unipotent quasigroup over Z 6 .
- (B) Those identities which involve two elements on one or both sides of the equality sign:
Lemma 2.3 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents a Stein third groupoid ( Z n , P ) over Z n if and only if a (1 + b + ) + c x b ( 2 + c 2 ) + y (2 bc -1) = 0 for all x, y ∈ Z n .
## Proof
P x, y ( ) satisfies the Stein third law ⇔ P P x, y , P [ ( ) ( y, x )] = y ⇔ a (1 + b + ) + c x b ( 2 + c 2 ) + y (2 bc -1) = 0 as required.
Lemma 2.4 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents a Stein third quasigroup ( Z n , P ) over Z n if and only if a (1 + b + )+ c x b ( 2 + c 2 ) + y (2 bc -1) = 0 and ( b, n ) = ( c, n ) = 1 for all x, y ∈ Z n .
## Proof
This is proved by using Lemma 2.3 and Theorem 1.2
Theorem 2.4 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents a Stein third groupoid ( Z n , P ) over Z n if and only if b 2 + c 2 ≡ 0(mod n ) , 2 bc ≡ 1(mod n ) and a = 0 .
## Proof
This is proved by using Lemma 2.3.
Theorem 2.5 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents a Stein third quasigroup ( Z n , P ) over Z n if and only if b 2 + c 2 ≡ 0(mod n ) , 2 bc ≡ 1(mod n ) and a = 0 .
## Proof
This is proved by using Lemma 2.4.
/negationslash
Theorem 2.6 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z p such that a = 0 . P x, y ( ) represents a Stein third groupoid ( Z p , P ) over Z p if and only if b 2 + c 2 ≡ 0(mod ) p and 2 bc ≡ 1(mod ) p .
## Proof
This is proved by using Lemma 2.3.
/negationslash
Theorem 2.7 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z p such that a = 0 . P x, y ( ) represents a Stein third quasigroup ( Z p , P ) over Z p if and only if b 2 + c 2 ≡ 0(mod ) p and 2 bc ≡ 1(mod ) p .
## Proof
This is proved by using Lemma 2.4.
Example 2.3 P x, y ( ) = 2 x +3 y is a linear bivariate polynomial over Z 5 . ( Z 5 , P ) is a Stein third groupoid over Z 5 .
Example 2.4 P x, y ( ) = 2 x +3 y is a linear bivariate polynomial over Z 5 . ( Z 5 , P ) is a Stein third quasigroup over Z 5 .
- (C) Those identities which involve three distinct elements on one or both sides of the equality sign:
Lemma 2.5 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents an Abel-Grassman groupoid ( Z n , P ) over Z n if and only if ( x -z )( b -c 2 ) = 0 for all x, z ∈ Z n .
## Proof
P x, y ( ) satisfies the Abel-Grassman law ⇔ P x, P [ ( y, z )] = P z, P [ ( y, x )] ⇔ P x, a ( + by + cz ) = P z, a ( + by + cx ) ⇔ a + bx + ( c a + by + cz ) = a + bz + ( c a + by + cx ) ⇔ ( x -z )( b -c 2 ) = 0 as required.
Lemma 2.6 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents an Abel-Grassman quasigroup ( Z n , P ) over Z n if and only if ( x -z )( b -c 2 ) = 0 and ( b, n ) = ( c, n ) = 1 . for all x, y, z ∈ Z n .
## Proof
This is proved by using Lemma 2.5 and Theorem 1.2
Theorem 2.8 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents an Abel-Grassman groupoid ( Z n , P ) over Z n if and only if c 2 ≡ b (mod n ) .
## Proof
This is proved by using Lemma 2.5.
Theorem 2.9 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents an Abel-Grassman quasigroup ( Z n , P ) over Z n if and only if c 2 ≡ b (mod n ) and ( b, n ) = ( c, n ) = 1 .
## Proof
This is proved by using Lemma 2.6.
Example 2.5 P x, y ( ) = 2 + 4 x +2 y is a linear bivariate polynomial over Z 6 . ( Z 6 , P ) is an Abel-Grassman groupoid over Z 6 .
Example 2.6 P x, y ( ) = 2 + 4 x +2 y is a linear bivariate polynomial over Z 5 . ( Z 5 , P ) is an Abel-Grassman quasigroup over Z 5 .
- (D) Those involving four elements:
Lemma 2.7 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents an external medial groupoid ( Z n , P ) over Z n if and only if w b ( 2 -c 2 )+ ( z c 2 -b 2 ) = 0 for all w, z ∈ Z n .
## Proof
P x, y ( ) satisfies the external medial law ⇔ P P w,x , P [ ( ) ( y, z )] = P P z, x , P [ ( ) ( y, w )] ⇔ a + ( b a + bw + cx ) + c a ( + by + cz ) = a + ( b a + bz + cx ) + c a ( + by + cw ) ⇔ w b ( 2 -c 2 ) + z c ( 2 -b 2 ) = 0 as required.
Lemma 2.8 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents an external medial quasigroup ( Z n , P ) over Z n if and only if w b ( 2 -c 2 ) + z c ( 2 -b 2 ) = 0 and ( b, n ) = ( c, n ) = 1 for all w, z ∈ Z n .
## Proof
This is proved by using Lemma 2.7 and Theorem 1.2.
Theorem 2.10 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . ( Z n , P ) represents an external medial groupoid over Z n if and only if b 2 ≡ c 2 (mod n ) .
## Proof
This is proved by using Lemma 2.7.
Theorem 2.11 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . ( Z n , P ) represents an external medial quasigroup over Z n if and only if b 2 ≡ c 2 ( mod n ) and ( b, n ) = ( c, n ) = 1 .
## Proof
This is proved by using Lemma 2.8 and Theorem 1.2.
Example 2.7 P x, y ( ) = 4 + 2 x +2 y is a linear bivariate polynomial over Z 6 . ( Z 6 , P ) is an external medial groupoid over Z 6 .
Example 2.8 P x, y ( ) = 2 + 8 x + y is a linear bivariate polynomial over Z 9 . ( Z 9 , P ) is an external medial quasigroup over Z 9 .
- (E) Those involving left or right inverse elements:
Lemma 2.9 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents a cross inverse property groupoid ( Z n , P ) over Z n if and only if a bc ( -1) + x b c ( 2 +1 -b -bc ) + cy bc ( -1) = 0 for all x, y ∈ Z n .
## Proof
P x, y ( ) satisfies the cross inverse property ⇔ P P x, y , x [ ( ) ρ )] = y ⇔ P a ( + bx + cy, x ρ ) = y ⇔ a + ( b a + bx + cy ) + cx ρ = y ⇔ a bc ( -1) + x b c ( 2 +1 -b -bc ) + cy bc ( -1) = 0 as required.
Lemma 2.10 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents a cross inverse property quasigroup ( Z n , P ) over Z n if and only if a bc ( -1)+ x b c ( 2 +1 - -b bc )+ cy bc ( -1) = 0 and ( b, n ) = ( c, n ) = 1 for all x, y, z ∈ Z n .
## Proof
This is proved by using Lemma 2.9 and Theorem 1.2
/negationslash
Theorem 2.12 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z p such that a = 0 . P x, y ( ) represents a CIP quasigroup ( Z p , P ) over Z p if and only if bc ≡ 1(mod ) p .
## Proof
This is proved by using Lemma 2.10.
/negationslash
Theorem 2.13 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n such that a = 0 and c is invertible in Z n . P x, y ( ) represents a CIP groupoid ( Z n , P ) over Z n if and only if bc ≡ 1(mod n ) .
## Proof
This is proved by using Lemma 2.9.
/negationslash
Theorem 2.14 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n such that a = 0 , c is invertible in Z n and ( b, n ) = ( c, n ) = 1 . P x, y ( ) represents a CIP quasigroup ( Z n , P ) over Z n if and only if bc ≡ 1(mod n ) .
## Proof
This is proved by using Lemma 2.10.
Theorem 2.15 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n . P x, y ( ) represents a CIP groupoid ( Z n , P ) over Z n if bc ≡ 1(mod n ) .
## Proof
This is proved by using Lemma 2.9.
Theorem 2.16 Let P x, y ( ) = a + bx + cy be a linear bivariate polynomial over Z n such that ( b, n ) = ( c, n ) = 1 . P x, y ( ) represents a CIP quasigroup ( Z n , P ) over Z n if bc ≡ 1(mod n ) .
## Proof
This is proved by using Lemma 2.10.
Example 2.9 P x, y ( ) = 2 + 4 x +4 y is a linear bivariate polynomial over Z 5 . ( Z 5 , P ) is a cross inverse property groupoid over Z 5 .
Example 2.10 P x, y ( ) = 3+4 x +4 y is a linear bivariate polynomial over Z 5 . ( Z 5 , P ) is a cross inverse property quasigroup over Z 5 .
16
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
| S/N | NAME | G | Q | Z n | Z p | HYPO | N AND S | EXAMPLE |
|-------|---------------------|--------|--------|--------|----------|--------|----------------------------------------------------------|------------------|
| 1 | Idempotent | /check | | /check | | | b + c = 1, a = 0 | 5 x +2 y, Z 6 |
| 2 | Unipotent | /check | | /check | | | b + c = 0 | 2 +4 x +2 y, Z 6 |
| 2 | Unipotent | | /check | /check | | | b + c = 0,( b, n ) = ( c, n ) = 1 | 2 +5 x + y , Z 6 |
| 3 | Commut | /check | | /check | | | b = c | 1 +4 x +4 y, Z 6 |
| 3 | Commut | | /check | /check | | | b = c ,( b, n ) = ( c, n ) = 1 | 1 +5 x +5 y, Z 6 |
| 4 | Sade Right | /check | | | /check | a = 0 | b = - 1 | 2 +6 x +4 y, Z 7 |
| 4 | Sade Right | | /check | | /check | a = 0 | b = - 1 | 1 +5 x +4 y, Z 7 |
| 5 | Sade Left | /check | | | /check | a = 0 | c = - 1 | 2 +4 x +5 y, Z 7 |
| 5 | Sade Left | | /check | | /check | a = 0 | c = - 1 | 2 +5 x +5 y, Z 7 |
| 6 | Right Alternative | /check | | | /check | a = 0 | b = c = 1 | 3 + x + y, Z 7 |
| 6 | Right Alternative | | /check | | /check | a = 0 | b = c = 1 | 3 + x + y, Z 7 |
| 7 | Left Alternative | /check | | | /check | a = 0 | b = c = 1 | 2 + x + y, Z 7 |
| 7 | Left Alternative | | /check | | /check | a = 0 | b = c = 1 | 2 + x + y, Z 7 |
| 8 | Medial Alternative | /check | | | /check | a = 0 | b = c | 2 +4 x +4 y, Z 7 |
| 8 | Medial Alternative | /check | | | /check | b = c | b + c = 1 | 2 +4 x +2 y, Z 5 |
| 8 | Medial Alternative | | /check | | /check | a = 0 | b = c | 2 +4 x +4 y, Z 7 |
| 8 | Medial Alternative | | /check | | /check | b = c | b + c = 1 | 2 +4 x +2 y, Z 7 |
| 9 | Right Semi Symmetry | /check | | | /check | a = 0 | b = c = - 1 | 2 +4 x +4 y, Z 5 |
| 9 | Right Semi Symmetry | /check | | /check | | a = 0 | bc = 1 , c 2 = - b | 5 x +2 y, Z 9 |
| 9 | Right Semi Symmetry | | /check | | /check | a = 0 | b = c = - 1 | 2 +4 x +4 y, Z 5 |
| 9 | Right Semi Symmetry | | /check | /check | | a = 0 | bc = 1 , c 2 = - b | 5 x +2 y, Z 9 |
| 10 | Left Semi | /check | | | /check | a = 0 | b = c = - 1 | 3 +4 x +4 y, Z 5 |
| 10 | Left Semi | /check | | /check | | a = 0 | b = 1 , b 2 = - c | x +9 y, Z 10 |
| 10 | Symmetry | | /check | | /check | a = 0 | b = c = - 1 | 3 +4 x +4 y, Z 5 |
| 10 | Left Semi | | /check | /check | | a = 0 | b = 1 , b 2 = - c | x +9 y, Z 10 |
| 11 | Stein First | /check | | | /check | a = 0 | b = c | 3 +4 x +4 y, Z 5 |
| 11 | Stein First | | /check | | /check | a = 0 | b = c | 2 +4 x +4 y, Z 5 |
| 12 | Stein Second | /check | | | /check | a = 0 | b = c | 3 +4 x +4 y, Z 5 |
| 12 | Stein Second | | /check | | /check | a = 0 | b = c | 2 +4 x +4 y, Z 5 |
| 13 | Schroder Second | /check | | /check | | | b 2 + c 2 = 1 , 2 bc = 0 , a = 0 | 2 x +3 y, Z 6 |
| 13 | Schroder Second | | /check | /check | | | 2 + c 2 = 1 , 2 bc = 0 , a = 0 , ( b, n ) = ( c, n ) = 1 | ? |
| 13 | Schroder Second | /check | | /check | a | = 0 | b + c = - 1 , b 2 + c 2 = 1 , 2 bc = 0 | ? |
| 13 | Schroder Second | | /check | | /check a | = 0 | b + c = - 1 , b 2 + c 2 = 1 , 2 bc = 0 | ? |
17
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
| 14 | Stein Third | /check | | /check | | | b 2 + c 2 = 0 , 2 bc = 1 , a = 0 | ? |
|------|----------------------|----------|--------|----------|--------|------------------|--------------------------------------------------------|------------------|
| 14 | Stein Third | | /check | /check | | | b, n ) = ( c, n ) = 1 , b 2 + c 2 = 0 , 2 bc = 1 , a = | ? |
| 14 | Stein Third | /check | | | /check | a = 0 | b 2 + c 2 = 0 , 2 bc = 1 , | 3 +2 x +4 y, Z 5 |
| 14 | Stein Third | | /check | | /check | a = 0 | b 2 + c 2 = 0 , 2 bc = 1 , | 2 +2 x +4 y, Z 5 |
| 15 | Associative | /check | | | /check | a = 0 | b = c = 1 | 2 + x + y, Z 6 |
| 15 | Associative | | /check | | /check | a = 0 | b = c = 1 | 2 + x + y, Z 6 |
| 16 | Slim | /check | | /check | | a = 0 , c invert | bc = 0 , c = 1 | ! |
| 16 | Slim | | /check | /check | | a = 0 , c invert | bc = 0 , c = 1 , ( b, n ) = ( c, n ) = 1 | ? |
| 17 | Cyclic Associativity | /check | | /check | | | b = c = 1 | 3 + x + y, Z 6 |
| 17 | Cyclic Associativity | | /check | /check | | | b = c = 1 , ( b, n ) = ( c, n ) = 1 | 3 + x + y, Z 6 |
| 18 | Right Permutability | /check | | /check | | | b = 1 | 1 + x +5 y, Z 6 |
| 18 | Right Permutability | | /check | /check | | | b = 1 , ( b, n ) = ( c, n ) = 1 | 1 + x +5 y, Z 6 |
| 19 | Left Permutability | /check | | /check | | | c = 1 | 1 +5 x + y, Z 6 |
| 19 | Left Permutability | | /check | /check | | | c = 1 , ( b, n ) = ( c, n ) = 1 | 3 +5 x + y, Z 6 |
| 20 | Abel Grassman | /check | | /check | | | c 2 = b | 2 +4 x +2 y, Z 6 |
| 20 | Abel Grassman | | /check | /check | | | c 2 = b, ( b, n ) = ( c, n ) = 1 | 2 +4 x +2 y, Z 9 |
| 21 | Commuting | /check | | | /check | a = 0 | b = c = 1 | 1 + x + y, Z 7 |
| 21 | Product | | /check | | /check | a = 0 | b = c = 1 | 1 + x + y, Z 7 |
| 22 | Dual Comm Product | /check | | | /check | a = 0 | b = c = 1 | 1 + x + y, Z 7 |
| 22 | Dual Comm Product | | /check | | /check | a = 0 | b = c = 1 | 1 + x + y, Z 7 |
| 23 | Right Transitivity | /check | | | /check | a = 0 | b = 1 , c = - 1 | 2 + x +6 y, Z 7 |
| 23 | Right Transitivity | | /check | | /check | a = 0 | b = 1 , c = - 1 | 2 + x +6 y, Z 7 |
| 24 | Left | /check | | | /check | a = 0 | b = - 1 , c = 1 | 2 +6 x + y, Z 7 |
| | Transitivity | | /check | | /check | a = 0 | b = - 1 , c = 1 | 2 +6 x + y, Z 7 |
| 25 | Schweitzer | /check | | /check | | b, c invert | b = 1 , c = - 1 | 2 + x +5 y, Z 6 |
| 25 | Schweitzer | | /check | /check | | b, c invert | b = 1 , c = - 1 , ( b, n ) = ( c, n ) = 1 | 2 + x +5 y, Z 6 |
| 25 | Schweitzer | /check | | | /check | a = 0 | b = 1 , c = - 1 | 3 + x +6 y, Z 7 |
| 25 | Schweitzer | | /check | | /check | a = 0 | b = 1 , c = - 1 | 3 + x +6 y, Z 7 |
| 26 | Dual of Schweitzer | /check | | /check | | b, c invert | b = 1 , c = - 1 | 2 + x +5 y, Z 6 |
| 26 | Dual of Schweitzer | | /check | /check | | b, c invert | b = 1 , c = - 1 , ( b, n ) = ( c, n ) = 1 | 2 + x +5 y, Z 6 |
| 26 | Dual of Schweitzer | /check | | | /check | a = 0 | b = 1 , c = - 1 | 3 + x +6 y, Z 7 |
| 26 | Dual of Schweitzer | | /check | | /check | a = 0 | b = 1 , c = - 1 | 3 + x +6 y, Z 7 |
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
| 27 | Right Self Distributive | /check | | | /check | | c = 1 - b, a = 0 | 3 x +5 y, Z 7 |
|------|---------------------------|----------|--------|--------|----------|-------------------------|-------------------------|------------------|
| 27 | Right Self Distributive | | /check | | /check | | c = 1 - b, a = 0 | 3 x +5 y, Z 7 |
| 28 | Left Self | /check | | | /check | | c = 1 - b, a = 0 | 3 x +5 y, Z 7 |
| 28 | Distributive | | /check | | /check | | c = 1 - b, a = 0 | 3 x +5 y, Z 7 |
| 29 | Right | /check | | /check | | b, c invert | b = c, 2 b 2 = b | ? |
| 29 | Abelian | | /check | /check | | b, c invert | b = c, 2 b 2 = b | ? |
| 29 | Distributivity | /check | | /check | | a = 0 | b = c, 2 b 2 = b | ? |
| 29 | | | /check | /check | | a = 0 | b = c, 2 b 2 = b | ? |
| 29 | | /check | | | /check | a = 0 | b = c, 2 b = 1 | 2 +3 x +3 y, Z 5 |
| 29 | | | /check | | /check | a = 0 | b = c, 2 b = 1 | 2 +3 x +3 y, Z 5 |
| 30 | Left | /check | | /check | | b, c invert | b = c, 2 b 2 = b | |
| 30 | Abelian | | /check | /check | | b, c invert | b = c, 2 b 2 = b | ? |
| 30 | Distributivity | /check | | /check | | a = 0 | b = c, 2 b 2 = b | ? |
| 30 | | | /check | /check | | a = 0 | b = c, 2 b 2 = b | ? |
| 30 | | /check | | | /check | a = 0 | b = c, 2 b = 1 | 2 +3 x +3 y, Z 5 |
| 30 | | | /check | | /check | a = 0 | b = c, 2 b = 1 | 2 +3 x +3 y, Z 5 |
| 31 | Bol | /check | | /check | | | b = c = 1 | 2 + x + y, Z 6 |
| 31 | Moufang | | /check | /check | | ( b, n ) = ( c, n ) = 1 | b = c = 1 | 2 + x + y, Z 6 |
| 32 | Dual Bol | /check | | /check | | | b = c = 1 | 2 + x + y, Z 6 |
| 32 | Moufang | | /check | /check | | ( b, n ) = ( c, n ) = 1 | b = c = 1 | 2 + x + y, Z 6 |
| 33 | Moufang | /check | | | /check | | b = c = 1 , a = 0 | x + y, Z 5 |
| 33 | | | /check | | /check | | b = c = 1 , a = 0 | x + y, Z 5 |
| 34 | R Bol | /check | | | /check | a = 0 | b 2 = 1 , b = c = 1 | 2 + x + y, Z 7 |
| 34 | | | /check | | /check | a = 0 | b 2 = 1 , b = c = 1 | 2 + x + y, Z 7 |
| 34 | | /check | | | /check | - 1 = b = c | b 2 = 1 , c = 1 , a = 0 | 8 x + y, Z 63 |
| 34 | | | /check | | /check | - 1 = b = c | b 2 = 1 , c = 1 , a = 0 | 8 x + y, Z 63 |
| 35 | L Bol | /check | | | /check | a = 0 | c 2 = 1 , b = c = 1 | 2 + x + y, Z 7 |
| 35 | | | /check | | /check | a = 0 | c 2 = 1 , b = c = 1 | 2 + x + y, Z 7 |
| 35 | | /check | | | /check | - 1 = b = c | c 2 = 1 , b = 1 , a = 0 | x +8 y, Z 63 |
| 35 | | | /check | | /check | - 1 = b = c | c 2 = 1 , b = 1 , a = 0 | x +8 y, Z 63 |
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
| 36 | RC 4 | /check | | | /check | a = 0 | c = b 2 = 1 | 8 x + y, Z 63 |
|------|--------|----------|--------|---------------|----------|---------------------------------------------------|---------------------------------------------------|-----------------------------------|
| 36 | RC 4 | | /check | | /check | a = 0 | c = b 2 = 1 | 8 x + y, Z 63 |
| 36 | | /check | | /check | | a = 0 , b, c invert | c = b 2 = 1 | 8 x + y, Z 63 |
| 36 | | | /check | /check | | a = 0 , b, c invert | c = b 2 = 1 , ( b, n ) = ( c, n ) = 1 | 8 x + y, Z 63 |
| 36 | | /check | | /check | | | b = - 1 , c = 1 | 2 +5 x + y, Z 6 |
| 36 | | | /check | /check | | | b = - 1 , c = 1 , ( b, n ) = ( c, n ) = 1 | 2 +5 x + y, Z 6 |
| 37 | LC 4 | /check | | | /check | a = 0 | b = c 2 = 1 | x +8 y, Z 63 |
| 37 | LC 4 | | /check | | /check | a = 0 | b = c 2 = 1 | x +8 y, Z 63 |
| 37 | LC 4 | /check | | /check | | a = 0 , b, c invert | b = c 2 = 1 | x +3 y, Z 8 |
| 37 | LC 4 | | /check | /check | | a = 0 , b, c invert | b = c 2 = 1 , ( b, n ) = ( c, n ) = 1 | x +4 y, Z 15 |
| 37 | LC 4 | /check | | /check | | | b = - 1 , c = 1 | 2 +5 x + y, Z 6 |
| 37 | LC 4 | | /check | /check | | | b = - 1 , c = 1 , ( b, n ) = ( c, n ) = 1 | 2 +5 x + y, Z 6 |
| 38 | RC 1 | /check | | | /check | a = 0 | c = b 2 = 1 | 8 x + y, Z 63 |
| 38 | RC 1 | | /check | | /check | a = 0 | c = b 2 = 1 | 8 x + y, Z 63 |
| 38 | RC 1 | /check | | /check | | a = 0 , b, c invert | c = b 2 = 1 | 8 x + y, Z 63 |
| 38 | RC 1 | /check | /check | /check | | a = 0 , b, c invert | c = b 2 = 1 , ( b, n ) = ( c, n ) = 1 | 8 x + y, Z 63 |
| 38 | RC 1 | | | /check | | | b = - 1 , c = 1 | 2 +5 x + y, Z 6 |
| 38 | RC 1 | | /check | /check | | | b = - 1 , c = 1 , ( b, n ) = ( c, n ) = 1 | 2 +5 x + y, Z 6 |
| 39 | LC 1 | /check | | | /check | a = 0 , c = 1 | c = - 1 | 3 x +6 y, Z 7 |
| 39 | LC 1 | | /check | | /check | a = 0 , c = 1 | c = - 1 | 3 x +6 y, Z 7 |
| 39 | LC 1 | /check | | /check | | a = 0 , c = 1 , c invert | c = - 1 | 5 x +5 y, Z 6 |
| 40 | LC 3 | /check | /check | /check /check | | a = 0 , c = 1 , c invert | c = - 1 , ( b, n ) = ( c, n ) = 1 c = 1 , b = - 2 | 5 x +5 y, Z 6 3 +4 x + y, Z 6 |
| 40 | LC 3 | | /check | /check | | | c = 1 , b = - 2 , ( b, n ) = ( c, n ) = 1 | 2 +5 x + y, Z 7 |
| 41 | RC 3 | /check | | /check | | | c = 1 , b = - 2 | 3 +4 x + y, Z 6 |
| 41 | RC 3 | | /check | /check | | | c = 1 , b = - 2 , ( b, n ) = ( c, n ) = 1 | 2 +5 x + y, Z 7 |
| 42 | C-Law | /check | | | /check | a = 0 | b = c = - 1 | 4 x +4 y, Z 5 |
| 42 | C-Law | | /check | | /check | a = 0 | b = c = - 1 | 4 x +4 y, Z 5 |
| 42 | C-Law | /check | /check | /check /check | | a = 0 , b = 1 , b, c inv a = 0 , b = 1 , b, c inv | b = c = - 1 b = c = - 1 , ( b, n ) = ( c, n ) = 1 | 3 +5 x +5 y, Z 6 3 +5 x +5 y, Z 6 |
| 43 | LIP | /check | | | /check | a = 0 | c 2 = b 2 = bc = 1 | |
| 43 | LIP | | /check | | /check | a = | 2 2 | ? |
| 44 | RIP | /check | | | /check | | | ? |
| 44 | RIP | | | | | 0 | c = b = bc = 1 | |
| | | | | | | a = 0 | c 2 = b 2 = bc = 1 | ? |
| | | | /check | | /check | a = 0 | c 2 = b 2 = bc = 1 | ? |
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
| 45 | 1st Right CIP | /check | | | /check | a = 0 | bc = 1 | 2 +3 x +4 y, | Z 11 |
|------|-----------------|----------|--------|--------|----------|---------------|----------------------------------|----------------|--------|
| 45 | 1st Right CIP | | /check | | /check | a = 0 | bc = 1 | 2 +3 x +4 y, | Z 11 |
| 45 | 1st Right CIP | /check | | /check | | a = 0 , c inv | bc = 1 | 3 +3 x +3 y, | Z 8 |
| 45 | 1st Right CIP | | /check | /check | | a = 0 , c inv | bc = 1 , ( b, n ) = ( c, n ) = 1 | 3 +3 x +3 y, | Z 8 |
| 46 | 2nd Right CIP | /check | | /check | | | bc = 1 | 3 +3 x +3 y, | Z 8 |
| 46 | 2nd Right CIP | | /check | /check | | | bc = 1 , ( b, n ) = ( c, n ) = 1 | 3 +3 x +3 y, | Z 8 |
| 47 | 1st Left CIP | /check | | | /check | a = 0 | bc = 1 | 2 +3 x +4 y, | Z 11 |
| 47 | 1st Left CIP | | /check | | /check | a = 0 | bc = 1 | 2 +3 x +4 y, | Z 11 |
| 47 | 1st Left CIP | /check | | /check | | a = 0 , b inv | bc = 1 | 3 +3 x +3 y, | Z 8 |
| 47 | 1st Left CIP | | /check | /check | | a = 0 , b inv | bc = 1 , ( b, n ) = ( c, n ) = 1 | 3 +3 x +3 y, | Z 8 |
| 48 | 2nd Left CIP | /check | | /check | | | bc = 1 | 3 +3 x +3 y, | Z 8 |
| 48 | 2nd Left CIP | | /check | /check | | | bc = 1 , ( b, n ) = ( c, n ) = 1 | 3 +3 x +3 y, | Z 8 |
| 49 | R AAIP | /check | | | /check | bc + b = 1 | b = c | 2 +4 x +4 y, | Z 11 |
| 49 | R AAIP | | /check | | /check | bc + b = 1 | b = c | 2 +4 x +4 y, | Z 11 |
| 49 | R AAIP | /check | | | /check | c = b | b + bc = 1 | 2 +3 x + y, | Z 5 |
| 49 | R AAIP | | /check | | /check | c = b | b + bc = 1 | 2 +3 x + y, | Z 5 |
| 50 | L AAIP | /check | | | /check | bc + b = 1 | b = c | 2 +4 x +4 y, | Z 11 |
| 50 | L AAIP | | /check | | /check | bc + b = 1 | b = c | 2 +4 x +4 y, | Z 11 |
| 50 | L AAIP | /check | | | /check | c = b | b + bc = 1 | 2 +3 x + y, | Z 5 |
| 50 | L AAIP | | /check | | /check | c = b | b + bc = 1 | 2 +3 x + y, | Z 5 |
| 51 | R AIP | /check | | /check | | | | a + bx + cy, | Z n |
| 51 | R AIP | | /check | /check | | | | a + bx + cy, | Z n |
| 51 | R AIP | /check | | | /check | | | a + bx + cy, | Z n |
| 51 | R AIP | | /check | | /check | | | a + bx + cy, | Z n |
| 52 | L AIP | /check | | /check | | | | a + bx + cy, | Z n |
| 52 | L AIP | | /check | /check | | | | a + bx + cy, | Z n |
| 52 | L AIP | /check | | | /check | | | a + bx + cy, | Z n |
| 52 | L AIP | | /check | | /check | | | a + bx + cy, | Z n |
| 53 | R SAIP | /check | | /check | | | | a + bx + cy, | Z n |
| 53 | R SAIP | | /check | /check | | | | a + bx + cy, | Z n |
| 53 | R SAIP | /check | | | /check | | | a + bx + cy, | Z n |
| 53 | R SAIP | | /check | | /check | | | a + bx + cy, | Z n |
21
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
/negationslash
| 54 | L SAIP | /check | | /check | | | | a + bx + cy, | Z n |
|------|----------|----------|--------|----------|--------|--------------------|----------------------------------|----------------|-------|
| 54 | L SAIP | | /check | /check | | | | a + bx + cy, | Z n |
| 54 | L SAIP | /check | | | /check | | | a + bx + cy, | Z n |
| 54 | L SAIP | | /check | | /check | | | a + bx + cy, | Z n |
| 55 | R WIP | /check | | | /check | a = 0 , c 2 = 0 | bc = 1 | 3 x +5 y, Z 7 | |
| 55 | R WIP | | /check | | /check | a = 0 , c 2 = 0 | bc = 1 | 3 x +5 y, | Z 7 |
| 55 | R WIP | /check | | /check | | a = 0 , c inv | bc = 1 | 3 x +4 y, | Z 6 |
| 55 | R WIP | | /check | /check | | a = 0 , c inv | bc = 1 , ( b, n ) = ( c, n ) = 1 | ? | |
| 55 | R WIP | /check | | /check | | a = 0 , bc + b = 1 | bc = 1 | ? | |
| 55 | R WIP | | /check | /check | | a = 0 , bc + b = 1 | bc = 1 , ( b, n ) = ( c, n ) = 1 | ? | |
| 56 | L WIP | /check | | | /check | a = 0 , b 2 = 0 | bc = 1 | 3 x +5 y, | Z 7 |
| 56 | L WIP | | /check | | /check | a = 0 , b 2 = 0 | bc = 1 | 3 x +5 y, | Z 7 |
| 56 | L WIP | /check | | /check | | a = 0 , b inv | bc = 1 | 3 x +4 y, Z | 6 |
| 56 | L WIP | | /check | /check | | a = 0 , b inv | bc = 1 , ( b, n ) = ( c, n ) = 1 | ? | |
| 56 | L WIP | /check | | /check | | a = 0 , bc + c = 1 | bc = 1 | ? | |
| 56 | L WIP | | /check | /check | | a = 0 , bc + c = 1 | bc = 1 , ( b, n ) = ( c, n ) = 1 | ? | |
| 57 | E l | /check | /check | | | | | a + bx + cy, | Z n |
| 57 | E l | | /check | /check | | | | a + bx + cy, | Z n |
| 57 | E l | /check | | | /check | | | a + bx + cy, | Z n |
| 57 | E l | | /check | | /check | | | a + bx + cy, | Z n |
| 58 | E r | /check | /check | | | | | a + bx + cy, | Z n |
| 58 | E r | | /check | /check | | | | a + bx + cy, | Z n |
| 58 | E r | /check | | | /check | | | a + bx + cy, | Z n |
| 58 | E r | | /check | | /check | | | a + bx + cy, | Z n |
| 59 | Right F | /check | /check | | | | | a + bx + cy, | Z n |
| 59 | Right F | | /check | /check | | | | a + bx + cy, | Z n |
| 59 | Right F | /check | | | /check | | | a + bx + cy, | Z n |
| 59 | Right F | | /check | | /check | | | a + bx + cy, | Z n |
| 60 | Left F | /check | /check | | | | | a + bx + cy, | Z n |
| 60 | Left F | | /check | /check | | | | a + bx + cy, | Z n |
| 60 | Left F | /check | | | /check | | | a + bx + cy, | Z n |
| 60 | Left F | | /check | | /check | | | a + bx + cy, | Z n |
/negationslash
Table 1: A Table for the Characterization of Varieties of Groupoids and Quasigroups Generated by P x, y ( ) over Z n
| 61 | Medial | /check | /check | | | | | a + bx + cy, Z | n |
|------|--------------------|----------|----------|--------|--------|---------|----------------------------------|------------------|-----|
| 61 | Medial | | /check | /check | | | | a + bx + cy, | Z n |
| 61 | Medial | /check | | | /check | | | a + bx + cy, | Z n |
| 61 | Medial | | /check | | /check | | | a + bx + cy, | Z n |
| 62 | Specialized Medial | /check | /check | | | | | a + bx + cy, | Z n |
| 62 | Specialized Medial | | /check | /check | | | | a + bx + cy, | Z n |
| 62 | Specialized Medial | /check | | | /check | | | a + bx + cy, | Z n |
| 62 | Specialized Medial | | /check | | /check | | | a + bx + cy, | Z n |
| 63 | First | /check | | | /check | | b = c | 2 +4 x +4 y, | Z 7 |
| 63 | Rectangle | | /check | | /check | | b = c | 2 +4 x +4 y, | Z 7 |
| 63 | Rectangle | /check | | /check | | c inv | b = c | 2 +4 x +4 y, | Z 6 |
| 63 | Rectangle | | /check | /check | | c inv | b = c, ( b, n ) = ( c, n ) = 1 | 2 +4 x +4 y, | Z 6 |
| 64 | Second Rectangle | /check | | | /check | | b = - c | 2 +4 x +4 y, | Z 7 |
| 64 | | | /check | | /check | | b = - c | 2 +4 x +4 y, | Z 7 |
| 64 | | /check | | /check | | b inv | b = - c | 2 +4 x +4 y, | Z 6 |
| 64 | | | /check | /check | | b inv | b = - c, ( b, n ) = ( c, n ) = 1 | 2 +4 x +4 y, | Z 6 |
| 65 | C i , i = 1 - 6 | | /check | | /check | | b = c | 3 +5 x +5 y, Z | 7 |
| 66 | CM i , i = 1 - 14 | | /check | | /check | b = - c | b = c | 3 +5 x +5 y, Z | 7 |
Remark 2.1 A summary of the results on the characterization of groupoids and quasigroups generated by P x, y ( ) is exhibited in Table 1. In the table ' G ' stands for 'groupoid', ' Q ' stands for 'quasigroup', 'HYPO' stands for 'hypothesis', 'N AND S' stands for 'necessary and sufficient condition(s)'. Cells with question marks mean examples could not be gotten.
## References
- [1] R. H. Bruck (1966), A survey of binary systems , Springer-Verlag, Berlin-G¨ttingeno Heidelberg, 185pp.
- [2] O. Chein, H. O. Pflugfelder and J. D. H. Smith (1990), Quasigroups and loops : Theory and applications , Heldermann Verlag, 568pp.
- [3] J. Dene and A. D. Keedwell (1974), Latin squares and their applications , the English University press Lts, 549pp.
- [4] E. G. Goodaire, E. Jespers and C. P. Milies (1996), Alternative loop rings , NHMS(184), Elsevier, 387pp.
- [5] T. G. Jaiyeola (2009), A study of new concepts in smarandache quasigroups and loops , ProQuest Information and Learning(ILQ), Ann Arbor, USA, 127pp.
- [6] A. D. Keedwell (2008), When is it hard to show that a quasigroup is a loop? , Comment. Math. Carolin., Vol. 49, No. 2, 241-247.
- [7] A. D. Keedwell (2009), Realizations of Loops and Groups defined by short identities , Comment. Math. Carolin., Vol. 50, No. 3, 373-383.
- [8] A. D. Keedwell and V. A. Shcherbacov (2003), Construction and properties of ( r, s, t ) -inverse quasigroups I , Discrete Math. 266, 275-291.
- [9] A. D. Keedwell and V. A. Shcherbacov (2004), Construction and properties of ( r, s, t ) -inverse quasigroups II , Discrete Math. 288, 61-71.
- [10] M. K. Kinyon and J. D. Phillips (2004), Axioms for trimedial quasigroups , Comment. Math. Univ. Carolinae 45, 287-294.
- [11] M. K. Kinyon and J. D. Phillips (2002), A note on trimedial quasigroups , Quasigroups and related systems 9, 65-66.
- [12] M. K. Kinyon and J. D. Phillips (2005), Rectangular loops and rectangular quasigroups , Comput. Math. Appl. 49, 1679-1685.
- [13] O. U. Kirnasovsky (1995), Linear isotopes of small order groups , Quasigroups and Related Systems, Vol. 2, 51-82.
- [14] V. Krcadinac, V. Volenec (2005), A class of quasigroups associated with a cubic Pisot number , Quasigroups and related systems, 13, 2, 269-280.
- [15] K. Kunen (1996), Quasigroups, Loops and Associative Laws , J. Alg.185, 194-204.
- [16] K. Kunen (1996), Moufang Quasigroups , J. Alg. 183, 231-234.
- [17] R. A. Mollin, C. Small (1987), On permutation polynomials over finite fields , Internat. J. Math. and Math. Sci. Vol. 10, No 3, 535-544.
- [18] D. C. Murdoch (1939), Quasigroups which satisfy certain generalized associative laws , Amer. J. Math. 61, 2, 509-522.
- [19] H. O. Pflugfelder (1990), Quasigroups and loops : Introduction , Sigma series in Pure Math. 7, Heldermann Verlag, Berlin, 147pp.
- [20] M. J. Pelling, D. G. Rogers (1979), Stein quasigroups II: Algebraic aspects , Bull. Austral. Math. Soc. 20, 321-344.
- [21] M. Polonijo (2005), On medial-like identities , Quasigroups and Related Systems 13, 281-288.
- [22] R. L. Rivest (2001), Permutation polynomials Modulo 2 w , Finite Fields and Their Applications 7, 287-292.
- [23] L. Rudolf, G. L. Mullen (1988), When does a polynomial over a finite field Permute the elements of the field? ,The American Mathematical Monthly, Vol. 95, No. 3, 243-246.
- [24] L. Rudolf, G. L. Mullen (1993), When does a polynomial over a finite field Permute the elements of the field? II , The American Mathematical Monthly, Vol. 100, No. 1, 71-74.
- [25] L. V. Sabinin (1999), Smooth quasigroups and loops , Kluver Academic Publishers, Dordrecht, 249pp.
- [26] J. D. H. Smith (2007), An introduction to quasigroups and their representations , Taylor and Francis Group, LLC.
/negationslash
- [27] G. R. Vadiraja Bhatta and B. R. Shankar (2009), Permutation Polynomials modulo n , n = 2 w and Latin Squares , International J. Math. Combin. 2, 58-65.
- [28] W. B. Vasantha Kandasamy (2002), Smarandache loops , Department of Mathematics, Indian Institute of Technology, Madras, India, 128pp.
| null |
[
"Emmanuel Ilojide",
"Temitope Gbolahan Jaiyeola",
"O. O. Owojori"
] |
2014-08-01T21:46:13+00:00
|
2014-08-01T21:46:13+00:00
|
[
"math.GR",
"20N02, 20NO5"
] |
Varieties of groupoids and quasigroups generated by linear-bivariate polynomials over ring Z_n
|
Some varieties of groupoids and quasigroups generated by linear-bivariate
polynomials $P(x,y)=a+bx+cy$ over the ring $\mathbb{Z}_n$ are studied.
Necessary and sufficient conditions for such groupoids and quasigroups to obey
identities which involve one, two, three (e.g. Bol-Moufang type) and four
variables w.r.t. $a$, $b$ and $c$ are established. Necessary and sufficient
conditions for such groupoids and quasigroups to obey some inverse properties
w.r.t. $a$, $b$ and $c$ are also established. This class of groupoids and
quasigroups are found to belong to some varieties of groupoids and quasigroups
such as medial groupoid(quasigroup), F-quasigroup, semi automorphic inverse
property groupoid(quasigroup) and automorphic inverse property
groupoid(quasigroup).
|
1408.0992v1
|
## Tunable asymmetric magnetoimpedance effect in ferromagnetic NiFe/Cu/Co films
E. F. Silva, 1 M. Gamino, 2 A. M. H. de Andrade, 2 M. A. Corrˆ ea, 1 M. V´ azquez, 3 and F. Bohn 1, ∗
1 Departamento de F´ ısica Te´ orica e Experimental,
Universidade Federal do Rio Grande do Norte, 59078-900 Natal, RN, Brazil de F´ ısica,
Universidade Federal do Rio Grande de Sul, 91501-970 Porto Alegre, RS, Brazil
3 Instituto de Ciencia de Materiales de Madrid, CSIC, 28049 Madrid, Spain
(Dated: August 6, 2014)
We investigate the magnetization dynamics through the magnetoimpedance effect in ferromagnetic NiFe/Cu/Co films. We observe that the magnetoimpedance response is dependent on the thickness of the non-magnetic Cu spacer material, a fact associated to the kind of the magnetic interaction between the ferromagnetic layers. Thus, we present an experimental study on asymmetric magnetoimpedance in ferromagnetic films with biphase magnetic behavior and explore the possibility of tuning the linear region of the magnetoimpedance curves around zero magnetic field by varying the thickness of the non-magnetic spacer material, and probe current frequency. We discuss the experimental magnetoimpedance results in terms of the different mechanisms governing the magnetization dynamics at distinct frequency ranges, quasi-static magnetic properties, thickness of the non-magnetic spacer material, and the kind of the magnetic interaction between the ferromagnetic layers. The results place ferromagnetic films with biphase magnetic behavior exhibiting asymmetric magnetoimpedance effect as a very attractive candidate for application as probe element in the development of auto-biased linear magnetic field sensors.
PACS numbers: 75.40.Gb, 75.30.Gw, 75.60.-d
Keywords: Magnetic systems, Magnetization dynamics, Magnetoimpedance effect, Ferromagnetic films
The magnetoimpedance effect (MI), known as the change of the real and imaginary components of electrical impedance of a ferromagnetic conductor caused by the action of an external static magnetic field, is commonly employed as a tool to investigate ferromagnetic materials. For a general review on the effect, we suggest the Ref. 1 . In recent years, the interest for this phenomenon has grown considerably not only for its contribution to the understanding of fundamental physics associated to magnetization dynamics , but also due to the 2 possibility of application of materials exhibiting magnetoimpedance as probe element in sensor devices for lowfield detection 3 . In this sense, the sensitivity and linearity as a function of the magnetic field are the most important parameters in the practical application of magnetoimpedance effect for magnetic sensors . 4 Experiments have been carried out in numerous magnetic systems, including ribbons 5-7 , sheets 8 , wires 9-13 , and, magnetic films 14-16,16,17,17-23 . However, although soft magnetic materials are highly sensitive to small field variations at low magnetic fields, due to magnetization process most of them essentially have nonlinear MI behavior around zero magnetic field, which prevents a simple straightforward derivation of an appropriate signal for sensor applications 23,24 .
The shift of the sensor operational region and the leading of the linear MI behavior at around zero magnetic field can be obtained primarily by applying a bias field or an electrical current to the ordinary MI element 24 . However, this approach proved to be disadvantageous from the practical and technological point of view, mainly due to energetic consumption. Recently, it has been shown that materials exhibiting asymmetric magnetoimpedance (AMI) effect arise as promising alternative with poten- tial of application, opening possibilities for the use of this kind of materials for the development of auto-biased linear magnetic field sensors. For these materials, the asymmetric effects are obtained by inducing an asymmetric static magnetic configuration, usually done by magnetostatic interactions 13,24,25 or exchange bias 4,16,23,26,27 .
For ferromagnetic films, the primary AMI results have been measured for exchange biased multilayers 16,23,27 . Theory and experiment agree well for MI curves shifted by the exchange bias field, following the main features of the magnetization curve, as well as it is verified that the linear region of AMI curves can be tuned to around zero just by modifying the angle between applied magnetic field and exchange bias field, or changing the probe current frequency. On the other hand, another promising possibility of AMI material resides in films presenting biphase magnetic behavior, with hard and soft ferromagnetic phases intermediated by a non-magnetic layer acting together.
In this work, we investigate the magnetoimpedance effect in ferromagnetic NiFe/Cu/Co films. We observe that the MI response is dependent on the thickness of the non-magnetic Cu spacer material, a fact associated to the kind of the magnetic interaction between the ferromagnetic layers. Here we show that the linear region of the asymmetric magnetoimpedance curves in these films is experimentally tunable by varying the thickness of the non-magnetic spacer material, and probe current frequency. The results place ferromagnetic films with biphase magnetic behavior exhibiting asymmetric magnetoimpedance effect as a very attractive candidate for application as probe element in the development of autobiased linear magnetic field sensors.
For this study, we produce Ni 81 Fe 19 (25
2
Instituto
nm)/Cu( t Cu )/Co(50 nm) ferromagnetic films, with t Cu = 0, 1 5, . 3, 5, 7, and 10 nm. The films are deposited by magnetron sputtering from targets of nominally identical compositions onto glass substrates, with dimensions of 8 × 4 mm . 2 A buffer Ta layer is deposited before the NiFe layer to reduce the roughness of the substrate, as well as a cap Ta layer is inserted after the Co layer in order to avoid oxidation of the sample. The deposition is carried out with the following parameters: base vacuum of 10 -8 Torr, deposition pressure of 2 0 mTorr with a 99 99% pure Ar at 32 sccm . . constant flow, and DC source with power of 150 W for the deposition of the NiFe and Co layers, while 100 W for the Cu and Ta layers. During the deposition, the substrate rotates at constant speed to improve the film uniformity, and a constant magnetic field of 2 kOe is applied perpendicularly to the main axis of the substrate in order to induce a magnetic anisotropy and define an easy magnetization axis. X-ray diffraction results, not shown here, calibrate the deposition rates and verify the Co(111) and NiFe(111) preferential growth of all films. Magnetization curves are obtained with a vibrating sample magnetometer, measured along and perpendicular to the main axis of the films, to verify the quasi-static magnetic behavior. Magnetization dynamics is investigated through MI measurements obtained using a RF-impedance analyzer Agilent model E 4991, with E 4991 A test head connected to a microstrip in which the sample is the central conductor. Longitudinal MI measurements are performed by acquiring the real R and imaginary X parts of the impedance Z over a wide range of frequencies, from 0 1 GHz up to 3 0 GHz, with . . 0 dBm (1 mW) constant power applied to the sample, characterizing the linear regime of driving signal, and magnetic field varying between ± 300 Oe, applied along the main axis of the sample. Detailed information on the MI experiment is found in Refs. 21,22 . In order to quantify the sensitivity and MI performance as a function of the frequency, we calculate the magnitude of the impedance change at the low field range ± 6 Oe using the expression 23
$$\frac { | \Delta Z | } { | \Delta H | } = \frac { | Z ( H = 6 \, O e ) - Z ( H = - 6 \, O e ) | } { 1 2 }. \quad ( 1 ) \quad \text{$\text{ and }$} \quad \text{ever},$$
Here, we consider the absolute value of ∆ , since the Z impedance around zero field can present positive or negative slopes, depending on the sample and measurement frequency. In particular, it is verified that | ∆ Z / | | ∆ H | is roughly constant at least for a reasonable low field range.
Figure 1 shows the quasi-static magnetization curves for selected films, measured with the external in-plane magnetic field applied along and perpendicular to the main axis. When analyzed as a function of the the thickness of the non-magnetic Cu spacer material, it is observed an evolution of the shape of the magnetization curves, indicating the existence of a critical thickness range, ∼ 3 nm, which splits the films in groups according the magnetic behavior. For films with t Cu below 3 nm, the NiFe and Co layers are ferromagnetically coupled. The angular dependence of the magnetization curves indicates an uniaxial in-plane magnetic anisotropy, induced by the magnetic field applied during the deposition process, and oriented perpendicularly to the main axis. Despite the similar magnetic behavior, the film with 1 5 nm-. thick Cu layer (not shown) has slightly higher coercive field if compared to the one for the film without spacer material, possibly associated to the increase of the whole sample disorder due to the non-formation of a regular complete Cu layer. The film with t Cu = 3 nm, within the critical Cu thickness range, presents an intermediate magnetic behavior, with smaller magnetic permeability, characterized by the first evidences of a small plateau in the measurement perpendicular to the main axis, and the appearance of magnetization regions associated to distinct anisotropy constants of the NiFe and Co layers. Films with t Cu above the critical thickness range exhibit a biphase magnetic behavior. The two-stage magnetization process is characterized by the magnetization reversion of the soft NiFe layer at low magnetic field, followed by the reversion of the hard Co layer at higher field. None substantial difference between the magnetization curves measured for films with t Cu > 3 nm is verified. In principle, the biphase magnetic behavior suggests that the ferromagnetic layers are uncoupled. The easy magnetization axis remains perpendicular to the main axis of the substrate, as expected. The weaker anisotropy induction and increase of hysteretic losses are primarily related to the roughness of the interfaces and lack of homogeneity of the Cu layer arisen as its thickness is raised 18 .
It is well-known that quasi-static magnetic properties play a fundamental role in the dynamic magnetic response and are reflected in the MI behavior 23 . The shape and amplitude of the magnetoimpedance curves are strongly dependent on the orientation of the applied magnetic field and ac current with respect to the magnetic anisotropies, magnitude of the external magnetic field, and probe current frequency, as well as are directly related to the main mechanisms responsible for the transverse magnetic permeability changes: skin effect and ferromagnetic resonance (FMR) effect 23,28,29 . However, magnetoimpedance effect can also provide a further insights on the nature of the interactions governing the magnetization dynamics and energy terms affecting the transverse magnetic permeability.
Regarding the MI results, Fig. 2 shows the MI curves, at the selected frequency of 0 75 GHz, for the films with . different thicknesses t Cu of the non-magnetic Cu spacer material. All samples exhibit a double peak behavior for the whole frequency range, a signature of the perpendicular alignment of the external magnetic field and ac current with the easy magnetization axis. An interesting feature related to the MI behavior of the NiFe/Cu/Co films resides in the amplitude and position of the peaks with the thickness of the non-magnetic Cu spacer material, and the probe current frequency.
Films with t Cu < 3 nm present the well-known sym-
M/M
1.0
0.5
0.0
-0.5
-1.0
1.0
0.5
0.0
-0.5
-1.0
-100
-50
0
50
100
H (Oe)
-100
-50
0
50
100
H (Oe)
FIG. 1: Representative normalized quasi-static magnetization curves for selected NiFe/Cu/Co films with different thicknesses of the non-magnetic Cu spacer material, obtained with the in-plane magnetic field applied along ( ‖ ) and perpendicular ( ⊥ ) to the main axis. Films with t Cu below 3 nm present behavior similar to that verified for the film with t Cu = 1 5 nm, while the . films with t Cu above the critical thickness range have behavior identical to the one observed for the film with t Cu = 7 nm. The film with t Cu = 3 nm is within the critical Cu thickness range and have an intermediate behavior.
metric magnetoimpedance behavior for anisotropic systems. The MI curves have the double peak behavior, symmetrical at aroud H = 0, with peaks with roughly the same amplitude. For frequencies up to ∼ 0 85 GHz, . the position of the peaks remains unchanged close to the anisotropy field, indicating that the skin effect is the main responsible by the changes of transverse magnetic permeability governing the magnetization dynamics. For frequencies above this value, not presented here, besides the skin effect, the FMR effect also becomes an important mechanism responsible for variations of the MI effect, a fact evidenced by the displacement of the peaks position toward higher fields as the frequency is increased. The contribution of the FMR effect to Z is also verified using the method described by Barandiar´n a et al. 30 , and previously employed by our group 17 .
On the other hand, films with t Cu ≥ 3 nm present noticeable asymmetric magnetoimpedance effect. The asymmetric behavior is assigned by two characteristic features: shift of the MI curve in field, depicted by the asymmetric position of the peaks, and asymmetry in shape, evidenced by the difference of amplitude of the peaks. Figure 3 shows the evolution of the MI curves, at selected frequencies between 0 5 GHz and 3 0 GHz, for . . the film with t Cu = 7 nm, as an example of the experimental result obtained for the ferromagnetic films with t Cu above 3 nm. Here, it is important to notice that the presented MI behavior is acquired when the field goes from negative to positive values. However, the MI curves are acquired over a complete magnetization loop and present hysteretic behavior. In particular, when the field goes from positive to negative values, the MI behavior is reversed.
From Fig. 3, regarding the position of the peaks, since the skin effect commands the dynamical behavior, the peaks remain invariable at the low frequency range. For this film, the peak at negative field values is located at ∼ -4 Oe, while the one at positive fields is at ∼ +30 Oe. For the other films with t Cu > 3 nm and biphase magnetic behavior, the peak at positive field is placed at similar value, although the location of the peak at
FIG. 2: The MI curves at frequency of 0 75 GHz measured . for the films with different thicknesses of the non-magnetic Cu spacer material t Cu . The MI curves are acquired over a complete magnetization loop and present hysteretic behavior. Here, we show just part of the curve, when the field goes from negative to positive values, to make easier the visualization of the whole MI behavior.
negative field present dependence with t Cu , as will be discussed. With respect to the amplitude of the peaks at low frequencies, for all films with t Cu > 3 nm, the MI behavior exhibits an asymmetric two-peak behavior, with the peak at negative field being with higher amplitude than the peak at positive field. As a signature of the emergence of the FMR effect, the displacement of the peak at negative field begins at ∼ 0 6 GHz, while the .
s
3 nm
s
M/M
7 nm
FIG. 3: Evolution of the experimental MI curves for selected frequencies for the ferromagnetic biphase fil with t Cu = 7 nm. Similar results are obtained for all the ferromagnetic films with t Cu above 3 nm and biphase magnetic behavior. We show just part of the curve, when the field goes from negative to positive values.
position of peak at positive field starts changing at ∼ 1 1 . GHz. Above ∼ 1 5 GHz, strong skin and FMR effects are . responsible by the MI variations. At this high frequency range, the asymmetry still remains in the portion of the impedance curve around the anisotropy fields. However, the displacement of the peaks toward higher fields suppress the impedance peak asymmetry, in position and amplitude, resulting is symmetric peaks around H = 0 with same amplitude. For the film with t Cu = 3 nm with intermediate magnetic behavior, similar features are observed respectively at ∼ 0 75 GHz, . ∼ 1 1 GHz, and . ∼ 2 0 . GHz.
The most striking finding in the dynamic magnetic response resides in the asymmetry of the MI curves measured for the films with biphase magnetic behavior. It is important to notice that the magnetoimpedance response is nearly linear for low magnetic field values, and the shape of the Z curves depends on the thickness of the non-magnetic Cu spacer material and probe current frequency. As a consequence, the best response can be tuned by playing with both parameters. Figure 4 shows the frequency spectrum of impedance variations between ± 6 Oe, as defined by Eq. (1), for each film, indicating the sensitivity around zero field.
From the figure, we verify that the films split in different groups according the sensitivity around zero field.
FIG. 4: Frequency spectrum of impedance variations between ± 6 Oe for the films with different thicknesses of the nonmagnetic Cu spacer material t Cu , indicating the sensitivity around zero field. Notice the kind of saturation effect observed as the t Cu increases above 3 nm, related to the amplitude and frequency at which the maximum impedance change is reached.
It is important to notice that each one is related to a given magnetic behavior, verified through the magnetization curves. Films with t Cu < 3 nm have the largest sensitivity values at ∼ 1 0 GHz, the film with . t Cu = 3 nm at ∼ 0 9 GHz, while the ones with . t Cu > 3 nm at ∼ 0 75 . GHz. For all of them, the sensitivity peak is found to be at frequencies just after the FMR effect starts appearing, and while the MI peaks are still placed close to the anisotropy fields. The highest sensitivity is observed for the films with t Cu > 3 nm, reaching ∼ 8 mΩ/Oe, and seems to be insensitive to the thickness of the nonmagnetic spacer material. In this situation, the AMI curves have nearly linear behavior at low magnetic field values, and the slope of the linear region at nearly zero field is negative, due to the shape of the MI curve.
Our results raise an interesting issue on the behavior of the peaks in the magnetoimpedance effect and the energy terms affecting the transverse magnetic permeability. Generally, our films consist of two ferromagnetic layers, with distinct anisotropy field strengths, intermediated by non-magnetic spacer material. We interpret our experimental data as a result caused by the competition between two types of magnetic interactions between ferromagnetic layers: exchange coupling between touching ferromagnetic phases, and long-range dipolarlike or magnetostatic coupling 31 . In particular, they are strongly dependent on the thickness of the spacer material, and the action of each one affect in different ways the MI behavior.
If both ferromagnetic layers are quasi-saturated, where there are no walls with wall stray fields, the coupling should adjust the magnetization of the two ferromagnetic layers parallel to each other. For t Cu < 3 nm, the strong coupling is caused by the exchange interaction between
touching ferromagnetic layers and through pinholes in the non-magnetic spacer, and the whole sample behaves as a single ferromagnetic layer 18 . In this sense, we confirm the expected symmetric magnetoimpedance behavior of single anisotropic systems.
For t Cu > 3 nm, the Cu layer is completelly filled 18 , and the nature of the coupling is magnetostatic. If the ferromagnetic layers were completely uncoupled, one could expect multiple peak MI behavior, associated to the anisotropy fields of each different layers, around ± 30 Oe and ± 10 Oe. This behavior is not verified here, indicating that the AMI can not be explained assuming independent reversal of the NiFe and Co layers. Thus, the asymmetry arises as a result of the magnetostatic coupling between the ferromagnetic layers. The origin of the magnetostatic coupling is ascribed to the hard Co magnetic phase in terms of an effective bias field, induced by divergences of magnetization mainly due to roughness in the interfaces and limits of the sample 24 , that must be taken into account as a contributor to the transverse permeability. The field penetrates the non-magnetic spacer layer and results in a torque on the magnetization of the opposite layer. It is important to point out that the anisotropy field of the hard Co layer is considerable larger than the soft NiFe layer, the reason why this asymmetric behavior is not verified in traditional multilayers.
In this sense, the main features of the asymmetric magnetoimpedance verified in films with biphase magnetic behavior can be explained through the effective interaction between the ferromagnetic layers. The influence on the soft NiFe layer is dependent on the magnetic state of the hard Co layer, as well as on the thickness of the Cu layer spacer. The difference of amplitude of the peaks is understood from the magnetic saturated state in terms of the orientation of the two layers. The peak at negative field is higher than that in positive field, since the magnetization of the soft NiFe layer is parallel to the magnetization of the hard Co layer and consequently to the magnetostatic field, as well as to the external field. Since the magnetization of the NiFe layer is reverted as the field is increased, the sense of this magnetization with respect to the magnetostatic field is modified, and this form closes the magnetic flux, resulting a lower peak 24 . Similar dependance with the orientation of magnetizations has already been verified in field-annealed Co-based amorphous ribbons . When the MI measurement is ana4 lyzed for decreasing magnetic field, the reverted behavior is observed, with the higher and smaller peaks at positive and negative fields, respectively, since the sense of the magnetization of the hard Co layer is the opposite.
By employing MI measurements, it is possible to estimate the effective coupling strenght between the NiFe and Co layers. This can be done by considering the location of the peaks in the MI curves at the low frequencies. Figure 5 shows the magnetic field values in which the impedance peaks are located, for different thicknesses of the non-magnetic Cu spacer material, at the low frequency range. The position of the impedance peak at
FIG. 5: Magnetic field values in which the impedance peaks are located, at 0 5 GHz, for the films with different thicknesses . of the non-magnetic Cu spacer material t Cu . Notice that the films with t Cu ≤ 1 5 nm present double peak behavior, sym-. metrical at aroud H = 0.
negative field presents a noticeable dependence with t Cu . In particular, it is verified a reduction of the field value where the peak is located as t Cu is increased, corroborating the assumption of a magnetostatic origin of the coupling between the ferromagnetic layers. In this sense, we interpret the reduction as an indication of the decrease of the bias field intensity acting on the soft NiFe layer as the Cu thickness is increased. On the other hand, the peak at positive field is located at ∼ 30 Oe, except for the sample without the non-magnetic spacer material. The constancy in the peak location, irrespective of t Cu and field value of the Co reversion for each sample, suggests that this value corresponds to a intrinsic feature of the ferromagnetic Co layer, since it presents similar thicknesses for all samples. Thereby, we understand it as the magnitude of the bias field induced by the hard Co layer.
In conclusion, we have investigated the magnetoimpedance effect in ferromagnetic NiFe/Cu/Co films and observed the dependence of the MI curves, in particular, amplitude and position of the peaks, with the thickness of the non-magnetic Cu spacer material. We have verified that the MI response of these films can be taylored by the kind of magnetic interaction between the ferromagnetic layers. In this sense, the coupling between the layers is usefull to develop materials with asymmetric MI. From the results, we have observed the crossover between two distinct magnetic behavior, associated to distict kind of the magnetic behavior between the ferromagnetic layers, exchange interaction and magnetostatic coupling, as the Cu thickness is altered crossing through t Cu = 3 nm. Thus, we have tuned the linear region of the asymmetric magnetoimpedance curves around zero magnetic field by varying the thickness of the non-magnetic spacer material, and probe current frequency. The highest sensitivity is observed for the films with t Cu > 3 nm,
reaching ∼ 8 mΩ/Oe, and seems to be insensitive to the thickness of the non-magnetic spacer material. These results extend the possibilities for application of ferromagnetic films with asymmetric magnetoimpedance as probe element for the development of auto-biased linear magnetic field sensors, placing films with biphase magnetic behavior as promissing candidates to optimize the MI performance.
## Acknowledgments
The authors thank Vivian Montardo Escobar for the fruitful discussions. The research is supported by the
- ∗ Electronic address: [email protected]
- 1 M. Knobel, M. V´ azquez, and L. Kraus, Handbook of Magnetic Materials (Elsevier Science B. V., Amsterdam, 2003), (ed. K. H. Buschow)Ch. 5, 1 - 69.
- 2 A. Yelon, D. Menard, M. Britel, and P. Ciureanu, Appl. Phys. Lett. 69 , 3084 (1996).
- 3 H. Chiriac, M. Tibu, A.-E. Moga, and D. D. Herea, J. Magn. Magn. Mater. 293 , 671 (2005).
- 4 C. G. Kim, K. J. Jang, D. Y. Kim, and S. S. Yoon, Appl. Phys. Lett. 75 , 2114 (1999).
- 5 R. L. Sommer and C. L. Chien, Phys. Rev. B 53 , R5982 (1996).
- 6 K. R. Pirota, L. Kraus, M. Knobel, P. G. Pagliuso, and C. Rettori, Phys. Rev. B 60 , 6685 (1999).
- 7 F. Amalou and M. A. M. Gijs, J. Appl. Phys. 95 , 1364 (2004).
- 8 M. Carara, A. Gundel, M. N. Baibich, and R. L. Sommer, J. Appl. Phys 84 , 3792 (1998).
- 9 R. S. Beach and A. E. Berkowitz, Appl. Phys. Lett. 64 , 3652 (1994).
- 10 J. Vel´zquez, M. V´zquez, D.-X. Chen, and A. Hernando, a a Phys. Rev. B 50 , 16737 (1994).
- 11 D. P. Makhnovskiy, L. V. Panina, and D. J. Mapps, Appl. Phys. Lett. 77 , 121 (2000).
- 12 M. V´ azquez, M. Knobel, M. L. Sanchez, R. Valenzuela, and A. P. Zhukov, Sens. Actuators A 59 , 20 (1997).
- 13 G. V. Kurlyandskaya et al. , J. Appl. Phys. 87 , 4822 (2000).
- 14 R. L. Sommer and C. L. Chien, Appl. Phys. Lett. 67 , 3346 (1995).
- 15 A. D. C. Viegas et al. , J. Appl. Phys. 101 , 033908 (2007).
Brazilian agencies CNPq (Grants No. 471302/2013-9, No. 310761/2011-5, No. 555620/2010-7), CAPES, and FAPERN (Grant Pronem No. 03 2012). / M.A.C. and F.B. acknowledge financial support of the INCT of Space Studies.
- 16 R. B. da Silva et al. , Appl. Phys. Lett. 94 , 042501 (2009).
- 17 M. A. Corrˆ ea et al. , J. Phys. D: Appl. Phys. 43 , 295004 (2010).
- 18 M. S. Marques et al. , Thin Solid Films 520 , 2173 (2012).
- 19 G. Kurlyandskaya et al. , J. Magn. Magn. Mater. 242-245 , 291 (2002).
- 20 G. V. Kurlyandskaya et al. , J. Phys.: Cond. Mat. 16 , 6561 (2004).
- 21 M. A. Corrˆ ea et al. , J. Phys. D: Appl. Phys. 41 , 175003 (2008).
- 22 A. M. H. de Andrade, M. A. Corra, A. D. C. Viegas, F. Bohn, and R. L. Sommer, J. Appl. Phys. 115 , 103908 (2014).
- 23 R. B. da Silva et al. , Appl. Phys. Lett. 104 , 102405 (2014).
- 24 J. Torrej´n, M. V´zquez, and L. V. Panina, J. Appl. Phys. o a 105 , 033911 (2009).
- 25 J. Sinnecker, A. de Araujo, R. Piccin, M. Knobel, and M. V´ azquez, J. Magn. Magn. Mater. 295 , 121 (2005).
- 26 N. A. Buznikov, C. Kim, C.-O. Kim, and S.-S. Yoon, Appl. Phys. Lett. 85 , 3507 (2004).
- 27 C. Garc´ ıa, J. M. Florez, P. Vargas, and C. A. Ross, Appl. Phys. Lett. 96 , 232501 (2010).
- 28 R. L. Sommer and C. L. Chien, Appl. Phys. Lett. 67 , 857 (1995).
- 29 M. A. Corrˆ ea et al. , J. Appl. Phys. 110 , 093914 (2011).
- 30 J. M. Barandiaran, A. Garcia-Arribas, and D. de Cos, J. Appl. Phys. 99 , 103904 (2006).
- 31 T. C. Schulthess and W. H. Butler, J. Appl. Phys. 87 , 5759 (2000).
|
10.1063/1.4895708
|
[
"E. F. Silva",
"M. Gamino",
"A. M. H. de Andrade",
"M. A. Corrêa",
"M. Vázquez",
"F. Bohn"
] |
2014-08-05T14:55:17+00:00
|
2014-08-05T14:55:17+00:00
|
[
"cond-mat.mtrl-sci",
"cond-mat.mes-hall"
] |
Tunable asymmetric magnetoimpedance effect in ferromagnetic NiFe/Cu/Co films
|
We investigate the magnetization dynamics through the magnetoimpedance effect
in ferromagnetic NiFe/Cu/Co films. We observe that the magnetoimpedance
response is dependent on the thickness of the non-magnetic Cu spacer material,
a fact associated to the kind of the magnetic interaction between the
ferromagnetic layers. Thus, we present an experimental study on asymmetric
magnetoimpedance in ferromagnetic films with biphase magnetic behavior and
explore the possibility of tuning the linear region of the magnetoimpedance
curves around zero magnetic field by varying the thickness of the non-magnetic
spacer material, and probe current frequency. We discuss the experimental
magnetoimpedance results in terms of the different mechanisms governing the
magnetization dynamics at distinct frequency ranges, quasi-static magnetic
properties, thickness of the non-magnetic spacer material, and the kind of the
magnetic interaction between the ferromagnetic layers. The results place
ferromagnetic films with biphase magnetic behavior exhibiting asymmetric
magnetoimpedance effect as a very attractive candidate for application as probe
element in the development of auto-biased linear magnetic field sensors.
|
1408.0993v2
|
## Quantum advantage for distributed computing without communication
/suppress L. Czekaj, 1, 2 M. Paw/suppressowski, l 1,2 T. V´ ertesi, 3 A. Grudka, 4 M. Horodecki, 1, 2 and R. Horodecki 5, 2
1 Faculty of Mathematics, Physics and Informatics, Gda´sk University, 80-952 Gda´sk,Poland n n 2
Institute for Nuclear Research, Hungarian Academy of Sciences, H-4001 Debrecen, P.O. Box 51, Hungary 4 Faculty of Physics, Adam Mickiewicz University, Umultowska 85, 61-614 Pozna´, Poland n 5 Faculty of Applied Physics and Mathematics, Gda´sk University of Technology, 80-952 Gda´sk, Poland n n
National Quantum Information Centre in Gda´sk, n 81-824 Sopot, Poland 3
Understanding the role that quantum entanglement plays as a resource in various information processing tasks is one of the crucial goals of quantum information theory. Here we propose a new perspective for studying quantum entanglement: distributed computation of functions without communication between nodes. To formalize this approach, we propose identity games . Surprisingly, despite of no-signaling, we obtain that non-local quantum strategies beat classical ones in terms of winning probability for identity games originating from certain bipartite and multipartite functions. Moreover we show that, for majority of functions, access to general non-signaling resources boosts success probability two times in comparison to classical ones, for number of outputs large enough.
Since famous Shor algorithm [1], a basic branch of quantum information science is devoted to search for quantum advantage in computing. In particular a lot of effort was devoted to distributed computation - mostly in terms of communication complexity. A strictly related domain is a huge 'industry' of Bell inequalities, which are actually an instance of distributed computation.
distributed computation of total function, if no communication between nodes is allowed? We answer this question introducing and studying Identity games (in short Id games , the origin of the name will be clarified later) which may be viewed as special type of Bell inequalities [4]. The setup is depicted in FIG. 1 and described with more details in further part of the paper.
So far, the paradigm of distributed computation assumed communication between the nodes. On the one hand the communication was directly quantified and the cost of communication was measured as in communication complexity problems (see [2] for review). On the other hand some global inputs/outputs processing was performed by a referee [3]. Here we focus on computation without communication.
Namely, we want to address the following question: is quantum mechanics superior to classical theory regarding
FIG. 1: Distributed computing without communication (Id games) Players share non-signaling resource, here understood as a correlations p /vector ( y /vector | x ). Each player receives private message x k ∈ { 0 1 , , . . . , m i - } 1 from referee. Messages are distributed according to the uniform probability distribution. Based on the input x k and his part of shared resource, player k computes the output y k ∈ { 0 1 , , . . . , m o - } 1 . No communication is allowed between the players during the game. Players have to compute some total function f , i.e. they win the game if their outputs fulfill /vector y = f /vector ( x ).
We should here stress the difference between the problem we pose and two existing results/approaches. The first one is 'guess your neighbour input' (GYNI) game [5]. GYNI game may be viewed as a distributed computation without communication with global constraints imposed on the input set. In contrast, in our approach, we consider all inputs. Secondly, in [6] the communication complexity problem for two parties was translated into scenario, where the goal of one of the parties is to compute some function without communication, but conditioned on an event that other party will not abort. This conditioning is a form of communication, while in our case there is no communication whatsoever.
Intuitively, computing of some non-trivial distributed function requires signaling. Hence there should not be a difference not only between quantum and classical theories, but also non-signaling theory should not give advantage. As an example we may refer to GYNI game where all inputs are allowed. In that case there is no advantage for non-signaling theories.
In this paper we show that quantum systems are superior to classical ones for bipartite and multipartite scenarios. We find that, for the simplest bipartite systems (i.e. each player has binary input and output) non-signaling theory does not offer advantage. However, if we move to more complicated bipartite setups, we obtain that quantum strategies beat classical ones.
Before we present our results, we discuss how the approach proposed in this paper relates to other frameworks for studying non-locality, in particular to Bell inequalities.
The value of Bell expression is the average probability that some function f /vector ( x ) of parties' inputs /vector x is equal to some function g /vector ( y ) of their outputs /vector y [22]. The most well
studied type of Bell inequalities are so-called XOR games [8]. There, the function g is simply XOR of the outcomes of all parties. Celebrated CHSH [9] inequality is the most well known example of this type. The main advantage of XOR games is that they allow to change a more complex function g into computing XOR's. As we said, the goal of distributed computation is to reach a point when no further processing is required. Because in this case g is just the identity it is why we call the inequalities of this type Id games (i.e. in the case of total functions).
Strangely this case has been left almost untouched with only one, as mentioned above, exception which is GYNI game [5]. As contrary to GYNI game where partial function is computed, here we focus on the total functions f . To our knowledge there are no examples of Bell inequalities of type of Id games for total functions.
In analogy to communication complexity problems [2], these Id games may by viewed as functional games where only one output value /vector y is a valid answer for given input /vector x . This is in contrast to the relational games (e.g. XOR game) where more than one answer is correct for given input /vector x .
As in communication complexity problems, Id games are intended to capture an advantage from using nonlocal resources in distributed computation. In contrast with communication complexity, we do not measure advantage in terms of communication cost. Actually, no communication between parties is allowed in our approach. We focus on the probability that players return correct output.
This paper is structured as follows: Firstly we formally define Id games and resources that can be used to play them. Then we show various examples with quantum advantage for bipartite scenario and discuss the gap between classical and non-signaling resources. Next we move to tripartite scenario. At the end we show that for a vast majority of the functions the classical resources are practically useless while the no-signalling ones allow for strictly better results. We conclude the paper with a discussion of our results and directions for future research.
In the Method section we present numerical examples, some statistics describing number of nontrivial Id games and other details.
## RESULTS
## Identity games
We study Bell inequalities in terms of nonlocal games. We consider n -player games. Each player k receives as an input a private message from referee x k ∈ { 0 1 , , . . . , ( m i -1) } and returns to him an output y k ∈ { 0 1 , , . . . , ( m o -1) } . m i and m o denote numbers of possible inputs and outputs respectively. Inputs are distributed according to some joint probability distribution q /vector ( x ) on vector /vector x ∈ { 0 1 , , . . . , ( m i -1) } n consisting of messages sent to each player. The goal for players is to achieve maximal value of functional called winning probability:
$$\omega = \sum _ { \vec { x }, \vec { y } } \delta ( f ( \vec { x } ) = g ( \vec { y } ) ) q ( \vec { x } ) p ( \vec { y } | \vec { x } ), \quad \quad ( 1 ) \\ \intertext { m a t / \lambda \, m a t / \lambda \, m a t \, m a t \, m a t \, m a t \, m a t \, m a t \, m a t }$$
where f /vector ( x ) and g /vector ( y ) are some functions of inputs and outputs and define the game, p /vector ( y /vector | x ) is conditional probability distribution describing the strategy.
In the preparation step, players are informed on the input distribution q /vector ( x ) and functions f and g . Then they establish common strategy p /vector ( y /vector | x ) based on the shared resources (classical, quantum, non-signaling). During the game, no communication is allowed between the players. Hence they make their decision on the base of their own inputs and shared resources. We will denote maximal winning probability for classical, quantum and nonsignaling resources as ω ∗ cl , ω ∗ q , ω ∗ ns respectively. The game is non-trivial if ω ∗ cl < ω ∗ ns .
In this paper we look for non-trivial Id games which are defined by the requirement that g is an identity function. Furthermore we consider only the case where g /vector ( x ) = 1 /m n i . That leads to winning probability in the form:
$$\omega = \sum _ { \vec { x }, \vec { y } } \delta ( f ( \vec { x } ) = \vec { y } ) \frac { 1 } { m _ { i } ^ { n } } p ( \vec { y } | \vec { x } ). \quad \quad ( 2 )$$
## Strategies
We describe strategies used by players in terms of generalized probabilistic theories. Players share n -partite resources. Each player perform some measurement x k on his part of the resource. The measurements are performed simultaneously and their results are distributed according to joint conditional probability p /vector ( y /vector | x ). The following theories, characterized by the conditions imposed on the p /vector ( y /vector | x ), are important for our purposes: (i) classical, (ii) quantum and (iii) non-signaling. In classical theory p /vector ( y /vector | x ) is a mixture of classical local probabilities according to unknown hidden parameter λ . Two party resource p /vector ( y /vector | x ) is of the form
$$\Phi _ { c l } ( y _ { 1 }, y _ { 2 } | x _ { 1 }, x _ { 2 } ) = \sum _ { \lambda } p ( \lambda ) p ( y _ { 1 } | x _ { 1 }, \lambda ) p ( y _ { 2 } | x _ { 2 }, \lambda ). \quad ( 3 ) \\ \Gamma _ { c l } \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \label l \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \attempties$$
For quantum theory, conditional probability distribution has to be reproduced by local measurements (described by projectors P (1) x,y , P (2) x,y ) performed on the shared quantum state ρ 12 :
$$p _ { q } ( y _ { 1 }, y _ { 2 } | x _ { 1 }, x _ { 2 } ) = \text{tr} \left [ \rho _ { 1 2 } P _ { x _ { 1 }, y _ { 1 } } ^ { ( 1 ) } \otimes P _ { x _ { 2 }, y _ { 2 } } ^ { ( 2 ) } \right ]. \quad ( 4 ) \\ \text{In non-signaling theories, the only condition is that}$$
In non-signaling theories, the only condition is that p /vector ( y /vector | x ) cannot signal between players. In formal way this condition is expressed for player 1 as
$$\overset {, \in } { \text{ each } } \quad \forall \sum _ { x _ { 2 }, x _ { 2 } ^ { \prime } } p _ { n s } ( y _ { 1 }, y _ { 2 } | x _ { 1 }, x _ { 2 } ) = \sum _ { y _ { 2 } } p _ { n s } ( y _ { 1 }, y _ { 2 } | x _ { 1 }, x _ { 2 } ^ { \prime } ) \ \ ( 5 )$$
and similarly for player 2.
To identify functions which lead to non-trivial Id games, we grouped the functions into equivalence classes invariant under local operations or players reordering. Namely we treat functions f 1 and f 2 as equivalent if f 2 might be obtained from f 1 by composition of the following operations: (i) input relabelling; (ii) output relabelling; (iii) output conditioning on local input: player k returns as an output h k ( y , x k k ); (iv) players reordering. These operations enabled for significant reduction of problem complexity. Then for one representative function from each equivalence class we obtained ω ∗ cl and ω ∗ ns by proper optimisation of (2): (i) p cl ( /vector y /vector | x ) is convex combination of extreme strategies which are easy to enumerate, we calculate ω ∗ cl by maximisation of (2) over that set; (ii) ω ∗ ns was obtained from linear programming with constraints imposed by no signaling conditions.
For non-trivial Id games we analysed performance of quantum strategies. Since we were not able to provide analytical results, we focused on upper bounding ω ∗ q numerically. For this purpose we used semidefinite programming (SDP) an approach introduced in [10, 11]. Beside the upper bound, for some cases we provided also concrete examples of quantum strategy offering advantage over classical ones. For all considered bipartite scenarios these advantage is equal to upper bound obtained from SDP.
In the rest of the paper, for simplicity, we take binary inputs ( m o = 2). The results obtained in this and following section can be generalized to any m o > 2 in a straightforward way.
## Two-player games
In the scenarios with binary input (i.e. m ( k ) i = 2; m o = 2), there are 4 4 = 256 functions, however none of them leads to non-trivial Id game.
Three inputs per player - The simplest setup where we can find non-trivial Id games and quantum advantage is the case where each player obtains one of the three input symbols ( m ( k ) i = 3; m o = 2). There are 4 9 = 262144 different functions in that setup which reduce to 2162 equivalence classes. There are 256 equivalence classes for which ω ∗ cl < ω ∗ ns . Together they contain 196992 functions. For all of these classes ω ∗ cl = 4 9 / ≈ 0 4444 and . ω ∗ ns = 1 2. /
We analyzed some of these classes in more detail looking for quantum strategies which may have advantage over classical ones. Interestingly we found that nonsignaling strategies winning the games are equivalent to using PR-Boxes [12] and an explicit example of optimal quantum strategy corresponds to the CHSH Bell experiment [9].
We calculated upper bound for ω q using 2-nd level hierarchies of SDP [10]. The highest bound is about 5 2% . better than classical strategy and reads 0 4675. . In comparison non-signaling strategy has 12 5% advantage. On .
the other hand we found that for some classes the bound obtained from SDP is equal (with respect to the numerical precision) to ω ∗ cl .
The example of function with the highest SDP bound is:
$$\begin{smallmatrix} x _ { 2 } \langle x _ { 1 } ( y _ { 2 }, y _ { 1 } ) & 0 & 1 & 2 \\ 0 & 0, 0 & 0, 0 & 0, 0 \\ 1 & 0, 0 & 1, 1 & 1, 1 \\ 2 & 0, 1 & 0, 1 & 1, 1 \end{smallmatrix}$$
Each cell refers to the output of function for given input. First and second player's inputs are in the columns and rows respectively.
Optimal classical strategy for this game is obtained when both players return 0 all the time. Optimal nonsignaling strategy is analogical to using PR-Box:
(6)
Quantum strategy which is optimal for this game (i.e. its attains the SDP bound [10]) looks as follows: Player 1 outputs deterministically 0 when he gets x 1 = 0 as an input. Otherwise he relabels his input { 1 } → { 1 } { , 2 } → { 0 } and plays XOR game with optimal quantum strategy. Player 2 performs analogically with relabeling { 1 } → { 0 } { , 2 } → { 1 } . Note that, when one of the players return 0 deterministically, another one returns 0 or 1 completely random. Success probability ω q for this strategy is (1 + 3 2 + √ 2+2 2 ) / 9 ≈ 0 4675. .
$$\text{mer} \text{-} \\ \text{eside} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{--} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{\text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--}
\text{for} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{-} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{-} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{-} \text{-} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{---} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{\text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--} \text{--}
\text{from} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{In} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{\text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in}
\text{from} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text1\text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in}\text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in}\
\text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from} \text{from}
\text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to}
\text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \text{to} \$$
However, if we consider another Id game (see Methods section for details), we find that the quantum maximum ω ∗ q is attained by using non-maximally entangled twoqubit states.
On the other hand, Id games turn out to be useful as dimension witnesses as well, i.e., they are able to witness Hilbert space dimension [14]. In an Id game given in Methods section the players must share at least three dimensional component spaces for maximum quantum violation ω ∗ q to happen.
We also found that the SDP bound is equal to classical limit for functions which were symmetric according to players exchange. As an example we present the following symmetric function:
$$\begin{smallmatrix} x _ { 2 } \langle x _ { 1 } ( y _ { 2 }, y _ { 1 } ) & 0 & 1 & 2 \\ 0 & 0, 0 & 0, 0 & 1, 0 \\ 1 & 0, 0 & 1, 1 & 1, 1 \\ 2 & 0, 1 & 1, 1 & 0, 0 \end{smallmatrix}$$
An optimal non-signaling strategy is analogical to us-
ing PR-box:
/negationslash
$$p ( y _ { 1 }, y _ { 2 } | x _ { 1 }, x _ { 2 } ) = \begin{cases} 1 / 2 & \text{if $y_{1}\oplus y_{2}=\left(x_{2}=2\wedge x_{1}\neq 2\right)$} & \text{over} \\ 0 & \text{otherwise} \end{cases} \quad \text{where}$$
(7)
The above feature that quantum strategies do not offer advantage over classical ones turns out to be true if the players receives four inputs each (instead of three). However, we found counterexample for the case of five inputs per player (see Methods).
Four inputs per player -Let us now move to the scenario where each of the two players may receive 4 inputs ( m ( k ) i = 4; m o = 2). This scenario is computationally too costly to classify all 4 16 different functions and make an exhaustive search for quantum violations. However, it is worthy to highlight a few particular Id games for which we find ω ∗ cl < ω ∗ q < ω ∗ ns :
- (i) Addition game. Let us define the Id game with the following function
$$2 y _ { 2 } + y _ { 1 } = x _ { 1 } + x _ { 2 } \bmod 4, \quad \quad ( 8 ) \quad \text{opt}.$$
where x , x 1 2 are assumed to take values in { 0 1 2 3 , , , } whereas y , y 1 2 ∈ { 0 1 , } . The /vector y = ( y , y 2 1 ) above encodes in two bits the result of adding two base-4 integers x 1 and x 2 in a modulo 4 arithmetic. For this game, we have the success probabilities, ω ∗ cl = 3 8 = 0 375, / . ω ∗ q = (2 + √ 2) / 8 ≈ 0 4268, and . ω ∗ ns = 1 2. / Hence, ω ∗ q and ω ∗ ns beat the classical limit ω ∗ cl by about 13 81% and . 33 33%, respectively. . Both values represent considerable improvement over the 3-input case discussed previously. The quantum maximum ω ∗ q ≈ 0 4268 is attained by us-. ing a maximally entangled two-qubit state and co-planar measurements which are presented in the Methods section.
- (ii) A facet-defining game. It turns out that there exists a 4-input Id game, which defines a facet of the Bell local polytope in the scenario of four binary inputs per party. This game is equivalent to the Bell inequality I 6 4422 defined by the paper of Brunner and Gisin [16]. See Methods section for an explicit construction of this game.
## Three-player games
Here we consider three-player games with binary inputs ( m ( k ) i = m o = 2). There are 8 8 = 16777216 different functions which reduce to 5876 equivalence classes. We found 68 equivalence classes (with 34176 functions together) for which ω ∗ cl < ω ∗ ns . According to classification of extremal non-signaling strategies [17], for most of equivalence classes, ω ∗ ns is achieved by decomposable strategies (i.e. strategies which may be decomposed into PR-box on 2 parties and local deterministic box on the remaining party). More details on the classification of winning strategies are presented in the Methods section.
Here we focus on one game where we can also provide an example of quantum strategy with advantage over classical one. The function f for this game may be written as:
$$y _ { 1 } \ & = \ ( \bar { x } _ { 1 } \wedge \bar { x } _ { 2 } ) \oplus \bar { x } _ { 3 } \quad \quad \ ( 9 ) \\ u _ { \infty } \ & = \ \bar { r } _ { \infty } \quad \quad \ \ ( 1 0 \);$$
$$y _ { 2 } \, = \, \bar { x } _ { 3 } \, & & ( 1 0 )$$
$$y _ { 3 } \ = \ 0.$$
In the optimal classical strategy all players return 0 all the time which leads to ω ∗ cl = 0 375. . Optimal nonsignaling strategy decomposes into PR-box shared between player 1 and 2 and deterministic strategy used by player 3 (he always returns 0). For that strategy we have ω ∗ ns = 0 5. .
Inspired by decomposability of optimal non-signaling strategy, we propose the following quantum strategy: player 1 and player 2 apply quantum strategy with maximal success probability for XOR game, i.e. the one optimizing y 2 ⊕ y 1 = ¯ x 2 ∧ ¯ . x 1 Player 3 uses deterministic strategy: he returns 0 all the time. Success probability achieved by quantum strategy for 2-player XOR game is cos 2 π 8 . It is easy to see from (9)-(11) that y 2 ⊕ y 1 = ¯ x 2 ∧ ¯ . x 1 Only one of two possible outputs winning XOR games is valid for the discussed function, and hence for this quantum strategy we get ω q = 1 2 cos 2 π 8 = 0 42677 . > ω ∗ cl . This value may be compared with the bound obtained from SDP which reads 0 42683 (1 + . AB hierarchy of [10, 11]).
## Generic advantage of Id games with multiple outcomes
We now argue, that for any number of players, for large enough number of outputs, the no-signaling theories beat the classical ones generically. Specifically, let us define as M cl ( ω ) ( M ns ( ω )) to be the number of functions for n parties, with m i = m o = m , for which the probability of successful implementation within classical (nonsignaling) theory is ω . We show that for any number of parties n the ratio M cl (2 1 -n ) M ns (2 1 -n ) goes to zero for increasing m . The proof is deferred to the Methods section.
## METHODS
## Two-player games
Id game using partially entangled states - We show an Id game which allows higher winning probability using partial entangled states than maximally entangled states of any dimension. The game is as follows:
$$\frac { x _ { 2 } \langle x _ { 1 } ( y _ { 2 }, y _ { 1 } ) } { 0 } \begin{vmatrix} 0 & 1 & 2 \\ 0, 1 & 1, 1 & 1, 0 \\ 1 \\ 2 \end{vmatrix} 0, 0 & 0, 1 & 1, 1 \\ 0, 1 & 1, 0 & 0, 1 \end{vmatrix}$$
with ω ∗ cl = 4 9 and / ω ∗ nl = 1 2. /
Using maximally entangled states (of any dimension), ω ∗ q + = 4 0178 . / 9. This value has been certified by using the SDP method introduced in Section 4.2 of [18]. However, the maximum using general quantum resources is ω ∗ q = 4 1224 . / 9, which saturates the SDP upper bound of [10]. In fact, ω ∗ q can be attained by using a partially entangled two-qubit state. Hence, the winning probability attainable with maximally entangled states (of any dimension) is strictly smaller than the one using nonmaximally entangled qubits.
Id game as a dimension witness -We present here an Id game which allows higher quantum violation if more than two-dimensional systems are considered. Hence, this Id game also gives an example to a dimension witness [14]: Maximum quantum violation does not happen in two dimensional systems. The players have to conduct measurements on at least three dimensional systems for maximum quantum violation to happen. The game is as follows:
$$\frac { x _ { 2 } \langle x _ { 1 } ( y _ { 2 }, y _ { 1 } ) } { 0 } \begin{vmatrix} 0 & 1 & 2 \\ 0, 1 & 1, 1 & 1, 0 \\ 1 \\ 2 \end{vmatrix} 0, 1 & 1, 1 & 1, 1 \\ 0, 1 & 1, 0 & 1, 0 \end{vmatrix}$$
with ω ∗ cl = 4 9 and / ω ∗ nl = 1 2. / Actually, no quantum violation can be observed for two dimensional systems, however, ω ∗ q = 4 1547005 . / 9 by using three dimensional systems. This value is certified by SDP hierarchy [10] as well.
Id game performing addition -The function defining the game can be written as
$$2 y _ { 2 } + y _ { 1 } = x _ { 1 } + x _ { 2 } \bmod 4, \quad \ \ ( 1 2 ) \quad ^ { \text{is $\epsilon$} } _ { \text{$f^{6}$} }$$
where x , x 1 2 take values in { 0 1 2 3 , , , } and y , y 1 2 ∈ { 0 1 , } , which is represented by the table
$$\begin{smallmatrix} x _ { 2 } \langle x _ { 1 } ( y _ { 2 }, y _ { 1 } ) & 0 & 1 & 2 & 3 \\ 0 & 0, 0 & 0, 1 & 1, 0 & 1, 1 \\ 1 & 0, 1 & 1, 0 & 1, 1 & 0, 0 \\ 2 & 1, 0 & 1, 1 & 0, 0 & 0, 1 \\ 3 & 1, 1 & 0, 0 & 0, 1 & 1, 0 \end{smallmatrix}$$
This can be further written as the following Bell functional:
$$I _ { \text{add} } = \frac { - I _ { 2 } ^ { ( 0, 1 ; 0, 1 ) } + I _ { 2 } ^ { ( 2, 3 ; 0, 1 ) } + I _ { 2 } ^ { ( 0, 1 ; 2, 3 ) } - I _ { 2 } ^ { ( 2, 3 ; 2, 3 ) } + 1 6 } { 6 4 },$$
/negationslash
A facet-defining Id game -The function to be considered is as follows:
with the CHSH game [9]: I ( i,j m,n ; ) 2 = -〈 A B i m 〉 + 〈 A B i n 〉 + 〈 A B j m 〉 + 〈 A B j n 〉 , where 〈 A B i m 〉 = p a ( = b i, m | ) -p a ( = b i, m | ), where a, b take values in { 0 1 , } . Due to Tsirelson [13], the quantum maximum for I 2 is 2 √ 2, from which an upper bound of (4(2 √ 2) + 16) / 64 = (2+ √ 2) / 2 ≈ 0 4268 for . I add easily follows. Interestingly, this upper bound can be obtained by an explicit quantum strategy. Just take the qubit observables A 0 = σ x , A 1 = σ z , B 0 = ( σ x -σ z ) / 2, and B 1 = ( - √ -σ x σ z ) / 2 along with the Bell state | φ + 〉 = ( 00 | 〉 + 11 ) | 〉 / 2 (where σ , σ x z refer to Pauli matrices). This provides 2 √ 2 for the quantity -I (0 1;0 1) , , 2 in Eq. (13). Then choose the rest of the observables as A 2 = -A 0 , A 3 = -A 1 , B 2 = -B 0 , and B 3 = -B 1 . These choices maximize the other three CHSH quantities as well saturating the upper bound ω ∗ q = (2 + √ 2) / 2 ≈ 0 4268 for . I add .
| x 2 \ x 1 ( y 2 , y 1 ) | 0 | 1 | 2 | 3 |
|---------------------------|-----|-----|-----|-----|
| 0 | 0,1 | 1,0 | 0,0 | 1,0 |
| 1 | 0,1 | 1,1 | 0,1 | 1,1 |
| 2 | 0,0 | 1,1 | 0,0 | 1,0 |
| 3 | 0,0 | 1,0 | 1,0 | 0,0 |
This table translates to the following Bell functional:
$$I _ { \text{facet} } = \frac { I _ { 4 4 2 2 } ^ { 6 } + 1 6 } { 6 4 }, \quad \quad \quad ( 1 4 )$$
where
$$\text{as} \quad & I _ { 4 4 2 2 } ^ { 6 } = & - I _ { 2 } ^ { ( 1, 0 ; 1, 0 ) } + I _ { 2 } ^ { ( 1, 0 ; 3, 2 ) } + I _ { 2 } ^ { ( 2, 3 ; 1, 0 ) } + I _ { 2 } ^ { ( 2, 3 ; 3, 2 ) } \\ & \quad - 2 \langle A _ { 3 } B _ { 0 } \rangle - 2 \langle A _ { 3 } B _ { 2 } \rangle + 2 \langle A _ { 2 } \rangle + 2 \langle A _ { 3 } \rangle \quad ( 1 5 )$$
is equivalent up to input/output relabellings with the I 6 4422 expression listed in the appendix of Ref. [16].
The maximum quantum violation is attained by using the Bell state | φ + 〉 and observables A i = /vector u i · /vector σ and B i = /vector v i · /vector σ , i = 0 1 2 3, where , , , /vector σ = ( σ , σ x y , σ z ) is the vector of Pauli matrices, and Alice and Bob's respective Bloch vectors /vector u i , /vector v i are given as follows:
$$\dots \dots - \dots \sigma \cdot \dots \dots \dots \\ \vec { u } _ { 0 } & = \left ( \sqrt { 1 - p ^ { 2 } }, 0, - p \right ) \\ \vec { u } _ { 1 } & = \left ( - \sqrt { 1 - p ^ { 2 } - \frac { 1 2 } { 1 5 } }, \sqrt { \frac { 1 2 } { 1 5 } }, - p \right ) \\ \vec { u } _ { 2 } & = \left ( 2 \sqrt { 1 - p ^ { 2 } - \frac { 1 2 } { 6 0 } }, \sqrt { \frac { 1 2 } { 6 0 } }, 2 p \right ) \\ \vec { u } _ { 3 } & = - \vec { u } _ { 2 },$$
and
$$\vec { v } _ { 0 } & = ( 0, 0, 1 ) & \frac { \omega _ { u } ^ { \kappa _ { u } } } { 0. 2 } \\ \vec { v } _ { 1 } & = \left ( - \frac { \sqrt { 5 } } { 3 }, 0, - \frac { 2 } { 3 } \right ) & \frac { 0. 2 } { 3 } \\ \vec { v } _ { 2 } & = \left ( \sqrt { \frac { 8 } { 9 } - q ^ { 2 } }, - q, \frac { 1 } { 3 } \right ) & \frac { 0. 3 } { 3 } \\ \vec { v } _ { 3 } & = \left ( \sqrt { \frac { 5 } { 9 } - q ^ { 2 } }, q, - \frac { 2 } { 3 } \right ), & \frac { 0. 5 } { ( 1 7 ) } \\ \end{pmatrix}, & \\ \, = \, 0. 4 0 8 2 4 8 \text{ and } q \, = \, 0. 7 3 0 2 9 7. \ \text{ With these set-} \quad \frac { \mathbb { I } } { \sim }$$
with p = 0 408248 and . q = 0 730297. . With these settings, we get ω ∗ q = 0 403093 for . I facet , which agrees with the upper bound on level 1 + AB of the SDP hierarchy. Note that the Bloch vectors (both for Alice and Bob) span the full three-dimensional space. Actually, the measurements attaining ω ∗ q cannot be brought to a co-planar form and consequently require the use of complex numbers.
Symmetric Id game with quantum advantage -We present here a symmetric Id game for five inputs per player which offers quantum advantage against classical strategies.
We found that the quantum maximum is ω ∗ q = 10 2950849 . / 25 (certified by SDP [10] as well) by performing measurements on a 2-qubit singlet state. The classical bound, on the other hand, is ω ∗ cl = 10 25. /
## Three-player games
Here we provide more detailed description of the numerical results obtained for 3-player Id games. We provide some statistics for ω ∗ cl and ω ∗ ns and classification of optimal non-signaling strategies according to [17] for these classes where ω ∗ ns > ω ∗ cl .
Number of equivalence classes for which given ω ∗ cl is obtained:
$$\frac { \omega _ { c l } ^ { * } } { 0. 2 5 } \left | \# \, \\ 0. 3 7 5 \right | ^ { 4 5 } _ { 2 3 }$$
Number of equivalence classes for which given ω ∗ ns is obtained:
The difference between ω ∗ ns and ω ∗ cl with the number of equivalence classes where it occurs:
| ∆ a = ω ∗ ns - | ω ∗ cl # |
|------------------|------------|
| 0.025 | 1 |
| 0.03125 | 1 |
| 0.041667 | 11 |
| 0.05 | 1 |
| 0.0625 | 51 |
| 0.083333 | 1 |
| 0.125 | 2 |
Similar list for relative differences between ω ∗ ns and ω ∗ cl :
| ∆ r = ω ∗ ns /ω ∗ cl - 1 | # |
|----------------------------|-----|
| 0.1 | 1 |
| 0.125 | 1 |
| 0.166667 | 21 |
| 0.166668 | 11 |
| 0.2 | 1 |
| 0.25 | 30 |
| 0.333332 | 1 |
| 0.333333 | 2 |
We do not present statistics for SDP bound of ω ∗ q since numerical inaccuracy make it hard to group them into classes.
Since non-signaling strategies which achieve ω ∗ ns for given f are extremal points of non-signaling polytope, we may apply classification from [17] to them. We found that optimal strategies belong to the following classes:
| class | # |
|---------|-----|
| 2 | 53 |
| 19 | 1 |
| 25 | 6 |
| 29 | 6 |
| 31 | 1 |
| 33 | 1 |
There may be also strategies from the other classes since we just checked to which class belongs the strategy obtained from linear programming. Strategies from class 2 may be decomposed into PR-box on 2 parties and local deterministic box on the remaining party. Strategies from classes 25 and 29 are optimal for GYNI game.
Decomposable strategies were discussed in the main part of the paper. Here we give an example of optimal strategy which belongs to class 25.
The function of the game is given by equations:
$$y _ { 1 } \, = \, \bar { x } _ { 3 } \, & & ( 1 8 ) \quad \text{cal}$$
$$y _ { 2 } \, = \, \bar { x } _ { 3 } & & ( 1 9 ) \quad \text{do} \\ y _ { 2 } \, = \, \bar { x } _ { 3 }$$
$$y _ { 3 } \, = \, \left ( \bar { x } _ { 3 } \wedge \bar { x } _ { 1 } \right ) \vee \left ( x _ { 3 } \wedge \bar { x } _ { 2 } \right ) \, \quad \ \ ( 2 0 ) \, \begin{array} { c } \text{Mo} \\ M _ { c } \end{array}$$
Classical and non-signaling strategies optimal for that game achieve ω ∗ cl = 1 4, / ω ∗ ns = 1 3 respectively which / give the gap ∆ a = 1 12, ∆ / r = 1 3. SDP bound for / ω ∗ q is 0 260746 .
In optimal classical strategy once again all players return 0 all the time while optimal non-signaling strategy has the form (we bolded the entries which win the game):
Generic advantage of Id games for no-signalling theories
| x 3 x 2 x 1 \ y 3 y 2 y 1 | 000 | 001 | 010 | 011 | 100 | 101 | 110 | 111 |
|-----------------------------|-------|-------|-------|-------|-------|-------|-------|-------|
| 000 | 1/3 | 0 | 1/3 | 0 | 0 | 0 | 0 | 1/3 |
| 001 | 1/3 | 0 | 0 | 1/3 | 0 | 0 | 1/3 | 0 |
| 010 | 1/3 | 0 | 1/3 | 0 | 0 | 0 | 0 | 1/3 |
| 011 | 1/3 | 0 | 0 | 1/3 | 0 | 0 | 1/3 | 0 |
| 100 | 0 | 0 | 1/3 | 1/3 | 1/3 | 0 | 0 | 0 |
| 101 | 0 | 0 | 1/3 | 1/3 | 1/3 | 0 | 0 | 0 |
| 110 | 1/3 | 0 | 0 | 1/3 | 0 | 0 | 1/3 | 0 |
| 111 | 1/3 | 0 | 0 | 1/3 | 0 | 0 | 1/3 | 0 |
We will now argue, that for any number of players, for large enough number of outputs, the no-signaling theories beat the classical ones generically. Specifically, let us define as M cl ( ω ) ( M ns ( ω )) to be the number of functions for n parties, with m i = m o = m , for which probability of successful implementation within classical (nonsignaling) theory is ω . We will show:
For any number of parties n the ratio M cl (2 1 -n ) M ns (2 1 -n ) goes to zero for increasing m .
To prove the claim, let us first note that, for any function f , non-signalling theories allow ω ns to be at least 2 1 -n (so that M ns is actually the number of all functions). This is achieved by a box defined as follows. Let ( f , ..., f 1 n ) = f ( x , ..., x 1 n ). Define the probability distribution of the box to be
/negationslash
$$P ( y _ { 1 }, \dots, y _ { n } | x _ { 1 }, \dots, x _ { n } ) = \begin{cases} \begin{array} { c c } 2 ^ { 1 - n } & \oplus _ { k = 1 } ^ { n } y _ { k } = \oplus _ { k = 1 } ^ { n } f _ { k } \\ 0 & \oplus _ { k = 1 } ^ { n } y _ { k } \neq \oplus _ { k = 1 } ^ { n } f _ { k } \end{array} & \stackrel { \dashrightarrow \cdots } { \text{We } \varepsilon } \\ \approx 2 \, \text{tin} \end{cases}$$
It is straightforward to check that this box is nonsignalling and the probability that /vector y = f ( x , ..., x 1 n ) is exactly 2 1 -n for any input. =
Let us now turn to classical theory. Obviously for any function we can have ω cl ≥ 2 -n . It is realized by a strategy in which all parties return random outputs regardless of their inputs. There are M = 2 nm n possible functions f . This number comes from the fact that for each function we need to specify the outcome for each of n players for any combination of m n possible inputs. We want to show that for a vast majority of functions f the classical resources (shared randomness and local computation) do not allow to reach a significantly higher probability. More precisely, our aim is to find (an upper bound on) M ω c ( cl ) - the number of functions with the average success probability of the best classical strategy larger or equal to particular ω cl .
One nice property of the classical average success probability is that there exists optimal strategy which is deterministic. (cf. [19]) - namely each party applies function y k to his input x k , k = 1 , . . . , n . Therefore, we can limit ourselves to considering only such strategies.
To enumerate all functions with success probability ω cl we need log M ω c ( cl ) bits. One of the possible ways of enumeration is first to describe the optimal strategies. To do so one needs to describe all the functions y k ( x k ). There are 2 m functions for every k , which gives 2 mn sets of n different functions. Then one can calculate the outputs of this strategy B = ( y 1 ( x 1 ) , ..., y n ( x n )). To fully characterize f one can simply take the description of the classical strategy, which requires nm bits, and append to it a list of numbers /vector b /vector ( x ) = f /vector ( x ) /circleminus B /vector ( x ). Where /circleminus is bitwise subtraction modulo m . However, each of numbers /vector b /vector ( x ) is equal to 0 whenever the classical strategy succeeds, which happens with probability ω cl . There are m n numbers /vector b which need to be specified and each of them has 2 n -letter alphabet, but because 0 will appear with probability ω cl the entropy of the whole set of /vector b 's is at most h ∗ ( ω cl ) m n , where h ∗ ( ω cl ) = h ω ( cl ) -(1 -ω cl ) log(2 n -1). h is Shannon's binary entropy and h ∗ is the highest entropy a variable with 2 n -letter alphabet can have if one of the letters has probability ω cl . It can be considered a generalization of h . h ∗ ( ω cl ) is equal to n only if ω cl = 2 -n and is strictly smaller otherwise.
This encoding uses
$$M ^ { \prime } = m n + h ^ { * } ( \omega _ { c l } ) m ^ { n } \text{ \quad \ \ } ( 2 2 )$$
bits. Since any encoding must require at least log M ω c ( cl ), then M ′ ≥ log M ω c ( ∗ cl ), which gives an upper bound on M ω c ( ∗ cl ). An important implication is that for any ω cl > 2 -n the ratio M ω c ( cl ) M goes to 0 as m goes to infinity. This means that as the size of the input grows the fraction of functions where the classical resources provide a non-negligible advantage over producing random outputs goes to zero which proves the claim.
We actually proved more: for a majority of the functions non-signalling resources can achieve a success rate ≈ 2 times higher than the classical ones.
## DISCUSSION
In this work, we addressed the following question: is quantum mechanics superior to classical theory regarding distributed computation of total function, if no communication between nodes is allowed? In order to answer this question we introduced Id games. The motivation for the studies of these games comes from the fact that they are instances of a complete nonlocal computation. In the case of any type of distributed computation any exclusions of some possible combinations of inputs for the parties are difficult to justify. Therefore, the functions g that we consider are total functions. Moreover, function g is defined such that only one output value /vector y is a valid answer for a given input /vector x . This research has also been partially inspired by the results of [5] where the authors
- [1] P. W. Shor, Polynomial-Time Algorithms for Prime Factorization and Discrete Logarithms on a Quantum Computer , SIAM J. Sci. Statist. Comput. 26 1484 (1997). Preprint at http://arXiv.org/abs/quant-ph/9508027 (1995).
- [2] H. Buhrman, R. Cleve, S. Massar, R. de Wolf, Nonlocality and Communication Complexity , Rev. Mod. Phys. 82 665 (2010). Preprint at http://arXiv.org/abs/0907.3584 (2009).
- [3] N. Linden, S. Popescu, A. J. Short, A. Winter, Quantum Nonlocality and Beyond: Limits from Nonlocal Computation , Phys. Rev. Lett. 99 , 180502 (2007). Preprint at http://arXiv.org/abs/quant-ph/0610097 (2006).
- [4] N. Brunner, D. Cavalcanti, S. Pironio, V. Scarani, S. Wehner, Bell nonlocality , Rev. Mod. Phys. 86 , 419 (2014).
- [5] M. Almeida, J.-D. Bancal, N. Brunner, A. Acin, N. Gisin, and S. Pironio, Guess Your Neighbors Input: A Multipartite Nonlocal Game with No Quantum Advantage , Phys. Rev. Lett. 104 , 230404 (2010).
- [6] I. Kerenidis, S. Laplante, V. Lerays, J. Roland, D. Xiao, Lower bounds on information complexity via zerocommunication protocols and applications , In 53rd Annual IEEE Symposium on Foundations of Computer Science (FOCS'12), pages 500-509 (2012).
- [7] D. Collins, N. Gisin, A relevant two qubit Bell inequality inequivalent to the CHSH inequality , J. Phys. A: Math. Gen. 37 , 1775-1787 (2004).
- [8] R. Cleve, P. Hoyer, B. Toner, and J. Watrous, Consequences and limits of nonlocal strategies , Proc. 19th IEEE Conf. on Computational Complexity, 236-249 (2004).
- [9] J. F. Clauser , M. A. Horne, A. Shimony, R. A. Holt, Proposed Experiment to Test Local Hidden-Variable Theories , Phys. Rev. Lett. 23 , 880 (1969).
- [10] M. Navascues, S. Pironio, A. Acin, Bounding the Set of Quantum Correlations , Phys. Rev. Lett. 98 , 010401
study a particular family of inequalities that is similar to Id games but defined for partial functions (i.e. there is a promise on the players' inputs). The results from [5] have been very useful in the studies of the foundations of the quantum theory [20, 21] and we believe that Id games will too.
## Acknowledgements
This work was supported by the ERC AdG grant QOLAPS, EC grant RAQUEL (323970) and National Science Centre project Maestro DEC2011/02/A/ST2/00305. M.P was supported by NCN grant 2013/08/M/ST2/00626 and FNP TEAM programme.
(2007).
- [11] M. Navascues, S. Pironio, A. Acin, A convergent hierarchy of semidefinite programs characterizing the set of quantum correlations , New J. Phys. 10 , 073013 (2008).
- [12] S. Popescu and D. Rohrlich, Quantum nonlocality as an axiom , Found. Phys. 24 , 379-385 (1994).
- [13] B. S. Tsirelson, Quantum generalizations of Bell's inequality , Lett. Math. Phys., 4 93-100, (1980).
- [14] N. Brunner, S. Pironio, A. Acin, N. Gisin, A. A. M´thot, e V. Scarani, Testing the Dimension of Hilbert Spaces , Phys. Rev. Lett. 100 , 210503 (2008).
- [15] M. Navascu´ es, G. de la Torre, T. V´rtesi, e Characterization of Quantum Correlations with Local Dimension Constraints and Its Device-Independent Applications , Phys. Rev. X 4 , 011011 (2014).
- [16] N. Brunner, N. Gisin, Partial list of bipartite Bell inequalities with four binary settings , Phys. Lett. A 372 , 3162 (2008).
- [17] S. Pironio, J.-D. Bancal, V. Scarani, Extremal correlations of the tripartite no-signaling polytope , J. Phys. A 44 , 065303 (2011).
- [18] B. Lang, T. V´ ertesi, M. Navascu´ es, Closed sets of correlations: answers from the zoo , preprint at http://arXiv.org/abs/1402.2850 (2014).
- [19] A. Fine, Hidden Variables, Joint Probability, and the Bell Inequalities , Phys. Rev. Lett. 48 , 291 (1982).
- [20] R. Gallego, L. E. Wurflinger, A. Acin, M. Navascues, Quantum Correlations Require Multipartite Information Principles , Phys. Rev. Lett. 107 , 210403 (2011).
- [21] T. Fritz, A. B. Sainz, R. Augusiak, J. B. Brask, R. Chaves, A. Leverrier, A. Ac´ ın, Local orthogonality as a multipartite principle for quantum correlations , Nat. Commun. 4 , 2263 (2013).
- [22] In some instances g may depend also on /vector x or the average weighted [7].
|
10.1103/PhysRevA.92.032122
|
[
"L. Czekaj",
"M. Pawlowski",
"T. Vertesi",
"A. Grudka",
"M. Horodecki",
"R. Horodecki"
] |
2014-08-05T14:55:30+00:00
|
2014-10-02T17:19:58+00:00
|
[
"quant-ph"
] |
Quantum advantage for distributed computing without communication
|
Understanding the role that quantum entanglement plays as a resource in
various information processing tasks is one of the crucial goals of quantum
information theory. Here we propose a new perspective for studying quantum
entanglement: distributed computation of functions without communication
between nodes. To formalize this approach, we propose identity games.
Surprisingly, despite of no-signaling, we obtain that non-local quantum
strategies beat classical ones in terms of winning probability for identity
games originating from certain bipartite and multipartite functions. Moreover
we show that, for majority of functions, access to general non-signaling
resources boosts success probability two times in comparison to classical ones,
for number of outputs large enough.
|
1408.0994v2
|
## Generalizing Block LU Factorization: A Lower-Upper-Lower Block Triangular Decomposition with Minimal Off-Diagonal Ranks
François Serre and Markus Püschel
Department of Computer Science, ETH Zurich
## Abstract
We propose a novel factorization of a non-singular matrix P , viewed as a 2 × 2 -blocked matrix. The factorization decomposes P into a product of three matrices that are lower block-unitriangular, upper block-triangular, and lower block-unitriangular, respectively. Our goal is to make this factorization 'as block-diagonal as possible' by minimizing the ranks of the off-diagonal blocks. We give lower bounds on these ranks and show that they are sharp by providing an algorithm that computes an optimal solution. The proposed decomposition can be viewed as a generalization of the well-known Block LU factorization using the Schur complement. Finally, we briefly explain one application of this factorization: the design of optimal circuits for a certain class of streaming permutations.
## 1 Introduction
Given is a non-singular matrix P ∈ GL m n + ( K ) over a field K . We partition P as
$$P = \begin{pmatrix} P _ { 1 } & P _ { 2 } \\ P _ { 3 } & P _ { 4 } \end{pmatrix}, \ \text{ such that } P _ { 1 } \text{ is } m \times m.$$
We denote the ranks of the submatrices with p i = rk P i , i = 1 2 3 4 , , , . Matrices are denoted with capital letters and vector spaces with calligraphic letters.
If P 1 is non-singular, then a block Gaussian elimination uniquely decomposes P into the form:
$$P = \begin{pmatrix} I _ { m } & \\ L & I _ { n } \end{pmatrix} \begin{pmatrix} C _ { 1 } & C _ { 2 } \\ & C _ { 4 } \end{pmatrix},$$
where I m denotes the m × m identity matrix. The rank of L = P P 3 -1 1 is equal to p 3 , and C 4 is the Schur complement of P 1 . Conversely, if such a decomposition exists for P , then P 1 is non-singular. This block LU decomposition has several applications including computing the inverse of P [2], solving linear systems [4], and in the theory of displacement structure [6]. The Schur complement is also used in statistics, probability and numerical analysis [11, 3].
Analogously, the following decomposition exists if and only if P 4 is non-singular:
$$P = \begin{pmatrix} C _ { 1 } & C _ { 2 } \\ & C _ { 4 } \end{pmatrix} \begin{pmatrix} I _ { m } \\ R & I _ { n } \end{pmatrix}.$$
This decomposition is again unique, and the rank of R is p 3 .
In this article, we release the restrictions on P 4 and P 1 and propose the following decomposition for a general P ∈ GL m n + ( K ) :
$$P = \begin{pmatrix} I _ { m } & \\ L & I _ { n } \end{pmatrix} \begin{pmatrix} C _ { 1 } & C _ { 2 } \\ & C _ { 4 } \end{pmatrix} \begin{pmatrix} I _ { m } & \\ R & I _ { n } \end{pmatrix},$$
where in addition we want the three factors to be 'as block-diagonal as possible,' i.e., that rk L +rk C 2 + rk R is minimal.
## 1.1 Lower bounds
The following theorem provides bounds on the ranks of such a decomposition:
Theorem 1. If a decomposition (3) exists for P ∈ GL m n + ( K ) , it satisfies
$$\text{rk} \, C _ { 2 } = p _ { 2 },$$
$$\text{rk} \, L & \geq n - p _ { 4 },$$
$$\text{rk} \, R \geq m - p _ { 1 },$$
$$\text{rk} \, R + \text{rk} \, L \geq p _ { 3 }.$$
In particular, the rank of C 2 is fixed and we have:
$$\ r k \, R + \ r k \, L \geq \max ( p _ { 3 }, n + m - p _ { 4 } - p _ { 1 } ).$$
We will prove this theorem in Section 4. Next, we assert that these bounds are sharp.
## 1.2 Optimal solution
The following theorem shows that the inequality (8) is sharp:
Theorem 2. If P ∈ GL m n + ( K ) , then there exists a decomposition (3) that satisfies
$$\ r k \, R + \ r k \, L = \max _ { \substack { \ r k \, L = n - p _ { 4 }. \\ \ r k \, L = n - p _ { 4 }. } } \, \ a n d$$
Additionally, such a decomposition can be computed with O (( m + n ) 3 ) arithmetic operations.
We prove this theorem in Section 5 when p 3 ≤ m + n -p 4 -p 1 , and in Section 6 for the case p 3 > m + n -p 4 -p 1 . In both cases, the proof is constructive and we provide a corresponding algorithm (Algorithms 3 and 4). Theorem 2 and the corresponding algorithms are the main contributions of this article.
Two cases. As illustrated in Figure 1, two different cases appear from inequality (8). If p 3 ≤ m + n -p 4 -p 1 , bound (7) is not restrictive, and the optimal pair (rk L, rk R ) is unique, and equals ( n -p , m 4 -p 1 ) . In the other case, where p 3 > m + n -p 4 -p 1 , bound (7) becomes restrictive, and several optimal pairs (rk L, rk R ) exist.
Figure 1: Possible ranks for L and R . On the left graph, p 3 < m + n -p 4 -p 1 . On the right graph, p 3 > m + n -p 4 -p 1 . The dot shows the decomposition provided by Theorem 2.
Example. As a simple example we consider the special case
$$P = \begin{pmatrix} & P _ { 2 } \\ P _ { 3 } \end{pmatrix}, \quad \text{with } n = m.$$
In this case P , 3 P 2 are non-singular and neither (1) nor (2) exists. Theorem 1 gives a lower bound of rk R +rk L ≥ 2 n , which implies that both R and L have full rank. Straightforward computation shows that for any non-singular L ,
$$P = \begin{pmatrix} I _ { n } & \\ L & I _ { n } \end{pmatrix} \begin{pmatrix} L ^ { - 1 } P _ { 3 } & P _ { 2 } \\ & - L P _ { 2 } \end{pmatrix} \begin{pmatrix} I _ { n } & \\ - ( L P _ { 2 } ) ^ { - 1 } P _ { 3 } & I _ { n } \end{pmatrix}$$
is an optimal solution. This also shows that the optimal decomposition (3) is in general not unique.
## 1.3 Flexibility
The following theorem adds flexibility to Theorem 2 and shows that a decomposition exists for any Pareto-optimal pair of non-diagonal ranks that satisfies the bounds of Theorem 1:
Theorem 3. If P ∈ GL m n + ( K ) and ( l, r ) ∈ N 2 satisfies l ≥ n -p , 4 r ≥ m -p 1 , and
$$r + l = \max ( p _ { 3 }, n + m - p _ { 4 } - p _ { 1 } ),$$
then P has a decomposition (3) with rk L = l and rk R = r .
In the case where p 3 ≤ m + n -p 4 -p 1 , the decomposition produced by Theorem 2 has already the unique optimal pair (rk L, rk R ) = ( n -p , m 4 -p 1 ) . In the other case, we will provide a method in Section 7 to trade between the rank of R and the rank of L , until bound (6) is reached. By iterating this method over the decomposition obtained in Theorem 2, decompositions with various rank tradeoffs can be built.
Therefore, it is possible to build decomposition (3) for any pair (rk L, rk R ) that is a Pareto optimum of the given set of bounds. As a consequence, if f : N 2 → R is weakly increasing in both of its arguments, it is possible to find a decomposition that minimizes f (rk L, rk R ) . Examples include min(rk L, rk R ) , max(rk L, rk R ) , rk L +rk R , rk L · rk R or √ rk 2 L +rk 2 R .
Generalization of block LU factorization. In the case where P 4 is non-singular, Theorem 2 provides a decomposition that satisfies rk L = 0 . In other words, it reduces to the decomposition (2). Using Theorem 3, we can obtain a similar result in the case where P 1 is non-singular. Since in this case m -p 1 = 0 , it is possible to choose r = 0 , and thus obtain the decomposition (1).
## 1.4 Equivalent formulations
Lemma 1. The following decomposition is equivalent to decomposition (3) , with analogous constraints for the non-diagonal ranks:
$$\ddot { \cdot } \cdot \dots \dots \dots \\ P = \begin{pmatrix} I _ { m } & \\ L _ { 3 } & L _ { 4 } \end{pmatrix} \begin{pmatrix} C _ { 1 } & C _ { 2 } \\ & I _ { n } \end{pmatrix} \begin{pmatrix} I _ { m } & \\ R _ { 3 } & R _ { 4 } \end{pmatrix}$$
In this case, the minimization of the non-diagonal ranks is exactly the same problem as in (3) . However, an additional degree of freedom appears: any non-singular n × n matrix can be chosen for either L 4 or R 4 .
It is also possible to decompose P into two matrices, one with a non-singular leading principal submatrix L 1 and the other one with a non-singular lower principal submatrix R 4 :
$$P = \begin{pmatrix} L _ { 1 } & L _ { 2 } \\ L _ { 3 } & L _ { 4 } \end{pmatrix} \begin{pmatrix} R _ { 1 } & R _ { 2 } \\ R _ { 3 } & R _ { 4 } \end{pmatrix}$$
Once again, the minimization of rk L 3 +rk R 3 is the same problem as in (3) . The two other non-diagonal blocks satisfy rk L 2 +rk R 2 ≥ p 2 .
Proof. The lower non-diagonal ranks are invariant through the following steps:
(3) ⇒ (9) . If P has a decomposition (3), a straightforward computation shows that:
$$P = \begin{pmatrix} I _ { m } & \\ L & C _ { 4 } \end{pmatrix} \begin{pmatrix} C _ { 1 } & C _ { 2 } \\ & I _ { n } \end{pmatrix} \begin{pmatrix} I _ { m } & \\ R & I _ { n } \end{pmatrix},$$
which has the form of decomposition (9).
(9) ⇒ (10) . If P has a decomposition (9), the multiplication of the two left factors leads to formulation (10). In fact, L 1 = C 1 and R 4 are both non-singular.
(10) ⇒ (3) . If P has a decomposition (10), then using (1) on the left factor, and (2) on the right factor, and multiplying the two central matrices leads to formulation (3).
## 1.5 Related work
Schur complement. Several efforts have been made to adapt the definition of Schur complement in the case of general P 1 and P 4 . For instance, it is possible to define an indexed Schur complement of another non-singular principal submatrix [11], or use pseudo-inverses [1] for matrix inversion algorithms.
Alternative block decompositions. A common way to handle the case where P 1 is singular is to use a permutation matrix B that reorders the columns of P such that the new principal upper submatrix is non-singular [11]. Decomposition (1) then becomes:
$$P = P B B ^ { T } = \begin{pmatrix} I _ { m } & \\ L & I _ { n } \end{pmatrix} \begin{pmatrix} C _ { 1 } & C _ { 2 } \\ & C _ { 4 } \end{pmatrix} B ^ { T }$$
However, B needs to swap columns with index ≥ m ; thus B T does not have the required form considered in our work.
One can modify the above idea to choose B such that B -1 has the shape required by decomposition (3):
$$P \ & = \ \begin{pmatrix} P _ { 1 } & P _ { 2 } \\ P _ { 3 } & P _ { 4 } \end{pmatrix} \begin{pmatrix} I _ { m } \\ - R & I _ { n } \end{pmatrix} \begin{pmatrix} I _ { m } \\ R & I _ { n } \end{pmatrix} \\ & = \ \begin{pmatrix} P _ { 1 } - R P _ { 2 } & P _ { 2 } \\ P _ { 3 } - R P _ { 4 } & P _ { 4 } \end{pmatrix} \begin{pmatrix} I _ { m } \\ R & I _ { n } \end{pmatrix} \\. \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \end{pmatrix}$$
Then the problem is to design R such that P 1 -RP 2 is non-singular and rk( P 3 -RP 4 ) + rk R is minimal. This basic idea is used in [9], where, however, only rk R is minimized, which, in general, does not produce optimal solutions for the problem considered here.
Finally, our decomposition also shares patterns with a block Cholesky decomposition, or the Block LDL decomposition, in the sense that they involve block uni-triangular matrices. However, the requirements on P and the expectations on the decomposition are different.
## 2 Application: Optimal Circuits for Streaming Permutations
The original motivation for considering our decomposition (3) was an important application in the domain of hardware design. Many algorithms in signal processing and communication consist of alternating computation and reordering (permutation) stages. Typical examples include fast Fourier transforms [10]; one example (a so-called Pease FFT) is shown in Fig. 2 for 16 data points. When mapped to hardware, permutations could become simple wires. However, usually, the data is not available in one cycle, but streamed in chunks over several cycles. Implementing such a 'streaming' permutation in this scenario on an application-specific integrated circuit (ASIC) or on a field programmable gate array (FPGA) becomes complex, as it now requires both logic and memory [7, 9]. It turns out that for an important subclass of permutations called 'linear,' the design of an optimal circuit (i.e., one with minimal logic) is equivalent to solving (3) in the field F 2 .
Next, we provide a few more details on this application starting with the necessary background information. However, we only provide a sketch; a more complete treatment can be found in [9].
## 2.1 Linear permutations
For 0 ≤ i < 2 m n + , we denote with i b the associated (bit) vector in F m n + 2 that contains the radix2 digits of i , with the most significant digit on top. For instance, for m + n = 4 , we have
$$\mu _ { \ } ^ { \ } ^ { \ } ^ { \ } ^ { \ } ^ { \ } _ { b } = \begin{pmatrix} 1 \\ 1 \\ 0 \\ 0 \end{pmatrix} \text{ and } 7 _ { b } = \begin{pmatrix} 0 \\ 1 \\ 1 \\ 1 \end{pmatrix}.$$
Figure 2: Dataflow of a Pease FFT on 16 points from right (input) to left (output). Computing stages, consisting of DFTs on two points and pointwise multiplications (not shown), alternate with permutations.
Figure 3: Sketch of an implementation of the bit reversal permutation on 2 3 elements streamed on two ports. The dataset enters within 4 cycles (right), and is retrieved within 4 cycles (left).
Any invertible ( m + n ) × ( m + n ) matrix P over F 2 induces a permutation π P ( ) on { 0 , . . . , 2 m n + - } 1 that maps i to j , where j b = P · i b .
For example, if we define
$$V _ { m + n } = \begin{pmatrix} 1 & & \\ \vdots & \ddots & \\ 1 & & 1 \end{pmatrix},$$
↦
then π V ( 3 ) is the mapping 0 → 0 , 1 → 1 2 , → 2 3 , → 3 4 , → 7 → 4 , 5 → 6 → 5 . More generally, π V ( m n + ) is the permutation that leaves the first half of the elements unchanged, and that reverts the second half.
↦
↦
↦
↦
↦
↦
↦
The mapping π : GL m n + ( F 2 ) → S 2 m n + is a group-homomorphism, and its range is called the group of linear permutations [8, 5]. This group contains many of the permutations used in signal processing and communication algorithms, including stride permutations, bit-reversal, Hadamard reordering, and the Grey code reordering.
## 2.2 Streamed linear permutations (SLP)
We want to implement a circuit that performs a linear permutation on 2 m n + points. If we assume that this circuit has 2 n input (and output) ports, this means that the dataset has to be split into 2 m parts that are fed (streamed) as input over 2 m cycles. Similarly, the permuted output will be produced streamed over 2 m cycles. As an example consider Fig. 3 in which 2 n = 2 and 2 m = 4 .
With this convention, the element with the index i = c · 2 m + p arrives as input in the c th cycle at the p th port. Particularly, the upper m bits of i b are the bit representation c b of the input cycle c , while the lower n bits are p b . For example, if m = 3 and n = 2 , then the element indexed with
$$1 4 _ { b } = \begin{pmatrix} 0 \\ 1 \\ 1 \\ 0 \end{pmatrix} = \begin{pmatrix} 3 _ { b } \\ 2 _ { b } \end{pmatrix}$$
Figure 4: Two hardware blocks to implement particular types of SLPs. On the left, an array of 4 RAM banks. Write and read addresses are used to delay differently elements. This block can be used to implement SLPs that do not permute across ports. On the right is an example of a switching network consisting of 4 2 × 2 -switches. Switching networks can be used to implement SLPs that do not permute across time.
will arrive during the third cycle on the second port.
Thus, it is natural to block the desired linear permutation
$$P = \begin{pmatrix} P _ { 1 } & P _ { 2 } \\ P _ { 3 } & P _ { 4 } \end{pmatrix}$$
such that P 1 is m × m . This implies that the element that arrives on port p during the c th cycle has to be routed to the output port P c 3 b + P p 4 b at output cycle P c 1 b + P p 2 b .
P 3 has a particular role here, as it represents 'how the routing between the different ports must vary during time', and directly influences the complexity of the implementation. For example, if P 3 = 0 , then the output is always P p 4 b without variation during time.
## 2.3 Implementing SLPs on hardware
Two special cases of SLPs can be directly translated to a hardware implementation. The first kind are the permutations that only permute across time, i.e., that do not change the port number of the elements. Thus, they satisfy P c 3 b + P p 4 b = p b for all c and p , and therefore have the form
$$P = \begin{pmatrix} P _ { 1 } & P _ { 2 } \\ & I _ { n } \end{pmatrix}.$$
These SLPs can be implemented using an array of 2 n blocks of RAMs as shown in Fig. 4(a). Conversely, SLPs that only permute across the ports within each cycle have the form
$$P = \begin{pmatrix} I _ { m } & \\ P _ { 3 } & P _ { 4 } \end{pmatrix}.$$
They do not require memory and can be implemented using a network of p 3 · 2 n -1 2 × 2 -switches as was shown in [9]. This result, combined with decomposition (9) and Theorem 2 yields an immediate corollary:
Theorem 4. For a given P , the associated SLP can be implemented using a RAM bank array (Fig. 4(a)) framed by two switching networks (Fig. 4(b)) with a total of max( p , n 3 + m -p 4 -p 1 ) · 2 n -1 2 × 2 -switches.
Our Algorithms 3 and 4, introduced later, provide an efficient method to compute this architecture.
## 3 Preliminaries
In this section, we will prove some basic lemmas that we will use throughout this article.
## 3.1 Properties of the blocks of an invertible matrix
In this subsection, we derive some direct consequences of the invertibility of P on the range and the nullspace of its submatrices.
Lemma 2. The following properties are immediate from the structure of P :
$$\ker P _ { 4 } \cap \ker P _ { 2 } = \{ 0 \}$$
$$\ker P _ { 3 } \cap \ker P _ { 1 } = \{ 0 \}$$
$$\lim P _ { 4 } + \lim P _ { 3 } = \mathbb { K } ^ { n }$$
$$\lim P _ { 2 } + \lim P _ { 1 } = \mathbb { K } ^ { m }$$
$$P _ { 3 } ( \ker P _ { 1 } ) \cap P _ { 4 } ( \ker P _ { 2 } ) = \{ 0 \}$$
$$P _ { 1 } ( \ker P _ { 3 } ) \cap P _ { 2 } ( \ker P _ { 4 } ) = \{ 0 \}$$
Proof. We prove here equation (15). If x ∈ ker P 1 and y ∈ ker P 2 satisfy P x 3 = P y 4 , we have:
$$\begin{pmatrix} P _ { 1 } & P _ { 2 } \\ P _ { 3 } & P _ { 4 } \end{pmatrix} \cdot \begin{pmatrix} x \\ - y \end{pmatrix} = \begin{pmatrix} P _ { 1 } x - P _ { 2 } y \\ P _ { 3 } x - P _ { 4 } y \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix}$$
Since P is non-singular, x = y = 0 , as desired.
These equalities yield the dimensions of the following subspaces:
## Corollary 1.
$$\dim P _ { 2 } ( \ker P _ { 4 } ) = \dim \ker P _ { 4 } = n - p _ { 4 }$$
$$\dim P _ { 1 } ( \ker P _ { 3 } ) = \dim \ker P _ { 3 } = m - p _ { 3 }$$
$$\dim P _ { 4 } ( \ker P _ { 2 } ) = \dim \ker P _ { 2 } = n - p _ { 2 }$$
$$\dim P _ { 3 } ( \ker P _ { 1 } ) = \dim \ker P _ { 1 } = m - p _ { 1 }$$
$$\dim \lim P _ { 4 } \cap \lim P _ { 3 } = p _ { 4 } + p _ { 3 } - n$$
$$\dim \lim P _ { 2 } \cap \lim P _ { 1 } = p _ { 2 } + p _ { 1 } - m$$
Proof. We prove equation (17). Since, by (11), ker P 4 ∩ ker P 2 = { 0 } , the dimension of the image of ker P 4 under P 2 has the same dimension as ker P 4 , which is n -p 4 .
Equation (21) is a consequence of equation (13) and of the rank-nullity theorem.
## 3.2 Algorithms on linear subspaces in matrix form
The algorithms we present in this article heavily rely on operations on subspaces of K m . To make the representation of these algorithms more practical for implementation, we introduce a matrix representation for subspaces and formulate the subspace operations needed in this paper on this representation.
We represent a linear subspace as an associated matrix 1 whose columns form a basis of this subspace. In other words, if A is a subspace of K m of dimension n , then we represent it using a m × n matrix A such that im A = A . In this case, and only in this case, we will use the notation 〈 A 〉 = im A to emphasize that the columns of A form a linear independent set.
With this notation we can formulate subspace computations as computations on their associated matrices as explained in the following. To formally emphasize this correspondence, these operations on matrices will carry the same symbol as the subspace computation (e.g., ∩ ) they represent augmented with an overline (e.g., ∩ ). The operations are collected in Table 1. All algorithms in this paper are written as sequences of these matrix operations. Because of this, we implemented the algorithms by first designing an object oriented infrastructure that provides these operations. Then we could directly map the algorithms, as they are formulated, to code.
1 We allow the existence of matrices with 0 column to represent the trivial subspace of K m .
Table 1: We summarize matrix operators to perform basic operations on subspaces. M is a general matrix, while A and B are matrices with m rows that represent the subspaces A and B ; i.e. A = 〈 A 〉 and B = 〈 B 〉 . Inspection of these routines shows that they all can be implemented with O m ( 3 ) runtime.
| Operation related to subspaces | Associated matrix operation | Correspondence |
|----------------------------------|-------------------------------|-------------------------------------|
| Kernel of a matrix M | ker M | 〈 ker M 〉 = ker M |
| Direct sum of subspaces A , B | ( A B ) | ( A B ) = A⊕B |
| Intersection of subspaces A , B | A ∩ B | 〈 A ∩ B 〉 = A∩B |
| Complement of a subspace B in A | A glyph[circleminus] B | 〈 A glyph[circleminus] B 〉 ⊕B = A |
Direct sum of two subspaces. If 〈 A , 〉 〈 B 〉 ≤ K m are two subspaces, then the direct sum can be computed by concatenating the two matrices: 〈 A 〉 ⊕ 〈 B 〉 = 〈 ( A B ) 〉 .
Null space of a matrix. Gaussian elimination can be used to compute the null space of a given m × n matrix M . Indeed, if the reduced column echelon form of the matrix ( M I n ) is blocked into the form ( M 1 M 3 M 4 ) , where M 1 has m rows and no zero column, then ker M = 〈 M 4 〉 . We denote this computation with ker M = M 4 .
Intersection of two subspaces. For two subspaces im A, im B ≤ K m , the intersection can be computed using the Zassenhaus algorithm. Namely, if the reduced column echelon form of ( A B A ) is blocked into the form ( C 1 C 3 C 4 ) , where C 1 and C 4 have m rows and no zero column, then im A ∩ im B = 〈 C 〉 . We denote this computation with A ∩ B = C .
4
4
Complement of a subspace within another one. Let im A ≤ im B ≤ K m . Then, a complement 〈 C 〉 of im A in im B , i.e., a space that satisfies 〈 C 〉 ⊕ im A = im B , can be obtained as described in Algorithm 1 . We denote this operation with 2 C = B glyph[circleminus] A .
## Algorithm 1 Complement of a subspace within another
Input:
Two matrices
A
and
B
with
m
rows
Output: A matrix A glyph[circleminus] B such that 〈 A glyph[circleminus] B 〉 ⊕ im B = im A
C ← A m × 0 matrix for each column vector b of B do
$$\inf _ { \substack { \ r k \left ( A \substack { \cdot \\ C \leftarrow ( C \right ) b ) } } } ^ { \cdot \cdot \cdot \cdot } \right ) > r k \left ( A \substack { \cdot \\ C \right ) \text{ then} }$$
end if end for
return
C
## 3.3 Double complement
Lemma 3. Let C be a finite-dimensional vector space and let A B ≤ C , with dim A ≥ dim B . Then, there exists a space S ≤ C such that:
$$\begin{cases} \ \mathcal { S } \oplus \mathcal { A } = \mathcal { C } \\ \ \mathcal { S } \cap \mathcal { B } = \{ 0 \} \end{cases}$$
Proof. We first consider the case where C = A + B . We denote with P (resp. Q ) a complement of A∩B in A (resp. in B ).
2 This algorithm can be implemented to run in a cubic arithmetical time by keeping a reduced column echelon form of ( A C ) , which makes it possible to check the condition within the loop in a squared arithmetical time.
We show first that P∩Q = 0 { } . Let v ∈ P ∩Q . As P ≤ A and Q ≤ B , we have v ∈ A∩B . Therefore, v ∈ P ∩ A ∩ B = 0 { } , as desired.
We now denote with b = { b 1 , . . . , b p } (resp. b ′ = { b ′ 1 , . . . , b ′ q } ) a basis of P (resp. Q ), implying q ≤ p . Considering w = { b 1 + b ′ 1 , . . . , b q + b ′ q } , the following holds:
- i) w is linear independent: if { α , . . . , α 1 q } ∈ K q is such that ∑ α w i i = 0 , then ∑ α b i i = ∑ α b i ′ i . As ∑ α b i i ∈ P and ∑ α b i ′ i ∈ Q , it comes that ∑ α b i ′ i = ∑ α b i i = 0 . It follows that for all i , α i = 0 , yielding the result.
- ii) 〈 w 〉 ∩ A = { 0 } : If v ∈ 〈 w 〉 ∩ A , then there exists { α , . . . , α 1 q } ∈ K q such that v = ∑ α i ( b i + b ′ i ) ∈ A . It implies ∑ α b i ′ i ∈ A . As the left hand side is in Q , it comes that ∑ α b i ′ i = 0 . It follows that for all i , α i = 0 , yielding the result.
- iii) 〈 w 〉 ∩ B = { 0 } : Same proof as above.
Then, as dim 〈 w 〉 = q = dim C dim A S , = 〈 w 〉 satisfies the desired conditions.
In the general case, where C > A + B , we use the same method, and simply add a complement S ′ of A + B in C to the solution.
Algorithm 2 uses the method used in this proof to compute a basis of S , given A B , and C . Note that if dim A = dim B , then S is a complement of both A and B in C .
## Algorithm 2 'Double complement' algorithm (Lemma 3)
Input: A B , and C , such that im A, im B ≤ im C and rk A ≥ rk B
Output: A matrix S such that 〈 S 〉 ⊕ im A = im C and 〈 S 〉 ∩ im B = 0 { } .
P ← A glyph[circleminus] ( A ∩ B )
Q ← B glyph[circleminus] ( A ∩ B )
P ′ ← P truncated such that P ′ and Q have the same size
S ′ ← C glyph[circleminus] ( A B )
$$\text{return } ( \grave { S } ^ { \prime } \ P ^ { \prime } + Q )$$
## 4 Proof of Theorem 1
We start with an auxiliary result that asserts that a decomposition of the form (3) is characterized by L .
Lemma 4. Decomposition (3) exists if and only if L is chosen such that P 4 -LP 2 is non-singular. In this case,
$$\ r k \, R = \ r k ( P _ { 3 } - L P _ { 1 } ).$$
Proof. We have:
$$\begin{pmatrix} I _ { m } & \\ L & I _ { n } \end{pmatrix} ^ { - 1 } P = \begin{pmatrix} P _ { 1 } & P _ { 2 } \\ P _ { 3 } - L P _ { 1 } & P _ { 4 } - L P _ { 2 } \end{pmatrix}$$
This matrix can be uniquely decomposed as in (2) if and only if P 4 -LP 2 is non-singular, and we have the desired value for rk R .
Now we start the actual proof of Theorem 1. If we assume that decomposition (3) exists for P , then Lemma 4 yields
$$P = \begin{pmatrix} I _ { m } & \\ L & I _ { n } \end{pmatrix} \cdot \begin{pmatrix} P _ { 1 } - P _ { 2 } ( P _ { 4 } - L P _ { 2 } ) ^ { - 1 } ( P _ { 3 } - L P _ { 1 } ) & P _ { 2 } \\ & & P _ { 4 } - L P _ { 2 } \end{pmatrix} \cdot \\ & \begin{pmatrix} I _ { m } & \\ ( P _ { 4 } - L P _ { 2 } ) ^ { - 1 } ( P _ { 3 } - L P _ { 1 } ) & I _ { n } \end{pmatrix}$$
It follows:
- · (4) is obvious from (25).
- · K n = im( P 4 -LP 2 ) ≤ im P 4 +im L . Thus, n ≤ p 4 +rk L , which yields (5).
- · im( P 3 -LP 1 ) = ( P 3 -LP 1 )( K n ) ≥ ( P 3 -LP 1 )(ker P 1 ) = P 3 (ker P 1 ) . (6) now follows from (20) and (23).
- · (7) is a direct computation:
$$\begin{array} { r c l } p _ { 3 } & = & \text{rk} ( P _ { 3 } - L P _ { 1 } + L P _ { 1 } ) \\ & \leq & \text{rk} ( P _ { 3 } - L P _ { 1 } ) + \text{rk} ( L P _ { 1 } ) \\ & \leq & \text{rk} \, R + \text{rk} \, L. \end{array}$$
## 5 Proof of Theorem 2, case p 3 ≤ m + n -p 4 -p 1
In this section, we provide an algorithm to construct an appropriate decomposition, in the case where p 3 ≤ m + n -p 4 -p 1 (Figure 1 left). This means that, using Lemma 4, we have to build a matrix L that satisfies
$$\begin{cases} P _ { 4 } - L P _ { 2 } \text{ is non-singular}, \\ \text{rk} \, L = n - p _ { 4 }, \\ \text{rk} ( P _ { 3 } - L P _ { 1 } ) = m - p _ { 1 }. \end{cases}$$
## 5.1 Sufficient conditions
We first derive a set of sufficient conditions that ensure that L satisfies the two following properties: P 4 -LP 2 non singular (Lemma 5) and rk( P 3 -LP 1 ) = m -p 1 (Lemma 6).
Lemma 5. If rk L = n -p 4 and im P 4 ⊕ LP 2 (ker P 4 ) = K n , then P 4 -LP 2 is non-singular.
Proof. We denote with U a complement of ker P 4 in K n , i.e., K n = ker P 4 ⊕ U . This implies im P 4 = P 4 ( K n ) = P 4 ( U ) . Now, let im P 4 ⊕ LP 2 (ker P 4 ) = K n . Hence, dim LP 2 (ker P 4 ) = n -p 4 = rk L from which we get im L = LP 2 (ker P 4 ) . In particular, LP 2 ( U ) ≤ im L = LP 2 (ker P 4 ) . Further,
$$\begin{array} { r c l } \text{im} ( P _ { 4 } - L P _ { 2 } ) & = & ( P _ { 4 } - L P _ { 2 } ) ( \mathcal { U } \oplus \ker P _ { 4 } ) \\ & = & ( P _ { 4 } - L P _ { 2 } ) ( \mathcal { U } ) + L P _ { 2 } ( \ker P _ { 4 } ) \end{array}$$
As LP 2 ( U ) ≤ LP 2 (ker P 4 ) and im P 4 ∩ LP 2 (ker P 4 ) = { 0 } , we have:
$$\begin{array} { r c l } \text{im} ( P _ { 4 } - L P _ { 2 } ) & = & P _ { 4 } ( \mathcal { U } ) + L P _ { 2 } ( \ker P _ { 4 } ) \\ & = & \text{im} P _ { 4 } + L P _ { 2 } ( \ker P _ { 4 } ) \\ & = & \mathbb { K } ^ { n } \end{array}$$
as desired.
Lemma 6. If, for every vector v of im P 1 , L satisfies Lv ∈ P P 3 -1 1 ( { v } ) , then rk( P 3 -LP 1 ) = m -p 1 .
Proof. Let Lv ∈ P P 3 -1 1 ( { v } ) for all v ∈ im P 1 . If u ∈ K m , we have:
$$( P _ { 3 } - L P _ { 1 } ) u \ \ & = \ \ P _ { 3 } u - L P _ { 1 } u \\ & \in \ \ & P _ { 3 } u - P _ { 3 } P _ { 1 } ^ { - 1 } ( \{ P _ { 1 } u \} ) \\ & \in \ \ & P _ { 3 } u - P _ { 3 } ( u + \ker P _ { 1 } ) \\ & \in \ \ & P _ { 3 } ( \ker P _ { 1 } )$$
Therefore, im( P 3 -LP 1 ) = P 3 (ker P 1 ) . Thus, rk P 3 -LP 1 = m -P 1 .
The following lemma summarizes the two previous results:
Lemma 7. Let Y be a complement of im P 4 and T a complement of P 1 (ker P 3 ) in im P 1 . If
$$\begin{cases} \text{im} \, L = \mathcal { Y }, \\ L \cdot P _ { 2 } ( \ker P _ { 4 } ) = \mathcal { Y }, \\ L \cdot v \in P _ { 3 } P _ { 1 } ^ { - 1 } ( \{ v \} ), \forall v \in \mathcal { T }, \\ L \cdot P _ { 1 } ( \ker P _ { 3 } ) = \{ 0 \}, \end{cases}$$
then L is an optimal solution 3 .
## 5.2 Building L
We now build a matrix L that satisfies the previous set of sufficient conditions. For all v in a complement T of P 1 (ker P 3 ) , L has to satisfy Lv ∈ P P 3 -1 1 ( { v } ) . We first show that, given a suitable domain and image, it is possible to build a bijective linear mapping that satisfies this property.
Lemma 8. Let P 1 (ker P 3 ) ⊕T = im P 1 and P 3 (ker P 1 ) ⊕V = im P 3 . If
$$\mathcal { F } = P _ { 1 } ^ { - 1 } ( \mathcal { T } ) \cap P _ { 3 } ^ { - 1 } ( \mathcal { V } ),$$
then the mapping f : T → V such that for all v ∈ F , f ( P v 1 ) = P v 3 is well defined and is an isomorphism.
Proof. We prove the lemma by first considering two functions f 1 and f 2 that are P 1 and P 3 restricted to F as shown in the diagram. We show that both are isomorphisms. Then f = f 2 ◦ f -1 1 is the desired function.
↦
↦
We begin with the surjectivity of f 1 . Let x ∈ T . As T ≤ im P 1 , there exists a vector v such that P v 1 = x . The coset v +ker P 1 is obviously a subset of P -1 1 ( T ) . Additionally, its image under P 3 , the coset P 3 ( v +ker P 1 ) = P v 3 + P 3 ker P 1 contains a unique representative P v 3 f in V , since im P 3 = P 3 (ker P 1 ) ⊕V . Therefore, v f ∈ P -1 3 ( V ) , and thus v f ∈ F and f 1 ( v f ) = x, as desired.
We now prove that f 1 is injective. Let v ∈ ker f 1 ≤ F . We have P v 3 ∈ V . Since v ∈ ker P 1 , P v 3 ∈ P 3 ker P 1 . Since P 3 (ker P 1 ) ∩ V = { 0 } , P v 3 = 0 and thus v ∈ ker P 3 . Equation (12) shows that v = 0 , as desired. Thus, f 1 is bijective.
↦
The proof that f 2 : F → V , v → P v 3 is bijective is analogous. It follows that f = f 2 ◦ f -1 1 is the desired isomorphism.
As explained below, we now build a matrix L that matches the conditions listed in Lemma 7. As they involve two spaces that may not be in a direct sum, P 2 (ker P 4 ) and a complement of P 1 (ker P 3 ) in im P 1 , some precautions must be taken.
3 The proposed set of sufficient conditions is stronger than necessary; if we replace the last condition with L P · 1 (ker P 3 ) ≤ P 3 (ker P 1 ) , if and only if holds.
We first construct the image Y of L . It must be a complement of im P 4 and must contain a complement Y 1 of P 3 (ker P 1 ) in im P 3 . As p 3 ≤ m + n -p 4 -p 1 , m -p 1 ≥ p 4 + p 3 -n , and we have dim( P 3 (ker P 1 )) ≥ dim(im P 4 ∩ im P 3 ) . Therefore, we can use the Lemma 3 to build a space Y 1 such that:
$$\begin{cases} \mathcal { Y } _ { 1 } \oplus P _ { 3 } ( \ker P _ { 1 } ) = \text{im} \, P _ { 3 }, \\ \mathcal { Y } _ { 1 } \cap \text{im} \, P _ { 4 } \cap \text{im} \, P _ { 3 } = \{ 0 \}. \end{cases}$$
We then complete Y 1 to form a complement Y of im P 4 .
We now decompose K m the following way:
$$\underbrace { \chi _ { 1 } \oplus \overbrace { \chi _ { 2 } } } _ { P _ { 2 } ( \ker P _ { 4 } ) } \oplus \chi _ { 3 } \oplus P _ { 1 } ( \ker P _ { 3 } ) } \oplus \chi _ { 4 } \ = \ \mathbb { K } ^ { m }$$
We define X 2 = P 2 (ker P 4 ) ∩ im P 1 . X ∩ 2 P 1 (ker P 3 ) = { 0 } according to equation (16). Then, we define X 3 as a complement of P 1 (ker P 3 ) ⊕ X 2 in im P 1 and X 1 as a complement of X 2 in P 2 (ker P 4 ) . X 4 is defined as a complement of X ⊕X ⊕X ⊕ 1 2 3 P 1 (ker P 3 ) .
Finally, we build L through the associated mapping, itself defined using a direct sum of linear mappings defined on the following subspaces of K m :
- · We use Lemma 8 to construct a bijective linear mapping f from T = X ⊕X 2 3 over Y 1 . By definition, for all v ∈ T , f verifies f ( v ) ∈ P P 3 -1 1 ( { v } ) . Furthermore, as f is bijective, its restriction on X 2 is itself bijective over f ( X 2 ) .
- · We complete this bijective linear mapping with g , a bijective linear mapping between X 1 and a complement Y 2 of f ( X 2 ) in Y . This way, the restriction of f ⊕ g on X ⊕X 1 2 = P 2 (ker P 4 ) is bijective over Y .
- · We consider the mapping h that maps P 1 (ker P 3 ) ⊕X 4 to { 0 } .
$$\begin{array} { c c c } \mathcal { X } _ { 1 } & & \mathcal { X } _ { 2 } \oplus \mathcal { X } _ { 3 } & & P _ { 1 } ( \ker P _ { 3 } ) \oplus \mathcal { X } _ { 4 } \\ \Big \downarrow g & & \Big \downarrow f & & \Big \downarrow h \\ \mathcal { Y } _ { 2 } & & \mathcal { Y } _ { 1 } & & \{ 0 \} \end{array}$$
The matrix associated with the linear mapping f ⊕ g ⊕ h satisfies all the conditions of Lemma 7, and is therefore an optimal solution.
This method is summarized in Algorithm 3, which allows us to construct a solution for Theorem 2. Its key part is the construction of a basis of F , which uses a generalized pseudo-inverse P † 1 (resp. P † 3 ) of P 1 (resp. P 3 ), i.e., a matrix verifying P P P 1 † 1 1 = P 1 (resp. P P P 3 † 3 3 = P 3 ). This algorithm is a main contribution of this article.
Inspection of this algorithm shows that its arithmetic cost is O (( m + n ) 3 ) .
## 5.3 Example
We illustrate our algorithm with a concrete example. Motivated by our main application (Section 2) we choose as base field K = F 2 . For m = 4 and n = 3 , we consider the matrix
$$P = \begin{pmatrix} P _ { 1 } & P _ { 2 } \\ P _ { 3 } & P _ { 4 } \end{pmatrix} = \begin{pmatrix} 1 & 1 & & 1 \\ 1 & 1 & & 1 & 1 \\ \frac { 1 & 1 & 1 & 1 \\ 1 & 1 & & 1 & 1 \\ 1 & 1 & & 1 & 1 \end{pmatrix}.$$
.
## Algorithm 3 Constructing L (Theorem 2), case p 3 ≤ m + n -p 4 -p 1
Input: mn , and P ∈ GL m n + ( K ) such that p 3 ≤ m + n -p 4 -p 1 Output: L
$$\begin{array} { l } Y _ { 1 } \leftarrow \text{Alg. 2 with } A = P _ { 3 } \cdot \overline { \ker } P _ { 1 }, \, B = P _ { 4 } \sqrt { } \cap P _ { 3 }, \, C = P _ { 3 } \\ Y \leftarrow \left ( Y _ { 1 } & I _ { m } \, \overline { \ominus } \, ( Y _ { 1 } & P _ { 4 } ) \right ) \end{array}$$
$$Y \leftarrow \left ( Y _ { 1 } \ I _ { m } \, \overline { \ominus } \left ( Y _ { 1 } \ P _ { 4 } \right ) \right )$$
X
2
←
(
P
2
·
ker
P
4
)
∩
P
1
X
X
X
F
Y
3
1
←
←
4
←
←
2
P
1
(
P
glyph[circleminus]
2
·
I
m
(
glyph[circleminus]
ker
←
Y
(
(
P
ker
1
P
(
P
P
1
1
glyph[circleminus]
(
P
P
3
·
·
4
†
1
(
ker
P
)
glyph[circleminus]
P
2
·
(
ker
3
X
)
2
·
ker
(
X
2
P
1
X
P
2
4
)
X
3
))
P
†
1
)
)
∩
·
X
2
ker
∩
F
P
3
))
L
R
←
(
P
1
·
F
X
1
P
1
·
ker
P
3
X
4
)
L L ← ( P 3 · F Y 2 Z ) , where Z is a zero filled matrix such that L L has the same number of columns as L R return L L · L -1 R
We observe p 3 = 3 ≤ m + n -p 4 -p 1 = 4 + 3 -1 -3 . Therefore, we can use Algorithm 3 to compute a suitable L .
The first step is to compute Y 1 . We have:
$$P _ { 3 } \cdot \overline { \ker } \, P _ { 1 } = \begin{pmatrix} 1 \\ 0 \\ 1 \end{pmatrix} \text{ and } P _ { 4 } \, \overline { \cap } \, P _ { 3 } = \begin{pmatrix} 0 \\ 1 \\ 0 \end{pmatrix}.$$
Using Algorithm 2, we get
$$Y _ { 1 } = \begin{pmatrix} 1 & 1 \\ 1 & 0 \\ 1 & 0 \end{pmatrix}.$$
Then, we complete it to form a complement of im P 4 :
$$Y = \begin{pmatrix} 1 & 0 \\ 0 & 1 \\ 0 & 1 \end{pmatrix}.$$
The next step computes the different domains:
$$X _ { 2 } = \begin{pmatrix} 1 \\ 1 \\ 0 \\ 0 \end{pmatrix}, \, X _ { 3 } = \begin{pmatrix} 1 \\ 0 \\ 0 \\ 0 \end{pmatrix}, \, X _ { 1 } = \begin{pmatrix} 0 \\ 0 \\ 0 \\ 1 \end{pmatrix} \text{ and } X _ { 4 } = ( ).$$
To compute F , we need pseudo-inverses of P 1 and P 2 :
$$P _ { 1 } ^ { \dagger } = \begin{pmatrix} 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 1 & 0 & 1 & 0 \end{pmatrix} \text{ and } P _ { 3 } ^ { \dagger } = \begin{pmatrix} 0 & 1 & 1 \\ 0 & 0 & 1 \\ 0 & 0 & 0 \\ 1 & 1 & 1 \end{pmatrix}.$$
We then obtain:
Then, we compute Y 2 :
$$F = \begin{pmatrix} 0 & 0 \\ 1 & 0 \\ 0 & 1 \\ 1 & 0 \end{pmatrix}.$$
$$\mathcal { Y } _ { 2 } = \begin{pmatrix} 1 \\ 0 \\ 0 \end{pmatrix}.$$
(
)
P
†
3
·
Y
1
)
Now we can compute L . With
$$L _ { R } = \begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 1 \\ 0 & 0 & 0 & 1 \\ 0 & 0 & 1 & 0 \end{pmatrix}, L _ { L } = \begin{pmatrix} 1 & 1 & 1 & 0 \\ 1 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \end{pmatrix}$$
we get
$$L = L _ { L } \cdot L _ { R } ^ { - 1 } = \begin{pmatrix} 1 & 1 & 1 & 1 \\ 1 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \end{pmatrix}.$$
The final decomposition is now obtained using (25):
This decomposition satisfies rk L = 2 and rk R = 1 , thus matching the bounds of Theorem 1. If we consider the application presented in Section 2, this decomposition provides a way to implement in hardware the permutation associated with P on 128 elements, arriving in chunks of 8 during 16 cycles. This yields an implementation consisting of a permutation network of 4 2 × 2 -switches, followed by a block of 8 RAM banks, followed by another permutation network with 8 2 × 2 -switches.
## 6 Proof of Theorem 2, case p 3 ≥ m + n -p 4 -p 1
In this case, the third inequality in Theorem 1 is restrictive (Figure 1 right). Using again Lemma 4, we have to build a matrix L satisfying:
$$\begin{cases} P _ { 4 } - L P _ { 2 } \text{ is non-singular} \\ \text{rk} \, L = n - p _ { 4 } \\ \text{rk} ( P _ { 3 } - L P _ { 1 } ) = p _ { 4 } + p _ { 3 } - n \end{cases}$$
As in the previous section, we will first provide a set of sufficient conditions for L and then build it.
## 6.1 Sufficient conditions
The set of conditions that we will derive in this subsection will be slightly more complex than in the previous section, as we cannot reach the intrinsic bound of rk( P 3 -LP 1 ) . Particularly, we cannot use Lemma 6 directly.
Lemma 9. If W is such that W⊕ im P 4 = K n and T is such that:
$$\begin{cases} \ \mathcal { T } \cap P _ { 1 } ( \ker P _ { 3 } ) = \{ 0 \} \\ \ \mathcal { T } \leq \lim P _ { 1 } \\ \ \dim \mathcal { T } = n - p _ { 4 }, \end{cases}$$
then if L satisfies
$$\begin{cases} \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{\ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text
\end{cases}$$
then L is a solution 4 that verifies rk L = n -p 4 and rk( P 3 -LP 1 ) = p 4 + p 3 -n .
Proof. Let L be a matrix that satisfies the conditions above. Using Lemma 5 as before, we get that rk L = n -p 4 and P 4 -LP 2 invertible.
Now, with the definition of T , we can define a dimension p 4 + p 3 + p 1 -m -n space T ′ such that im P 1 = P 1 (ker P 3 ) ⊕T ⊕T ′ . Then, we define a matrix L ′ such that:
$$\begin{cases} \begin{array} { c } L ^ { \prime } \cdot v \in L v - P _ { 3 } P _ { 1 } ^ { - 1 } ( \{ v \} ), \, \text{for all } v \in \mathcal { T } ^ { \prime } \\ L ^ { \prime } \cdot v = 0, \, \text{for all } v \, \text{in } P _ { 1 } ( \ker P _ { 3 } ) \oplus \mathcal { T } \\ L ^ { \prime } \cdot v = 0, \, \text{for all } v \, \text{in a complement of} \, \text{im} P _ { 1 }, \end{array} \end{cases}$$
L ′ is therefore a rank p 4 + p 3 + p 1 -m -n matrix such that for all v ∈ im P , 1 ( L -L v ′ ) ∈ P P 3 -1 1 ( { v } ) . We apply Lemma 6 on L -L ′ and get:
$$\begin{array} { r c l } \text{rk} ( P _ { 3 } - L P _ { 1 } ) & = & \text{rk} ( P _ { 3 } - ( L - L ^ { \prime } ) P _ { 1 } - L ^ { \prime } P _ { 1 } ) \\ & \leq & \text{rk} ( P _ { 3 } - ( L - L ^ { \prime } ) P _ { 1 } ) + \text{rk} ( L ^ { \prime } P _ { 1 } ) \\ & \leq & \text{dim} P _ { 3 } ( \ker P _ { 1 } ) + \text{rk} L ^ { \prime } \\ & \leq & p _ { 4 } + p _ { 3 } - n, \end{array}$$
as desired.
## 6.2 Building L
We will build a matrix L that matches the conditions listed in Lemma 9. As before, we consider the image Y of L first. We will design it such that it is a complement of im P 4 , and that is contained in a complement Y ′ of P 3 (ker P 1 ) in im P 3 . Using p 3 ≥ m + n -p 4 -p 1 and Lemma 3, we can get a space Y such that:
$$\begin{cases} \ \mathcal { Y } \oplus ( \text{im} \, P _ { 4 } \cap \text{im} \, P _ { 3 } ) = \text{im} \, P _ { 3 } \\ \ \mathcal { Y } \cap P _ { 3 } ( \ker P _ { 1 } ) = \{ 0 \} \end{cases}$$
This space satisfies Y ⊕ im P 4 = im P 3 + im P 4 = K n , and can be completed to a complement Y ′ of P 3 (ker P 1 ) in im P 3 . Note that we will use Y ′ only to define f ; the image of L will be Y .
Now, as before, we build L through the associated mapping, itself defined using a direct sum of linear mappings defined on the same subspaces of K m as in Section 5.2.
- · We use Lemma 8 to construct a first bijective linear mapping f ′ between T ′ = X ⊕ X 2 3 and Y ′ . As f ′ is bijective, we can define T = f ′-1 ( Y ) and f = f ′ T . Thus, T satisfies the properties in Lemma 9 and L the condition for all v ∈ T , Lv ∈ P P 3 -1 1 ( { v } ) .
- · Then, we consider a complement X ′ 1 of T ∩X 2 in P 2 (ker P 4 ) , a complement Y ′ 2 of f ( T ∩X 2 ) in Y and a bijective linear mapping g between X ′ 1 and Y ′ 2 . This way, the restriction of f ⊕ g on P 2 (ker P 4 ) is bijective over Y .
The rest of the algorithm in similar to the previous case:
- · We consider the mapping h that maps P 1 (ker P 3 ) to { 0 } .
- · To complete the definition of L , we take a mapping h ′ between a complement X ′ 4 of X ′ 1 ⊕ T ⊕ P 1 (ker P 3 ) and { 0 } .
4 If we replace the last condition with L · P 1 (ker P 3 ) ≤ P 3 (ker P 1 ) , this set of conditions is actually equivalent to having an optimal solution L that satisfies rk L = n -p 4 .
·
The matrix associated with the mapping f ⊕ g ⊕ h ⊕ h ′ satisfies all the conditions of Lemma 9, and is therefore an optimal solution.
Algorithm 4 summarizes this method, and allows to construct a solution for Theorem 2, in the case where p 3 > m + n -p 4 -p 1 . This algorithm is a main contribution of this article.
## Algorithm 4 Constructing L (Theorem 2), case p 3 > m + n -p 4 -p 1
Input: mn P , , ∈ GL m n + ( K ) such that p 3 > m + n -p 4 -p 1 Output: L
Y ← Alg. 2 with A = P 4 ∩ P 3 , B = P 3 · ker P 1 and C = P 3
X
2
←
(
P
2
·
ker
P
4
)
∩
P
1
X
F
X
X
3
←
←
(
′
1
←
4
Y
′
2
←
←
P
1
glyph[circleminus]
ker
P
(
P
2
I
(
P
1
·
m
Y
1
·
P
ker glyph[circleminus]
glyph[circleminus]
(
(
X
P
3
·
ker
P
†
3
(
1
P
·
4
′
1
(
X
)
glyph[circleminus]
P
F
2
((
1
∩
P
1
·
(
F
P
†
1
X
2
X
)
3
·
))
F
)
P
∩
1
X
2
∩
(
ker
X
2
·
ker ker
P
)
P
3
)
P
1
))
L
R
←
(
P
1
·
F
X
′
1
P
1
·
ker
P
3
X
4
)
L L ← ( P 3 · F Y ′ 2 Z ) , where Z is a zero filled matrix such that L L has the same number of columns as L R return L L · L -1 R
As in the previous case, this algorithm has an arithmetic cost cubic in m + n .
## 6.3 Example
We now consider, for K = F 2 , m = 4 and n = 3 , the matrix
$$\Lambda _ { \Lambda } = - \frac { \Lambda _ { \Lambda } } { 2 }, \, \mathbb { m } - \frac { \Lambda _ { \Lambda } } { 2 }, \, \mathbb { m } - \mathfrak { m }, \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \
P = \begin{pmatrix} P _ { 1 } & P _ { 2 } \\ P _ { 3 } & P _ { 4 } \end{pmatrix} = \begin{pmatrix} 1 & 1 & 1 & 1 & 1 \\ 1 & 1 & 1 & 1 & 1 & 1 \\ 1 & 1 & 1 & 1 & 1 & 1 \\ 1 & 1 & 1 & 1 & 1 \\ 1 & 1 & 1 & 1 & 1 \end{pmatrix} \,. \\ \Lambda _ { \Lambda } \,. \quad \Lambda _ { \Lambda } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m } \, \mathfrak { m }.$$
We observe p 3 = 3 > m + n -p 4 -p 1 = 4 + 3 -2 -3 . Therefore, we use Algorithm 4 to compute a suitable L .
The first step is to compute Y . We have:
$$P _ { 3 } \cdot \overline { \ker } \, P _ { 1 } = \begin{pmatrix} 0 \\ 1 \\ 1 \end{pmatrix} \text{ and } P _ { 4 } \, \overline { \cap } \, P _ { 3 } = \begin{pmatrix} 0 \\ 1 \\ 0 \end{pmatrix}.$$
Using Algorithm 2, we get
$$Y _ { 1 } = \begin{pmatrix} 1 & 1 \\ 1 & 0 \\ 1 & 0 \end{pmatrix}.$$
Then, we complete it to form a complement of im P 4 :
$$Y = \begin{pmatrix} 1 & 0 \\ 1 & 0 \\ 0 & 1 \end{pmatrix}.$$
)
3
P
†
3
·
Y
)
The next step computes the different domains:
$$X _ { 2 } = ( \, ) \text{ and } X _ { 3 } = \begin{pmatrix} 1 & 0 \\ 0 & 1 \\ 1 & 0 \\ 0 & 0 \end{pmatrix}.$$
To compute F , we need pseudo-inverses of P 1 and P 2 :
$$P _ { 1 } ^ { \dagger } = \begin{pmatrix} 0 & 0 & 0 & 0 \\ 0 & 1 & 0 & 1 \\ 0 & 0 & 1 & 1 \\ 0 & 1 & 0 & 0 \end{pmatrix} \text{ and } P _ { 3 } ^ { \dagger } = \begin{pmatrix} 1 & 1 & 0 \\ 0 & 0 & 0 \\ 1 & 0 & 1 \\ 0 & 1 & 0 \end{pmatrix}$$
and get
$$F = \begin{pmatrix} 1 \\ 1 \\ 0 \\ 0 \end{pmatrix}.$$
Next we compute the remaining subspaces that depend on F :
$$X _ { 1 } ^ { \prime } = \begin{pmatrix} 0 \\ 1 \\ 1 \end{pmatrix}, \, X _ { 4 } = \begin{pmatrix} 1 \\ 0 \\ 0 \end{pmatrix}, \, \text{and} \, Y _ { 2 } ^ { \prime } = \begin{pmatrix} 1 \\ 0 \\ 1 \end{pmatrix}.$$
Noe we can compute L . With
$$L _ { R } = \begin{pmatrix} 1 & 0 & 1 & 1 \\ 1 & 1 & 0 & 0 \\ 1 & 1 & 1 & 0 \\ 0 & 1 & 1 & 0 \end{pmatrix}, L _ { L } = \begin{pmatrix} 1 & 1 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 1 & 1 & 0 & 0 \end{pmatrix}$$
we get
$$L = L _ { L } \cdot L _ { R } ^ { - 1 } = \begin{pmatrix} 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \end{pmatrix}.$$
The final decomposition is obtained using (25):
$$P = \begin{pmatrix} 1 & & \mathcal { \Psi } & \mathcal { \Psi } & \mathcal { \Psi } \end{pmatrix} \\ \begin{pmatrix} 1 & & \mathcal { \Psi } & \mathcal { \Psi } & \mathcal { \Psi } \end{pmatrix} \\ \begin{pmatrix} 1 & & \mathcal { \Psi } & \mathcal { \Psi } & \mathcal { \Psi } \end{pmatrix} \\ \begin{pmatrix} 1 & & \mathcal { \Psi } & \mathcal { \Psi } & \mathcal { \Psi } & \mathcal { \Psi } \end{pmatrix} \cdot \begin{pmatrix} 1 & & 1 & 1 & 1 \\ 1 & & 1 & 1 & 1 \\ 1 & & 1 & 1 & 1 \\ 1 & & & 1 & 1 \\ & & & 1 & 1 & 1 \\ & & & 1 & 1 \\ \end{pmatrix} \\ \begin{pmatrix} 1 & & & \mathcal { \Psi } & \mathcal { \Psi } & \mathcal { \Psi } & \mathcal { \Psi } & \mathcal { \Psi } \end{pmatrix}. \\ \begin{pmatrix} 1 & & & 1 & 1 \\ 1 & & 1 & 1 \\ & & & 1 & 1 \\ & & & 1 & 1 \\ \end{pmatrix} \\ \end{pmatrix} \\ \text{composition satisfies rk L = 1 and rk R = 2, thus matching the bounds of Theorer} \\ \text{re, if we consider the application of Section 2. The decomposition shows that we can}$$
This decomposition satisfies rk L = 1 and rk R = 2 , thus matching the bounds of Theorem 1. As before, if we consider the application of Section 2. The decomposition shows that we can implement in hardware the permutation associated with P on 128 elements, arriving in chunks of 8 during 16 cycles through a permutation network of 8 2 × 2 -switches, followed by a block of 8 RAM banks, followed by another permutation network with 4 2 × 2 -switches.
Figure 5: L ′ trades a rank of L for a rank of R on the associated decomposition.
## 7 Rank exchange
The solution built in Section 6.2 satisfies rk L = n -p 4 and rk P 3 -LP 1 = p 4 + p 3 -n . In this section, we will show that it is possible to construct a solution for all possible pairs (rk L, rk P 3 -LP 1 ) matching the bounds in Theorem 1. First, we will construct a rank 1 matrix L ′ that will trade a rank of L for a rank of P 3 -LP 1 (i.e., rk( L + L ′ ) = 1 + rk L , rk( P 3 -( L + L P ′ ) 1 ) = rk( P 3 -LP 1 ) -1 and P 4 -( L + L P ′ ) 2 is non-singular) (see Figure 5). This method can then be applied several times, until rk( P 3 -LP 1 ) reaches its own bound, m -p 1 .
We assume that L satisfies the following conditions:
$$\begin{cases} \ P _ { 4 } - L P _ { 2 } \text{ is non-singular} \\ \ r k \, L + \ r k ( P _ { 3 } - L P _ { 1 } ) = p _ { 3 } \\ \ r k ( P _ { 3 } - L P _ { 1 } ) > m - p _ { 1 } \end{cases}$$
As in the previous sections, we first formulate sufficient conditions on L ′ , before building it.
## 7.1 Sufficient conditions
We now define C = P 1 -P 2 ( P 4 -LP 2 ) -1 ( P 3 -LP 1 ) .
Lemma 10. If z ∈ K m satisfies z / ∈ ker( P 3 -LP 1 ) + ker P 1 and L ′ satisfies:
glyph[negationslash]
$$\begin{cases} \quad \cdot \quad \cdot \quad \cdot \\ \quad \begin{cases} \text{rk} L ^ { \prime } = 1 \\ L ^ { \prime } P _ { 1 } \ker ( P _ { 3 } - L P _ { 1 } ) = \{ 0 \} \\ L ^ { \prime } P _ { 1 } z = ( P _ { 3 } - L P _ { 1 } ) z \\ L ^ { \prime } C z \neq 0 \end{cases} \end{cases}$$
Then L + L ′ is an optimal solution to our problem that satisfies rk( P 3 -( L + L P ′ ) 1 ) = rk( P 3 -LP 1 ) -1 . Let
Proof. We first prove that P 4 -( L + L P ′ ) 2 = ( I -L P ′ 2 ( P 4 -LP 2 ) -1 )( P 4 -LP 2 ) is non singular. x ∈ ker( I -L P ′ 2 ( P 4 -LP 2 ) -1 ) . x satisfies:
$$x - L ^ { \prime } P _ { 2 } ( P _ { 4 } - L P _ { 2 } ) ^ { - 1 } x = 0.$$
Therefore, x ∈ im L ′ . As rk L ′ = 1 , ∃ λ ∈ K , x = λ P ( 3 -LP 1 ) z = λL P z ′ 1 . It comes that
$$\lambda L ^ { \prime } P _ { 1 } z - \lambda L ^ { \prime } P _ { 2 } ( P _ { 4 } - L P _ { 2 } ) ^ { - 1 } ( P _ { 3 } - L P _ { 1 } ) z = 0.$$
glyph[negationslash]
Finally, λL Cz ′ = 0 , which implies, as L Cz ′ = 0 , λ = 0 , as desired.
Wenow prove that rk( P 3 -( L + L P ′ ) 1 ) = rk( P 3 -LP 1 ) -1 . Wehave already ker( P 3 -( L + L P ′ ) 1 ) ≤ ker( P 3 -LP 1 ) as L P ′ 1 ker( P 3 -LP 1 ) = { 0 } . We also have ( P 3 -( L ′ + L P ) 1 ) z = 0 . As z / ∈ ker( P 3 -LP 1 ) + ker P 1 , ker( P 3 -( L + L P ′ ) 1 ) ≥ ker( P 3 -LP 1 ) ⊕〈 〉 z . Therefore,
$$\dim \ker ( P _ { 3 } - ( L + L ^ { \prime } ) P _ { 1 } ) \geq 1 + \dim \ker ( P _ { 3 } - L P _ { 1 } ),$$
as desired.
## 7.2 Building L ′
Lemma 11. ker( P 3 -LP 1 ) ∩ ker P 1 = { 0 }
$$P r o o f. \text{ This is a consequence of } ( 2 4 ) \colon \text{ the block column } \begin{pmatrix} P _ { 1 } \\ P _ { 3 } - L P _ { 1 } \end{pmatrix} \text{ has full rank.}$$
Thus, we have dim(ker( P 3 -LP 1 ) ⊕ ker P 1 ) = 2 m -p 1 -rk( P 3 -LP 1 ) < m . Decomposition (25) shows that C is non-singular, and we have dim C -1 P 1 ker( P 3 -LP 1 ) = dim P 1 ker( P 3 -LP 1 ) = dim ker( P 3 -LP 1 ) = m -rk( P 3 -LP 1 ) . Using Lemma 3, we can build a space Z such that:
$$\begin{cases} \ \mathcal { Z } \oplus \ker ( P _ { 3 } - L P _ { 1 } ) \oplus \ker P _ { 1 } = \mathbb { K } ^ { m } \\ \ \mathcal { Z } \cap C ^ { - 1 } P _ { 1 } \ker ( P _ { 3 } - L P _ { 1 } ) = \{ 0 \} \end{cases}$$
We can now pick a nonzero element z ∈ Z and build a corresponding L ′ :
- · If Cz ∈ P 1 ker( P 3 -LP 1 ) ⊕〈 P z 1 〉 : We take a complement A of P 1 ker( P 3 -LP 1 ) ⊕〈 P z 1 〉 and build L ′ such that:
$$\begin{cases} \ L ^ { \prime } P _ { 1 } z = ( P _ { 3 } - L P _ { 1 } ) z \\ \ L ^ { \prime } ( P _ { 1 } \ker ( P _ { 3 } - L P _ { 1 } ) \oplus \mathcal { A } ) = \{ 0 \} \end{cases}$$
glyph[negationslash]
glyph[negationslash]
glyph[negationslash]
We have L Cz ′ = 0 . In fact, Cz can be uniquely decomposed in the form k + λP z 4 , where k ∈ P 1 ker( P 3 -LP 1 ) and λ ∈ K . As z / C ∈ -1 P 1 ker( P 3 -LP 1 ) , λ = 0 . Then, L Cz ′ = L k ′ + L λP z ′ 4 = 0 + λ P ( 3 -LP 1 ) z = 0 .
- · If Cz / P ∈ 1 ker( P 3 -LP 1 ) ⊕〈 P z 1 〉 : The vector a = Cz -P z 1 is outside of P 1 ker( P 3 -LP 1 ) ⊕〈 P z 1 〉 . Therefore, it is possible to build a complement A of P 1 ker( P 3 -LP 1 ) ⊕ 〈 P z 1 〉 that contains a . Then, we build L ′ as before:
$$\begin{cases} \ L ^ { \prime } P _ { 1 } z = ( P _ { 3 } - L P _ { 1 } ) z \\ \ L ^ { \prime } ( P _ { 1 } \ker ( P _ { 3 } - L P _ { 1 } ) \oplus \mathcal { A } ) = \{ 0 \} \end{cases}$$
glyph[negationslash]
As in the previous case, we have L Cz ′ = L a ′ + L P z ′ 1 = 0 + ( P 3 -LP 1 ) z = 0 .
In both cases, the matrix L ′ we built satisfies the conditions of Lemma 10. Therefore, L + L ′ is the desired solution.
Algorithm 5 summarizes this method, and allows to build a new optimal solution from a pre-existing one, with a different trade-off. This algorithm is a main contribution of this article.
## Algorithm 5 Exchanging ranks between L and R (Theorem 3)
Input: P and a solution L such that rk( P 3 -LP 1 ) > m -p 1 Output: A new optimal solution L with a rank incremented by 1
K
←
ker(
P
3
-
LP
1
)
C
←
P
1
-
P
2
(
P
4
-
LP
2
)
-
1
(
P
3
-
LP
1
)
$$Z \leftarrow \text{Alg.} \text{ } 2 \text{ with } A = \left ( K \not \, \overline { \ker } P _ { 1 } \right ), \, \overset { i } { B } = C ^ { - 1 } P _ { 1 } K \text{ and } C = I _ { m }$$
z
←
first column of
Z
if
Cz
∈
(
P
1
·
K
P z
A
←
I
m
glyph[circleminus]
P
1
·
1
)
(
K
then
z
)
else
$$\begin{smallmatrix} a \leftarrow ( C - P _ { 1 } ) z \\ A \leftarrow ( I _ { m } \overline { \ominus } ( P _ { 1 } K \ \ P _ { 1 } z \ \ a ) \ \ a ) \end{smallmatrix}$$
end if
L ′ R ← ( P z 1 P K 1 A ) L ′ L ← ( ( P 3 -LP 1 ) z F ) , where F is a zero filled matrix such that L ′ L has the same number of columns as L ′ R
$$\text{return } L + L ^ { \prime } _ { L } \cdot L ^ { \prime - 1 } _ { R }$$
## 7.3 Example
To illustrate Algorithm 5, we continue the example of Section 6.3, and the matrix L that we found. We have:
$$K = \begin{pmatrix} 1 & 0 \\ 0 & 1 \\ 0 & 0 \\ 0 & 0 \end{pmatrix} \text{ and } C = \begin{pmatrix} 1 & & 1 \\ 1 & & & 1 \\ 1 & 1 & & \end{pmatrix}.$$
Using Algorithm 2, we get
$$Z = z = \begin{pmatrix} 0 \\ 0 \\ 1 \\ 0 \end{pmatrix}.$$
$$\text{As } C z \, = \, \begin{pmatrix} 0 \\ 0 \\ 1 \\ 0 \end{pmatrix} \text{ is not included in } ( P _ { 1 } K \, \ P _ { 1 } z ) \, = \, \begin{pmatrix} \overset { \cdot \, \cdot } { \cdot } & 0 \\ 0 & 1 & 0 \\ 1 & 0 & 0 \\ 0 & 0 & 1 \end{pmatrix}, \, \text{we compute } A \, \text{as a complement of}$$
$$\begin{pmatrix} \mathbb { V } / \\ ( P _ { 1 } K & P _ { 1 } z ) \text{ that contains } a = \begin{pmatrix} 1 \\ 0 \\ 0 \\ 0 \end{pmatrix} \colon \end{pmatrix}$$
$$\mathcal { A } = \left \langle \begin{pmatrix} 1 \\ 0 \\ 0 \end{pmatrix} \right \rangle.$$
Now, we compute L ′ , using:
$$L ^ { \prime } _ { R } = \begin{pmatrix} 1 & 0 & 1 & 1 \\ 0 & 1 & 0 & 0 \\ 1 & 0 & 1 & 0 \\ 0 & 1 & 1 & 0 \end{pmatrix}, \, L ^ { \prime } _ { L } = \begin{pmatrix} 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \end{pmatrix}, \, \text{and} \, L ^ { \prime } = \begin{pmatrix} 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 1 & 1 & 1 \end{pmatrix}.$$
Finally, we get the new decomposition, using, as usual, Equation (25):
As expected, the left off-diagonal rank has increased by one, while the right one has decreased by one. The two different decompositions that we now have for P cover all the possible tradeoffs that minimize the off-diagonal ranks.
## 8 Conclusion
In this paper, we introduced a novel block matrix decomposition that generalizes the classical block-LU factorization. A 2 × 2 -blocked invertible matrix is decomposed into a product of three matrices: lower block unitriangular, upper block triangular, and lower block unitriangular matrix, such that the sum of the off-diagonal ranks are minimal. We provided an algorithm that computes an optimal solution with an asympotic number of operations cubic in the matrix size. We note that we implemented the algorithm for finite fields, for rational numbers, for Gaussian rational numbers and for exact real arithmetic for validation. For a floating point implementation, numerical issues may arise.
The origin of the considered decomposition, as we explained, is in the design of optimal circuits for a certain class of streaming permutations that are very relevant in practice. However, we believe that because of its simple and natural structure, the matrix decomposition is also of pure mathematical interest. Specifically, it would be interesting to investigate if the proposed decomposition is a special case of a more general problem that involves, for example, finer block structures.
## References
- [1] D. Carlson, E. Haynsworth, and T. Markham. A generalization of the Schur complement by means of the Moore-Penrose inverse. SIAM Journal on Applied Mathematics , 26(1):169-175, 1974.
- [2] E. Chow and Y. Saad. Approximate inverse techniques for block-partitioned matrices. SIAM Journal on Scientific Computing , 18(6):1657-1675, 1997.
- [3] Richard W. Cottle. Manifestations of the Schur complement. Linear Algebra and its Applications , 8(3):189 - 211, 1974.
- [4] James W. Demmel, Nicholas J. Higham, and Robert S. Schreiber. Block LU factorization. Numerical Linear Algebra with Applications , 2(2):173-190, 1992.
- [5] Jacques Lenfant and Serge Tahé. Permuting data with the Omega network. Acta Informatica , 21(6):629-641, 1985.
- [6] Victor Pan. Structured matrices and polynomials: unified superfast algorithms . Springer Science & Business Media, 2001.
- [7] Keshab K. Parhi. Systematic synthesis of DSP data format converters using life-time analysis and forward-backward register allocation. IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing , 39(7):423-440, Jul 1992.
- [8] Marshall C. Pease. The indirect binary n-cube microprocessor array. IEEE Transactions on Computers , 26(5):458-473, May 1977.
- [9] Markus Püschel, Peter A. Milder, and James C. Hoe. Permuting streaming data using RAMs. Journal of the ACM , 56(2):10:1-10:34, 2009.
- [10] R. Tolimieri, M. An, and C. Lu. Algorithms for Discrete Fourier Transforms and Convolution . Springer, 2nd edition, 1997.
- [11] Fuzhen Zhang. The Schur Complement and its Applications , volume 4. Springer, 2006.
|
10.1016/j.laa.2016.07.020
|
[
"François Serre",
"Markus Püschel"
] |
2014-08-05T15:04:34+00:00
|
2015-10-06T12:00:57+00:00
|
[
"math.RA",
"15A23"
] |
A Lower-Upper-Lower Block Triangular Decomposition with Minimal Off-Diagonal Ranks
|
We propose a novel factorization of a non-singular matrix $P$, viewed as a
$2\times 2$-blocked matrix. The factorization decomposes $P$ into a product of
three matrices that are lower block-unitriangular, upper block-triangular, and
lower block-unitriangular, respectively. Our goal is to make this factorization
"as block-diagonal as possible" by minimizing the ranks of the off-diagonal
blocks. We give lower bounds on these ranks and show that they are sharp by
providing an algorithm that computes an optimal solution. The proposed
decomposition can be viewed as a generalization of the well-known Block LU
factorization using the Schur complement.
|
1408.0995v2
|
## A Genus Two Curve Related to the Class Number One Problem
by Viet Kh. Nguyen 1
` la m´moire de V. A. Iskovskikh a e
## Abstract
We give another solution to the class number one problem by showing that imaginary quadratic fields Q ( √ -d ) with class number h ( -d ) = 1 correspond to integral points on a genus two curve K 3 . In fact one can find all rational points on K 3 . The curve K 3 arises naturally via certain coverings of curves: K 3 → K 6 , K 1 → K 2 with K 2 : y 2 = 2 x x ( 3 -1) denoting the Heegner curve, also in connection with the so-called Heegner-Stark covering K 1 → K s .
- 1. Introduction. In his famous Disquisitiones Arithmeticae (Article 303) Gauss hypothetically believed that there are only nine imaginary quadratic fields Q ( √ -d ) with class number h ( -d ) = 1, namely with d = 1 2 3 7 11 19 43 67 163. , , , , , , , , It is known nowadays as the Baker-Heegner-Stark theorem.
In the early 1980's V. A. Iskovskikh organized a seminar devoted to a beautiful approach of Heegner who was substantially first to give a complete proof of the theorem of BakerHeegner-Stark based on an approach using modular functions ([2], cf. also [1], [6], [4], [7], [3], [5]). More precisely, assuming d ≡ 3 (mod 8) (the main case), Heegner showed that every such a field corresponds to an integral point on the so-called Heegner curve:
$$y ^ { 2 } & = 2 x ( x ^ { 3 } - 1 )$$
In the higher class number case the statement reads as every field Q ( √ -d ), d ≡ 3 (mod 8) corresponds to an integral point of degree h ( -d ) on K 2 , provided 3 /notbar d . One may have a remark that the ' exceptional ' field Q ( √ -3) (corresponding to the point (0 , 0) on K 2 ) is the only case with 3 | d such that j -invariant is a cube (namely equal to zero). On the contrary to the case of higher class numbers one sees, for example if h ( -d ) = 2, three fields Q ( √ -51), Q ( √ -123) and Q ( √ -267) have no representatives on K 2 along this line of arguments.
The aim of this note is to show that for our purpose one can involve another genus two curve K 3 :
$$\frac { ( \mathcal { X } _ { 3 } ) \ 8 x ^ { 8 } - 3 2 x ^ { 6 } y + 4 0 x ^ { 4 } y ^ { 2 } + 6 4 x ^ { 5 } - 1 6 x ^ { 2 } y ^ { 3 } - 1 2 8 x ^ { 3 } y + y ^ { 4 } + 4 8 x y ^ { 2 } + 9 6 x ^ { 2 } - 3 2 y - 2 4 = 0 } { \mu \Phi \Phi \Phi \Phi \Phi \Phi \Phi }$$
1 IMC Institute & Fpt University
2010 Mathematics Subject Classification. Primary 11G30, 11R29, 11R11.
realized at that time by the author.
The main result of the note is
Theorem A. The curve K 3 has the following integral points
$$\begin{smallmatrix} ( 3, 6 ), & ( - 1, 2 ), & ( 1, 6 ), & ( 3, 1 4 ), & ( 7, 2 6 ), & ( - 1 7, 1 5 0 ) \\ \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \end{smallmatrix}$$
corresponding to the fields respectively. Besides it has three 'extraneous' integral points: ( -1 , -2) , ( -3 6) , , (1 , 2) (in fact (1 , 2) is a double point on K 3 ) plus two more rational points ( -9 17 , 6 289 ) , ( -155 79 , 42486 6241 ) . All the listed are the only rational points on K 3 .
$$\ n r e s p r o n u a n y & \, \omega \ u v e \ j u v e \\ & \, \mathbb { Q } ( \sqrt { - 3 } ), \, \mathbb { Q } ( \sqrt { - 1 1 } ), \, \mathbb { Q } ( \sqrt { - 1 9 } ), \, \mathbb { Q } ( \sqrt { - 4 3 } ), \, \mathbb { Q } ( \sqrt { - 6 7 } ), \, \mathbb { Q } ( \sqrt { - 1 6 3 } ) \\ \intertext { c o n c t i a l l o l } \, \underset { \mathcal { D } a o d i a l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l } \, \underset { \mathcal { D } a o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l od i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o dd i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d i a l l o d j u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e\ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e \ u v e$$
As a corollary one obtains a new solution to Gauss' ' tenth discriminant problem '.
Nowadays with the aid of computational tools one can easily compute the genus of K 3 . A surprising observation is the fact that K 3 is birational to a genus 2 curve K s studied independently by Heegner and Stark ([2], [6]).
Acknowledgments. I would like to thank V. A. Abrashkin for a nice talk on the class number problem at the seminar mentioned in the Introduction. This paper is dedicated to the 75th birthday of my teacher V. A. Iskovskikh with profound admiration and memory.
- 2. The curve K 3 . Let us first recall some known results from the Weber-HeegnerBirch theory ([10], [2], [1], [3]). We shall use Heegner's notation σ for the Schl¨fli functions a of Stufe 48 which are defined by
$$\sigma ( \tau ) = q ^ { - \frac { 1 } { 2 4 } } \, \prod _ { n = 1 } ^ { \infty } \, \left ( 1 + q ^ { 2 n - 1 } \right ) ; \quad \sigma _ { 1 } ( \tau ) = q ^ { - \frac { 1 } { 2 4 } } \, \prod _ { n = 1 } ^ { \infty } \, \left ( 1 - q ^ { 2 n - 1 } \right ) ;$$
$$\sigma _ { 2 } ( \tau ) = \sqrt { 2 } \ q ^ { \frac { 1 } { 1 2 } } \ \prod _ { n = 1 } ^ { \infty } \ ( 1 + q ^ { 2 n } ) ;$$
with q = e iπτ , Im( τ ) > 0. Then the 24 th powers σ 24 , -σ 24 1 , -σ 24 2 are the roots of
$$U ^ { 3 } - 4 8 U ^ { 2 } + ( 7 6 8 - j ) U - 4 0 9 6 = 0$$
where j is the modular invariant.
We assume throughout d is positive ≡ 3 (mod 8) and put ω = 1 2 (3 + √ -d ). Then -σ 24 2 ( ω ) is real and positive. Let φ denote the 3 rd power σ 3 2 ( ω ) up to a suitable 16 th root of unity such that φ is real and positive. Then φ 8 is an algebraic integer satisfying equation (2 1) with . j = j ( ω ) and by Weber's results φ , 4 φ 2 are algebraic integers of degree 3 h ( -d ). Furthermore Weber's conjecture proved in [1] tells us that indeed √ 2 φ is integral of degree 3 h ( -d ), so it satisfies a cubic equation
$$W ^ { 3 } - 2 a _ { 3 } W ^ { 2 } + 2 b _ { 3 } W - 8 = 0$$
with a , b 3 3 integral of degree h ( -d ). A computation in conjunction with (2 1) shows that . ( a , b 3 3 ) gives rise to an integral point of degree h ( -d ) on K 3 . In particular if h ( -d ) = 1 we get the first statement of Theorem A.
If we argue in the same way with a cubic equation for φ 2
$$T ^ { 3 } - 2 a _ { 2 } T ^ { 2 } + 2 b _ { 2 } T - 8 = 0$$
we get a relation between a , b 2 2 describing a curve K 6 which is birationally isomorphic to the Pell conics
$$\int _ { \cdot } \left ( \frac { k } { 2 } - ( a _ { 2 } + 1 ) \right ) ^ { 2 } - 2 \left ( \frac { a _ { 2 } + 1 } { 2 } \right ) ^ { 2 } = 1$$
where b 2 = k a ( 2 -1) + 2.
In view of (2 2), (2 3) the covering . . K 3 → K 6 can be given in the following explicit form
$$a _ { 2 } = a _ { 3 } ^ { 2 } - b _ { 3 }, \quad 2 b _ { 2 } = b _ { 3 } ^ { 2 } - 8 a _ { 3 }$$
from which may find interesting relations with the Pell sequence.
Remark 2.1. By virtue of (2 4) one sees that . a 3 satisfies the following equation
$$z ^ { 4 } - 2 a _ { 2 } z ^ { 2 } - 8 z + ( a _ { 2 } ^ { 2 } - 2 b _ { 2 } ) = 0$$
Therefore (2 3) is nothing but Euler's resolvent for (2 5). . .
Example 2.2. As noted in the Introduction we illustrate our approach with some values d multiple of 3. By using the Tables from [10] and [9] one sees that the fields Q ( √ -51), Q ( √ -123), Q ( √ -267) give rise to three real quadratic points
$$\Psi _ { 1 } \nu _ { 2 } \dashrightarrow \dashrightarrow \Psi _ { 2 } \nu _ { 3 } \nu _ { 4 } \dashrightarrow \dashrightarrow \dashrightarrow \Psi _ { 3 } \dashrightarrow \Psi _ { 4 } \nu _ { 5 } \dashrightarrow \\ ( - 1, 8 + 2 \sqrt { 1 7 } ), \quad ( 4 + \sqrt { 4 1 }, 4 0 + 6 \sqrt { 4 1 } ), \quad ( - 1 0 - \sqrt { 8 9 }, 3 1 0 + 3 2 \sqrt { 8 9 } ) \\ \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
on K 3 respectively.
3. The curve K s . If we assume in addition 3 /notbar d , then the 8 th powers σ 8 , -σ 8 1 , -σ 8 2 are the roots of
$$V ^ { 3 } - \gamma _ { 2 } V - 1 6 = 0 \\ \vdots \dots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots$$
where γ 2 is the Weber function of Stufe 3 such that γ 3 2 = . j
In the notation of 2. let ϕ be a cubic root of φ that is real and positive. Then as above ϕ 2 , √ 2 ϕ are algebraic integral of degree 3 h ( -d ). From (3 1) and a cubic equation for . ϕ 2
$$Z ^ { 3 } - 2 \alpha _ { 2 } Z ^ { 2 } + 2 \beta _ { 2 } Z - 2 = 0$$
with α , β 2 2 integral of degree h ( -d ) we come to the equation ( K 2 ) of Heegner's curve with x = α 2 , y = β 2 -2 α 2 2 .
Now writing a cubic equation for √ 2 ϕ
$$S ^ { 3 } - 2 \alpha _ { 3 } S ^ { 2 } + 2 \beta _ { 3 } T - 4 = 0$$
with α , β 3 3 integral of degree h ( -d ) one gets a covering K 1 → K 2 considered in [2], [6] which can be given by
$$\alpha _ { 2 } & = \alpha _ { 3 } ^ { 2 } - \beta _ { 3 }, \quad 2 \beta _ { 2 } = \beta _ { 3 } ^ { 2 } - 4 \alpha _ { 3 } \\ \cdot \quad \cdot \quad \ e \quad \alpha \varepsilon \cdot \quad \ e \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{} \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{ } \text{$$
Hence the defining equation for K 1 is as follows
$$( \mathcal { K } _ { 1 } ) \quad \ 8 x ^ { 8 } - 3 2 x ^ { 6 } y + 4 0 x ^ { 4 } y ^ { 2 } + 3 2 x ^ { 5 } - 1 6 x ^ { 2 } y ^ { 3 } - 6 4 x ^ { 3 } y + y ^ { 4 } + 2 4 x y ^ { 2 } + 2 4 x ^ { 2 } - 8 y = 0$$
Despite a similarity with equation ( K 3 ) this is a curve of genus 9.
Proposition 3.1. The curve K 1 has exactly 6 integral points (0 , 0) , (1 , 2) , ( -1 0) , , (0 , 2) , ( -1 2) , , (2 , 6) which are in 1 - √ 1 correspondence with class number one fields Q ( √ -3) , Q ( √ -11) , Q ( √ -19) , Q ( √ -43) , Q ( -67) , Q ( √ -163) .
The proof follows immediately from (3 4) and [2]. . The fact that these integral points exhaust all rational points on K 1 follows from Heegner-Stark results ([2], [6]) on rational points on curve K s
$$i ^ { 2 } + 4 z + 1 )$$
$$w ^ { 2 } & = 2 z ( z ^ { 4 } + 4 z ^ { 3 } - 2 z ^ { 2 } + 4 z + 1 )$$
The covering K 1 → K s is given by
$$z = \frac { \beta _ { 3 } } { \alpha _ { 3 } ^ { 2 } } - 1 ; \ \ w = 4 ( z - 2 ) y - 2 ( 3 z ^ { 2 } - 2 z - 1 ), \ \ y = \frac { 1 } { \alpha _ { 3 } ^ { 3 } }$$
It should be noted that Heegner used a little different normalization ( cf. [2], formulae (151) -(152)). Furthermore in order to enumerate all rational points of K s one may apply the argument of [2] in solving equation ( K 2 ) in integers. A precise statement is as follows.
Theorem 3.2. The only rational points on the curve K s are (0 , 0) , ( ± 4 1) , , ( ± 4 , -1) , ( ± 7 4 , 1 2 ) , ( ± 14 2) , .
Theorem 3.3. The curves K 3 and K s are birationally isomorphic.
Since K 3 is hyperelliptic of genus 2 one may use van Hoeij's algorithm ([8]) and a computing program ( e.g. Maple) to write down its Weierstrass normal form which turns out to be the equation ( K s ). The corresponding birational isomorphism K s → K 3 is given by
$$x = - \frac { z ^ { 4 } + 8 z ^ { 3 } + 2 w z + 1 8 z ^ { 2 } + 6 w - 3 } { z ^ { 4 } + 4 z ^ { 3 } - 2 z ^ { 2 } - 1 2 z + 1 }$$
where
$$x = - \frac { z ^ { 4 } + 8 z ^ { 3 } + 2 w z + 1 8 z ^ { 2 } + 6 w - 3 } { z ^ { 4 } + 4 z ^ { 3 } - 2 z ^ { 2 } - 1 2 z + 1 } \\ y = \frac { 2 \ P _ { 8 } ( z, w ) } { ( z ^ { 4 } + 4 z ^ { 3 } - 2 z ^ { 2 } - 1 2 z + 1 ) ^ { 2 } } \\ \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \colon \\ \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \\ \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots$$
$$P _ { 8 } ( z, w ) & = ( 2 z ^ { 5 } + 1 0 z ^ { 4 } + 3 6 z ^ { 3 } + 6 8 z ^ { 2 } + 1 0 z - 3 0 ) w \\ & \quad - z ^ { 8 } + 6 0 z ^ { 6 } + 1 9 2 z ^ { 5 } + 8 2 z ^ { 4 } - 1 2 8 z ^ { 3 } + 1 7 2 z ^ { 2 } + 6 4 z + 7. \\. & \quad.. & \quad.. & \quad..$$
Theorem A now follows from Theorems 3.2 and 3.3.
We note that the inverse birational isomorphism K 3 → K s has more complicated formulae for ( z, w ). For completeness we include them here in the following form
$$z = 1 - \frac { 4 x ^ { 3 } - 4 x y - y ^ { 2 } + 4 x + 4 } { 2 x ^ { 4 } + 2 x ^ { 3 } - 3 x ^ { 2 } u - 2 x u + 6 x - u + 2 }$$
where
$$& z = 1 - \frac { 4 x ^ { 3 } - 4 x y - y ^ { 2 } + 4 x + 4 } { 2 x ^ { 4 } + 2 x ^ { 3 } - 3 x ^ { 2 } y - 2 x y + 6 x - y + 2 } \\ w = - \frac { 2 \ P _ { 1 2 } ( x, y ) } { ( x - 1 ) ( x ^ { 2 } + 1 ) ( x ^ { 2 } - 2 x - 1 ) ( x ^ { 2 } + 2 x + 3 ) ( x + 1 ) ^ { 5 } }$$
P 12 ( x, y ) = 4 x 12 +252 x 11 -24 x 10 y +156 x 10 -622 x y 9 +15 x y 8 2 +440 x 9 -514 x y 8 + +322 x y 7 2 -x y 6 3 +1256 x 8 -708 x y 7 +288 x y 6 2 -21 x y 5 3 +1536 x 7 -620 x y 6 +310 x y 5 2 -19 x y 4 3 +1344 x 6 -716 x y 5 +64 x y 4 2 -20 x y 3 3 +440 x 5 -640 x y 4 +22 x y 3 2 -7 x y 2 3 -12 x 4 -316 x y 3 -8 x y 2 2 -3 xy 3 -124 x 3 -140 x y 2 -6 xy 2 -y 3 -92 x 2 -22 xy + y 2 -16 x +2 y.
Remark 3.4. There is a map K 2 → K 6 ([2]) given by
$$a _ { 2 } = 4 \alpha _ { 2 } ^ { 3 } - 6 \alpha _ { 2 } \beta _ { 2 } + 3 ; \ \ b _ { 2 } = 4 \beta _ { 2 } ^ { 3 } - 1 2 \alpha _ { 2 } \beta _ { 2 } + 6$$
and similarly a map K 1 → K 3
$$a _ { 3 } = 2 \alpha _ { 3 } ^ { 3 } - 3 \alpha _ { 3 } \beta _ { 3 } + 3 ; \ \ b _ { 3 } = \beta _ { 3 } ^ { 3 } - 6 \alpha _ { 3 } \beta _ { 3 } + 6$$
which turned out quite useful in the computation process.
## References
- [1] Birch, B. J., Weber's class invariants , Mathematika 16 (1969), 283-294.
- [2] Heegner, K., Diophantische Analysis und Modulfunktionen , Math. Z. 56 , Heft 3 (1952), 227-253.
- [3] Schertz, R., Complex multiplication , Cambridge University Press, 2010.
- [4] Serre, J.-P., Lectures on the Mordell-Weil theorem , Vieweg, 1989.
- [5] Shafarevich, I. R., The tenth discriminant problem , Algebra i Analiz, 25 :4 (2013), 260-277 (in Russian).
- [6] Stark, H. M., Class numbers of complex quadratic fields , Modular Functions of One Variable I, Ed.: Kuyk W., Lect. Notes Math. 320 (1973), 153-174.
- [7] Stark, H. M., The Gauss class-number problems , Clay Math. Proc., 7 (2007), 247-256.
- [8] van Hoeij, M., An algorithm for computing the Weierstrass normal form of hyperelliptic curves , arXiv:math/0203130, www.math.fsu.edu/ ∼ hoeij/algcurves.html
- [9] Watson, G. N., Singular moduli (5), Proc. London Math. Soc., 42 (1937), 377-397.
- [10] Weber, H., Elliptische Functionen und algebraische Zahlen , Braunschweig, 1891; and Lehrbuch der Algebra , 2-te Auflage, Bd 3, Braunschweig, 1908.
IMC Institute & Fpt University Permanent: 8 Ton That Thuyet, My Dinh, Tu Liem, Hanoi, Vietnam e-mail: [email protected], [email protected]
| null |
[
"Viet K. Nguyen"
] |
2014-08-05T15:10:39+00:00
|
2014-11-26T09:37:57+00:00
|
[
"math.AG",
"Primary 11G30, 11R29, 11R11"
] |
A Genus Two Curve Related to the Class Number One Problem
|
We give another solution to the class number one problem by showing that
imaginary quadratic fields $\Q(\sqrt{-d})$ with class number $h(-d)=1$
correspond to integral points on a genus two curve $\mscrK_3$. In fact one can
find all rational points on $\mscrK_3$. The curve $\mscrK_3$ arises naturally
via certain coverings of curves:\ $\mscrK_3\rg\mscrK_6$,\
$\mscrK_1\rg\mscrK_2$\ with $\mscrK_2\colon y^2=2x(x^3-1)$ denoting the Heegner
curve, also in connection with the so-called Heegner-Stark covering
$\mscrK_1\rg\mscrK_s$.
|
1408.0997v1
|
## Brane Tilings as On-shell Diagrams
Alvaro M. Alhambra - Supervisor: Yang-Hui He Department of Physics, University of Oxford
August 6, 2014
## Abstract
A new way of computing scattering amplitudes in a certain very important QFT ( N = 4 SYM) has recently been developed, in which an algebraic structure called the positive Grassmannian plays a very important role. The mathematics of the positive Grassmannian involve, among other things, bipartite graphs, which also appear in the formulation of a certain class of conformal field theories that are currently being generalized into Bipartite Field Theories (BFT). The fact that the same structures appear in two such different realms of physics suggests a deeper connection between the two that is yet to be fully unveiled. Here we explore that potential connection by looking at the graphs of a certain class of BFTs, the brane tilings, in terms of the new mathematics developed for the computation of the amplitudes. This way we produce a set of data that will hopefully be useful in the development of that connection.
## 1 Introduction
In particle physics, the most important processes we are usually able to compute are scattering ones, in which a set of incoming free particles (meaning particles with definite momentum, represented by eigenstates of a quadratic Hamiltonian, with no interaction terms) interact in a region of spacetime and form a possibly different set of outgoing free particles. Understanding these amounts to calculating the so-called S -matrix, which maps the initial and final states by giving the amplitudes of the transitions
$$\mathcal { A } ( i \rightarrow f ) = \langle f | \, S \, | i \rangle \,. \quad \quad ( 1 ) \quad ^ { 1 } I$$
There is a very well-understood procedure for calculating these amplitudes, with origin in the work of Feynman, Schwinger, Dyson and others during the 60s, based on the famous Feynman diagrams , which is now part of any introductory course in QFT [1]. It enables the computation of the amplitudes to all orders in perturbation theory, although in practice only the first few leading orders are calculated.
Still, this method, although originally aimed towards simplifying the calculations, does get very complicated quickly for an increasing number of particles, even if the final answers are usually very simple. This fact suggests that there may be more straightforward ways of computing the perturbative S matrix, in which aspects not obvious in the standard description are manifest, and viceversa. It is even thought that such method will bring a different way of looking at QFT that might shed some light into some of the big open problems in physics, by creating an alternative method of computing local physical observables, the amplitudes [2-4].
A research program for finding such a potential new formulation has been very actively developed in the past decade with very rapid success. A new and 'very unexpected' 1 picture [3,5] has very recently been developed for the scattering amplitudes of a very important gauge theory, N = 4 Super Yang-Mills, widely present in the context of string theory, and in particular in the AdS/CFT correspondence [6-9], and it is hoped to be generalized to more complicated theories in the future. Remarkably, this new way of computing amplitudes moves its main focus away from locality and unitarity, ubiquitous in the Feyman diagram formalism and quantum mechan-
1 In the words of Edward Witten.
ics in general, in exchange for making a different symmetry of the S matrix manifest, the so-called Yangian symmetry. A brief account of how this line of work has progressed in the past times is given in Section 2.
A key in the development of the understanding of this new formulation was finding that a certain algebraic object plays a fundamental role in this story, the positive Grassmannian [5]. This object was being studied by mathematicians [10-12] almost at the same time as physicists realized of its importance, making it a wonderful instance of an unexpected encounter between physics and mathematics. Section 3 this report will be devoted to explaining some of these new mathematics and the way in which they are relevant to physics.
The mathematics of the positive Grassmannian naturally involve bipartite graphs , which are a kind of graphs with black and white vertices which are only joined to those of the opposite color. It turns out that these objects also play a crucial role in the classification of some field theories with certain symmetry properties, the bipartite field theories (BTFs) [13] that, as much as the already mentioned N = 4 SYM does, make an appearance in the context of string theory and the AdS/CFT correspondence [14]. These are explained in Section 4.
Here, in Section 5, and with the help of a software package specific for working with positive geometry and the positive Grassmannian [15], we compute the list of bipartite graphs on a 2dimensional torus that produce consistent BFTs (so-called brane tilings ) [16] in terms of the mathematics of the positive Grassmannian (the full list is shown in Appendix F). The aim of this is to help develop a link between scattering amplitudes and BFTs that will hopefully be expanded and better understood in the future. Some effort is currently being made in sheding some light in this so far speculative connection [13,17,18], and the work shown in this report intends to be a small contribution in this direction.
## 2 Scattering amplitudes
The details of N = 4 SYM are beyond the scope of this report, but it is worth giving some details of the amplitudes of the theory. The sort of amplitudes we will be talking about are contributions to some order of the so-called T matrix 〈 f | T i | 〉 = 〈 f | S -1 | i 〉 in N = 4 SYM, which is a supersymmetric SU N ( ) gauge theory of massless particles. The leading-order amplitudes will be the tree-level amplitudes, while higher order ones will be loop amplitudes. In particular, the objects to calculate are the color-stripped amplitudes A n (1 , ..., n ) of n outgoing particles (given QFT's crossing symmetry we can make them all outgoing) defined as
$$\inf _ { \text{math-} \\ \mathcal { e d t o } } } \mathcal { A } _ { n } = \sum _ { \sigma } \text{Tr} ( T _ { \sigma ( 1 ) } \dots T _ { \sigma ( n ) } ) \mathcal { A } ( \sigma ( 1 ), \dots, \sigma ( n ) ), \ ( 2 )$$
where T i are the generators of the SU N ( ) algebra corresponding to certain colors of the particles and σ are the n -permutations with different cyclic ordering (so that the sum includes all different permutations up to cyclic rotations - note that this is the way to sum over all possible traces of n generators). Because we are in a supersymmetric theory it is more natural to talk about superamplitudes , in which the helicities of the outgoing particles (excitations of superfields ) do not need to be specified during calculations [19]. Instead, our amplitude will contain δ -functions in the so-called ˜ η I Grassmann variables (despite the name, these are not related to Grassmannians) which we can later integrate out, and depending in they way in which we decide to do this we obtain amplitudes corresponding to different combinations of helicities. For example, from the same superamplitude we can obtain the amplitude of scattering of two + 1 2 fermions and two +1 gluons or the amplitude of three +1 gluons and a scalar. We give more details about supersymmetry and superamplitudes in N = 4 in Appendix A.
## 2.1 Towards the positive Grassmannian
Some great progress was already being made in amplitudes from the early 90s, mainly by Bern, Dixon, Kosower and others [20, 21], but what triggered this particular line of study of amplitudes, in terms of twistors and recursion relations, was Witten's 2003 paper [22]. There, he noted that the scattering amplitudes of N = 4 SYM are dual to solutions of a certain string theory when the kinematical data is represented
in twistor space [23, 24]. This inspired a further finding of crucial importance: the BCFW recursion relations [25, 26]. In 2005, Britto, Cachazo, Feng and Witten himself found a recursive way of building scattering amplitudes in gauge theories at tree level (in processes with no loops, or internal particles over which we need to integrate) from a sum of combinations of lower-point amplitudes (BCFW terms, these findings are explained further in Appendix B). Soon after, the language of the BCFW was translated into to twistor diagrams [27], forecasting how important twistor geometry and twistor diagrams (later to be identified as on-shell diagrams) would be.
In the BCFW construction of amplitudes, the various terms added are separately non-local (due to some so-called 'spurious poles'), but when adding them these poles magically cancel into some manifestly local amplitude, and hence we see the first sign of locality being an emergent feature, as opposed to a manifest one as it occurs in the Feynman diagrams language.
The next breakthrough came during 2007, when a new remarkable symmetry [28, 29] was found for amplitudes of N = 4 in the planar limit (which we explain in Section 3.1). This dual superconformal symmetry [30] was seen to merge with another well-known symmetry of the theory, the superconformal symmetry, into an infinitedimensional symmetry (meaning an infinite number of generators can be built out of the generators of the two symmetries together) named a Yangian symmetry [31], which is better understood in twistor space [32]. Remarkably, the BCFW terms that make up an amplitude were found to be separately Yangian invariant.
Meanwhile, an idea was conjectured that involved calculating amplitudes using the Grassmannian for the first time (not yet the positive Grassmannian) [33, 34]. It was found that the residues of an integrand using the Grassmannian naturally yields terms that are invariant under Yangian symmetry [35, 36], the way BCFW terms are. In particular, this integrand generates the so-called leading singularities , which are algebraic functions of external momenta with the right symmetry properties that, when grouped together in certain ways, give different representations of scattering amplitudes (such as, for in- stance, the BCFW representation) [37]. Still, this was far from well understood, as it was not known how it should produce the right combination of leading singularities.
It was soon found [38] that, when writing the right BCFW terms of certain amplitudes with a new set of kinematical variables (the momentum twistor variables), and in such a way that we get a local answer, the result looks like a polytope 2 in that twistor space, hinting towards a possible geometrical understanding of the Grassmannian integral [39]. This new space of variables allowed for the natural generalization of BCFW relations from tree-level to all-loop orders [40, 41], from which it was found that the integrand for any amplitude could be written in terms of a set of variables α i such that they appear in the integration measure as
$$\prod _ { i } d ( \log ( \alpha _ { i } ) ) = \prod _ { i } \frac { d \alpha _ { i } } { \alpha _ { i } }. \quad \quad ( 3 )$$
This important integration measure and some work preceding it is further explained in Appendix B. Hence, what was known by 2010/11 was that the amplitudes could be thought of as a sum of residues of an integrand over that d(log( α i )) form that generates leading singularities with the right symmetry properties (the most important being the Yangian symmetry), and that the result will probably look like 'some volume of some polytope' made of a sum of such leading singularities that has to yield a local, unitary answer.
The turning point ocurred with the publication of [5] in December of 2012, where the positive Grassmannian comes into scene. Before then there were very good hints as to how the integrand should be written and how the leading singularities produced should be added [37-40], but suddenly physicists realized that what they were really doing was an integration over subsets or cells of the positive Grassmannian , not the Grassmannian itself, and that within the mathematics of the positive Grassmannian there is a well-defined procedure to write the integrand as a function of the already mentioned α i variables. These variables have a natural interpretation in this framework when thinking of the cells in terms
2 A generalization of a polygon to n dimensions.
of permutations and graphs, which represent the way in which small amplitudes are glued together into higher-order ones. We now proceed to define the Grassmannian, its positive stratification into cells and the way in which this mysterious integrand over them is generated through the gluing of simple 3-particle amplitudes.
## 3 The positive Grassmannian and its positroid stratification
The Grassmanian Gr( k, n ) is the set of k -planes through the origin in R n , represented as a k × n matrix where the rows are a basis of the k -plane. This means that the action of the general linear group GL k ( ) has no effect on the particular k -plane, so we can gauge-fix the matrix to have the identity as one of its minors. In this way, any one element of Gr(2 4) can be, for example ,
$$C = \begin{pmatrix} 1 & 0 & c _ { 1 } & c _ { 2 } \\ 0 & 1 & c _ { 3 } & c _ { 4 } \end{pmatrix}, \quad c _ { i } \in \mathbb { R }. \quad ( 4 ) \quad \text{Fol}$$
We shall also define an orthogonal complement C ⊥ (an ( n -k ) × n matrix), such that C C · ⊥ = 0. The positive Grassmannian Gr + ( k, n ) is then simply the subset of the Grassmannian whose elements have all their k × k minors with positive determinant, and, as we discuss in detail in Appendix C, constitutes a generalization of the space of the inside of a polygon. It turns out that it is possible to map the so-called positroid cells of Gr + ( k, n ) (a certain kind of subsets) to permutations and to bipartite graphs (as will be discussed in detail below), and that in such map we produce C matrices representing those cells, which contain some α i variables. These will produce the integrand we are looking for [5], which is
$$\overset { \wedge } { \prod _ { i } } \frac { \text{d} \alpha _ { i } } { \alpha _ { i } } \delta ^ { k \times 4 } ( C \cdot \tilde { \eta } ) \delta ^ { k \times 2 } ( C \cdot \tilde { \lambda } ) \delta ^ { 2 \times ( n - k ) } ( \lambda \cdot C ^ { \perp } ). \quad \text{With } t \\ ( 5 ) \underset { \text{ we can} } { \text{and its} }$$
Here, C is the k × n matrix of a cell (in the amplitude, n will be the total number of particles, and k is the number of negative helicity ones), λ and ˜ λ are 2 × n matrices ( λ ....λ 1 n ) representing, respectively, the left-handed and right-handed spinors that describe the momenta of the outgoing par- ticles (remember that we do not need to specify helicities, the only input are the 4-momenta), and ˜ is a 4 η × n matrix with the superspace variables described in Section 2 and Appendix A. The last of these δ -functions is found via the auxiliary integral
$$\delta ^ { 2 ( n - k ) } ( \lambda \cdot C ^ { \perp } ) = \int \mathrm d ^ { 2 \times k } \rho \, \delta ^ { 2 \times n } ( \rho \cdot C - \lambda ). \ ( 6 )$$
The twistorial representation of momenta is ubiquitous in this line of work, but here for simplicity we write the integrand in terms of the more familiar spinors.
The form (5), once properly written and integrated, and with potential IR and UV divergences taken care of (when possible), will produce any leading singularity, but we first need to i) be able to write C from the corresponding on-shell diagrams (to be explained in 3.1) and ii) know the right contour of integration, which amounts to identifying the leading singularities we want with the right on-shell diagrams.
Following [5], we will proceed to explain how the first of these points is done, involving the mathematics that will later be connected to the bipartite field theories. We will not explain the second point here, which has only been fully developed as late as December 2013 [3, 42]. It is worth mentioning, however, that, at least at tree-level, knowing how to do this entails a geometrical understanding of amplitudes as polytopes in a space where our kinematical variables are, instead of spinors, momentum twistor variables [19]. Such polytopes are the so-called amplituhedrons [3, 42], and their inside is mapped by their corresponding positroid cells (this is explained further in Appendix C).
## 3.1 On-shell diagrams, graphs and permutations
With the BCFW recursion method for tree-level and its generalization to all loops [40] we see that we can express the amplitudes as sums of combinations of smaller amplitudes (the leading singularities), and such combinations are represented by objects called on-shell diagrams . These onshell diagrams will be bicolored graphs and, because we are only dealing with amplitudes in the planar limit, these graphs will in principle be pla-
nar, meaning that they can be drawn on a plane, with no edges intersecting . 3 A detailed description of the correspondence between on-shell diagrams and the full amplitudes is beyond the scope of this report, but for our purposes it will be enough to keep in mind that a tree amplitude is built as a sum of one or more leading singularities generated from certain diagrams, in a way consistent with locality and unitarity, and that loop amplitudes are generally just written in terms of the integrand, as their final form presents a number of additional difficulties.
The smallest amplitudes we can have, and the ones out of which we build our on-shell diagrams, are 3-point amplitudes, which we will represent as trivalent white vertices when the total helicity of the particles is +1 and trivalent black vertices when it is -1, as these are the only two possible sums of helicities that give a non-zero amplitude. They are:
.
And the corresponding amplitudes are
$$A _ { 3 } ^ { ( 1 ) } = \frac { \delta ^ { 1 \times 4 } ( [ 2 3 ] \tilde { \eta } _ { 1 } + [ 3 1 ] \tilde { \eta } _ { 2 } + [ 1 2 ] \tilde { \eta } _ { 3 } ) } { [ 1 2 ] [ 2 3 ] [ 3 1 ] } \delta ^ { 2 \times 2 } ( \lambda \cdot \tilde { \lambda } ) \quad \stackrel { \dots } { t }$$
for the white vertex and
$$A _ { 3 } ^ { ( 2 ) } = \frac { \delta ^ { 2 \times 4 } ( \lambda _ { 1 } \tilde { \eta _ { 1 } } + \lambda _ { 2 } \tilde { \eta _ { 2 } } + \lambda _ { 3 } \tilde { \eta _ { 3 } } ) } { \langle 1 2 \rangle \langle 2 3 \rangle \langle 3 1 \rangle } \delta ^ { 2 \times 2 } ( \lambda \cdot \tilde { \lambda } )$$
for the black vertex. The [ ij ] and 〈 ij 〉 are Lorentz invariant products of, respectively, left handed λ i and right-handed ˜ λ i spinors. It turns out that the form of these amplitudes is fixed by Poincar´ e invariance only [5,19], and will hence be our ideal building blocks. Because we are dealing with superamplitudes, no helicities need to be specified in the external legs of the graph.
To build more complicated on-shell diagrams we need to join many of these vertices together into a graph. For example, the k = 2 4-particle amplitude (denoted as A MHV 4 ) is generated from
3 Outside of the context of on-shell diagrams, the planar limit is also defined as the approximation of amplitudes to first order in 1 /N , with N the number of colors of the theory.
the diagram [5]
.
(7)
It turns out that, in an on-shell diagram, there are certain internal operations that leave its physical information unchanged. These are
- · Merge: we may put two adjacent vertices of the same color together. This is the condition that enables us to make the graphs always bipartite.
- · Square move: we can see that the 4particle amplitude graph is cyclically invariant, which means that it can be rotated with no effect on the final permutation. This is of course still the case if such 4-particle graph is embedded within a bigger one. The rotated graph can then always be made bipartite again via merging.
- · Bubble deletion: in which we eliminate two vertices joined together by two edges. As opposed to the other two, it reduces the number of faces by one, meaning (as we will see later) that the number of α i is also decreased. In any case, this is not a problem in terms of the integrand; a bubble deletion means that a d(log α i ) is factored out of the differential form, and then upon taking a residue around α i = 0 we get the form of the reduced diagram. A graph in which not all possible bubbles have been deleted is a reducible graph, and they are needed in the
There is a way to characterize the set of graphs with the same physical information, which is via a permutation σ a ( ). This is a bijective map of the integers 1 ...n to itselfs which can be represented as { σ (1) , ..., σ ( n ) } . In any on-shell diagram/bicolored graph, we choose any outer node a (from 1 to n ) and start a path in which we turn left in any white node and right in any black node. This path will end in a possibly different outer node from 1 to n , and that number will be σ a ( ). It is guaranteed that if two paths have different starting point they will also have different finishing points, and viceversa, so our permutation is always well-defined for planar bicolored graphs. It is easy to see that none of the operations described above change the associated permutation. An example of the path prescription is shown in Fig.1.
The permutation already tells us the total number of negative helicity particles k of that on-shell diagram, which will correspond to the number of rows in the Grassmannian cell. If we decorate a permutation σ ∗ ( a ) by assigning σ a ( ) = σ ∗ ( a ) + n when σ ∗ ( a ) < a (as shown in the example in Fig. 1), then the number of elements bigger than n in that permutation (the 'decorated' ones) will be k . This way, the map is not from 1 ...n into itself but to 1 .... 2 n . The outer nodes that are decorated in the permutation will be called sources , while the others are sinks . Again, an example is given in Fig.1.
## 3.2 The BCFW bridge
The canonical operation from which we construct our on-shell diagrams from simpler ones is the addition of a BCFW bridge. If we have two adjacent external legs, then such a bridge corresponds to the operation which changes σ b ( ) by an adjacent permutation of the two nodes involved in the bridge (we can also allow the two nodes of the bridge to have some self-identified outer nodes inbetween). We assign a variable α i to the new edge, such that the momenta entering the rest of the graphs through those two legs is changed to
$$\tilde { \text{ferent} } \quad & \lambda _ { a } \to \lambda _ { \hat { a } } = \lambda _ { a } \quad \text{ and } \quad \lambda _ { b } \to \lambda _ { \hat { b } } = \lambda _ { b } + \alpha _ { i } \lambda _ { a } \\ \tilde { \text{inish-} } \quad & \tilde { \lambda } _ { a } \to \tilde { \lambda } _ { \hat { a } } = \tilde { \lambda } _ { a } - \alpha _ { i } \tilde { \lambda } _ { b } \quad \tilde { \lambda } _ { b } \to \tilde { \lambda } _ { \hat { b } } = \tilde { \lambda } _ { b } \\ \text{raphs.} \quad & \tilde { \eta } _ { a } \to \tilde { \eta } _ { \hat { a } } = \tilde { \eta } _ { a } - \alpha _ { i } \tilde { \eta } _ { b } \quad \tilde { \eta } _ { b } \to \tilde { \eta } _ { \hat { b } } = \tilde { \eta } _ { b } \quad \text{ (8)} \\ \infty \, d \infty \quad \pi \, \pi \, \cdots \quad \epsilon \quad \pi \, \pi \, \cdots \quad \epsilon \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \quad \pi \, \cdots \ \end{$$
The addition of a BCFW bridge to the diagram changes the corresponding integrand f 0 to
$$f ( \dots ; \lambda _ { a }, \widetilde { \lambda } _ { a }, \widetilde { \eta } _ { a } ; \lambda _ { b }, \widetilde { \lambda } _ { b }, \widetilde { \eta } _ { b } ; \dots ) \quad \quad ( 9 ) \\ = \frac { d \alpha _ { i } } { \alpha _ { i } } f _ { 0 } ( \dots ; \lambda _ { \hat { a } }, \widetilde { \lambda } _ { \hat { a } }, \widetilde { \eta } _ { \hat { a } } ; \lambda _ { \hat { b } }, \widetilde { \lambda } _ { \hat { b } }, \widetilde { \eta } _ { \hat { b } } ; \dots ). \\ \quad \text{$\mathbb{m} \cdot \dots$} \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdots \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cd.$$
Which, given the form of (5), corresponds to an additional d(log α i ) factor and a change in the columns c i of the matrix C of the form
$$c _ { b } \rightarrow c _ { b } + ( - 1 ) ^ { q } \alpha _ { i } c _ { a } \,. \quad \quad ( 1 0 )$$
Where q in the number of columns c s between the two involved in the bridge such that σ s ( ) = s + n (the number of sources that are self-identified and that are between the two outer nodes involved in the construction of the bridge). We can see that if C ∈ Gr + ( k, n ) before the bridge, the operation (10) leaves this positivity intact as long as α i > 0. For this to occur we need the extra ( -1) q factor.
The fact that the bridge corresponds to an adjacent permutation in σ a ( ) means we can decompose any on-shell diagram into the bridges that built it, by looking at how undoing such adjacent permutations leads to a decoration of the identity { 1 ....n } . There is not a unique way of doing this, but for convenience here we will be
Figure 1: A graph and its path prescription corresponding to the decorated permutation σ = { 3 4 6 5 7 , , , , } , or the undecorated permutation σ ∗ = { 3 4 1 5 2 , , , , } , hence the corresponding C ∈ Gr + (2 , 5). Here 1, 2 and 4 are sinks , and 3 and 5 are sources .
using the lexicographic BCFW decomposition, which is:
Lexicographic BCFW-Bridge Decomposition: Given a permutation σ (which is not a decoration of the identity, in which case the decomposition is already done) we decompose it as ( ac ) ◦ σ ′ where 1 ≤ a < c ≤ n is the lexicographically-first pair 4 (as chosen by our decomposition convention) separated only by self-identified legs b (such that σ b ( ) = b mod( n )) and for which σ a ( ) < σ c ( ), and where σ ′ is the permutation with the adjacent transposition of the ( ac ) pair 'undone'. The process is repeated until σ ′ is a decoration of the identity.
This essentially amounts to an ordered way of decomposing σ a ( ) into a well-defined series of adjacent permutations that lead to the identity. Examples of this are given in Section 3.3 and Appendix E, where we also show the graph construction explicitly.
4 The lexicographically first order is analogous to alphabetical order in which the numbers are the letters of the alphabet. This way, the adjacent decomposition of (23) will precede (34) or (24), but not (12).
A caveat is in order here. We mentioned that the bubble-deletion procedure takes loopintegrands to lower dimensional integrands by the removal of α i , and that this leaves the permutation invariant. This means the permutation characterizes a whole family of non-reduced graphs which have a reduced graph in common, and hence the BCFW construction procedure only produces the reduced graph given by that permutation. The main consequence for the computation of amplitudes is that, currently, loop integrands cannot be written by combinatorial means only (as they require non-reduced integrands). In any case, this will not be important for our means, as the graphs we will be dealing with are found to be always fully reduced.
## 3.3 From the bipartite graph to the positroid cell
We are now ready to explain the full process of how to map an on-shell diagram (a bipartite graph) to a cell of Gr + ( k, n ), along with the example corresponding to graph 2.3b in our list of bipartite field theories (explained in Section 5). The steps are:
- · Identify the permutation σ associated with the graph via the zig-zag paths, and decorate it such that k of its elements are bigger than n . In our case the graph shown below has a decorated permutation σ a ( ) = { 2 5 6 9 7 10 , , , , , } , from an undecorated σ a ( ) = { 2 5 6 3 1 4 , , , , , } .
- · Decompose the permutation via the BCFW procedure to find an ordered path of adjacent permutations. These correspond to operations of the form of (10). The result of this is a decoration of the identity σ ∗ with
k decorated terms, each of which will be labeled by j ∈ { 1 ...k } starting from the left. In our example:
↦
↦
| Pair ( ac ) | 1 ↓ | 2 ↓ | 3 ↓ | 4 ↓ | 5 ↓ | 6 ↓ | BCFW shift |
|-------------------------------------|---------------|---------------|---------------|----------------|---------------|---------------------|-------------------------------------------------------------------------------------------------------------------------|
| (1 2) (1 3) (3 4) (4 5) (1 4) (4 6) | 2 5 6 6 6 7 7 | 5 2 2 2 2 2 2 | 6 6 5 9 9 9 9 | 9 9 9 5 7 6 10 | 7 7 7 7 5 5 5 | 10 10 10 10 10 10 6 | c 2 → c 2 + α 6 c 1 c 3 → c 3 + α 5 c 1 c 3 → c 3 + α 4 c 3 c 5 → c 5 + α 3 c 4 c 4 → c 4 - α 2 c 1 c 6 → c 6 + α 1 c 4 |
↦
↦
Here the permutated pairs are highlighted in blue and the elements of the decorated identity are in gray. Our decoration of the identity is σ ∗ = 7 2 9 10 5 6 . { , , , , , }
- · Given σ ∗ , we write a k × n matrix whose only non-zero entries are ones at the i -th column and j -th row, where i is such that σ ∗ ( ) i = n + i and j is as was just defined. This is
$$\left ( \begin{array} { c c c c c } 1 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 1 & 0 & 0 \end{array} \right ).$$
↦
- · Starting with this matrix, we perform the list of operations of the BCFW decomposition in reverse order, such that our first one would be c 6 → c 6 + α c 1 4 and the last one c 2 → c 2 + α c 6 1 . This yields the final C matrix we use in our integrand (5), which is
↦
$$\left ( \begin{array} { c c c c } 1 & \alpha _ { 6 } & \alpha _ { 5 } & - \alpha _ { 2 } & - \alpha _ { 2 } \alpha _ { 3 } & 0 \\ 0 & 0 & 1 & \alpha _ { 4 } & 0 & 0 \\ 0 & 0 & 0 & 1 & \alpha _ { 3 } & \alpha _ { 1 } \end{array} \right ).$$
It must be said, however, that the relation between the matrix and the permutation is much deeper than this procedure suggests. As we explain in Appendix D, the permutation encodes the mutual dependence of the adjacent columns c i .
From this construction we can see that our integrand will have d α i variables, where d + 1 is the number of faces of the graph. This number d will be the dimension of the cell C . But the δ -functions of the integrand (5) only impose 2 n -4 constraints, given that we have 2 n δ 's minus the 4 we need for conservation of momentum (amplitudes always have a factor of the form δ 4 ( ∑ P µ i )).
↦
↦
This means that our integrand only yields a leading singularity of a tree amplitude in cases where d = 2 n -4. If d < 2 n -4 we have that there are constraints on the external momenta of particles, and if d > 2 n -4 not all of the α i are specified. In the latter case these extra degrees of freedom may also correspond to the loop momenta in the amplitudes, which need to later be integrated over.
Hence, in general, the correct statement is that on-shell diagrams map not to leading singularities but to integrands only, all of which are in any case part of the general picture of positive geometry. The integral from such integrands will then in most cases be a purely formal object, much more work is needed to compute obsevable amplitudes (some of which is explained in, among others, [5, 19, 40, 42]). In our case, when computing our graphs from BFTs we find that only exceptionally do these map to leading singularities.
With this construction method and the rest of the explanations above we have only scratched the surface of the depths of positive geometry and planar bicolored graphs. Appendices C, D, and E expand on these issues, and the more thorough explanation of positive geometry and plabic graphs can be found in [5] and the references therein. In any case, we now have enough tools to compute the bipartite graphs from the BFTs; in the next section we proceed to explain where the ones considered here are taken from.
## 4 Bipartite field theories
Following the AdS/CFT conjecture [7], the quest for conformal field theories with well-defined AdS duals led to the study of quiver gauge theories , with N = 1 [43]. These were found to have some remarkable combinatorial properties [44] that were better undestood in terms of brane tilings , which are bipartite graphs defined on a 2D torus [45,46] (unlike the planar ones we have so far been dealing with). This class is now understood as a subset of the bipartite field theories [13, 47], in which the study of CFTs defined through a graph has been extended to Riemann surfaces beyond a torus [17,18,48]. In these theories, the graph defines the Lagrangian interaction
terms of the field theory by the following rules:
- · Internal faces: each face is associated a SU N ( ) gauge group.
- · Internal edges: each edge is assigned a certain chiral multiplet (a supersymmetric field representing all possible helicities). These will have a local gauge symmetry given by the product of the groups associated to the two faces it separates.
- · Nodes: each node will be a superpotential term in the Lagrangian given by the product of all the fields whose edges have an endpoint in that node. We assign plus signs to white nodes and minus signs to black nodes.
Graphs with boundaries, such as planar ones in a disk, need extra care and have additional procedures to account for the external faces and nodes [17]. The full development of such rules for boundaries will be a key in understanding the relevance of positive geometry (where currently only planar graphs with boundaries are understood) to the general BFTs. In any case, the rules above are all that is needed for a brane tiling to define a conformal field theory.
An algorithm to generate all possible such brane tilings was presented in [16], where a full list of theories with at most 6 superpotential terms is produced (in the language of graphs, a total of 6 nodes). Here we take all the consistent graphs from that list and look at them in terms of the positive geometry explained in Section 3 by constructing their corresponding positroid cells. The full list of brane tilings and their cells is shown in Appendix F. For example, one of the brane tilings directly taken from that work, listed as 2.3, is
with a superpotential given by
$$H _ { I } = - X _ { 1 2 } ^ { 1 }. X _ { 2 3 } ^ { 2 }. X _ { 3 1 } ^ { 1 } + X _ { 1 2 } ^ { 2 }. X _ { 2 3 } ^ { 2 }. X _ { 3 1 } ^ { 2 } \quad ( 1 1 ) \\ + \phi _ { 1 }. X _ { 1 2 } ^ { 1 }. X _ { 2 3 } ^ { 1 }. X _ { 3 1 } ^ { 1 } - \phi _ { 1 }. X _ { 1 2 } ^ { 2 }. X _ { 2 3 } ^ { 1 }. X _ { 3 1 } ^ { 2 }.$$
Where the X and φ are the 4 different superfields (each with its particular gauge symmetries) and each term is given by one of the nodes in the fundamental region. We see that the relevant part of the graph is defined as the fundamental region (the red square), and the toric embedding is represented by writing it within a lattice of such regions. Going to the end of a fundamental region in a 2D torus takes you directly to the opposite end. If we write it as a lattice, instead of going to that opposite end we reach the opposite end of the next fundamental region; the two spaces (lattice and torus) are hence mathematically equivalent.
The main reason why it is believed that the positive Grassmannian framework may also be important here is because the operations described in 3.1 also have a clear meaning in the context of these theories. Namely, a square move links a BFT with its Seiberg dual 5 [49], and the merging of adjacent vertices is related to integrating out mass terms of the Lagrangian (essentially the opposite procedure of a Higgs symmetry breaking). The bridge removal operation also has a natural understanding in the BFTs, as it is briefly described in Appendix C. All these coincidences may indicate that, at least for planar BFTs with boundaries, a form of an associated permutation may be useful in their classification, and that positive geometry may play an important role in their future understanding. Also, the procedures described above for generating Grassmannian cells are currently being extended to non-planar graphs [18, 50], and it is hoped that such new mathematics will generate progress in the BFTs side of the story.
Anatural thing to do from here is hence to map these brane tilings into on-shell diagrams and see how much information can we extract from their positive geometry. This is the subject of the next section.
5 Seiberg duality is a remarkable property of N = 1 theories which makes pairs of them equivalent at low energies.
## 5 The positive geometry of brane tilings
## 5.1 Generation of the diagrams from the tilings
The present difficulty in mapping brane tilings and on-shell diagrams is that while the former (at least, the ones considered here) are defined on a torus, the relevant graphs in the context of the positive Grassmannian are planar and with boundaries. This means there is an ambiguity in translating the fundamental region from ones to others, and the map from brane tilings to corresponding on-shell diagrams is not well-defined. Indeed, on-shell diagrams must at least have a boundary to have any physical meaning at all, which we need to artificially assign to brane tilings. For the example above, one can place the fundamental region such that it contains either
With permutation { 2 4 8 11 6 9 13 15 , , , , , , , } and positroid cell
$$\left ( \begin{array} { c c c c c c c c } 1 & \alpha _ { 7 } & \alpha _ { 6 } & 0 & - \alpha _ { 3 } & - \alpha _ { 3 } \alpha _ { 4 } & 0 & 0 \\ 0 & 0 & 1 & \alpha _ { 5 } & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 1 & \alpha _ { 4 } & \alpha _ { 2 } & 0 \\ 0 & 0 & 0 & 0 & 0 & 1 & \alpha _ { 1 } \end{array} \right ), \quad ( 1 2 ) \quad 5. 2$$
or, by moving the fundamental region in a different way
,
with permutation { 2 5 6 9 7 10 , , , , , } and cell
$$\begin{pmatrix} 1 & \alpha _ { 6 } & \alpha _ { 5 } & - \alpha _ { 2 } & - \alpha _ { 2 } \alpha _ { 3 } & 0 \\ 0 & 0 & 1 & \alpha _ { 4 } & 0 & 0 \\ 0 & 0 & 0 & 1 & \alpha _ { 3 } & \alpha _ { 1 } \end{pmatrix}. \quad ( 1 3 )$$
This is a particularly illustrative example because, even if these two on-shell diagrams come from the same brane tiling, they appear to be completely different. They do not have the same n , nor the same k , nor the same number of dimensions. One caveat we must mention is that there is not a fixed way of assigning the labels { 1 ....n } to the outer edges (there is a cyclic redundancy), but for consistency we will choose them such that the final permutation is the lexicographically minimal.
For this case it is easy to see there are no more possibilities, but, as the graphs complicate, the number of different cells we can find increases significantly. Beyond the method of moving the cell boundary around the graph, spanning all the ways for the nodes to combine within it (as we have shown in this example), we have so far been unable to find an algorithm to produce all the possible ways of making this translation.
In any case, a list with all the consistent graphs in [16] with the different on-shell diagrams it can contain (if not all, at least the ones we have been able to find), and with their associated permutations and cells, has been produced and can be found in Appendix F. Although the list is not exhaustive, we hope the data presented here will help to develop the link between BFTs and scattering amplitudes.
## 5.2 Features of the diagrams
The different graphs from each brane tiling have some common features. It is easy to see from the ways in which the boundary of our diagrams are defined that, if n I is the number of internal edges and n the external ones (also, the number of particles and number of columns of C ), n +2 n I is the same for all diagrams coming from one given tiling. We also have Euler's identity for planar graphs,
$$n _ { F } + n _ { V } = 1 + n _ { I } + n, \text{ \quad \ \ } ( 1 4 )$$
where n V and n F are the number of vertices and of faces, and the fact that the number of faces is
related to the dimension of the cell, ( n F = d +1). Putting them together we get that the number
$$d + n _ { V } + n _ { I } = n + 2 n _ { I } \quad \ \ ( 1 5 ) \ \frac { \text{emou} } { 1 \text{s} \vdots }$$
is also the same for all diagrams generated from the same tiling. All these, by construction, will have the same number of vertices, and hence, the number d + n I is also fixed. This can be easily checked in Appendix F. For instance, in our example 2.3 one graph has
and the other ones has
These similarities, however, are only at the level of the graphs. The permutations represented by the different diagrams do not in general have any apparent common features. The arbitrariness in the generation of the boundary means the zig-zag paths also have arbitrary endpoints, and this is reflected in the fact that their positive geometry is quite different (different structure and dimensions of C ).
As a way through this, one could think of defining such zig-zag paths to live in the 2D torus of the tilings in a way in which they will have no endpoints, just as loops within the torus. It is currently known that such paths are related to the moduli space of the corresponding field theories [13] (a space related to the string theory embedding of such theories, which is in this case a Calabi-Yau 3-fold), but beyond this it is unclear what further lessons can be learnt from them. Along these lines, an idea to explore could be whether these paths on the torus can be mapped to some algebraic structure like planar graphs map to cells in Gr + ( k, n ). This would perhaps require the development of new mathematics beyond the positive Grassmannian.
Another general feature of all the diagrams listed is that they do not in general have enough loops to generate just a leading singularity (meaning that the dimension of the cell d < 2 n -4), so they represent singularities that impose constraints in the external momenta of the particles. For example, the simplest diagram (1.1), given by
represents a leading singularity with k = 2 and n = 4 that, because it over-determines the α i , produces a constraint on external momenta. As mentioned in Section 3.1, to get the actual 4particle tree amplitude one needs the leading singularity generated from
(19)
which has one face (hence one more dimension), meaning 2 n -4 = d = 4. This adds a further difficulty in our identification of brane tilings with loop and tree amplitudes: in general, they are not diagrams relevant to actual physical processes.
Related to this there is a further common feature of brane tilings: all the graphs are fully reduced, in the sense that no bubbles can be deleted from them. This can be seen for our diagrams in Appendix F by checking that in those with more than one loop (graphs with one or no loops are straightforwardly reduced) cannot be made to have any bubbles via square moves and merges. We are able to state this because the BCFW construction procedure outlined here allows only for reduced graphs, and we observe that none of the diagrams listed appear to be any different (in particular, with less loops) than their brane tiling counterparts from [16]. Heuristically, the main reason for this feature is the low number of loops our diagrams have.
## 6 Conclusions and outlook
In the report we showed how to map bipartite graphs to cells of Gr + ( k, n ), and we explained how these cells appear in the integrand of the amplitudes of N = 4 SYM to any loop order. Here, we use this map to associate cells to the fundamental regions of brane tilings, for which we tabulate the possible moves of each fundamental region. We have seen that the on-shell diagrams produced, when coming from the same tiling, have some common features such as the same number of dimensions and internal lines, and that these are generally too low-dimensional to produce a leading singularity. We hope the data in Appendix F will be useful in establishing a meaningful correspondence between BFTs and amplitudes.
In this report we have attempted to explore this connection through the direct identification of theories with on-shell diagrams, but in order to proceed further we first need to understand:
- · i) how to properly generate boundaries for BFTs without it, as it is our case, and how to make physical sense of such a procedure.
- · ii) what does it mean, physically, to have a correspondence between a field theory and a scattering amplitude in a different theory. For example, it is obvious that arbitrarily complicated graphs are of relevance in scattering amplitudes of many-particle collisions [5], but it is less clear of what relevance arbitrarily complicated gauge theories are.
We mentioned that non-reduced graphs are necessary in the loop integrands (indeed, the removed α i variables can represent the loop momenta), but our identification of graphs in terms of permutations makes no distinction between reduced and non-reduced diagrams (given that bubble deletion leaves it invariant). This shows there is still work to do in the classification of non-reduced diagrams, as this limitation currently means the all-loop integrand cannot yet be determined by purely combinatorial methods [5]. In connection to BFTs, it would be interesting to see whether more complicated theories than the ones listed here, potentially with non-reduced graphs, may find the combinatorial description of non-reduced graphs useful [13].
The latest developments have led to a purely geometrical understanding of amplitudes in planar N = 4 SYM [3,42]. There are a number of ways to keep advancing in this direction. We have already mentioned the extension of this framework beyond the planar limit, for which, motivated by BFTs, some small progress has been made [18], despite the fact that the non-planar sector does not have the very important dual superconformal symmetry. An indication on how this may proceed is the fact that it has been shown that toric geometry is a very natural tool to study the topology of the positive Grassmannian [11, 18, 50], as the positroid cells can be mapped to certain toric varieties. Further lines of work, as suggested in [5], may aim to implement positive geometry in amplitudes beyond N = 4. In particular, amplitudes of theories with less to no supersymmetry can be represented by on-shell diagrams in which we define a perfect ordering , where the edges are assigned a direction representing the flow of helicities between the 3-point amplitudes of the diagram.
In the end, the aim is to obtain a general way of computing scattering amplitudes that makes no reference to locality or unitarity, but makes other symmetries manifest (in the present case, the Yangian symmetry). The ultimate motivation of this program [2,3] is that we may be able to calculate local observables (amplitudes) without making any reference to space-time, emergent through the final structure of such observables. Such a dual formulation would make no reference to concepts like Hilbert space, Lagrangians, path integrals,...and may be a hint towards a deeper theory of nature in a way similar to how the variational principle formulation of classical mechanics is closer to path-integral quantum mechanics than Newton's laws are [3,4].
The BFTs may play an important role in this story. The fact that the the same combinatorics appear in both places through bipartite graphs, and, specially, the fact that BFTs have a degree of SUSY ( N = 1) that may be described through on-shell diagrams [5] make them candidates for being next in the line for their amplitudes to be understood geometrically. However,
much progress is needed in elucidating how realistic this possibility is.
In any case, one does not have to go that far into the connection. It may be that there is not a strong physical link between amplitudes in N = 4 SYM and BFTs, but the new mathematics brought up by the study of those amplitudes, starting with the positive Grassmannian, may still be relevant in the analysis and classification of the BFTs. This is the line we mainly want to push with this work, where we explored a possible way of identifying brane tilings and cells of the Grassmannian.
The main difficulty of the analysis here is that the graphs relevant for scattering amplitudes always have a boundary, while the brane tilings do not. This is one of the motivations that has led to the development of the more general BFTs [13,47] (for graphs with boundaries), but the opposite line of development, the mathematics for graphs with no boundaries to help the understanding of brane tilings, are yet to be envisioned (were they possible at all). As mentioned in the previous section, one could hope, for example, that some algebraic structure similar to the positive Grassmannian may be linked with these, the way in which the positive Grassmannian links to BFTs on a disk. The coincidences explained at the end of Section 4, that square moves, merges, and edge deletion, have a meaning in both realms, may by themselves justify the pursuit of such aims.
There is still a lot of work to do if we fully want to unveil the role that BFTs play in the amplitudes research program and viceversa, but there are enough reasons to believe that insightful results will come with it.
## A Supersymmetry and superamplitudes
For supersymmetric theories with maximal supersymmetry, we can define a superspace with coordinates ˜ η A with the A going from 1 to N (4 in this case). These variables are said to be fermionic as opposed to standard bosonic spacetime coordinates because the generators of the Lie group follow anti-commutation relations instead of commutation ones. If we label the gen- erators Q A and their Hermitian conjugates ˜ Q A , these are
$$\{ Q _ { \alpha } ^ { A }, \tilde { Q } _ { \dot { \beta } B } \} = 2 \sigma _ { \alpha \dot { \beta } } ^ { \mu } P _ { \mu } \delta _ { B } ^ { A }. \quad \ \ ( 2 0 )$$
Where σ µ αβ ˙ are Pauli matrices, P µ is an associated 4-momentum and δ A B is a Kronecker delta. Given this, the superspace transformations act on states as U = exp( ˜ iη I Q I ), and supersymmetric Lagrangians are invariant under these transformations. The effect of the SUSY generators on the different fields and one-particle states is that of transforming them into a lower-helicity oneparticle state. Given that we have 4 of them in N = 4, if we start with a fundamental positive helicity gluon g + denoted as | Ω , then our theory 〉 will have four gluinos λ A of helicity (+1 / 2) (for each of the A indices), 6 scalars S AB (for each combination AB ), four ( -1 2) gluinos / λ ABC and 1 negative helicity gluon g -.
These are hence all the possible particles that can arise in our scattering processes. It turns out that SUSY enables us to write all different 16 states into a single chiral superfield as an expansion of the fermionic variables ˜ η A . This is
$$\int _ { \substack { \Omega = & g ^ { + } + \tilde { \eta } _ { A } \lambda ^ { A } - \frac { 1 } { 2! } \tilde { \eta } _ { A } \tilde { \eta } _ { B } S ^ { A B } \\ \intertext { g e } } \left ( \begin{array} { c } \Omega = & g ^ { + } + \tilde { \eta } _ { A } \lambda ^ { A } - \frac { 1 } { 2! } \tilde { \eta } _ { A } \tilde { \eta } _ { B } S ^ { A B } \\ & \left ( \end{array} \right ) \right ) \right ) \int _ { \substack { \Omega = & g ^ { + } + \tilde { \eta } _ { 1 } \tilde { \eta } _ { 2 } \tilde { \eta } _ { 3 } \tilde { \eta } _ { 4 } g ^ { - } }. \quad ( 2 1 ) \\ \text{by the one or the amount to close that will be out on one} \end{array} \right )$$
These are the superparticles that will be outgoing in our amplitudes, and the fact that we can write it like this means we do not have to specify helicities when writing our on-shell diagrams. These ˜ η A variables then show up for the amplitudes in δ -functions as shown in (5), and the number of these delta functions (4 × k ) determines the total number of negative helicity states k , which is also the number of rows in the Grassmannian. For example, k = 2 corresponds to the lowest non-zero amplitude, the MHV or maximally helicity violating . From that amplitude we can extract the amplitude with all g + and 2 g -, but we can also extract the amplitude for all g + except 4 scalars, or any other combination, provided that the total sum of the helicities is the same. The way this proceeds is via an integral. If we have the n -particle MHV amplitude A MHV n ( λ, λ, η ˜ ˜), then, for example, to get the amplitude where particles i and j are negative helicity gluons we integrate
the following way glyph[negationslash]
$$\dots \dots \dots \dots \circ \dots \nu \\ \mathcal { A } ( g _ { 1 } ^ { + } \dots g _ { i } ^ { - } \dots g _ { j } ^ { - } \dots g _ { n } ^ { + } ) & = & ( 2 2 ) \\ \int \prod _ { a = 1, \, a \neq i, j } ^ { n } \mathrm d ^ { 4 } \tilde { \eta } _ { a } \tilde { \eta } _ { a } ^ { 4 } \int \mathrm d ^ { 4 } \tilde { \eta } _ { i } \mathrm d ^ { 4 } \tilde { \eta } _ { j } \mathcal { A } _ { n } ^ { M H V } \\. \quad \. \quad \. \quad \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \ \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \ }$$
where the d 4 ˜ η a are the differentials over the 4dimensional superspace of each of the n particles. Any other combination of helicities will then include a different product of ˜ η a in the integrand.
## B Recursion relations, the treelevel BCFW scheme and the loop integrand
As mentioned in Section 2, a number of methods have been developed that help compute amplitudes from lower point ones. In these on-shell approaches, the idea is to explote the analytic properties of the amplitudes with the external momenta deformed and made complex. The most famous of them are the BCFW recursion relations at tree level [25,26], which we proceed to outline, following [19]. The deformation of momenta entails the introduction of a set of n complex-valued vectors r µ i added to the external momenta
$$\hat { p } _ { i } ^ { \mu } \equiv p _ { i } ^ { \mu } + z r _ { i } ^ { \mu }, \quad \text{ \quad \ \ } ( 2 3 ) \ \text{ she} \.$$
where z ∈ C and r µ i is a set of complex-valued vectors such that
- · ∑ n i =1 r µ i = 0
- · r i · r j = 0 for all i, j , including i = j
- · p i · r i = 0 for each i
Given this definition, this set has three important properties
- 1. Momentum conservation still holds: ∑ n i =1 ˆ p µ i = 0.
- 2. Shifted momenta are on-shell: ˆ p 2 i = 0.
- 3. If we define a non-trivial subset of momenta { p µ i } i ∈ I , and P µ I = ∑ i ∈ I p µ i then (defining R I = ∑ I r µ i )
$$\hat { P } _ { I } ^ { 2 } = P _ { I } ^ { 2 } + z 2 P _ { I } \cdot R _ { I } \quad \ \ ( 2 4 ) \ \text{whit}$$
$$\text{And hence, if } z _ { I } = - \frac { P _ { I } ^ { 2 } } { 2 P _ { I } \cdot R _ { I } } \\ \hat { P } _ { I } ^ { 2 } = - \frac { P _ { I } ^ { 2 } } { z _ { I } } ( z - z _ { I } ).$$
$$( 2 5 )$$
The claim now is that the deformed amplitude will have a set of poles at some z I , and that the residues at those poles can be written from lowerpoint amplitudes. If we consider the Feyman diagrams of the amplitude, the only place where such poles can appear is in the internal shifted propagators, which by conservation of momentum must be of the form 1 /P ˆ 2 I for some set I , and, from (25), these must be simple poles. It can be seen from the definition that the residue of that pole z I corresponds to that propagator going on-shell.
If we then consider the function A n ( z /z ) , then by Cauchy's theorem (the sum of the residues must be zero) we have that
$$\mathcal { A } _ { n } = - \sum _ { z _ { I } }$$
where A n is both the residue at z = 0 and the tree amplitude, the sum is over all different poles, and B n is the pole at z →∞ (which will in most cases be zero). From our explanation above, the residues at z I correspond to progators going onshell, meaning that the amplitude factorizes such that
$$\text{Res} _ { z = z _ { I } } \frac { A _ { n$$
And finally the whole amplitude can be expressed as the sum of such poles, made out of lower-point amplitudes ( A L and A R ) with one of its external particles having momentum P I
$$\text{or-} \quad \begin{array} { c } \mu \\ \mathcal$$
The BCFW recursion relations are a type of recursion relation in which the deformation goes through only two adjacent external legs in such a way that the spinor associated to two external particles are deformed by (exactly like in (8))
$$\underset { \tilde { \lambda } } { \text{rem} } \, \$$
which can be proved to be equivalent to a defor-
mation such as that given in (23). But this is only one method, there are alternative recursive procedures with different deformation prescriptions, such as the CSW construction [51]. It turns out that the different methods give different decompositions of the tree amplitudes which, in the language of the tree amplituhedron, correspond to different ways of triangulating it, to later add up the volume of all those triangles together [3]. The integrand described in this work (5) has a singularity structure such that the residues of the α i match the terms in the BCFW expansion (28).
For specific details and examples of how this works in practice we refer the reader to [19, 52], but in general this is enough to see how amplitudes are built from smaller ones via recursion relations, an idea that has culminated in the recursion relations for general planar on-shell diagrams outlined in Section 3.
We see that what we just described only works at tree level, where the poles are simple and well defined, and there is no room for either IR or UV divergences. This gets much more complicated for loop amplitudes (essentially, higher order terms of the perturbative S -matrix), for which some techniques started to be developed in the early 90's with the work of Bern, Dixon, Kosower and others [20, 21, 53, 54]. These generalized unitarity methods showed, among other things, that the analytic structure (the poles of the integrand associated to it) of loop amplitudes is much more rich and complex than at tree level. In particular, as mentioned in [5], the 1-loop amplitude for n = 4 is related to the corresponding tree amplitudes by (as seen first in [6])
$$\begin{array} {$$
where l is the 4-momentum of the loop and p i the 4-momenta of the particles. We can see that this is IR divergent when integrated (given the amount and degree of the poles), and hence the focus is placed on the integrand rather than the integral itself (as it is done in general for loop amplitudes in N = 4). The merit of the work [40] is that it was found that, in general, such loop integrands can be described as a (log( . α i )) if we make a certain change of variables (which is motivated by the description of kinematical data in momentum twistor space, following [27, 38]) and that, together with the analytic structure of the amplitudes at tree level, any amplitude at any loop-order has a corresponding integrand with a singularity structure given by some product of d (log( α i )). The final integrand is, of course, over the cells of the positive Grassmannian, as given by (5). In this particular case, the loop part of the integrand (30) can be written as
$$\mathfrak { \alpha } _ { i } \quad & \text{ the integrand $(30)$ can be written as} \\ \mathfrak { \cdot } \alpha _ { i } \quad & \int \frac { d ^ { 4 } l ( p _ { 1 } ^ { \mu } + p _ { 2 } ^ { \mu } ) ^ { 2 } ( p _ { 1 } ^ { \mu } + p _ { 3 } ^ { \mu } ) ^ { 2 } } { l ^ { 2 } ( l + p _ { 1 } ^ { \mu } ) ^ { 2 } ( l + p _ { 1 } ^ { \mu } + p _ { 2 } ^ { \mu } ) ^ { 2 } ( l - p _ { 4 } ^ { \mu } ) ^ { 2 } } = \\ \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \cdfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfkappa \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \kappa \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \frak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfj \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \vec { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \dag } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot }
\mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \chfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathf ] } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \\ \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot }
\mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathiota } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \cdot } \mathfrak { \$$
where l ∗ is defined as one of the two points nullseparated from all the external momenta (such that ( l ∗ -p µ i ) 2 = 0. This, in the context of the positive Grassmannian, is written as the simple form
$$\int \frac { d \alpha _ { 1 } } { \alpha _ { 1 } } \frac { d \alpha _ { 2 } } { \alpha _ { 2 } } \frac { d \alpha _ { 3 } } { \alpha _ { 3 } } \frac { d \alpha _ { 4 } } { \alpha _ { 4 } }. \quad \quad ( 3 2 )$$
This finding was one of the first indicators that a simple way of writing the integrand of the loop amplitudes existed, and ultimately led to the positive Grassmannian framework [5] described throughout this work.
## C Positive geometry and the inside of polytopes
If we have 3 masses c i in R 2 such that they form a triangle, the 'centre of mass'
$$\vec { y } = c _ { 1 } \vec { r } _ { 1 } + c _ { 2 } \vec { r } _ { 2 } + c _ { 3 } \vec { r } _ { 3 }, \text{ \quad \ \ } ( 3 3 )$$
will be a point in the inside of that triangle as long as the c i > 0, and by varying these we map the whole inside of the triangle. We can see this means ( c 1 , c 2 , c 3 ) ∈ Gr + (1 3). , We also note that the points where any c i = 0 correspond to the boundaries (the edges) of the triangle.
This can be extended easily to Gr + (1 , n ) if we instead of a triangle have n points forming a polytope, and the vectors r i do not have to be in any particular space for the inside of the polytope to make mathematical sense. Indeed, the physically
relevant space ( momentum twistor space ) will be the projective space CP 3 . It turns out that for the external points to have a well defined inside (meaning that they form a well-defined convex polytope) the matrix of the coordinates of those points must meet a condition of having all its cyclic minors positive [3].
Furthermore, such notion of the inside of a polytope can be generalized if we enlarge our notion of positivity with the positive Grassmannian. We generalize the condition c i > 0 to the condition that the minors of a k × n matrix C need to be positive. Indeed the polytopes whose volume yields the amplitude (the amplituhedrons) have an inside mapped by Gr + ( k, n ). We define the α i variables such that they span the different sectors or cells of Gr + ( k, n ), and, in analogy with the case of a triangle, we have that setting one of these to zero corresponds to moving along the boundary of such a polytope.
The volume of any of these objects can in general be obtained from an integrand with logarithmic singularities at its α i = 0 boundaries [3,19], which means that an integrand such as (5) may indeed yield a volume that in the right cases will correspond to a scattering amplitude.
For the case of scattering amplitudes at tree level, this means that, given an integrand (5) with a certain n and k of dimension 2 n -3, its various possible residues give 2 n -4-dimensional forms that can be integrated into leading singularities, and the right combination of such leading singularities (such that the sum is local and unitary) gives an amplitude A k n , which can be seen as a volume in momentum twistor space.
This boundary operation amounts to removing an edge in the graph (note that not all edges can be removed this way, only those coming from a BCFW decomposition). This operation also has a definite meaning in the context of BFTs, where it is identified as a higgsing . This means that the scalar field of the chiral multiplet represented by the edge (the superfield) acquires a non-zero vacuum expectation value [13].
## D The permutation as the label of the cells
The 'boundary' operation of setting α i = 0 explained in C is the opposite to the addition of a BCFW bridge that we explain in Section 3.2. Indeed, we see that the BCFW operation essentially complicates the dependence between adjacent columns of C . Given the construction procedure and the fact that the permutation includes all the information about a cell, one question we can ask is: how does the permutation encode the mutual dependence between the columns of the C matrix? It turns out that, if we define a rank function
$$r [ a ; b ] = \text{rank} ( c _ { a } \dots.. c _ { b } ), \text{ \quad \ \ } ( 3 4 )$$
Then, σ a ( ) is such that it is the smallest number with the property that
$$r [ a ; \sigma ( a ) ] = r [ a + 1 ; \sigma ( a ) ]. \quad \ \ ( 3 5 )$$
This also implies that r a [ ; σ a ( ) -1] > r a [ + 1; σ a ( ) -1] and that, reciprocally, a is the maximal column such that r a [ ; σ a ( )] = r a [ ; σ a ( ) -1]. A way to easily compute this permutation from C is via writing a table of ranks of the adjacent columns of C . For example, for the permutation
$$\{ 3, 7, 6, 1 0, 9, 8, 1 3, 1 2 \}$$
which has a corresponding matrix (in the lexicographic scheme, as defined in Section 3.2)
$$\begin{array} { c c c c c c c c c } \text{arious} \\ \text{forms} \\ \text{writes}, \\ \text{singu-} \\ \text{itarv} \end{array} \left ( \begin{array} { c c c c c c c c c } 1 & \alpha _ { 9 } & 0 & - \alpha _ { 5 } & - \alpha _ { 5 } \alpha _ { 6 } & 0 & 0 & 0 \\ 0 & 1 & \alpha _ { 8 } & \alpha _ { 7 } & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 1 & \alpha _ { 3 } + \alpha _ { 6 } & \alpha _ { 3 } \alpha _ { 4 } & 0 & - \alpha _ { 1 } \\ 0 & 0 & 0 & 0 & 1 & \alpha _ { 4 } & \alpha _ { 2 } & 0 \end{array} \right )$$
We can write a chart listing the combinations of adjacent vectors of different ranks
| consec. chains of colums | span(rank) |
|----------------------------|--------------|
| (1)(2)(3)(4)(5)(6)(7)(8) | P 0 (1) |
| (123)(34)(45)(56)(678)(81) | P 1 (2) |
| (3456)(56781)(81234) | P 2 (3) |
In rank r -1 there are some sets { c a -1 ...c σ a ( ) -1 } for which there is a corresponding set of rank r of the form { c a ...c σ a ( ) } . These r -1 subsets will be called the maximal planes', and are such that they are the smallest for which addition of one column to the right and to the left has the same effect in terms of rank. These are hence the sets
of columns which we need to read off the permutation. In the table above these maximal planes are highlighted in blue. We have left out many aspects of positive geometry and the combinatorial description of the positive Grassmannian due to space constraints, but the full discussion of these issues can be found in [5] and mathematical literature such as [10,11].
## E A further example of the lexicographic BCFW decomposition scheme
Here we show another example of the decomposition of the permutation, together with a careful description of the building of the graph step by step through BCFW bridges which will hopefully make the procedure clearer.
If we choose the decorated permutation σ = 3 7 6 10 9 8 13 12 , the decomposition, as defined { , , , , , , , } in Section 3.2 is given by
↦
↦
| τ | 1 ↓ | 2 ↓ | 3 ↓ | 4 ↓ | 5 ↓ | 6 ↓ | 7 ↓ | 8 ↓ | BCFW shift |
|-------------------------------------------------------|---------------------|----------------------------|---------------------|-------------------------|-----------------------|---------------------|-----------------------------|------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| (1 2) (2 3) (2 4) (4 5) (1 4) (5 6) (4 5) (5 7) (4 8) | 3 7 7 7 7 9 9 9 9 9 | 7 3 6 10 10 10 10 10 10 10 | 6 6 3 3 3 3 3 3 3 3 | 10 10 10 6 9 7 7 8 8 12 | 9 9 9 9 6 6 8 7 13 13 | 8 8 8 8 8 8 6 6 6 6 | 13 13 13 13 13 13 13 13 7 7 | 12 12 12 12 12 12 12 12 12 8 | c 2 → c 2 + α 9 c 1 c 3 → c 3 + α 8 c 2 c 4 → c 4 + α 7 c 2 c 5 → c 5 + α 6 c 4 c 4 → c 4 - α 5 c 1 c 6 → c 6 + α 4 c 5 c 5 → c 5 + α 3 c 4 c 7 → c 7 + α 2 c 5 c 8 → c 8 - α 1 c 4 |
↦
↦
In this table, the numbers permuted by the bridge are highlighted, and the ones corresponding to a decorated identity are in light grey.
If we construct this graph bridge by bridge via taking this decomposition backwards, the steps are the following (these correspond to all the adjacent permutations in the table above)
.
As we can see, each of these steps entails the creation of a bridge of the form shown in the first graph of Section 3.2. The final cell, created the way shown in Section 3.3, has a matrix C
$$\begin{pmatrix} 1 & \alpha _ { 9 } & 0 & - \alpha _ { 5 } & - \alpha _ { 5 } \alpha _ { 6 } & 0 & 0 & 0 \\ 0 & 1 & \alpha _ { 8 } & \alpha _ { 7 } & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 1 & \alpha _ { 3 } + \alpha _ { 6 } & \alpha _ { 3 } \alpha _ { 4 } & 0 & - \alpha _ { 1 } \\ 0 & 0 & 0 & 0 & 1 & \alpha _ { 4 } & \alpha _ { 2 } & 0 \end{pmatrix}.$$
Another more carefully worked example of how to construct the matrix is given in Section 3.3.
↦
↦
↦
↦
↦
## F The positroid cells of the list of brane tilings
In the next few pages we show the on-shell diagrams that we have taken from the list of brane tilings (bipartite graphs on a 2D torus) from [16]. In that work, all the consistent and inconsistent brane tilings for up to 6 superpotential terms (with 3 white and black vertices) are produced and listed. Here, we show the on-shell diagrams that have been produced from each of those tilings through the procedure explained in Section 5.1, and the corresponding positroid cells calculated as it is explained in Section 3.3.
To facilitate the comparison, the diagrams are listed under the same label as the brane tiling they come from in the original list such that all diagrams coming from tiling 3 2 in [16] are labeled 3 2 . . a , 3 2 . b ,... We show the different cells to which a boundary was assigned, with the dimensions of the cell ( n k , and d ), the permutation of the diagram and the representation of their positroid cells C . There is an arbitrariness as to how to assign the labels 1 .....n to the external particles, but for practical convenience and consistency we always choose the permutation with the lowest lexicographic order (analogous to the alphabetical order, but changing numbers for letters).
The first things we can see is that in general the diagrams do not have enough loops to reach the 2 n -4 dimensions necessary to produce just a leading singularity, and instead they produce a leading singularity with additional constraints on the external momenta. We also see that diagrams coming from the same tiling do have different parameters, complicating the analysis of these in terms of the positive geometry presented here. The rest of the discussion on the features of these graphs can be found in Section 5.2.
| | α 3 α 2 0 0 1 α 1 ) | 1 α 5 α 4 α 3 0 0 0 0 0 1 α 2 0 0 0 0 0 1 α 1 | α 5 0 0 0 0 0 1 α 4 0 0 0 0 0 0 1 α 3 0 0 0 0 0 1 α 2 α 1 | α 5 α 3 α 3 α 4 0 0 0 1 α 4 0 - α 1 0 0 1 α 2 0 | α 6 0 - α 4 0 0 0 1 α 5 0 0 0 0 0 0 1 α 3 0 0 0 0 0 1 α 2 α 1 | α 7 α 6 α 4 α 4 α 5 0 0 0 0 0 1 α 5 0 0 α 1 0 0 0 1 α 3 0 0 0 0 0 0 1 α 2 0 | 0 0 0 0 0 α 5 α 4 0 0 α 1 0 1 α 3 0 0 0 0 1 α 2 0 |
|------------|------------------------------|---------------------------------------------------------|---------------------------------------------------------------------|-------------------------------------------------------|-------------------------------------------------------------------------|---------------------------------------------------------------------------------------|-------------------------------------------------------------|
| Graph/Perm | ( 1 0 | | | | | | |
| | | | 2,4,9,11,7,13,8,14 } | | | { 2,3,8,7,9,13,14,12 } | |
| | | | α 6 0 0 0 | | { 2,4,7,11,9,13,8,14 } | | |
| | { 2,4,5,7 } | { 2,3,6,7,10,11 } | { 1 0 0 0 | { 2, 6, 5, 7, 10, 9 } | | | { 2,4,9,7,8,13,14,11 } α 7 0 0 0 |
| | | | | | α 7 0 0 0 | 1 0 0 0 | α 6 1 0 0 |
| Cell | | | | | | | |
| | | | | 1 0 0 | 1 0 0 0 | | 1 0 0 0 |
| # | (1 . 1 a ) n = 4 k = 2 d = 3 | (1 . 2 a ) n = 6 k = 3 d = 5 | (2 . 1 a ) n = 8 k = 4 | (2 . 1 b ) n = 6 k = 3 d = 5 | (2 . 2 a ) n = 8 k = 4 d = 7 | (2 . 2 b ) n = 8 k = 4 d = 7 | (2 . 2 c ) n = 8 k = 4 d = 7 |
| | | | d = 6 | | | | |
| Cell | 1 α 7 α 6 α 5 0 0 0 0 0 0 0 1 0 - α 2 - α 2 α 3 0 0 0 0 0 1 α 4 0 0 0 0 0 0 0 1 α 3 α 1 | 6 0 - α 3 - α 3 α 4 0 0 α 5 0 0 0 0 0 1 α 4 α 2 0 0 0 0 1 α 1 | 1 α 6 α 5 - α 2 - α 2 α 3 0 0 0 1 α 4 0 0 0 0 0 1 α 3 α 1 | 1 α 9 α 8 α 7 0 0 0 0 0 0 0 0 0 1 α 2 α 2 α 6 0 0 α 2 α 3 0 0 0 0 0 1 α 6 0 0 α 3 α 1 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 1 α 4 0 0 | 1 α 9 α 8 α 5 α 5 α 7 α 5 α 6 0 0 0 0 0 0 0 1 α 7 α 6 0 0 α 2 0 0 0 0 0 0 1 α 4 0 0 0 0 0 0 0 0 0 1 α 3 0 0 0 0 0 0 0 0 0 0 1 α 1 | α 7 - α 3 - α 3 α 5 - α 3 α 4 0 0 1 α 6 0 0 0 0 0 1 α 5 α 4 α 2 0 0 0 0 0 1 α 1 | 0 0 0 0 0 0 0 0 0 α 7 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 1 0 0 0 0 α 2 α 1 0 0 0 1 α 5 0 0 0 0 0 0 0 0 1 α α 0 0 |
|------------|-----------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------|-----------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------|
| Graph/Perm | 2,3,8,9,7,13,12,14 } | { 2, 4, 7, 6, 9, 11 } 1 α 7 α 0 0 1 0 0 0 0 0 0 | | 2,3,10,11,8,9,16,17,14,15 | 2,3,10,5,8,11,16,17,14,19 | 2,5,8,11,6,9,12,15 } 1 α 8 0 0 0 0 0 0 | 2,4,13,15,11,17,9,19,10,20,12,18 } 1 α 9 α 8 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 4 3 |
| # | (2 . 2 d ) n = 8 k = 4 d = 7 { | (2 . 3 a ) n = 8 k = 4 d = 7 | (2 . 3 b ) n = 6 k = 3 | (2 . 4 a ) n = 10 k = 5 d = 9 { | (2 . 5 a ) n = 10 k = 5 d = 9 { | . 5 b ) n = 8 k = 4 d = 8 { | (3 . 1 a ) n = 12 k = 6 d = 9 |
| | | | d = 6 | | | (2 | |
| Cell | 1 α 7 α 4 α 4 α 5 α 4 α 5 α 6 0 0 0 0 0 1 α 5 α 5 α 6 0 0 α 1 0 0 0 1 α 6 0 - α 2 0 0 0 0 0 1 α 3 0 0 | 1 α 6 0 - α 2 - α 2 α 3 0 0 1 α 5 α 4 0 0 0 0 0 1 α 3 α 1 | 1 α 7 α 5 α 5 α 6 0 0 0 0 0 0 1 α 6 α 4 0 - α 2 0 0 0 0 0 1 α 3 0 0 0 0 0 0 0 0 1 α 1 | 1 α 7 0 - α 4 - α 4 α 5 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 1 α 5 α 3 α 2 0 0 0 0 0 0 0 1 α 1 | 1 α 7 α 6 0 0 0 0 0 0 0 1 α 5 0 0 - α 2 0 0 0 0 1 α 4 α 3 0 0 0 0 0 0 0 0 1 α 1 | 1 α 10 α 9 0 - α 7 0 0 0 0 0 0 0 0 0 1 α 8 0 0 0 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 0 α 1 0 0 0 0 0 0 0 1 α 4 α 3 0 0 0 0 0 0 0 0 0 0 0 1 α 2 0 | 1 α 9 α 8 0 0 0 0 0 - α 2 0 0 0 1 0 - α 4 - α 4 α 7 - α 4 α 5 0 0 0 0 0 0 1 α 6 α 6 α 7 0 0 0 0 0 0 0 0 1 α 7 α 5 α 3 0 0 0 0 0 0 0 0 0 0 1 α 1 |
|------------|-------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Graph/Perm | { 2,8,7,6,9,13,12,11 } | { 3,6,5,8,7,10 } | { 2,6,4,9,8,13,11,15 } | { 2,6,5,11,9,8,12,15 } | { 2,5,9,8,6,12,11,15 } | { 2,4,7,15,13,17,12,9,11,20,22,18 } | { 2,8,10,7,6,14,13,15,11,19 } |
| # | (3 . 1 b ) n = 8 k = 4 d = 7 | (3 . 2 a ) n = 6 k = 3 d = 6 | (3 . 2 b ) n = 8 k = 4 d = 7 | (3 . 2 c ) n = 8 k = 4 d = 7 | (3 . 2 d ) n = 8 k = 4 d = 7 | (3 . 4 a ) n = 10 k = 5 d = 9 | (3 . 4 b ) n = 10 k = 5 d = 9 |
| Cell | 1 α 9 α 8 α 5 α 5 α 6 α 5 α 6 α 7 0 0 0 0 0 0 0 1 α 6 α 6 α 7 0 0 0 - α 1 0 0 0 0 1 α 7 0 0 α 2 0 0 0 0 0 0 1 α 4 0 0 0 0 0 0 0 0 0 1 α 3 0 0 | 1 α 9 α 8 α 7 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 α 1 0 0 0 0 0 1 α 4 0 - α 2 0 0 0 0 0 0 0 1 α 3 0 0 1 0 0 | α 9 α 8 0 0 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 α 1 0 0 0 0 1 α 5 0 - α 3 - α 2 0 0 0 0 0 0 1 α 4 0 0 0 | 1 α 10 0 0 0 0 0 0 - α 4 0 α 2 0 0 0 1 α 9 0 0 0 0 0 0 0 0 0 0 0 1 α 7 α 7 α 8 0 0 0 0 0 0 0 0 0 0 1 α 8 α 6 α 5 0 0 0 0 0 0 0 0 0 0 0 0 1 α 3 0 0 0 0 0 0 0 0 0 0 0 0 1 α 1 1 α 9 α 8 α 6 α 6 α 7 0 0 0 0 0 0 0 0 1 α 7 0 0 0 0 - α 1 0 0 0 0 1 0 - α 3 - α 3 α 4 0 0 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 1 α 4 α 2 0 | 1 α 9 α 8 0 0 0 0 0 0 0 0 0 1 α 7 α 5 α 5 α 6 0 0 0 - α 1 0 0 0 0 1 α 6 0 0 α 2 0 0 0 0 0 0 1 α 4 0 0 0 0 0 0 0 0 0 1 α 3 0 0 | 1 α 5 + α 7 α 5 α 6 - α 2 - α 2 α 3 0 0 1 α 6 α 4 0 0 0 0 0 1 α 3 α 1 |
|------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------|
| Graph/Perm | 2,3,10,9,8,11,16,17,15,14 } | 2,3,8,11,14,10,9,17,16,15 } | 2,7,11,13,10,8,16,9,15,14 } 2,10,7,15,6,16,8,17,12,21,13,23 } | 2,3,10,9,11,8,16,15,17,14 } | 2,4,11,9,8,10,16,17,15,13 } | { 3,5,6,8,7,10 } |
| # | (3 . 4 c ) n = 10 k = 5 d = 9 | (3 . 4 d ) n = 10 k = 5 d = 9 (3 . 5 a ) n = 10 k = 5 | d = 9 (3 . 5 b ) n = 12 k = 6 d = 10 | (3 . 5 c ) n = 10 k = 5 d = 9 | (3 . 5 d ) n = 10 k = 5 d = 9 | (3 . 6 a ) n = 6 k = 3 d = 7 |
| Cell | α 8 α 7 0 0 0 0 0 0 1 α 3 + α 6 α 3 α 5 α 3 α 4 0 0 0 0 1 α 5 α 4 α 2 0 0 0 0 0 0 1 α 1 | α 4 + α 6 ( α 4 + α 6 ) α 7 α 4 α 5 0 0 0 1 α 7 α 5 0 - α 2 0 0 0 1 α 3 0 0 0 0 0 0 1 α 1 | α 8 0 0 0 0 0 0 0 1 α 7 0 0 0 α 3 0 - α 1 0 1 α 6 0 0 0 0 0 0 0 1 α 5 α 4 0 0 0 0 0 0 0 0 1 α 2 0 | α 8 α 7 0 - α 3 - α 3 α 4 0 0 0 1 α 6 α 5 0 0 0 0 0 0 1 α 4 α 2 0 0 0 0 0 0 1 α 1 | α 7 α 8 0 0 0 α 3 0 0 α 8 0 - α 5 - α 4 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 1 α 2 0 0 0 0 0 0 1 α 1 | α 7 α 8 0 0 α 4 0 0 0 α 8 0 - α 5 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 1 α 3 0 0 0 0 0 0 1 α 2 α 1 | 0 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 1 α 6 - α 3 - α 3 α 4 0 α 1 0 0 1 α 5 0 0 0 0 0 0 1 α 4 α 2 0 |
|------------|---------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------|
| | 1 0 0 0 | 1 α 8 0 0 0 0 0 0 | 1 α 9 0 0 0 0 0 0 0 0 | 1 0 0 0 | 1 α 9 α 7 0 0 1 0 0 0 0 0 0 0 0 0 | 1 α 9 α 7 0 0 1 0 0 0 0 0 0 0 0 0 | 1 α 9 α 8 0 0 0 0 0 0 0 0 0 0 0 0 |
| Graph/Perm | { 2,5,9,8,6,11,12,15 } | { 2,6,4,8,9,13,11,15 } | | { 2,4,8,6,11,9,13,15 } | { 2,6,5,10,14,7,13,11,18,19 } | { 2,6,5,9,14,13,11,17,10,18 } | { 2,3,11,8,14,9,16,10,17,15 } |
| # | (3 . 6 b ) n = 8 k = 4 d = 8 | (3 . 6 c ) n = 8 k = 4 d = 8 | (3 . 6 d ) n = 10 k = 5 d = 9 | (3 . 6 e ) n = 8 k = 4 d = 8 | (3 . 6 f ) n = 10 k = 5 d = 9 | (3 . 6 g ) n = 10 k = 5 d = 9 | (3 . 6 h ) n = 10 k = 5 d = 9 |
| Cell | 1 α 9 α 8 0 - α 5 - α 5 α 6 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 0 1 α 6 α 4 0 0 0 0 0 0 0 0 0 1 α 3 0 - α 1 0 0 0 0 0 0 0 1 α 2 0 | 1 α 8 0 - α 3 - α 5 ( - α 3 - α 5 ) α 6 - α 3 α 4 0 0 0 0 1 α 7 0 0 0 0 0 0 0 1 α 6 α 4 α 2 0 0 0 0 0 0 0 1 α 1 | 1 α 9 α 8 α 7 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 1 α 5 0 0 - α 2 0 0 0 0 0 0 1 α 4 α 3 0 0 0 0 0 0 0 0 0 0 1 α 1 | 1 α 11 α 10 0 0 0 α 6 0 0 0 0 0 0 0 1 α 8 α 8 α 9 0 0 0 0 0 0 0 0 0 0 1 α 9 α 7 0 0 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 0 0 1 α 4 α 3 0 - α 1 0 0 0 0 0 0 0 0 0 1 α 2 0 | 1 α 11 α 10 α 9 0 0 0 0 0 0 0 0 0 0 0 1 α 8 0 0 0 0 0 0 0 0 0 0 0 1 α 7 α 6 0 0 0 0 - α 1 0 0 0 0 0 0 1 0 - α 3 - α 3 α 4 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 0 0 1 α 4 α 2 0 | 1 α 11 α 10 α 8 α 8 α 9 0 0 0 0 0 0 0 0 0 0 1 α 9 0 0 0 0 0 0 α 1 0 0 0 0 1 α 3 α 3 α 7 0 0 α 3 α 4 0 0 0 0 0 0 0 1 α 7 0 0 α 4 α 2 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 | 1 α 11 α 10 α 9 0 0 0 0 0 0 0 0 0 0 0 1 0 - α 2 - α 2 α 8 0 0 - α 2 α 5 - α 2 α 3 0 0 0 0 0 1 α 4 α 4 α 8 0 0 α 4 α 5 0 0 0 0 0 0 0 1 α 8 0 0 α 5 α 3 α 1 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 |
|------------|---------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Graph/Perm | { 2,4,9,13,6,11,15,10,18,17 } | { 2,6,5,11,8,9,12,15 } | { 2,3,7,11,14,10,8,16,15,19 } | 2,6,9,5,15,16,13,19,11,12,22,20 } | 2,3,6,13,16,11,12,10,20,19,21,17 } | 2,3,12,11,13,9,10,19,20,17,18,16 } | 2,3,12,13,11,9,10,19,20,17,16,18 } |
| # | (3 . 6 i ) n = 10 k = 5 d = 9 | (3 . 6 j ) n = 8 k = 4 d = 8 | (3 . 6 k ) n = 10 k = 5 d = 9 | (3 . 13 a ) n = 12 k = 6 d = 11 | (3 . 13 b ) n = 12 k = 6 d = 11 | (3 . 13 c ) n = 12 k = 6 d = 11 | (3 . 13 d ) n = 12 k = 6 d = 11 |
| Cell | 1 α 11 α 10 α 9 0 0 0 0 0 0 0 0 0 0 0 1 α 2 α 2 α 7 α 2 α 7 α 8 0 0 0 - α 2 α 3 0 0 0 0 0 1 α 7 α 7 α 8 0 0 0 - α 3 - α 1 0 0 0 0 0 1 α 8 0 0 α 4 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 | 1 α 11 α 10 0 0 0 0 0 0 0 0 0 0 0 1 α 9 α 8 0 0 0 0 0 0 α 1 0 0 0 0 1 α 3 α 3 α 7 0 0 α 3 α 4 0 0 0 0 0 0 0 1 α 7 0 0 α 4 α 2 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 | 1 α 12 α 11 α 10 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 9 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 8 α 7 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 α 2 α 2 α 4 α 2 α 3 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 4 α 3 α 1 | 1 α 12 α 11 α 8 α 8 α 10 α 8 α 9 0 0 0 0 0 0 0 0 0 0 0 1 α 10 α 9 0 0 0 0 0 0 α 2 0 0 0 0 0 0 1 0 0 0 0 0 - α 3 0 0 0 0 0 0 0 0 1 α 7 0 0 α 4 0 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 1 | 1 α 12 α 11 α 10 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 - α 1 0 0 0 0 1 α 9 α 6 α 6 α 8 α 6 α 7 0 0 0 0 0 0 0 0 0 0 0 1 α 8 α 7 0 0 α 3 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 2 0 | 1 α 12 α 11 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 2 α 2 α 8 α 2 α 8 α 10 α 2 α 8 α 9 0 0 0 0 α 2 α 3 0 0 0 0 0 1 α 8 α 8 α 10 α 8 α 9 0 0 0 0 α 3 α 1 0 0 0 0 0 1 α 10 α 9 0 0 α 5 0 0 0 0 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 4 0 0 1 α 12 α 11 α 8 α 8 α 10 α 8 α 9 0 0 0 0 0 0 0 0 0 0 0 1 α 10 α 9 0 0 0 0 0 0 α 2 0 0 0 0 0 0 1 α 4 0 0 0 α 4 α 5 0 0 0 0 0 0 0 0 0 1 0 0 0 α 5 α 3 0 0 0 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 1 |
|------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Graph/Perm | { 2,3,12,13,10,9,11,19,20,18,16,17 } | { 2,4,13,11,12,9,10,19,20,17,18,15 } | { 2,3,6,15,18,7,19,14,12,23,24,13,22,25 } | { 2,3,14,5,12,15,10,11,22,23,21,20,18,27 } 2,3,14,15,6,13,8,11,19,23,24,21,26,18 } | { 2,3,15,14,12,7,10,13,22,23,20,25,18,19 } | { { 2,3,14,5,12,15,11,10,22,23,20,21,18,27 } |
| # | (3 . 13 e ) n = 12 k = 6 d = 11 | (3 . 13 f ) n = 12 k = 6 d = 11 (3 . 13 g ) n = 14 | k = 7 d = 12 (3 . 13 h ) n = 14 k = 7 | d = 12 (3 . 13 i ) n = 14 k = 7 d = 12 | (3 . 13 j ) n = 14 k = 7 d = 12 | (3 . 13 k ) n = 14 k = 7 d = 12 |
| Cell | 1 α 12 α 11 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 2 α 2 α 8 α 2 α 8 α 10 α 2 α 8 α 9 0 0 0 0 α 2 α 3 0 0 0 0 0 1 α 8 α 8 α 10 α 8 α 9 0 0 0 0 α 3 α 1 0 0 0 0 0 1 α 10 α 9 0 0 α 5 0 0 0 0 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 4 0 0 | 1 α 10 α 9 0 0 α 4 + α 6 α 4 α 5 0 0 0 0 0 1 α 8 0 0 0 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 0 1 α 5 α 3 0 0 0 0 0 0 0 0 0 1 α 2 α 1 | 1 α 10 α 9 α 8 - α 4 - α 4 α 5 - α 4 α 5 α 6 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 1 α 5 α 5 α 6 0 - α 2 0 0 0 0 0 0 1 α 6 α 3 0 0 0 0 0 0 0 0 0 0 1 α 1 | 1 α 9 α 6 + α 8 α 6 α 7 - α 3 - α 3 α 4 0 0 0 0 1 α 7 α 5 0 0 0 0 0 0 0 1 α 4 α 2 0 0 0 0 0 0 0 1 α 1 | 1 α 9 α 8 - α 3 - α 5 ( - α 3 - α 5 ) α 6 - α 3 α 4 0 0 0 0 1 α 7 0 0 0 0 0 0 0 1 α 6 α 4 α 2 0 0 0 0 0 0 0 1 α 1 | 1 α 10 α 9 α 8 - α 5 - α 5 α 6 0 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 1 α 6 0 0 - α 2 0 0 0 0 0 0 1 α 4 α 3 0 0 0 0 0 0 0 0 0 0 1 α 1 | 1 α 11 α 10 α 9 α 7 α 7 α 8 0 0 0 0 0 0 0 0 0 0 1 α 8 0 0 0 0 - α 2 0 0 0 0 0 0 1 α 6 α 5 0 0 0 0 0 0 0 0 0 0 0 1 α 4 0 0 0 0 0 0 0 0 0 0 0 1 α 3 0 0 0 0 0 0 0 0 0 0 0 0 1 α 1 |
|------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Graph/Perm | 2,3,15,14,12,7,10,13,22,23,20,25,18,19 | { 2,5,7,13,14,9,11,16,10,18 } | { 2,3,8,10,14,7,11,16,15,19 } | { 2,4,6,8,11,9,13,15 } | { 2,5,6,11,8,9,12,15 } | { 2,3,7,10,14,11,8,16,15,19 } | { 2,3,4,12,7,13,10,18,20,21,17,23 } |
| # | (3 . 13 l ) n = 14 k = 7 d = 12 | (3 . 14 a ) n = 10 k = 5 d = 10 | (3 . 14 b ) n = 10 k = 5 d = 10 | (3 . 14 c ) n = 8 k = 4 d = 9 | (3 . 14 d ) n = 8 k = 4 d = 9 | (3 . 14 e ) n = 10 k = 5 d = 10 | (3 . 14 f ) n = 12 k = 6 d = 11 |
| Cell | 1 α 11 α 10 α 8 α 8 α 9 0 0 0 - α 4 0 0 0 0 0 0 1 α 9 0 - α 6 0 0 0 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 0 0 0 1 α 3 0 0 0 0 0 0 0 0 0 0 0 1 α 2 α 1 | 1 α 11 α 10 α 8 α 8 α 9 0 0 0 0 - α 3 0 0 0 0 0 1 α 9 0 - α 6 0 0 0 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 0 0 0 1 α 5 α 4 0 0 0 0 0 0 0 0 0 0 0 0 1 α 2 0 0 0 0 0 0 0 0 0 0 0 1 α 1 | 1 α 10 α 9 α 8 0 0 α 3 α 3 α 4 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 1 α 6 α 5 0 0 0 0 0 0 0 0 0 1 α 4 α 2 0 0 0 0 0 0 0 0 0 1 α 1 | 1 α 10 α 9 α 8 - α 5 - α 5 α 6 0 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 1 α 6 0 - α 3 0 0 0 0 0 0 0 1 α 4 0 0 0 0 0 0 0 0 0 0 1 α 2 α 1 | 1 α 10 α 9 0 0 0 0 0 0 0 0 0 1 α 6 + α 8 α 6 α 7 0 0 0 0 0 0 0 0 1 α 7 α 5 0 0 α 2 α 1 0 0 0 0 0 1 α 4 0 0 0 0 0 0 0 0 0 1 α 3 0 0 | 1 α 9 α 8 α 5 + α 7 α 5 α 6 0 0 0 0 0 0 1 α 3 + α 6 α 3 α 4 0 0 0 0 0 0 1 α 4 0 - α 1 0 0 0 0 0 1 α 2 0 | 1 α 9 α 8 - α 5 - α 5 α 6 α 2 α 2 α 3 0 0 0 1 α 7 0 0 0 0 0 0 0 1 α 6 α 4 0 0 0 0 0 0 0 1 α 3 α 1 |
|------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------|
| Graph/Perm | 2,3,8,6,11,17,16,19,13,21,12,22 } | 2,3,8,6,12,17,16,9,19,13,22,23 } | 2,3,6,10,14,8,15,11,17,19 } | 2,3,7,9,14,11,16,15,10,18 } | 2,5,11,8,13,9,16,17,10,14 } | { 2,3,7,8,9,12,14,13 } | { 2,5,7,11,8,12,9,14 } |
| # | (3 . 14 g ) n = 12 k = 6 d = 11 | (3 . 14 h ) n = 12 k = 6 d = 11 | (3 . 14 i ) n = 10 k = 5 d = 10 | (3 . 14 j ) n = 10 k = 5 d = 10 | (3 . 14 k ) n = 10 k = 5 d = 10 | (3 . 14 l ) n = 8 k = 4 d = 9 | (3 . 14 m ) n = 8 k = 4 d = 9 |
| Cell | 1 α 13 α 12 α 11 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 2 α 2 α 10 0 0 0 0 0 0 α 2 α 3 0 0 0 0 0 1 α 10 0 0 0 0 0 0 α 3 α 1 0 0 0 0 0 1 α 5 α 5 α 9 0 0 α 5 α 6 0 0 0 0 0 0 0 0 0 1 α 9 0 0 α 6 α 4 0 0 0 0 0 0 0 0 0 1 α 8 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 7 0 0 0 0 | 1 α 11 α 10 - α 6 - α 6 α 8 - α 6 α 7 0 0 0 0 0 0 1 α 9 0 0 0 0 0 0 0 0 0 1 α 8 α 7 - α 2 - α 2 α 4 - α 2 α 3 0 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 1 α 4 α 3 α 1 | 1 α 13 α 12 α 7 α 7 α 11 0 0 α 7 α 8 0 0 0 0 0 0 0 0 0 1 α 11 0 0 α 8 α 6 0 0 0 0 0 0 0 0 0 1 α 10 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 5 α 4 0 0 α 1 0 0 0 0 0 0 0 0 0 0 1 α 3 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 2 0 | 1 α 13 α 12 α 11 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 4 α 4 α 10 0 0 0 0 α 4 α 5 0 0 0 0 0 0 0 1 α 10 0 0 0 0 α 5 0 0 α 1 0 0 0 0 0 1 α 7 α 7 α 9 α 7 α 8 0 0 0 0 0 0 0 0 0 0 0 1 α 9 α 8 α 6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 3 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 2 0 | 1 α 12 α 11 α 10 0 0 0 0 0 0 0 0 0 0 0 1 α 4 + α 8 ( α 4 + α 8 ) α 9 0 0 α 4 α 5 0 0 0 0 0 0 0 1 α 9 0 0 α 5 0 - α 2 - α 1 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 0 1 α 3 0 0 | 1 α 12 α 11 α 8 α 8 α 10 α 8 α 9 0 0 0 0 0 0 0 0 0 1 α 10 α 5 + α 9 α 5 α 7 α 5 α 6 0 0 0 0 0 0 0 0 0 1 α 7 α 6 0 - α 3 0 0 0 0 0 0 0 0 0 1 α 4 0 0 0 0 0 0 0 0 0 0 0 0 1 α 2 0 0 0 0 0 0 0 0 0 0 0 1 α 1 1 α 11 α 7 + α 10 α 7 α 9 α 7 α 8 0 0 0 0 0 0 0 1 α 9 α 4 + α 8 α 4 α 6 α 4 α 5 0 0 0 0 0 0 0 1 α 6 α 5 α 3 0 0 0 0 0 0 0 0 0 1 α 2 0 0 0 0 0 0 0 0 0 1 α 1 |
|------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Graph/Perm | 2,3,14,15,12,13,10,11,22,23,20,21,18,19 } | { 2,5,10,13,8,11,16,9,14,17 } | 2,3,10,7,8,19,20,15,18,13,14,25,26,23 } | 2,3,14,15,10,13,8,9,20,21,18,25,26,19 } 2,3,10,13,8,11,18,19,16,21,12,17 } | { 2,3,12,5,7,13,9,16,20,18,22,23 } | { { 2,4,10,6,11,7,13,15,18,19 } |
| # | (3 . 26 a ) n = 14 k = 7 d = 13 | (3 . 27 a ) n = 10 k = 5 d = 11 | (3 . 27 b ) n = 14 k = 7 d = 13 | (3 . 27 c ) n = 14 k = 7 d = 13 (3 . 27 d ) n = 12 k = 6 d = 12 | (3 . 28 a ) n = 12 k = 6 d = 12 | (3 . 28 b ) n = 10 k = 5 d = 11 |
| Cell | 1 α 12 α 11 α 7 + α 9 ( α 7 + α 9 ) α 10 α 7 α 8 0 0 0 0 0 0 0 0 0 1 α 10 α 8 0 0 0 - α 3 0 α 1 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 0 0 1 α 4 0 0 0 0 0 0 0 0 0 0 0 0 1 α 2 0 | 1 α 12 α 5 + α 8 ( α 5 + α 8 ) α 11 0 - ( α 5 + α 8 ) α 9 - α 5 α 7 - α 5 α 6 0 0 0 0 0 0 1 α 11 0 - α 9 - α 7 - α 6 - α 4 0 0 0 0 0 0 1 α 10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 3 0 0 0 0 0 0 0 0 0 0 0 1 α 2 0 0 0 0 0 0 0 0 0 0 0 1 α 1 | 1 α 11 α 7 + α 10 α 7 α 9 α 7 α 8 0 0 0 0 0 0 0 1 α 9 α 4 + α 8 α 4 α 6 α 4 α 5 0 0 0 0 0 0 0 1 α 6 α 5 α 3 0 0 0 0 0 0 0 0 0 1 α 2 0 0 0 0 0 0 0 0 0 1 α 1 | 1 α 13 α 12 α 9 α 9 α 11 α 9 α 10 0 0 0 0 0 0 0 0 0 0 0 1 α 11 α 10 0 - α 7 0 0 0 0 0 0 0 0 0 0 0 1 α 8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 4 0 - α 2 - α 1 0 0 0 0 0 0 0 0 0 0 1 α 3 0 0 1 | α 13 α 12 α 8 + α 10 ( α 8 + α 10 ) α 11 α 8 α 9 0 0 0 0 0 0 0 0 0 0 0 1 α 11 α 9 0 0 0 0 0 0 - α 2 - α 1 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 α 4 0 0 0 0 0 0 0 0 0 0 1 α 6 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 α 3 0 0 1 α 12 α 11 α 6 + α 9 ( α 6 + α 9 ) α 10 0 - α 6 α 7 0 0 0 0 0 0 0 0 1 α 10 0 - α 7 0 0 0 α 2 α 1 0 0 0 0 1 α 8 0 0 0 0 0 0 0 0 0 0 0 0 1 α 5 0 0 0 0 0 0 0 0 0 0 0 1 α 4 0 0 0 0 0 0 0 0 0 0 0 1 α 3 0 0 | 1 α 12 α 11 α 10 0 0 0 0 0 0 0 0 0 0 0 1 α 9 0 0 0 0 0 0 0 0 0 0 0 1 α 8 0 0 0 0 0 0 0 0 0 0 0 1 0 - α 3 - α 5 ( - α 3 - α 5 ) α 6 - α 3 α 4 0 0 0 0 0 0 0 0 1 α 7 0 0 0 0 0 0 0 0 0 0 0 1 α 6 α 4 α 2 α 1 |
|------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Graph/Perm | 2,3,9,5,11,13,18,19,20,12,22,16 } 2,7,5,6,16,12,8,13,15,21,22,23 } | 2,3,9,5,11,13,18,19,20,12,22,16 } 2,7,5,6,16,12,8,13,15,21,22,23 } | { 2,4,10,6,11,7,13,15,18,19 } | 2,3,12,5,7,15,20,18,22,23,13,25,14,24 } | 2,3,12,5,13,15,20,10,22,23,21,25,14,18 } 2,3,10,6,11,17,13,19,20,21,12,16 } | 2,3,10,13,16,17,9,19,11,18,12,20 } |
| # | (3 . 28 c ) n = 12 k = 6 d = 12 | (3 . 28 d ) n = 12 k = 6 d = 12 | (3 . 28 e ) n = 10 k = 5 d = 11 | (3 . 28 f ) n = 14 k = 7 d = 13 (3 . 28 g ) | n = 14 k = 7 d = 13 (3 . 28 h ) n = 12 k = 6 d = 12 | (3 . 28 i ) n = 12 k = 6 d = 12 |
## References
- [1] M. E. Peskin and D. V. Schroeder, An Introduction to quantum field theory . Westview Press Inc, 1995.
- [2] N. Arkani-Hamed, F. Cachazo, and J. Kaplan, 'What is the Simplest Quantum Field Theory?,' JHEP , vol. 1009, p. 016, 2010.
- [3] N. Arkani-Hamed and J. Trnka, 'The Amplituhedron,' 2013 - arXiv:1312.2007 [hep-th].
- [4] N. Arkani-Hamed, 'The amplituhedron,' talk given at the SUSY 2013 conference, 30 August 2013.
- [5] N. Arkani-Hamed, J. L. Bourjaily, F. Cachazo, A. B. Goncharov, A. Postnikov, et al. , 'Scattering Amplitudes and the Positive Grassmannian,' 2012 - arXiv:1212.5605 [hep-th].
- [6] M. B. Green, J. H. Schwarz, and L. Brink, 'N = 4 yang-mills and n = 8 supergravity as limits of string theories,' Nuclear Physics B , vol. 198, no. 3, pp. 474 - 492, 1982.
- [7] J. M. Maldacena, 'The Large N limit of superconformal field theories and supergravity,' Adv.Theor.Math.Phys. , vol. 2, pp. 231-252, 1998.
- [8] S. Gubser, I. R. Klebanov, and A. M. Polyakov, 'Gauge theory correlators from noncritical string theory,' Phys.Lett. , vol. B428, pp. 105-114, 1998.
- [9] E. Witten, 'Anti-de Sitter space and holography,' Adv.Theor.Math.Phys. , vol. 2, pp. 253-291, 1998.
- [10] A. Postnikov, 'Total positivity, Grassmannians, and networks,' 2006 - arXiv:math/0609764 [math.CO].
- [11] L. W. A. Postnikov, D Speyer, 'Matching polytopes, toric geometry, and the non-negative part of the grassmannian,' Journal of Algebraic Combinatorics , vol. 30, pp. 173-191, 2009.
- [12] R. Marsh and K. Rietsch, 'Parametrizations of flag varieties,' Representation Theory of the American Mathematical Society , vol. 8, no. 9, pp. 212-242, 2004.
- [13] S. Franco, 'Bipartite Field Theories: from D-Brane Probes to Scattering Amplitudes,' JHEP , vol. 1211, p. 141, 2012.
- [14] Y.-H. He, 'Bipartita: Physics, Geometry and Number Theory,' The Proceedings of The XXIXth International Colloquium on Group-Theoretical Methods in Physics , 2012.
- [15] J. L. Bourjaily, 'Positroids, Plabic Graphs, and Scattering Amplitudes in Mathematica,' 2012 - arXiv:1212.6974 [hep-th].
- [16] J. Davey, A. Hanany, and J. Pasukonis, 'On the Classification of Brane Tilings,' JHEP , vol. 1001, p. 078, 2010.
- [17] S. Franco, D. Galloni, and R.-K. Seong, 'New Directions in Bipartite Field Theories,' JHEP , vol. 1306, p. 032, 2013.
- [18] S. Franco, D. Galloni, and A. Mariotti, 'The Geometry of On-Shell Diagrams,' 2013 -arXiv:1310.3820 [hep-th].
- [19] H. Elvang and Y.-t. Huang, Scattering Amplitudes . Lecture Notes, 2013.
| [20] | Z. Bern, L. J. Dixon, D. C. Dunbar, and D. A. Kosower, 'One loop n point gauge theory amplitudes, unitarity and collinear limits,' Nucl.Phys. , vol. B425, pp. 217-260, 1994. |
|--------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [21] | Z. Bern, L. J. Dixon, D. C. Dunbar, and D. A. Kosower, 'Fusing gauge theory tree amplitudes into loop amplitudes,' Nucl.Phys. , vol. B435, pp. 59-101, 1995. |
| [22] | E. Witten, 'Perturbative gauge theory as a string theory in twistor space,' Com- mun.Math.Phys. , vol. 252, pp. 189-258, 2004. |
| [23] | R. Penrose, 'Twistor algebra,' J.Math.Phys. , vol. 8, p. 345, 1967. |
| [24] | A. Ferber, 'Supertwistors and Conformal Supersymmetry,' Nucl.Phys. , vol. B132, p. 55, 1978. |
| [25] | R. Britto, F. Cachazo, and B. Feng, 'New recursion relations for tree amplitudes of gluons,' Nucl.Phys. , vol. B715, pp. 499-522, 2005. |
| [26] | R. Britto, F. Cachazo, B. Feng, and E. Witten, 'Direct proof of tree-level recursion relation in Yang-Mills theory,' Phys.Rev.Lett. , vol. 94, p. 181602, 2005. |
| [27] | A. P. Hodges, 'Twistor diagram recursion for all gauge-theoretic tree amplitudes,' 2005 arXiv:hep-th/0503060. |
| [28] | L. F. Alday and J. Maldacena, 'Gluon scattering amplitudes at strong coupling,' Journal of High Energy Physics , vol. 2007, no. 06, p. 064, 2007. |
| [29] | J. Drummond, J. Henn, V. Smirnov, and E. Sokatchev, 'Magic identities for conformal four- point integrals,' JHEP , vol. 0701, p. 064, 2007. |
| [30] | J. Drummond, 'Review of AdS/CFT Integrability, Chapter V.2: Dual Superconformal Sym- metry,' Lett.Math.Phys. , vol. 99, pp. 481-505, 2012. |
| [31] | J. Drummond and L. Ferro, 'Yangians, Grassmannians and T-duality,' JHEP , vol. 1007, p. 027, 2010. |
| [32] | L. Mason and D. Skinner, 'Dual Superconformal Invariance, Momentum Twistors and Grass- mannians,' JHEP , vol. 0911, p. 045, 2009. |
| [33] | N. Arkani-Hamed, F. Cachazo, C. Cheung, and J. Kaplan, 'The S-Matrix in Twistor Space,' JHEP , vol. 1003, p. 110, 2010. |
| [34] | N. Arkani-Hamed, F. Cachazo, C. Cheung, and J. Kaplan, 'A Duality For The S Matrix,' JHEP , vol. 1003, p. 020, 2010. |
| [35] | N. Arkani-Hamed, F. Cachazo, and C. Cheung, 'The Grassmannian Origin Of Dual Supercon- formal Invariance,' JHEP , vol. 1003, p. 036, 2010. |
| [36] | J. Drummond and L. Ferro, 'The Yangian origin of the Grassmannian integral,' JHEP vol. 1012, p. 010, 2010. |
| [37] | N. Arkani-Hamed, J. Bourjaily, F. Cachazo, and J. Trnka, 'Unification of Residues and Grass- mannian Dualities,' JHEP , vol. 1101, p. 049, 2011. |
| [38] | A. Hodges, 'Eliminating spurious poles from gauge-theoretic amplitudes,' JHEP , vol. 1305, p. 135, 2013. |
- [39] N. Arkani-Hamed, J. L. Bourjaily, F. Cachazo, A. Hodges, and J. Trnka, 'A Note on Polytopes for Scattering Amplitudes,' JHEP , vol. 1204, p. 081, 2012.
- [40] N. Arkani-Hamed, J. L. Bourjaily, F. Cachazo, S. Caron-Huot, and J. Trnka, 'The All-Loop Integrand For Scattering Amplitudes in Planar N=4 SYM,' JHEP , vol. 1101, p. 041, 2011.
- [41] R. H. Boels, 'On BCFW shifts of integrands and integrals,' JHEP , vol. 1011, p. 113, 2010.
- [42] N. Arkani-Hamed and J. Trnka, 'Into the Amplituhedron.' 2013 - arXiv:1312.7878 [hep-th].
- [43] B. Feng, A. Hanany, and Y.-H. He, 'D-brane gauge theories from toric singularities and toric duality,' Nucl.Phys. , vol. B595, pp. 165-200, 2001.
- [44] A. Hanany and K. D. Kennaway, 'Dimer models and toric diagrams,' 2005 -arXiv:hepth/0503149.
- [45] S. Franco, A. Hanany, K. D. Kennaway, D. Vegh, and B. Wecht, 'Brane dimers and quiver gauge theories,' JHEP , vol. 0601, p. 096, 2006.
- [46] K. D. Kennaway, 'Brane Tilings,' Int.J.Mod.Phys. , vol. A22, pp. 2977-3038, 2007.
- [47] S. Franco, D. Galloni, and A. Mariotti, 'Bipartite Field Theories, Cluster Algebras and the Grassmannian,' 2014 - arXiv:1404.3752 [hep-th].
- [48] S. Franco and A. Uranga, 'Bipartite Field Theories from D-Branes,' 2013 - arXiv:1306.6331 [hep-th].
- [49] N. Seiberg, 'Electric -magnetic duality in supersymmetric nonAbelian gauge theories,' Nucl.Phys. , vol. B435, pp. 129-146, 1995.
- [50] A. Amariti and D. Forcella, 'Scattering Amplitudes and Toric Geometry,' JHEP , vol. 1309, p. 133, 2013.
- [51] F. Cachazo, P. Svrcek, and E. Witten, 'MHV vertices and tree amplitudes in gauge theory,' JHEP , vol. 0409, p. 006, 2004.
- [52] B. Feng and M. Luo, 'An Introduction to On-shell Recursion Relations,' Front. Phys. , vol. ,2012,7, pp. :533-575, 5.
- [53] Z. Bern, L. J. Dixon, and D. A. Kosower, 'On-Shell Methods in Perturbative QCD,' Annals Phys. , vol. 322, pp. 1587-1634, 2007.
- [54] Z. Bern and Y.-t. Huang, 'Basics of Generalized Unitarity,' J.Phys. , vol. A44, p. 454003, 2011.
| null |
[
"Alvaro M. Alhambra"
] |
2014-08-05T15:17:37+00:00
|
2014-08-05T15:17:37+00:00
|
[
"hep-th",
"math-ph",
"math.MP"
] |
Brane Tilings as On-shell Diagrams
|
A new way of computing scattering amplitudes in a certain very important QFT
(N=4 SYM) has recently been developed, in which an algebraic structure called
the positive Grassmannian plays a very important role. The mathematics of the
positive Grassmannian involve, among other things, bipartite graphs, which also
appear in the formulation of a certain class of conformal field theories that
are currently being generalized into Bipartite Field Theories (BFT). The fact
that the same structures appear in two such different realms of physics
suggests a deeper connection between the two that is yet to be fully unveiled.
Here we explore that potential connection by looking at the graphs of a certain
class of BFTs, the brane tilings, in terms of the new mathematics developed for
the computation of the amplitudes. This way we produce a set of data that will
hopefully be useful in the development of that connection.
|
1408.0998v1
|
## The Case for a Mixed-Initiative Collaborative Neuroevolution Approach
Sebastian Risi, Jinhong Zhang, Rasmus Taarnby, Peter Greve, Jan Piskur, Antonios Liapis, Julian Togelius
IT University Copenhagen, 2300 Copenhagen, Denmark { sebr, jinh, reta, pgre, japi, anli, juto } @itu.dk
## Abstract
It is clear that the current attempts at using algorithms to create artificial neural networks have had mixed success at best when it comes to creating large networks and/or complex behavior. This should not be unexpected, as creating an artificial brain is essentially a design problem. Human design ingenuity still surpasses computational design for most tasks in most domains, including architecture, game design, and authoring literary fiction. This leads us to ask which the best way is to combine human and machine design capacities when it comes to designing artificial brains. Both of them have their strengths and weaknesses; for example, humans are much too slow to manually specify thousands of neurons, let alone the billions of neurons that go into a human brain, but on the other hand they can rely on a vast repository of common-sense understanding and design heuristics that can help them perform a much better guided search in design space than an algorithm. Therefore, in this paper we argue for a mixed-initiative approach for collaborative online brain building and present first results towards this goal.
and Stanley, 2011), a method that rewards novel behaviors instead of rewarding objective performance, has shown promise and significantly outperforms objective-based approaches in a variety of different domains. Other approaches are based on interactive evolutionary computation (IEC) methods, wherein the human user guides evolution by repeatedly choosing from a set of candidates. Woolley and Stanley (2011) showed that interactive evolution can help to discover artifacts which are very hard to evolve with traditional evolutionary approaches. Recently Woolley and Stanley (2014) combined IEC with novelty search, demonstrating that the approaches complement each other and together address some of the challenges that each method struggles with by itself (e.g. novelty search can get lost in large search spaces, interactive evolution is limited by user fatigue). In a related approach, Bongard and Hornby (2013) recently demonstrated that human input and objective-based search can also be combined synergistically to solve challenging robotic domains.
With around 200 billion neurons and 125 trillion synapses, the human brain is the most complex system known to exist. The brain's structural complexity, with intricate synaptic motifs that repeat throughout it, gives rise to our unique mental abilities. Additionally, the brains plasticity allows us to learn new abilities throughout our life (e.g. learning a new language) and to change our behavior based on past experience.
Therefore, creating similar artificial structures by recapitulating the process that created intelligence on earth, is an intriguing possibility. In this context, neuroevolution (i.e. evolving artificial neural networks (ANNs) via evolutionary algorithms) has shown promising results in a variety of different domains (Floreano et al., 2008; Stanley and Miikkulainen, 2004; Yao, 1999). However, these results still pale in comparison to the capabilities of natural brains. The reasons for this are manifold. Especially the problem of deceptive fitness landscapes (i.e. mutations increase fitness but actually lead further away from the final objective) has limited the scope of problems amenable to evolutionary algorithms.
Research in circumventing this problem has mainly focused on two different ideas. First, novelty search (Lehman
While these human in the loop approaches have shown promising results, we argue in this paper that they do not exploit the whole range of human intuition, ability to identify promising stepping stones, or collaborative problem-solving skills that could ultimately allow us to create more brain-like artificial structures. For example, collaborative games like Foldit hint at the power of crowdsourcing the human brain's natural abilities for specific tasks (e.g. pattern matching, spatial reasoning), which are hard to solve by computational approaches. However, in traditional IEC applications the role of the user is often reduced to solely judging the created artifacts and only 'nudging' evolution by deciding between a discrete choice of candidates. In other words, only the computer creates content (e.g. images, ANNs, etc.) and the role of the human is to guide evolution to content they prefer.
An approach that does require significantly more input from the user is a mixed-initiative process (Liapis et al., 2014), wherein both the computer and the human take turns in creating content. Yet while mixed-initiative based approaches have shown promise in the context of procedurally generated content for games (Liapis et al., 2013; Yannakakis
Figure 1: BrainCrafter Web Interface. Beta version online at: braincrafter.dk
et al., 2014), they have not yet been applied to the evolution of large-scale ANNs for complex tasks. However, the promise of such an approach is a system that can benefit from the different skillsets of a human and a computational creator. For example, while a human user might develop an intuition about promising domain-dependent network topographies, a computational method is likely more effective at fine-tuning specific synaptic weights. Therefore, we argue for a collaborative mixed-initiative approach in which both the computer and the collaborating human users can take initiative and propose changes to an evolving neural networks. That is, at any point the human can revert to just doing selection and let the evolutionary algorithm serve up new content to judge, or decide to jump in and have a more active role in the design of the ANN.
Our recent work has taken steps towards this goal by focusing on two parallel lines of research. First, before introducing a mixed-initiative approach it is useful to determine how good we as humans are at building complex ANNs for certain tasks (without evolution) and if collaborating with other users proves useful. Insights from this experiment should in turn provide useful clues about the strengths of human ANN design and most importantly, non-intuitive aspects of the design process we tend to struggle with (i.e. aspects which would benefit most from an assisting computational creator). In this context we recently introduced BrainCrafter (braincrafter.dk), which is an online application that allows the user to build ANNs for specific control programs (e.g. a robot that must traverse a maze) by adding neurons and connections in a drag & drop like fashion (Figure 1). While building ANNs the users can observe the resulting simulated robot behaviors in real-time, proving insights into the effects of different network modifications. BrainCrafter also allows users to collaborate by building on high-scoring solutions created by other people.
While the ongoing BrainCrafter experiment should pro-
Figure 2: Picture CPPN-Compiler User Interface. Beta version online at: rasmustaarnby.dk/thesis
vide insights into our abilities to collaboratively construct complex networks, more challenging domains will likely require the orchestrated effort of both human and computational creators. Therefore extending BrainCrafter to support the users' collaborative engineering efforts through a mixed-initiative approach is an important next goal. To allow seamlessly switching between interactive evolution and manual ANN design will require an ANN representation that (1) can be edited easily by the user on a local (e.g. individual neurons) and global level (e.g. overall network topography and topology), (2) is evolvable and compact, and (3) allows the system to produce meaningful suggestions based on the user's input.
In this context we are building on the a generative encoding called compositional pattern producing networks (CPPNs; Stanley (2007)) and the recently introduced concept of a CPPN-Compiler (Risi, 2013). CPPNs, which are based on principles of how natural organisms develop, allow the compact encoding of complex patterns, from two-dimensional images (Secretan et al., 2011) and threedimensional forms (Clune and Lipson, 2011) to large-scale ANNs (Stanley et al., 2009; Clune et al., 2011; Risi and Stanley, 2012). The idea behind the CPPN-Compiler is to allow the user to directly compile a high-level description of the desired starting structure into the CPPN itself. We are currently exploring the benefits of this approach through the collaborative interactive evolution of images (Figure 2), in which the user can draw a vector image and annotate it with important regularities like symmetry. The CPPNCompiler then compiles this high-level description into the CPPN itself. Since the compiled CPPN now directly embodies the annotated domain regularities (e.g. bilateral symmetric arms), the produced offspring show meaningful variations that nevertheless share common features. Considering the insight that CPPNs can also produce large-scale ANNs (Stanley et al., 2009) and can be modified to create
Figure 3: Mixed-initiative Neuroevolution Framework
certain neural topologies (ES-HyperNEAT; Risi and Stanley (2012)), opens up the intriguing possibility of a neural CPPN-Compiler. Our current efforts focus on such a compiler that will form the backbone for our mixed-initiative BrainCrafter application.
The envisioned collaborative mixed-initiative system is depicted in Figure 3. The users can collaboratively construct neural networks and annotate them with regularities (e.g. symmetry, repetition, etc.), which are used by the computational creator to construct the internal CPPN model and in turn propose meaningful variations to the user. At any point in the process the human can revert to just doing selection or decide to directly edit the ANN, wherein phenotypic edits are directly compiled back into the genotype. The promise of the proposed system is that it could allow a variety of tasks to be solved by many people online within a mixed-initiative environment, which have heretofore proven too difficult. We expect this project will profoundly impact the fields of ANN research and potentially also deepen our understanding of the way biological neural networks solve certain problems.
## References
Bongard, J. C. and Hornby, G. S. (2013). Combining fitness-based search and user modeling in evolutionary robotics. In Proceeding of the fifteenth annual conference on Genetic and evolutionary computation conference , pages 159-166. ACM.
Clune, J. and Lipson, H. (2011). Evolving three-dimensional objects with a generative encoding inspired by developmental biology. In Proceedings of the European Conference on Artificial Life, See http://EndlessForms. com .
Clune, J., Stanley, K. O., Pennock, R. T., and Ofria, C. (2011). On the performance of indirect encoding across the continuum of regularity. Evolutionary Computation, IEEE Transactions on , 15(3):346-367.
Floreano, D., D¨ urr, P., and Mattiussi, C. (2008). Neuroevolution: from architectures to learning. Evol. Intelligence , 1(1):47-62.
Lehman, J. and Stanley, K. O. (2011). Abandoning objectives: Evolution through the search for novelty alone. Evolutionary Computation , 19(2):189-223.
Liapis, A., Shaker, N., and Smith, G. (2014). Mixed-initiative approaches. In Shaker, N., Togelius, J., and Nelson, M. J., editors, Procedural Content Generation in Games: A Textbook and an Overview of Current Research . Springer.
Liapis, A., Yannakakis, G. N., and Togelius, J. (2013). Sentient sketchbook: Computer-aided game level authoring. In Proceedings of ACM Conference on Foundations of Digital Games (FDG) .
Risi, S. (2013). A compiler for cppns: Transforming phenotypic descriptions into genotypic representations. In 2013 AAAI Fall Symposium Series .
Risi, S. and Stanley, K. O. (2012). An enhanced hypercube-based encoding for evolving the placement, density, and connectivity of neurons. Artificial Life , 18(4):331-363.
Secretan, J., Beato, N., D'Ambrosio, D., Rodriguez, A., Campbell, A., Folsom-Kovarik, J., and Stanley, K. (2011). Picbreeder: A case study in collaborative evolutionary exploration of design space. Evolutionary Computation , 19(3):373-403.
Stanley, K. O. (2007). Compositional pattern producing networks: A novel abstraction of development. Genetic Programming and Evolvable Machines Special Issue on Developmental Systems , 8(2):131-162.
Stanley, K. O., D'Ambrosio, D. B., and Gauci, J. (2009). A hypercube-based indirect encoding for evolving large-scale neural networks. Artificial Life , 15(2):185-212.
Stanley, K. O. and Miikkulainen, R. (2004). Competitive coevolution through evolutionary complexification. JAIR , 21:63-100.
Woolley, B. G. and Stanley, K. O. (2011). On the deleterious effects of a priori objectives on evolution and representation. In Proceedings of the 13th annual conference on Genetic and evolutionary computation , pages 957-964. ACM.
Woolley, B. G. and Stanley, K. O. (2014). A novel humancomputer collaboration: Combining novelty search with interactive evolution. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO-2014) , New York, NY, USA. ACM.
Yannakakis, G. N., Liapis, A., and Alexopoulos, C. (2014). Mixedinitiative co-creativity. In Proceedings of the 9th Conference on the Foundations of Digital Games .
Yao, X. (1999). Evolving artificial neural networks. Proceedings of the IEEE , 87(9).
| null |
[
"Sebastian Risi",
"Jinhong Zhang",
"Rasmus Taarnby",
"Peter Greve",
"Jan Piskur",
"Antonios Liapis",
"Julian Togelius"
] |
2014-08-05T15:18:13+00:00
|
2014-08-05T15:18:13+00:00
|
[
"cs.NE"
] |
The Case for a Mixed-Initiative Collaborative Neuroevolution Approach
|
It is clear that the current attempts at using algorithms to create
artificial neural networks have had mixed success at best when it comes to
creating large networks and/or complex behavior. This should not be unexpected,
as creating an artificial brain is essentially a design problem. Human design
ingenuity still surpasses computational design for most tasks in most domains,
including architecture, game design, and authoring literary fiction. This leads
us to ask which the best way is to combine human and machine design capacities
when it comes to designing artificial brains. Both of them have their strengths
and weaknesses; for example, humans are much too slow to manually specify
thousands of neurons, let alone the billions of neurons that go into a human
brain, but on the other hand they can rely on a vast repository of common-sense
understanding and design heuristics that can help them perform a much better
guided search in design space than an algorithm. Therefore, in this paper we
argue for a mixed-initiative approach for collaborative online brain building
and present first results towards this goal.
|
1408.0999v2
|
## FiniteN corrections to Vlasov dynamics and the range of pair interactions
Andrea Gabrielli 1 2 , , Michael Joyce 3 4 , and Jules Morand 3 4 ,
1 ISC - CNR, UoS 'Sapienza', Dipartimento di Fisica, Universit` a 'Sapienza', Piazzale Aldo Moro 5, 00185 - Rome,Italy 2 INFM, Unit` a Roma 1, Dipartimento di Fisica, Universit` 'Sapienza', Piazzale Aldo Moro 5, 00185 - Rome,Italy a 3 UPMC Univ Paris 06, UMR 7585, LPNHE, F-75005, Paris, France and 4 CNRS IN2P3, UMR 7585, LPNHE, F-75005, Paris, France
## Abstract
We explore the conditions on a pair interaction for the validity of the Vlasov equation to describe the dynamics of an interacting N particle system in the large N limit. Using a coarse-graining in phase space of the exact Klimontovich equation for the N particle system, we evaluate, neglecting correlations of density fluctuations, the scalings with N of the terms describing the corrections to the Vlasov equation for the coarse-grained one particle phase space density. Considering a generic interaction with radial pair force F r ( ), with F r ( ) ∼ 1 /r γ at large scales, and regulated to a bounded behaviour below a 'softening' scale ε , we find that there is an essential qualitative difference between the cases γ < d and γ > d , i.e., depending on the integrability at large distances of the pair force. In the former case the corrections to the Vlasov dynamics for a given coarse-grained scale are essentially insensitive to the softening parameter ε , while for γ > d the amplitude of these terms is directly regulated by ε , and thus by the small scale properties of the interaction. This corresponds to a simple physical criterion for a basic distinction between long-range ( γ ≤ d ) and short range ( γ > d ) interactions, different to the canonical one ( γ ≤ d +1 or γ > d +1 ) based on thermodynamic analysis. This alternative classification, based on purely dynamical considerations, is relevant notably to understanding the conditions for the existence of so-called quasi-stationary states in long-range interacting systems.
PACS numbers: 05.70.-y, 05.45.-a, 04.40.-b
Interactions are canonically characterized as shortrange or long-range on the basis of the fundamental distinction which arises in equilibrium statistical mechanics between interactions for which the energy is additive and those for which it is non-additive (for reviews, see e.g. [14]). For a system of particles interacting via two body interactions with a pair potential V ( r ), the system is then long-range (or 'strongly long range' [4] ) if and only if V ( r ) decays at large distances slower than one over the separation r to the power of the spatial dimension d . In the last decade there has been considerable study of this class of interactions. One of the very interesting results about systems in this class which has emerged - essentially through numerical study of different models - is that, like for the much studied case of gravity in astrophysics, their dynamics leads, from generic initial conditions, to so-called quasi-stationary states: macroscopic non-equilibrium states which evolve only on timescales which diverge with particle number. Theoretically these states are interpreted in terms of a description of the system's dynamics by the Vlasov equation, of which they represent stationary solutions. Their physical realizations arise in numerous and very diverse systems, ranging from galaxies and 'dark matter halos' in astrophysics and cosmology (see e.g. [5]) to the red spot on Jupiter (see e.g. [6]), to laboratory systems such as cold atoms [7], and even to biological systems[8]. A basic question is whether the appearance of these out of equilibrium stationary states - and more generally the validity of the Vlasov equation to describe the system's dynamics
-applies to the same class of long-range interactions as defined by equilibrium statistical mechanics, or only to a sub-class of them, or indeed to a larger class of interactions. In short, in what class of systems can we expect to see these quasi-stationary states? Are they typical of long-range interactions as defined canonically? Or are they characteristic of a different class?
To answer these questions requires establishing the conditions of validity of the Vlasov equation, and specifically how such conditions depend on the two-body interaction. In the literature there are, on the one hand, some rigourous mathematical results establishing sufficient conditions for the existence of the Vlasov limit. It has been proven notably [9-11] that the Vlasov equation is valid on times scales of order ∼ logN times the dynamical time, for strictly bound pair potentials decaying at large separations r slower than r -( d -2) . On the other hand, both results of numerical study and various theoretical approaches, based on different approximations or assumptions, suggest that much weaker conditions are sufficient, and the timescales for the validity of the Vlasov equation can be much longer. In the much studied case of gravity, notably, a treatment originally introduced by Chandrasekhar, [12] and subsequently refined by other authors (see e.g. [13-16]) in which nonVlasov effects are assumed to be dominated by incoherent two body interactions gives a time scale ∼ N/logN times the dynamical time for the validity of the Vlasov equation, at least close to stationary solutions representing quasi-stationary states, and this in absence of a regu-
larisation of the singularity in the two-body potential. Theoretical approaches in the physics literature derive the Vlasov equation and kinetic equations describing corrections to it (for a review see[1, 17] either within the framework of the BBGKY hierarchy[18] or starting from the exact Klimontovich equation for the N body system [19, 20]). These approaches are both widely argued (see e.g. [1, 4, 17, 21, 22]), to lead generally to lifetimes of quasi-stationary states of order ∼ N times the dynamical time for any softened pair potential , except in the special case of spatially homogeneous quasi-stationary states one dimension.
In this article we address the question of the validity of Vlasov dynamics using an approach starting from the exact Klimontovich equation. Instead of considering, as is often done (see e.g. [1, 4, 17]), an average over an ensemble of initial conditions to define a smooth one particle phase density, we follow an approach (described e.g. in [23]) in which such a smoothed density is obtained by performing a coarse-graining in phase space. This approach gives the Vlasov equation for the coarse-grained phase space density when certain terms are discarded. We then study how the latter 'non-Vlasov' terms depend on the particle number N , and on the scales introduced by the coarse-graining. In particular we develop this study analyzing the dependence on the large and small distance behavior of the two body potential. Our analysis leading to the scaling behaviours of these terms is based only one very simple - but physically reasonable - hypothesis that we can neglect all correlations in the (microscopic) N body configurations other than those coming from the mean (coarse-grained) phase space density. The main physical result we highlight is that, under this simple hypothesis, the coarse-grained dynamics of an interacting N -particle system shows a very different dependence on the pair interaction at small scale depending on how fast the interaction decays at large distances: for interactions of which the pair force is integrable at large scales the coarse-grained N body dynamics is highly sensitive to how the potential is softened at much smaller scales, while for pair forces which are non-integrable the opposite is true. Correspondingly, while the Vlasov limit may be obtained for any pair interaction which is softened suitably at small scales, the conditions on the short-scale behaviour of the interaction are very different depending on whether its large scale behaviour is in one of of these two classes. This result provides a more rigorous basis for a 'dynamical classification' of interactions as long-range or short-range, which has been introduced on the basis of simple considerations of the probability distribution of the force on a random particle in a uniform particle distribution in [24], and found also in [25] to coincide with a classification based on the dependence on softening of collisional time scales using a generalisation of the analysis of Chandrasekhar for the case of gravity.
The article is organized as follows. In the next section we derive the equation for the coarse-grained phase space density and write in a simple form the non-Vlasov terms which our subsequent analysis focusses on. In section II we first explain our central hypothesis concerning the N -body dynamics, and then apply it to evaluate the statistical properties of the non-Vlasov terms. In the following section we then determine the scaling behaviours of these expressions, i.e., how they depend parametrically on the relevant parameters introduced, and in particular on the two parameters characterising the two body interaction - its large scale decay and the scale at which it is softened. In section IV we use these expressions to identify the dominant contributions to the non-Vlasov terms, which turn out to differ depending on how rapidly the interaction decays at large scales. In the following section we present more complete exact results for the one dimensional case and the comparison with a simple numerical simulation. We then summarize our results and conclusions, discussing in particular the central assumptions and the dependence of our findings on them.
## I. A VLASOV-LIKE EQUATION FOR THE COARSE-GRAINED PHASE SPACE DENSITY
In this section we summarize an approach used to justify the validity of the Vlasov equation for long-range interacting systems. The approach involves using a coarsegraining, in phase space, of the full N body dynamics and leads to an evolution equation for the coarse grained phase space density which consists of the Vlasov terms, plus additional terms. This equation, and the specific form of the non-Vlasov terms we derive, is the starting point for our analysis in the subsequent sections. We follow closely at the beginning the presentation and notation of [23].
We consider a d -dimensional system of N particles of identical mass m = 1 interacting only through the a generic two body force, denoting g x ( ) the force on a particle at x exerted by another one at the origin.
At any time t , the N particles have phase space positions which we denote { ( x i , v i ) } i =1 ..N , and the microscopic (or fine-grained, or Klimontovich) one particle phase-space density is simply the distribution
$$f _ { k } ( { \mathbf x }, { \mathbf v }, t ) = \sum _ { i = 1 } ^ { N } \delta ( { \mathbf x } - { \mathbf x } _ { i } ( t ) ) \delta ( { \mathbf v } - { \mathbf v } _ { i } ( t ) ) \,. \quad ( 1 )$$
Likewise the microscopic density in one particle coordinate space is
$$n _ { k } ( { \mathbf x } ) = \int f _ { k } d ^ { d } { \mathbf v } = \sum _ { i = 1 } ^ { N } \delta ( { \mathbf x } - { \mathbf x } _ { i } ) \quad \quad ( 2 )$$
The full evolution of the N body system can be written in the form of the so-called Klimontovich equation for the microscopic phase space density:
$$\frac { \partial f _ { k } } { \partial t } + \mathbf v \frac { \partial f _ { k } } { \partial \mathbf x } + \mathbf F [ n _ { k } ] ( \mathbf x ) \frac { \partial f _ { k } } { \partial \mathbf v } = 0. \quad \ \ ( 3 )$$
where
$$\mathbf F [ n _ { k } ] ( x ) = \int _ { \Omega } g ( x - x ^ { \prime } ) n _ { k } ( x ^ { \prime } ) d ^ { d } x ^ { \prime } = \sum _ { i = 1 } ^ { N } g ( x - x _ { i } ) \ ( 4 ) \quad \xi _ { \mathbf F } ( x$$
is the exact force at point x (due to all other particles). The only assumption made in deriving this equation from the equations of motion of the individual particles is that the force g x ( ) is bounded as x → 0.
Introducing a top-hat window function W ( z = z , . . . , z 1 d ),
$$W ( z _ { 1 }, \dots, z _ { d } ) = \begin{cases} \ 1, & \text{if} \, | z _ { i } | < \frac { 1 } { 2 }, \, \forall i = 1 \dots d \\ 0 & \text{otherwise} \end{cases} \quad ( 5 ) \quad \text{arour}$$
we define the coarse-grained phase space density:
$$f _ { 0 } ( \mathbf x, \mathbf v, t ) = & \int \frac { d ^ { d } \mathbf x ^ { \prime } } { \lambda _ { x } ^ { d } } \frac { d ^ { d } \mathbf v ^ { \prime } } { \lambda _ { v } ^ { d } } W \left ( \frac { \mathbf x - \mathbf x ^ { \prime } } { \lambda _ { x } } \right ) W \left ( \frac { \mathbf v - \mathbf v ^ { \prime } } { \lambda _ { v } } \right ) f _ { k } ( \mathbf x ^ { \prime }, \mathbf v ^ { \prime }, t ) } { ( 6 ) } \quad \begin{array} { c } \text{is the} \\ \text{act for} \\ \text{force} \end{array} \end{array} \right )$$
where λ x and λ v are the characteristic sizes of the coarsegraining cell in position and, respectively, velocity space. We will denote by C ( x v , ) the coarse graining cell centred at ( x v , ), which thus has a phase space volume λ d x λ d v , and
$$N _ { c } ( { \mathbf x }, { \mathbf v }, t ) = \lambda _ { x } ^ { d } \lambda _ { v } ^ { d } f _ { 0 } ( { \mathbf x }, { \mathbf v }, t ) \quad \quad ( 7 ) \quad \text{coa} _ { \text{ing} }$$
is the number of particles in C ( x v , ). We suppose that the coarse graining cell is always much smaller (in both real and velocity space) than the characteristic size of the system, but sufficiently large to contain a large number of particles, i.e., λ x glyph[lessmuch] L x and λ v glyph[lessmuch] L v where L x and L v the characteristic size of the system in coordinate and velocity space respectively, and
$$1 \ll N _ { c } ( { \mathbf x }, { \mathbf v }, t ) \ll N \,. \quad \quad ( 8 ) \quad _ { \ast \varepsilon } ^ { \text{ res} }$$
We define also the coarse-grained spatial density as
$$n _ { 0 } ( \mathbf x, t ) = \int d ^ { d } \mathbf v f _ { 0 } ( \mathbf x, \mathbf v, t ) = \int \frac { d ^ { d } x ^ { \prime } } { \lambda _ { x } } W \left ( \frac { \mathbf x - x ^ { \prime } } { \lambda _ { x } } \right ) n _ { k } ( \mathbf x ^ { \prime } ) \quad \text{relax} \quad \text{$\text{$num$,} \\ \text{$\text{$and$es}$} \\ \text{$\text{$\text{$finite}$} } \end{matrix}$$
By integrating Eq. (3) over the coarse-graining cell we obtain straightforwardly the following equation:
$$\frac { \partial f _ { 0 } } { \partial t } + \mathbf v \cdot \frac { \partial f _ { 0 } } { \partial \mathbf x } + \mathbf F _ { 0 } ( \mathbf x ) \cdot \frac { \partial f _ { 0 } } { \partial \mathbf v } = - \frac { \partial } { \partial \mathbf x } [ f _ { 0 } \, \xi _ { \mathbf v } ] - \frac { \partial } { \partial \mathbf v } [ f _ { 0 } \, \xi _ { \mathbf F } ] \ ( 1 0 )$$
where
$$\mathbf F _ { 0 } ( \mathbf x, t ) = \int _ { \Omega } \mathbf g ( \mathbf x - \mathbf x ^ { \prime } ) n _ { 0 } ( \mathbf x ^ { \prime }, t ) d ^ { d } \mathbf x ^ { \prime } \,, \quad ( 1 1 ) \quad \text{$\mathbf W$} \\ \text{$\text{ble $\mathfrak t$} }$$
i.e., the force at the point x due to the coarse-grained distribution (which we will identify as the mean-field force). Furthermore
$$\xi _ { v } ( { \mathbf x }, { \mathbf v }, t ) = \left [ \frac { 1 } { N _ { c } ( { \mathbf x }, { \mathbf v }, t ) } \sum _ { i \in \mathcal { C } ( { \mathbf x }, { \mathbf v } ) } { \mathbf v } _ { i } ( t ) \right ] - { \mathbf v } \quad ( 1 2 ) \quad \text{$\text{$two$} } \quad \text{$two$} \quad \text{$paid$}$$
where v i is the velocity of particle i , and
$$\mathcal { + } ( 4 ) \quad \xi _ { \mathbf F } ( \mathbf x, \mathbf v, t ) = \left [ \frac { 1 } { N _ { c } ( \mathbf x, \mathbf v, t ) } \sum _ { i \in \mathcal { C } ( \mathbf x, \mathbf v ) } \mathbf F _ { i } ( t ) \right ] - \mathbf F _ { 0 } ( \mathbf x, t ) \ ( 1 3 )$$
where F i is the exact force acting on the particle i , i.e., F i = F x ( i , t ) where
$$\mathbf F ( { \mathbf x }, t ) = \int _ { \Omega } g ( { \mathbf x } - { \mathbf x } ^ { \prime } ) n _ { k } ( { \mathbf x } ^ { \prime }, t ) d ^ { d } { \mathbf x } ^ { \prime } \,. \quad ( 1 4 )$$
Thus ξ v is the 'velocity fluctuation' in the cell ( x v , ) around the coarse-grained velocity v , i.e., difference between the arithmetic mean of particles velocities in the cell (i.e. the velocity of the center of mass) and the velocity at the centre of the coarse-graining cell, and ξ F ( x v , , t ) is the 'force fluctuation' around the coarse-grained force, i.e., the difference between the arithmetic mean of the exact forces acting on each particle in the cell (equal to the force on the centre of mass of the particles in the cell) and the force at the centre of the cell due to the coarsegrained particle distribution.
If the right hand side of Eq. (10) is set equal to zero, we obtain, given Eq. (11), the Vlasov equation for the coarse-grained phase space density f 0 ( x v , , t ). Establishing the validity of the Vlasov equation in an appropriate limit thus requires showing that the terms on the righthand side may indeed be taken to zero in this limit. For a real system, for which N is finite and the typical number of particles in a coarse-graining cell is finite, Eq. (10) is not closed for the coarse-grained phase space density, but rather coupled to the fine-grained density through the terms on the right-hand side. If it is possible to define the Vlasov limit for the system, these terms will represent at any finite (but large) N , small perturbations to the pure Vlasov evolution of the coarse-grained distribution associated with the 'graininess' of the system which, under suitable hypothesis, are responsible for the relaxation of the system to the thermodynamic equilibrium. In the rest of this article we focus on these terms, and establish their scaling with N (or, equivalently, as a function of the characteristic scales of the coarse-graining cell), given certain simplifying hypotheses.
## II. STATISTICAL EVALUATION OF THE FINITEN FLUCTUATING TERMS
We now focus on the two terms on the right-hand side of Eq. (10). Their direct evaluation is clearly impossible as in principle it requires knowledge of the full finegrained phase space density. We can, however, determine how they scale as a function of relevant parameters by using a statistical approach: given a coarse-grained distribution f 0 ( x v , ) we can consider the fine-grained distribution to be characterized by an ensemble of realizations of particle distributions having f 0 ( x v , ) as mean density. If we know the statistical properties of this ensemble, we
can then, in principle, calculate those of the fluctuating terms on the right-hand side of Eq. (10). In particular, we can then consider how the amplitudes of these statistical quantities depend on the relevant parameters.
We make here the most simple possible hypothesis about this ensemble for the fine-grained phase space density: we suppose that it corresponds to the ensemble of realizations of an inhomogeneous Poissonian point process with mean density given by the coarse-grained phase space density f 0 ( x v , ). In other words we assume that the particles are randomly distributed, without any correlation, inside each coarse-grained cell, with a mean density which varies from cell to cell following f 0 ( x v , ). The density-density correlations are thus fully described by the one point distribution f 0 ( x v , ), and all other correlations, of the fluctuations around this mean density, are neglected. Physically this means we retain the 'pure' finite N (discreteness) effects arising from the fluctuations of the mean density, but neglect the correlation of these fluctuations. Alternatively one can consider that we proceed by assuming complete ignorance of the distribution of particles below the coarse-graining scale, other than the information furnished about it by the coarse-grained density itself. This allows us in effect to close, at least in a statistical formulation, Eq. (10) for f 0 ( x v , ). Indeed this hypothesis is arguably the most natural one to make in seeking to obtain a criterion for the validity of the Vlasov equation which is based only on the coarse graining density f 0 ( x v , ) and the properties of the pair interaction itself.
Formally we can state our assumption to be that the relevant terms can be evaluated by considering an ensemble of realizations of a point process with the N particle probability distribution in phase space given by (see, e.g., [26])
$$\mathcal { P } _ { N } ( \mathbf x _ { 1 }, \mathbf v _ { 1 } ;.. ; \mathbf x _ { N }, \mathbf v _ { N } ) = \prod _ { i = 1 } ^ { N } \frac { f _ { 0 } ( \mathbf x _ { i }, \mathbf v _ { i } ) } { N } \quad ( 1 5 ) \quad \stackrel { \infty } { \mathcal { C } ( \mathbf x, } }$$
assuming the coarse-grained phase space density f 0 ( x v , ) to be a smooth function. In practice it is convenient to perform the calculation with finite coarse-grained cells in which f 0 ( x v , ) is fixed, and the particles are distributed randomly in each coarse-grained cell.
## A. Mean and variance of ξ v
We first evaluate the mean and variance of ξ v , as defined by Eq. (12), assuming now the defined properties of the ensemble of realizations. Given that the latter assigns randomly the velocities of the particles in the coarse-grained cell, which is centred at ( x v , ), we evidently have
$$\left \langle \frac { 1 } { N _ { c } ( { \mathbf x$$
where 〈 ... 〉 denotes the ensemble average. Therefore we have, as would be expected,
$$\langle \xi _ { \mathbf v } \rangle = 0 \,.$$
Calculating the variance of ξ v gives straightforwardly that
$$\langle \xi _ { \mathbf v } ^ { 2 } \rangle = \frac$$
where we can take N c = f 0 ( x v , ) λ λ d x d v to be the average number of particles in the cell, given that N c = 〈 N c 〉 + δN c where by hypothesis δN c ∼ √ N c glyph[lessmuch] N c 1 . Given that N c ( x v , ) is large, and v i are considered independent and identically distributed variables, ξ v is thus simply a Gaussian random variable with mean zero and variance given by Eq. (18).
## B. Mean and variance of ξ F
Let us evaluate now the first two moments of ξ F . To evaluate these quantities in the ensemble defined above, it is useful first to note that
$$\begin{array} {$$
where the i runs over the N c particles inside the coarsegraining cell C ( x v , ), and the index I over the other N -N c particles outside the cell . This is the case because we assume g ( -x ) = -g x ( ) (making the sum over all the mutual forces of the pairs of particles in the cell C ( x v , ) vanishes). As the individual pair forces depend only on the spatial positions of particles, we need only specify, to calculate the ensemble average of powers of the force, the probability distribution for the spatial positions of these points. Given the writing of the force in (19), in which the sum is performed separately over the particles inside and outside the coarse-grained cell C ( x v , ) considered, it is convenient to write the ensemble average in a similar form. As noted above we can take N c to be fixed and equal to its average value, n 0 ( x ) λ d x , up to negligible corrections of order 1 / √ N c . We can then write the N particle probability distribution as
$$\stackrel { \cdot } { \text{rties} } } _ { \text{$$
1 In a uniform Poisson process with mean density n 0 , the PDF P N,V ( ) of the number of particles N in a volume V is P N,V ( ) = ( n V o ) N e -n V 0 /N ! (see, e.g., [26])
where p ( x ) is the one-point probability distribution function of the spatial position of a particle given the condition that it is contained in the cell, and ˆ( p y ) is the onepoint probability distribution of the spatial position of a particle given the condition that it is outside the cell. Given that the N c particles are randomly distributed inside the cell, we have simply
$$p ( { \mathbf x } _ { i } ) = \begin{cases} \lambda _ { x } ^ { - d }, & { \mathbf x } \in \mathcal { S } ( { \mathbf x } ) \\ 0, & \text{otherwise} \end{cases} \quad \quad ( 2 1 ) \quad \end{cases}$$
where S ( x ) is the support of W ( x /λ x ), i.e., the 'stripe' in phase space with the same spatial coordinates as the phase space cell C ( x v , ) (see Fig. 1). Assuming the number of particles in the cell N c to be small compared to the total number of particles N s ( x v , ) in this stripe (which is itself small compared to the total number of particles N ), ˆ( p y ) can be approximated everywhere simply by the unconditional one point PDF for the spatial distribution obtained by integrating (15) over all but one space coordinate, i.e.,
$$\hat { p } ( \mathbf y ) = \frac { n _ { 0 } ( \mathbf y ) } { N } \quad \quad \quad ( 2 2 ) \quad \ \ t a i \, \\ \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
$$1. \ M e a n \ o f \, \xi _ { F }$$
Using (20) to calculate the ensemble average of the exact force exerted by all particles on those in a coarsegrained cell, we have
$$\text{grained cell, we have} \\ \langle \bar { F } ( x, \boldsymbol v, t ) \rangle & = ( N - N _ { c } ) \int _ { \Omega } d ^ { d } x ^ { \prime } \int _ { \Omega } d ^ { d } y ^ { \prime } p ( x ^ { \prime } ) \hat { p } ( \boldsymbol y ^ { \prime } ) g ( x ^ { \prime } - \boldsymbol y ^ { \prime } ) \text{ once of } t t } \\ & = \int _ { S ( \boldsymbol x ) } \frac { d ^ { d } x ^ { \prime } } { \lambda _ { x } ^ { d } } \int _ { \Omega } d ^ { d } \boldsymbol y ^ { \prime } n _ { 0 } ( \boldsymbol y ^ { \prime } ) \boldsymbol g ( x ^ { \prime } - \boldsymbol y ^ { \prime } ) } \\ & = \int _ { S ( \boldsymbol x ) } \frac { d ^ { d } x ^ { \prime } } { \lambda _ { x } ^ { d } } \boldsymbol F _ { 0 } ( \boldsymbol x ^ { \prime } ) \text{ (2)} \\ \text{Thus}$$
Thus
$$\langle \xi _ { \mathbf F } ( { \mathbf x }, { \mathbf v }, t ) \rangle = \int _ { \mathcal { S } ( { \mathbf x } ) } \frac { d ^ { d } { \mathbf x } ^ { \prime } } { \lambda _ { x } ^ { d } } [ { \mathbf F } _ { 0 } ( { \mathbf x } ^ { \prime } ) - { \mathbf F } _ { 0 } ( { \mathbf x } ) ] \,. \quad ( 2 4 )$$
Assuming that we can neglect the variation of the meanfield F 0 in the coarse-graining cell, we obtain
$$\langle \xi _ { \text{F} } ( \mathbf x, \mathbf v, t ) \rangle = 0$$
2. Variance of ξ F
We now calculate
$$\left \langle \xi _ { F } ^ { 2 } ( { \mathbf x }, { \mathbf v }, t ) \right \rangle = \left \langle \bar { \mathbf F } ^ { 2 } ( { \mathbf x }, t ) \right \rangle - { \mathbf F } _ { 0 } ^ { 2 } ( { \mathbf x }, t ) \quad ( 2 6 ) \quad \text{cs} _ { \colon m t \sim. }$$
We first break 〈 ξ 2 F ( x v , , t ) 〉 into two parts glyph[negationslash]
$$\langle \bar { \mathbf F } ^ { 2 } ( { \mathbf x }, t ) \rangle = \left \langle \frac { 1 } { N _ { c } ^ { 2 } } \sum _ { i } \mathbf F ^ { 2 } ( { \mathbf x } _ { i } ) \right \rangle + \left \langle \frac { 1 } { N _ { c } ^ { 2 } } \sum _ { i, j ; j \neq i } \mathbf F ( { \mathbf x } _ { i } ) \cdot \mathbf F ( { \mathbf x } _ { j } ) \right \rangle ^ { \text{give} } ^ { \text{} } \text{.}$$
FIG. 1. (Color online) Schema, in one dimension, of the coarse-graining of phase space, showing also the different regions into which the domain of integration is divided for our analysis of the integrals in the force variance: C is the coarsegrained cell with centre at ( x v , ), S (in white) is the 'stripe' enclosing all points in [ x -λ x 2 , x + λ x 2 ]; ∂ ε S is the region containing particles within a distance ε of the boundary of the stripe.
where all the sums are over the N c particles in the cell C ( x v , ). We will refer to the first term on the right hand side as the 'diagonal' contribution, and the second term as the 'off-diagonal' contribution to the variance: the first is the contribution to the variance due to the variance of the force on each particle of the cell, the second the contribution to the variance arising from the correlation of the forces on different particles in the cell.
Further we can use Eq. (19) to rewrite each term as two terms glyph[negationslash]
glyph[negationslash]
glyph[negationslash]
$$\text{two terms} ^ { \cdot \cdot \cdot \cdot } ( 2 3 ) \Big < \frac { 1 } { N _ { c } ^ { 2 } } \sum _ { i } \sum _ { I } g ^ { 2 } ( x _ { i } - x _ { I } ) \Big > \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^{\cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{mean} ^ { \cdot } \\ \text{median} } \text{one sums over the } N _ { c } \text{ particles inside}$$
glyph[negationslash]
where i, j again denote sums over the N c particles inside the coarse-grained cell and I, J over the N -N c particles outside the cell. Performing the ensemble average by integrating over the PDF (20), the result can be conveniently divided in two parts. The first and third terms give
$$\int _ { \Omega } \frac { d ^ { d } { \mathbf x } ^ { \prime } } { \lambda _ { x } ^ { d } } n _ { 0 } ( { \mathbf x } ^ { \prime } ) \int _ { \mathcal { S } ( { \mathbf x } ) } d ^ { d } { \mathbf y } \, { \mathbf g } ^ { 2 } ( { \mathbf y } ^ { \prime } - { \mathbf x } ^ { \prime } ) \quad \ \ ( 2 8 )$$
and
$$\left ( 1 - \frac { 1 } { N _ { c } ( { \mathbf x }, { \mathbf v } ) } \right ) \int _ { \Omega } \frac { d ^ { d } { \mathbf x } ^ { \prime } } { \lambda _ { x } ^ { d } } n _ { 0 } ( { \mathbf x } ^ { \prime } ) \left [ \int _ { \mathcal { S } ( { \mathbf x } ) } d ^ { d } { \mathbf y } ^ { \prime } g ( { \mathbf y } ^ { \prime } - { \mathbf x } ^ { \prime } ) \right ] ^ { 2 } \quad \text{the c} } \text{ on ho} \\ \text{non-trivial} \quad \text{Both the second and fourth term can be} \quad \text{it is e}$$
respectively. Both the second and fourth terms can be expressed purely in terms of the mean-field, as
$$\frac { 1 } { N _ { c } } \left ( 1 - \frac { 1 } { N } \right ) \int _ { \mathcal { S } ( \mathbf x ) } \frac { d ^ { d } x ^ { \prime } } { \lambda _ { x } ^ { d } } \mathbf F _ { 0 } ^ { 2 } ( \mathbf x ^ { \prime } ) \quad \quad ( 3 0 ) \quad \text{ing} _ { \text{for} }$$
and
$$\left ( 1 - \frac { 1 } { N _ { c } } \right ) & \left ( 1 - \frac { 1 } { N } \right ) \int _ { \mathcal { S } ( \mathbf x ) } \frac { d ^ { d } \mathbf x _ { 1 } } { \lambda _ { x } ^ { d } } \int _ { \mathcal { S } ( \mathbf x ) } \frac { d ^ { d } \mathbf x _ { 2 } } { \lambda _ { x } ^ { d } } \mathbf F _ { 0 } ( \mathbf x _ { 1 } ) \cdot \mathbf F _ { 0 } ( \mathbf x _ { 2 } ) \right ) \quad \varepsilon. \\ \text{constraint, and, }$$
respectively.
Assuming again that we can neglect the variation of the mean-field F 0 in the coarse-graining cell, we can perform the integrals in the last two expressions, and then obtain
$$& \text{the integrals in the last two expressions, and then obtain } \quad \begin{array} { c } N \stackrel { \alpha \infty } { \longrightarrow } \\ \left \langle \xi _ { F } ^ { 2 } ( x, v ) \right \rangle = \frac { 1 } { N _ { c } ( x, v ) } \int _ { \Omega } d ^ { d } x ^ { \prime } n _ { 0 } ( x ^ { \prime } ) \int _ { \mathcal { S } ( x, \lambda ) } ^ { d } y \lambda _ { x } ^ { - d } g ^ { 2 } ( y ^ { \prime } - x ^ { \prime } ) \end{array} \quad \begin{array} { c } \text{that tl} \\ \text{cedure} \\ } \\ + ( 1 - \frac { 1 } { N _ { c } ( x, v ) } ) \int _ { \Omega } d ^ { d } x ^ { \prime } n _ { 0 } ( x ^ { \prime } ) \left [ \int _ { \mathcal { S } ( x, \lambda ) } ^ { d } y ^ { \prime } \lambda _ { x } ^ { - d } g ( y ^ { \prime } - x ^ { \prime } ) \right ] ^ { 2 } \quad \begin{array} { c } \left \langle \xi _ { v } ^ { 2 } \right \rangle, \, 1 \\ \text{in ord} \\ \text{assum} \\ \text{scale} \, \text{o} \\ } \\ - \frac { 1 } { N } \mathbf F _ { 0 } ^ { 2 } ( x ) \quad & ( 3 2 ) \text{ \text{ \text{acrix} } \\ \text{Our analysis below will focus essentially on the first } \quad \text{space} \, \end{array} \,.$$
Our analysis below will focus essentially on the first two terms in this expression as they are those which describe the contribution to the fluctuating terms which are potentially sensitive to the small scale properties of the pair force g x ( ). We note that the first term comes from the 'diagonal' part of the variance, and more specifically it represents the contribution to the variance arising from the force on a single particle in the cell due to a particle outside the cell: we will thus refer to it as the two body contribution. The second integral in (32), on the other hand, arises from the 'off-diagonal' part of the variance, and more specifically it represents the contribution to the variance of the force on the cell due to the correlation of the force exerted on two particles inside the cell exerted by a particle outside the the cell: we will refer to it therefore as the three body contribution.
## III. PARAMETRIC DEPENDENCE OF THE FLUCTUATIONS
We now analyse the expressions we have obtained, focussing on how their value depends parametrically on the relevant parameters we have introduced, notably the number of particles in the system ( N ), its size ( L x , L v ), the number of particles in the coarse-graining cells ( N c ) and the coarse-graining scales ( λ x , λ v ). Further we will take the pair force to be given by g x ( ) = g | x | -γ ˆ for x γ < 0, and by
$$g ( x ) = g \begin{cases} \left | x \right | ^ { - \gamma } \hat { x }, & | x | \geq \varepsilon \\ \varepsilon ^ { - \gamma } \hat { x }, & | x | < \varepsilon \end{cases} \quad \ \ ( 3 3 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
for γ > 0, where g is the coupling constant (and g < 0 for the case of an attractive interaction). We will focus then on how the amplitude of the fluctuations of the force depend also on the exponent γ and the characteristic length ε at which the force is regularized at small scales. Indeed it is evident that we must introduce such a regularization of the pair force, as without it the integrals in (32) can be ill-defined. It is important to note that our essential scaling results are not dependent on the use of the specific form of the regularization in Eq. (33): what is necessary to obtain these results is only that the force be bounded above by a constant of order ε -γ below a scale of order ε .
## A. Mean field Vlasov limit
The mean-field Vlasov limit is formulated by taking N → ∞ at fixed system size, and scaling g ∝ N -1 so that the mean field F 0 remains fixed. Applying this procedure to the expressions we have obtained for 〈 ξ 2 v 〉 and 〈 ξ 2 v 〉 , both indeed converge to zero, for any non-zero ε . In order to obtain these expressions we have, as noted, assumed also that the mean-field does not vary on the scale of the coarse-graining cell. Assuming that the characteristic scale of variation of the coarse-grained phase space density, and mean field, is the system size L x , for a finite coarse-graining cell we expect corrections to our expressions due to the variation of these quantities on the scale λ x which are suppressed at least by λ /L x x relative to those we have calculated. The coarse-grained phase space density f 0 ( x v , , t ) thus indeed obeys the Vlasov equation in the usual formulation of this limit, when the size of the coarse-graining cell is taken to be negligible with respect to the system size.
We now study more closely the approach to the Vlasov limit as characterized by the scaling behaviour of the corrections to it. We assume that these are given by those of the statistical quantities we have calculated, i.e., we take
$$| \xi _ { \mathbf v } | \sim \sqrt { \langle \xi _ { \mathbf v } ^ { 2 } \rangle } \,, \ | \xi _ { \mathbf F } | \sim \sqrt { \langle \xi _ { \mathbf F } ^ { 2 } \rangle } \, \quad \quad ( 3 4 )$$
## B. Velocity fluctuations
Given the result (18) and that N c ∼ N L L d x d v λ λ d x d v we infer
$$| \xi _ { \mathbf v } | \sim \frac { 1 } { \sqrt { N } } \left ( \frac { L _ { x } } { \lambda _ { x } } \right ) ^ { d / 2 } \left ( \frac { L _ { v } } { \lambda _ { v } } \right ) ^ { d / 2 } \lambda _ { v } \quad \ \ ( 3 5 )$$
There is evidently no dependence on the pair force in this term. As already noted we can recover the Vlasov limit by taking N →∞ . Further we can see that this result remains valid for arbitrarily small (but non-zero) values of the ratio λ x L x so that variation of coarse-grained quantities on the coarse-grained scale can indeed be neglected.
## C. Force fluctuations
The scaling of the last term on the right-hand side (32) is already explicited, representing simply a fluctuation of the force about the mean field of order 1 / √ N times the mean field itself. In order to determine the dependences of the first two terms we need to analyse carefully that of the integrals. To do so we divide the domains of integration in the double integral into appropriate subdomains, isolating the region which may depends on the lower cutoff ε in the pair force.
## 1. 2 body contribution
We consider first the diagonal two body contribution, writing the double integral as
$$\text{writing the double integral as} & \\ & \left [ \int _ { \Omega / S / \partial _ { \varepsilon } \mathcal { S } } d ^ { d } x ^ { \prime } + \int _ { \partial _ { \varepsilon } \mathcal { S } } d ^ { d } x ^ { \prime } + \int _ { S / \partial _ { \varepsilon } \mathcal { S } } d ^ { d } x ^ { \prime } \right ] \otimes \\ & \left ( n _ { 0 } ( x ^ { \prime } ) \int _ { \mathcal { B } ( x ^ { \prime }, \varepsilon ) } d ^ { d } y ^ { \prime } \lambda _ { x } ^ { - d } g ^ { 2 } ( y ^ { \prime } - x ^ { \prime } ) \\ & + n _ { 0 } ( x ^ { \prime } ) \int _ { \mathcal { S } ( x ) / \mathcal { B } ( x ^ { \prime }, \varepsilon ) } d ^ { d } y \lambda _ { x } ^ { - d } g ^ { 2 } ( y ^ { \prime } - x ^ { \prime } ) \right ) \quad ( 3 6 ) \quad \Pr \, \text{crout} \\ \text{where in this context} \otimes \text{indicates the interaction onera} -$$
where in this context ⊗ indicates the integration operation on the terms in square brackets. As illustrated for the one dimensional case in Fig. 1, the integral over x ′ , over the whole of space (Ω), has been divided into three parts:
- · ∂ ε S : the set of points which are within a distance ε of the boundary of the stripe S . The volume of this region is of order ελ d -1 x .
- · S /∂ ε S : the set of points belonging to S but not belonging to ∂ ε S . For ε glyph[lessmuch] λ x , its volume is of order λ d x .
- · Ω / S /∂ ε S : the rest of space.
and the integral over y ′ , over the stripe S has been divided into two parts:
- · B ( x ′ , ε ): the set of points in S which are a distance of less than ε from the point x ′ . This region has a volume of order ε d .
- · S /B ( x ′ , ε ): the rest of S ; for ε glyph[lessmuch] λ x , its volume is of order λ d x .
We consider now one by one the terms in the integral written as in (36). The first term in the integration over x ′ excludes the region where the pair force is ε -dependent, and is over a volume of order L d x . Hence if we suppose n 0 ∼ const , it gives a contribution which scales as
$$\sim n _ { 0 } L _ { x } ^ { d - 2 \gamma } \quad \quad \quad ( 3 7 ) \quad p \varepsilon$$
For the second region of integration over x ′ , of volume of order ελ d -1 x , the region B ( x ′ , ε ) of the integration over y ′ gives a contribution of order ε -2 γ over a volume of order ε d , and thus
$$\sim n _ { 0 } ( \varepsilon \lambda _ { x } ^ { d - 1 } ) \varepsilon ^ { d } \lambda _ { x } ^ { - d } \varepsilon ^ { - 2 \gamma } \,. \quad \quad ( 3 8 )$$
In the region S ( x ) / B ( x ′ , ε ) on the other hand, of volume of order λ d x , we have
$$\sim n _ { 0 } ( \varepsilon \lambda _ { x } ^ { d - 1 } ) \lambda _ { x } ^ { - 2 \gamma }$$
In the third region of integration over x ′ , of volume of order λ d x , we obtain again a contribution of order ε -2 γ in the volume B ( x ′ , ε ), and thus
$$\sim n _ { 0 } \varepsilon ^ { d - 2 \gamma } \quad \quad \quad ( 4 0 )$$
while the second term region of the y ′ integration gives
$$\sim n _ { 0 } \lambda _ { x } ^ { d - 2 \gamma } \,.$$
## 2. 3 body contribution
Proceeding in the same manner we write the 3-body contribution to the variance of the force as
$$\text{contribution to the variance of the force as} \\ \left [ \int _ { \Omega / S / \partial _ { \epsilon } \mathcal { S } } d ^ { d } x ^ { \prime } + \int _ { \partial _ { \epsilon } \mathcal { S } } d ^ { d } x ^ { \prime } + \int _ { \mathcal { S } / \partial _ { \epsilon } \mathcal { S } } d ^ { d } x ^ { \prime } \right ] \otimes \\ \left ( n _ { 0 } ( x ^ { \prime } ) \int _ { \mathcal { B } ( x ^ { \prime }, \varepsilon ) } d ^ { d } \mathcal { Y } ^ { \prime } \lambda _ { x } ^ { - d } g ( \mathcal { Y } ^ { \prime } - x ^ { \prime } ) \\ + n _ { 0 } ( x ^ { \prime } ) \int _ { \mathcal { S } ( x, \lambda ) / \mathcal { B } ( x ^ { \prime }, \varepsilon ) } d ^ { d } \mathcal { Y } \lambda _ { x } ^ { - d } g ( \mathcal { Y } ^ { \prime } - x ^ { \prime } ) \right ) ^ { 2 } \\ \text{The first term in the information over $\mathcal{ }v^{\prime}$ or $\mathcal{v}$ or $\mathcal{v}$ or $\mathcal{v}$} \text{than} \,.$$
The first term in the integration over x ′ gives then a contribution
$$\sim n _ { 0 } L _ { x } ^ { d - 2 \gamma }$$
Compared to the 2-body integral, the analysis of the remaining parts is essentially the same, except for one important difference: as the integration over y ′ is over the vector pair force, the integral over y ′ is zero when integrated in a sphere around x ′ ; in particular the integration over B ( x ′ , ε ) vanishes when B ( x ′ , ε ) is fully contained in S . This is the case for the integration region S /∂ ε S , which therefore does not depend on ε and simply gives a contribution of order
$$\sim n _ { 0 } \lambda _ { x } ^ { d - 2 \gamma } \,.$$
For the second integration region in the integral over x ′ , of volume ∼ ελ d -1 x , B ( x ′ , ε ) is not fully contained in S and we have therefore a contribution from this part of the integration over y ′ of order ε λ d -d x ε -γ , while the second part, which does not depend on ε , gives a contribution
of order λ -γ x . Taking the square and multiplying by the volume of ∂ ε S , we obtain three terms:
$$\sim n _ { 0 } \varepsilon \lambda _ { x } ^ { d - 1 } \varepsilon ^ { 2 d } \lambda _ { x } ^ { - 2 d } \varepsilon ^ { - 2 \gamma } \quad \quad ( 4 3 ) \quad \text{the}$$
from the square of the first term,
$$\sim n _ { 0 } \varepsilon \lambda _ { x } ^ { d - 1 } \varepsilon ^ { d } \lambda _ { x } ^ { - d } \varepsilon ^ { - \gamma } \lambda _ { x } ^ { - \gamma } \quad \quad ( 4 4 ) \quad \text{we}$$
from the cross term, and
$$\sim n _ { 0 } \varepsilon \lambda _ { x } ^ { d - 1 } \lambda _ { x } ^ { - 2 \gamma } \quad \quad ( 4 5 ) \quad \ w \hbar$$
from the square of the second term.
## IV. FORCE FLUCTUATIONS ABOUT THE VLASOV LIMIT: DEPENDENCE ON ε
Gathering together the expressions derived above, we obtain, keeping only the leading divergence in ε in each of the 2-body and 3-body contributions,
$$& \text{or use $z$-\inftyy$ and $\sigma$-\inftyy$ continuous}, & \quad \Im F \cdot \\ & \langle \xi _ { F } ^ { 2 } ( x, v ) \rangle = & \frac { g ^ { 2 } } { N _ { c } } ( C _ { \Omega } n _ { 0 } L _ { x } ^ { d - 2 \gamma } + C _ { S } n _ { 0 } \lambda _ { x } ^ { d - 2 \gamma } + C _ { \partial S } n _ { 0 } \varepsilon ^ { d - 2 \gamma } ) \quad \text{This that $t$} \\ & + g ^ { 2 } ( 1 - \frac { 1 } { N _ { c } } ) \left ( C _ { \Omega } ^ { \prime } n _ { 0 } L _ { x } ^ { d - 2 \gamma } + C _ { S } ^ { \prime } n _ { 0 } \lambda _ { x } ^ { d - 2 \gamma } \quad \text{$sub-ca$} \\ & + C _ { \partial S } ^ { \prime } n _ { 0 } \varepsilon ^ { 2 d + 1 - 2 \gamma } \lambda _ { x } ^ { - ( d + 1 ) } \right ) - \frac { 1 } { N } | \mathbf F [ n _ { 0 } ] ( x ) | ^ { 2 } \quad \\ & \text{where all $C$ and $C$ or one another $N$ to that one has not}$$
where all C ∗ and C ′ ∗ are constants (Note that we have not included (44) because when it diverges, for γ > d +1, the term retained is indeed more rapidly divergent.)
Depending on the values of γ and d different terms dominate. We consider each case.
$$\text{A. } \text{ Case } \gamma < \frac { d } { 2 }$$
In this range there are no divergences as ε → 0, and for ε glyph[lessmuch] λ x glyph[lessmuch] L x , the dominant term from the two integrals is
$$g ^ { 2 } n _ { 0 } L _ { x } ^ { d - 2 \gamma } \sim \frac { 1 } { n _ { 0 } L _ { x } } g ^ { 2 } n _ { 0 } ^ { 2 } L _ { x } ^ { 2 } \sim \frac { 1 } { N } | \mathbf F _ { 0 } | ^ { 2 } \quad ( 4 7 ) \quad \text{(7)} \quad \text{(the 1)} \\ \text{Not} \quad \text{(a)}$$
and therefore we infer the scaling of the total force fluctuation is
$$| \xi _ { \mathbf F } | \sim \frac { 1 } { \sqrt { N } } | \mathbf F _ { 0 } |$$
As noted above we therefore obtain in this case the Vlasov limit taking N → ∞ with g ∼ 1 /N . Further we conclude that, at finite N , the fluctuations around the mean-field force are dominated by contributions coming from fluctuations of the density at the scale of the system size, which dominate those coming both from the scale λ x of the coarse-graining cell and those from the scale ε at which the pair force is regularized.
$$B. \ \text{ Case } \frac { d } { 2 } < \gamma < d + \frac { 1 } { 2 }$$
In this range of γ , there is a divergence at ε → 0 in the contribution coming from the 2-body term, while the 3-body term remains finite. Keeping only the dominant contributions to the two integrals when ε glyph[lessmuch] λ x glyph[lessmuch] L x , we obtain
$$\int _ { \pi } \langle \xi _ { \mathrm F } ^ { 2 } ( \mathbf x ) \rangle \sim C _ { \varepsilon } \frac { g ^ { 2 } n _ { 0 } ( \mathbf x ) } { N _ { c } ( \mathbf x, \mathbf v ) } \varepsilon ^ { d - 2 \gamma } + C _ { \lambda } ^ { \prime } g ^ { 2 } n _ { 0 } ( \mathbf x ) \lambda _ { x } ^ { d - 2 \gamma } \quad ( 4 9 )$$
where C ε and C ′ λ are constants.
Given the divergence in ε we see explicitly that in this case the Vlasov limit is obtained taking N → ∞ with g ∼ N -1 at finite non-zero ε , and this limit can only be defined if such a small scale regularisation of the pair force is introduced.
Using N c ( x v , ) ∝ f 0 ( x v , ) λ λ d x d v , we can write the dominant fluctuations as
$$\L a c h \quad \langle \xi _ { \text{F} } ^ { 2 } ( \mathbf x ) \rangle \sim g ^ { 2 } \frac { n _ { 0 } ( \mathbf x ) } { \lambda _ { x } ^ { d } } \left ( C _ { \varepsilon } \frac { \varepsilon ^ { d - 2 \gamma } } { f _ { 0 } ( \mathbf x, \mathbf v ) \lambda _ { v } ^ { d } } + C _ { \lambda } ^ { \prime } \lambda _ { x } ^ { 2 d - 2 \gamma } \right ) \ ( 5 0 )$$
This expression allows us to conclude, as anticipated, that there is a crucial difference between the following sub-cases: i) the range of γ in which the first term dominates, and ii) the range in which the second term dominates:
$$1. \ F o r \, \frac { d } { 2 } < \gamma < d \, \colon$$
In this case the exponent of λ x in the second term inside the brackets in (50) is positive, and therefore when we increase λ x at fixed ε (and fixed λ v ), this term dominates over the first one. More specifically when λ x glyph[greatermuch] ε f ( 0 ( x v , ) λ ε d v d ) -1 2 d -2 γ this term dominates, and
$$\langle \xi _ { F } ^ { 2 } \rangle \sim g ^ { 2 } n _ { 0 } ( { \mathbf x } ) \lambda _ { x } ^ { d - 2 \gamma }$$
Thus, even though the amplitude of the fluctuations depend on ε , and diverges as ε → 0, for a sufficiently large coarse-graining cell the fluctuations become in practice effectively insensitive to the value of ε , for a wide range of values which is such that the larger is λ x the smaller is the lower limit on ε , with the latter vanishing as λ x →∞ . Note that since
$$e u a s \, \underset { \left | \xi _ { \text{F} } \right | } { \sim } \frac { 1 } { \sqrt { n _ { 0 } ( \mathbf x ) \lambda _ { x } ^ { d } } } \left ( \frac { \lambda _ { x } } { L _ { x } } \right ) ^ { d - \gamma } | \mathbf F _ { 0 } ( \mathbf x ) | \\ \sim \frac { 1 } { \sqrt { N } } \left ( \frac { \lambda _ { x } } { L _ { x } } \right ) ^ { d / 2 - \gamma } | \mathbf F _ { 0 } | \\ \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \end{array}$$
we can also neglect the final term in (32).
$$2. \ F o r \, d < \gamma < d + \frac { 1 } { 2 } \, \colon$$
In this range it is instead the exponent of λ x in the second term inside the brackets in (50) which is negative,
and as a consequence it is now the first term which dominates when we make the coarse-graining scale λ x large. We have therefore
$$\langle \xi _ { \mathrm F } ^ { 2 } \rangle \sim g ^ { 2 } \frac { n _ { 0 } } { f _ { 0 } \lambda _ { x } ^ { d } \lambda _ { v } ^ { d } } \varepsilon ^ { d - 2 \gamma } \,. \quad \quad \ ( 5 2 ) \quad g ^ { 2 } \, _ { \mu }$$
Further as this can be rewritten as
$$| \xi _ { \mathbf F } | \sim \frac { 1 } { \sqrt { N N _ { c } } } \left ( \frac { \varepsilon } { L _ { x } } \right ) ^ { d / 2 - \gamma } | \mathbf F _ { 0 } | \, \quad \ \ ( 5 3 ) \quad + \varepsilon$$
it follows that for sufficiently small ε this is the dominant contribution to the fluctuations. In this range therefore the leading contribution to the force fluctuations is directly dependent on ε .
$$C. \ \ C a s e \ d + \frac { 1 } { 2 } < \gamma$$
In this case both the integrals giving the 2-body and 3-body contributions are divergent at small ε , but the dominant divergence comes from the latter giving
$$\langle \xi _ { F } ^ { 2 } ( { \mathbf x } ) \rangle \sim g ^ { 2 } n _ { 0 } ( { \mathbf x } ) \frac { \varepsilon ^ { 2 d + 1 - 2 \gamma } } { \lambda _ { x } ^ { d + 1 } } \quad \quad ( 5 4 ) \quad \text{ with}$$
This case is therefore like the previous case ( d < γ < d + 1 2 ): the dominant contribution to the fluctuations is divergent as ε → 0.
## V. EXACT ONE DIMENSIONAL CALCULATION AND NUMERICAL RESULTS
In the case of a one dimensional system, d = 1, it is possible to perform explicitly the integrals in (32) to obtain exactly the expression of the force fluctuations. In order to illustrate our main result above, we can compare the expressions we obtain, and in particular their leading scaling behaviours, with what is obtained directly by measuring the force fluctuations in cells on realizations of a homogeneous Poisson particle distribution. As we are interested primarily in the ε dependence of these fluctuations we consider solely the contribution to them from the part of the integral which is potentially sensitive to them. We can therefore write
$$\begin{array} { c } \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c}.\end{array} \end{array} \begin{array} { c } \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text`} \\ \langle \xi _ { \text{F} } ^ { 2 } ( \mathbf x ) \rangle _ { \text{S} } = \frac { n _ { 0 } } { N _ { c } } \int _ { - \frac { \lambda _ { \text{S} } } { 2 } } ^ { \frac { \lambda _ { \text{S} } } { 2 } } } ^ { \frac { \lambda _ { \text{S} } } { 2 } } d x ^ { \prime } \int _ { - \frac { \lambda _ { \text{S} } } { 2 } } ^ { \frac { \lambda _ { \text{S} } } { 2 } } d y ^ { \prime } \lambda _ { x } ^ { - 1 } g ( y ^ { \prime } - x ^ { \prime } ) \end{array} \begin{array} { c } \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \end{array} \end{array} \begin{array} { c } \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text`} \\ \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c.\end{array} } \end{array} \begin{array} { c } \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \vec{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \text{c} \end{array} \end{array}$$
where we use the subscript in 〈 ξ 2 F ( x ) 〉 S to indicate that this is the contribution to the force variance sourced by particles in the phase space 'stripe' S , and
$$g ( x ) = g \begin{cases} \frac { x } { | x | ^ { \gamma + 1 } }, & | x | \geq \varepsilon \\ \frac { x } { | x | } \varepsilon ^ { - \gamma } & | x | < \varepsilon \end{cases} \quad \text{(5.5)} \quad \text{and} \quad \text{the}$$
Integrating we obtain
$$\begin{array} { c } \int \text{dom} _ { \text{ large} } \\ \int \text{large} _ { \text{ the gap} } \\ ( 5 2 ) & g ^ { 2 } \frac { 1 } { N _ { c } } \frac { n _ { 0 } ( x ) } { \lambda _ { x } } \left [ \frac { 1 } { 1 - \gamma } \varepsilon ^ { 2 - 2 \gamma } - \frac { 4 } { 1 - 2 \gamma } \lambda _ { x } \varepsilon ^ { 1 - 2 \gamma } } \\ & + \frac { 1 } { ( 1 - 2 \gamma ) ( 1 - \gamma ) } \lambda _ { x } ^ { 2 - 2 \gamma } \right ] \\ ( 5 3 ) & + g ^ { 2 } ( 1 - \frac { 1 } { N _ { c } } ) \frac { n _ { 0 } ( x ) } { \lambda _ { x } ^ { 2 } } \left [ \frac { 2 \gamma ( 1 + 2 \gamma ) } { 2 ( 1 - \gamma ) ( 3 - 2 \gamma ) } \varepsilon ^ { 3 - 2 \gamma } \\ \int \text{inant} _ { \text{erefore} } \\ \int \text{is di} _ { \text{ is di} } \\ & - \frac { 4 } { ( 1 - \gamma ) ( 2 - \gamma ) ( 3 - \gamma ) } \varepsilon ^ { - \gamma } ( \lambda _ { x } ^ { 3 - \gamma } - ( \lambda _ { x } - \varepsilon ) ^ { 3 - \gamma } } \\ & + \frac { 4 } { ( 1 - \gamma ) ( 2 - \gamma ) } \varepsilon ^ { 1 - \gamma } ( \lambda _ { x } - \varepsilon ) ^ { 2 - \gamma } } \\ & - \frac { 4 } { ( 1 - \gamma ) ^ { 2 } } \left ( \frac { \lambda _ { x } } { 2 } \right ) ^ { 2 - 2 \gamma } \left ( \frac { \lambda _ { x } } { 2 - \varepsilon } \right ) ^ { \infty } _ { n = 0 } \frac { ( \gamma - 1 ) _ { n } } { 1 + 2 n } \left ( 1 - 2 \frac { \varepsilon } { \lambda _ { x } } \right ) ^ { 2 n } \\ \int \text{the} _ { \text{ the} } \\ & + \frac { 2 } { ( 1 - \gamma ) ^ { 2 } ( 3 - 2 \gamma ) } \lambda _ { x } ^ { 3 - 2 \gamma } \right ] \\ \\ ( 5 4 ) & \quad \text{where} \left ( x \right ) _ { n } \text{ is the Pochhammer symbol}, ( x ) _ { n } = \frac { \Gamma ( x ) } { \Gamma ( x ) } = \\ \int \text{$\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\quad $\ \end{array}$$
where ( x ) n is the Pochhammer symbol, ( x ) n = Γ( x + ) n Γ( x ) = x x ( +1) ... ( x + n -1).
Expanding this expression in the limit ε/λ → 0, and keeping only the leading ε -dependent terms, we obtain:
$$& \text{keeping only the leading } \varepsilon \text{-dependent terms, we obtain} \colon \\ & \langle \xi _ { F } ^ { 2 } ( \mathbf x ) \rangle _ { \mathcal { S } } = & ( \mathbf s ) \\ & g ^ { 2 } \frac { n _ { 0 } ( \mathbf x ) } { N _ { c } } \left [ - \frac { 4 } { 1 - 2 \gamma } \varepsilon ^ { 1 - 2 \gamma } + \frac { 1 } { ( 1 - 2 \gamma ) ( 1 - \gamma ) } \lambda _ { x } ^ { 1 - 2 \gamma } \right ] \\ \text{is} & \quad + g ^ { 2 } n _ { 0 } ( \mathbf x ) ( 1 - \frac { 1 } { N _ { c } } ) \left [ \frac { 2 \gamma ( 1 + 2 \gamma ) } { 2 ( 1 - \gamma ) ( 3 - 2 \gamma ) } \lambda _ { x } ^ { - 2 } \varepsilon ^ { 3 - 2 \gamma } \\ \text{In} & \quad + \frac { 4 } { ( 1 - \gamma ) ^ { 2 } } \lambda _ { x } ^ { - 1 - \gamma } \varepsilon ^ { 2 - \gamma } + C \lambda _ { x } ^ { 1 - 2 \gamma } \right ] \\ \text{...} & \quad \text{...} & \quad \text{...} & \quad \text{...}$$
where C is a constant depending on γ . Comparing this expression with (46) which we obtained for the d -dimensional case, we note that we indeed have agreement when we take d = 1. The term of order ε 2 -γ corresponds to the term of order ε d +1 -γ in (44) which we did not include in (46) because it is never the leading divergence.
We now compare these analytical expressions with those obtained from a direct numerical estimation of the same quantity in a Poissonian realization of a particle system. To do so we have distributed N = 72900 particles randomly in a 'phase space box' (cf. illustration in Fig. 1) of side 27 λ v (in velocity) and 9 λ x (in position). Thus there are 243 cells, containing, on average, 300 particles. We then calculate the exact force on each particle in the system due to all particles in the stripe to which it belongs, i.e., F ( x i ) = ∑ j ∈S g ( x i -x j ). For each cell we then average this quantity to get ¯ F = 1 N c ∑ i ∈C F k ( x i ), and we finally estimate the variance of this quantity over the 243 cells.
FIG. 2. (Color online) The variance of the force on particles in a phase space cell due to particles in the corresponding stripe, plotted as a function of ε , for the different values of γ . The crosses indicate the points obtained from the numerical simulation described in the text and dotted lines are the corresponding analytical results, Eq. (56). We use units in which λ x = 1 and g = 1.
Shown in Fig. 2 are the results of these numerical simulations compared with the theoretical results, Eq. (56), for a range of values of ε and different values of γ . The agreement is excellent in all cases. Further we have verified that, except in the region where ε approaches λ x , the results are in excellent agreement with the expression (57) for the small ε/λ x behaviour. Thus we find results completely in line with our determination of the scalings in this limit for the d -dimensional case, leading to the different behaviours described in Section IV: for γ = 0 25, in the range . γ < d/ 2, the force fluctuations are independent of ε ; for γ = 0 75 and . γ = 1 25, in the range . d/ 2 < γ < d/ 2 + 1, the divergent 2-body term ∼ ε 1 -2 γ dominates at small ε , but is then overtaken at larger ε by the flat behaviour of the the dominant ε -independent term in the 3-body term; for γ = 1 75 and . γ = 2 5, both . in the range d/ 2 + 1 < γ we have again at the smallest glyph[epsilon1] the behaviour ∼ ε 1 -2 γ from the 2-body term, but at larger glyph[epsilon1] instead the dominant behaviour ε -dependent term ∼ ε 3 -2 γ from the 3-body term.
## VI. DISCUSSION AND CONCLUSIONS
Let us now discuss further the physical significance of these results, and in particular how they justify the basic qualitative distinction between interactions as we have anticipated in the introduction. We have considered a generic N -particle system with Hamiltonian dynamics described by a two body interaction with a pair force ∼ 1 /r γ and regularized below a scale ε to a constant value. Introducing a coarse-graining in phase space, one obtains an equation for the coarse-grained phase space density with terms corresponding to the Vlasov equation, in addition to 'non-Vlasov' terms which are functionals of the microscopic phase density. We then made the hy- pothesis that the typical amplitude of these terms can be estimated by assuming this microscopic phase space density to be given by a realization of an inhomogeneous Poissonian point process, in which the mean density is specified by the coarse-grained phase space density but the density fluctuations are uncorrelated. In other words we neglect all contributions due to correlation of density fluctuations with respect to the ones due to the simple products of mean densities. Doing so we have determined the scalings of these latter terms as a function of the different scales in the problem, of γ and of the spatial dimension d . If the assumed hypothesis is valid (for large N ), a limit in which the Vlasov equation applies is simply any one in which the derived non-Vlasov terms go to zero while the Vlasov terms tend to fixed finite values. Our results show firstly that, for any interaction in the class we have considered, such a limit exists and can be defined in different ways: notably, by taking N to infinity at fixed values of the other scales, and in particular at fixed smoothing; or, alternatively, by taking the limit in which the coarse-graining scale, while always remaining small compared to the system size, is taken arbitrarily large compared to the other scales, and in particular the smoothing scale ε . However, when this limit is taken, there is a important difference between the case in which γ > d or γ < d : for the former case the dominant correction term to the Vlasov limit is always strongly dependent on ε , while in the latter case the dominant term, at fixed λ x , is independent of ε in a wide range whose lower limit vanishes as λ x diverges. This different leading dependence on ε corresponds to fluctuations which are sourced by quite different contributions in the two cases: for γ > d , the dominant fluctuation in the force on a coarse-graining cell comes from the contributions coming from forces on individual particles due to single particles which are 'closeby', i.e., within a radius ε : for γ < d , on the other hand, the dominant fluctuation comes from the coherent effect on two different particles in the cell coming from a single particle which is 'far-away', i.e. at a distance of order the size of the coarse-graining cell, or, even for γ < d/ 2, of order the size of the system. Given that these amplitudes are those of the terms describing the corrections to the Vlasov evolution due to the large, but finite, number of particles in a real system, this means that the dynamics of the particle system at a coarse-grained scale in the two cases is dominated by completely different contributions: by the particle distribution at the smallest scales when γ > d , and by the particle distribution at the coarse-graining scale or larger when γ < d . Thus in the latter case we can decouple the dynamics at the coarse-graining scale from that at smaller scales (the interparticle distance, the scale of particle size), while for γ > d we cannot do so. Or, in the language of renormalisation theory, the former admit a kind of universality in which the coarse-grained dynamics is insensitive to the form of the interaction below this scale, while the latter do not. This is a basic qualitative distinction between the
dynamics in these two cases, which corresponds naturally to what one call 'short-range' or 'long-range'. In order to distinguish it from the canonical distinction based on thermodynamical considerations, following [24], we can refer to γ > d as dynamically short-range , and γ < d as dynamically long-range 2 .
For what concerns quasi-stationary states the implications of this result and relevance of the classification are simple: for all cases one would expect that such states may exist (since the Vlasov limit exists) but the conditions for their existence, which requires that the time scales of their persistence be long compared to the system dynamical time, will be very different. For γ > d , their lifetime, which would be expected to be directly related to the amplitude of the non-Vlasov term, can be long only if the smoothing scale in the force is sufficiently large; in the case γ < d , their lifetime will be expected to be independent of ε . We note that this is precisely in line with results of analytical calculations based on the Chandrasekhar approach to estimation of the relaxation rate, and the results of numerical simulations of systems of this kind reported in [25].
These results are all, as we have emphasized, built on our central hypothesis that correlation of density fluctuations, associated with the finite particle number, may be neglected. To formulate it we must define a smooth mean phase space density by introducing a coarse-graining scale, which is assumed to be a 'mesoscopic scale': arbitrarily small compared to the system size, and yet large enough so that the phase space cell contains many particles. Our central hypothesis is not one of which we have proven the validity, but it is a consistent, simple and physically reasonable one, analogous to that of 'molecular chaos' in the derivation of the Boltzmann equation. It is also in line with the fact that when the Vlasov approximation is valid, the system dynamics is defined in terms only of the mean density and fluctuations due to higher order density correlations are considered subdominant. We note above all that it leads to conclusions and behaviours which are very reasonable physically, even in the cases we have not focussed on but to which our analysis can be applied. For example, if we consider a hard repulsive core interaction without a smoothing, i.e. let ε → 0 for γ large, we infer that the force fluctuation on a coarse-grained cell diverges and that there is thus no Vlasov limit. Indeed in this case an appropriate two body collision operator is required to take into account the effects of interactions between particles.
Finally we note that the ensemble we have assumed to describe the fine-grained phase space density is a realization of a Poisson process with mean density given
2 Or, alternatively, if one adopts the terminology advocated by [4], in which the thermodynamic distinction is made between 'strongly long-range' and 'weakly long-range', our classification could be described as a distinction between 'dynamically strongly long-range' and 'dynamically weakly long-range' interactions.
by the coarse-grained space. This implies that we include (Poissonian) fluctuations of particle number at all scales: indeed these (finite particle number) fluctuations are the source of the fluctuation terms in the dynamical equations which we have analysed. One could consider that it might be more appropriate physically to take, in estimating the fluctuations induced at a given coarsegrained density, the ensemble in which the particle number is constrained to be fixed in all coarse-graining cells. In other words we could average, given a coarse-grained phase space density, only over the configurations in which the particles, of fixed number in each cell, are distributed randomly within the cells. It is straightforward to verify, with the appropriate modification of the average, that doing so can only change our results for what concerns the large scale contributions to the force fluctuations: the diverging behaviours at small scales we have focussed on arise from the fact that there is a finite probability for a particle to be arbitrarily close to another one, and the local value of the density will at most modify the amplitude but not the parametric dependence of this term. On the other hand, our determinations of the parametric dependence of the contributions to the force fluctuations from the bulk will be expected to depend on how the particle fluctuations at larger scales are constrained 3 .
In future work we plan to explore the possibility of developing the approach used in this article to understand and describe further the effect of the 'non-Vlasov' collisional - terms on the evolution of a finite N system, i.e., to use it to develop a kinetic theory for the N particle system. In this context it would be interesting to determine whether existing kinetic equations such as that of Lenard-Balescu, or variants of it developed in the literature (see, e.g., [17, 21] for a discussion and references), could be derived in a different way from this starting point, or even potentially modified or formulated differently. In this respect we note that the results derived here already provide a better basis for many derivations of such equations (and in particular the Lenard-Balescu equation) which take as a starting point the assumption that the Vlasov equation applies in the large N limit (and then derive the kinetic equation for perturbations about it). It would be interesting also to clarify in particular the relation between this approach and that of Chandrasekhar which, as we have noted, when extended to a generic interaction has been shown [25] to give very consistent conclusions about the sensitivity of collisional relaxation of quasi-stationary states to the small scale regularisation. Likewise it would be interesting to try
3 A simple example is the case of one dimensional gravity, i.e., γ = 0 in d = 1. In this case Poissonian force fluctuations indeed diverge with system size [27] precisely as derived here. On the other hand, because the force is independent of distance, the force fluctuation in a cell coming from particles in other cells vanishes if the number of particles in these cells does not fluctuate.
to test directly in numerical simulations for the validity of our central hypothesis about correlations, and also characterize using analytical methods the robustness of our central conclusions to the existence of different weak
- [1] A. Campa, T. Dauxois, and S. Ruffo, Phys. Rep. 480 , 57 (2009)
- [2] T. Dauxois, S. Ruffo, E. Arimondo, and M. Wilkens, Lecture Notes in Physics Volume 602 (Springer Berlin / Heidelberg, 2002)
- [3] A. Campa, A. Giansanti, G. Morigi, and F. Sylos Labini, Dynamics and Thermodynamics of systems with long range interactions: theory and experiments , Vol. 970 (2008)
- [4] F. Bouchet, S. Gupta, and D. Mukamel, Physica A 389 , 4389 (2010)
- [5] J. Binney and S. Tremaine, Galactic Dynamics , 2nd ed. (Princeton University Press, 2008)
- [6] F. Bouchet and J. Sommeria, J. Fluid Mech. 464 , 165 (2002)
- [7] A. Olivetti, J. Barr´ e, B. Marcos, F. Bouchet, and R. Kaiser, Phys. Rev. Lett. 103 , 224301 (2009)
- [8] J. Sopik, C. Sire, and P.-H. Chavanis, Phys. Rev. E 72 , 26105 (2005)
- [9] W. Braun and K. Hepp, Commun. Math. Phys. 113 , 101 (1977)
- [10] H. Spohn, Large scale dynamics of interacting particles (Springer, 1991) ISBN 9783642843730
- [11] M. Hauray and P.-e. Jabin, Arch. Ration. Mech. Anal. 183 , 489 (2007), arXiv:arXiv:1107.3821v5
- [12] S. Chandrasekhar, Astrophys. Journal. 97 , 255 (1943)
- [13] M. Henon, Ann. d'Astrophysique 21 , 186 (1958)
- [14] P.-H. Chavanis, J. Stat. Mech. Theory Exp. 2010 ,
correlations of the density fluctuations.
We thank Bruno Marcos and Pascal Viot for very useful discussions.
P05019 (2010)
- [15] R. T. Farouki and E. E. Salpeter, Astrophys. J. 427 , 676 (1994)
- [16] R. T. Farouki and E. E. Salpeter, Astrophys. J. 253 , 512 (1982)
- [17] P. H. Chavanis, Eur. Phys. J. Plus 127 , 19 (Feb. 2012)
- [18] R. Balescu, Equilibrium and Non-Equilibrium Statistical Mechanics , Wiley-Interscience publication (John Wiley & Sons, 1975) ISBN 9780471046004
- [19] I. U. L. Klimontovich, The Statistical Theory of Nonequilibrium Processes in Plasma , International series of monographs in natural philosophy (M.I.T. Press, 1967)
- [20] D. R. Nicholson, Introduction to plasma theory (Cambridge Univ Press, 1983)
- [21] P.-H. Chavanis, Eur. Phys. J. Plus 128 , 106 (2013)
- [22] T. R. Filho, M. Amato, and A. Figueiredo, Phys. Rev. E 89 , 1 (2012), arXiv:1203.0082v1
- [23] T. Buchert and A. Dom´ ınguez, A&A 460 , 443 (2005)
- [24] A. Gabrielli, M. Joyce, B. Marcos, and F. Sicard, J. Stat. Phys. 141 , 970 (2010)
- [25] A. Gabrielli, M. Joyce, and B. Marcos, Phys. Rev. Lett. 105 , 210602 (2010)
- [26] A. Gabrielli, F. Labini, M. Joyce, and L. Pietronero, Statistical Physics for Cosmic Structures (Springer, 2006) ISBN 9783540269991
- [27] A. Gabrielli and M. Joyce, Phys. Rev. E 81 , 21102 (2010)
|
10.1103/PhysRevE.90.062910
|
[
"Andrea Gabrielli",
"Michael Joyce",
"Jules Morand"
] |
2014-08-05T15:21:25+00:00
|
2015-10-07T09:56:16+00:00
|
[
"cond-mat.stat-mech"
] |
Finite $N$ corrections to Vlasov dynamics and the range of pair interactions
|
We explore the conditions on a pair interaction for the validity of the
Vlasov equation to describe the dynamics of an interacting $N$ particle system
in the large $N$ limit. Using a coarse-graining in phase space of the exact
Klimontovich equation for the $N$ particle system, we evaluate, neglecting
correlations of density fluctuations, the scalings with $N$ of the terms
describing the corrections to the Vlasov equation for the coarse-grained one
particle phase space density. Considering a generic interaction with radial
pair force $F(r)$, with $F(r) \sim 1/r^\gamma$ at large scales, and regulated
to a bounded behaviour below a "softening" scale $\varepsilon$, we find that
there is an essential qualitative difference between the cases $\gamma < d$ and
$\gamma > d$, i.e., depending on the integrability at large distances of the
pair force. In the former case the corrections to the Vlasov dynamics for a
given coarse-grained scale are essentially insensitive to the softening
parameter $\varepsilon$, while for $\gamma > d$ the amplitude of these terms is
directly regulated by $\varepsilon$, and thus by the small scale properties of
the interaction. This corresponds to a simple physical criterion for a basic
distinction between long-range ($\gamma \leq d $) and short range ($\gamma >
d$) interactions, different to the canonical one ($\gamma \leq d +1$ or $\gamma
> d +1$ ) based on thermodynamic analysis. This alternative classification,
based on purely dynamical considerations, is relevant notably to understanding
the conditions for the existence of so-called quasi-stationary states in
long-range interacting systems.
|
1408.1000v3
|
## Estimating Renyi Entropy of Discrete Distributions
Jayadev Acharya 1 , Alon Orlitsky 2 , Ananda Theertha Suresh 2 , and Himanshu Tyagi 3
1
Massachusetts Institute of Technology ([email protected]) 2 University of California, San Diego ( { alon, asuresh } @ucsd.edu) 3 Indian Institute of Science ([email protected])
## Abstract
It was recently shown that estimating the Shannon entropy H ( p ) of a discrete k -symbol distribution p requires Q ( k /log k ) samples, a number that grows near-linearly in the support size. In many applications H ( p ) can be replaced by the more general R´ enyi entropy of order a , H a ( p . We determine the number of samples needed to estimate ) H a ( p ) for all a , showing that a < 1 requires a super-linear, roughly k 1/ a samples, noninteger a > 1 requires a near-linear k samples, but, perhaps surprisingly, integer a > 1 requires only Q ( k 1 -1/ a ) samples. Furthermore, developing on a recently established connection between polynomial approximation and estimation of additive functions of the form GLYPH<229> x f ( p x ) , we reduce the sample complexity for noninteger values of a by a factor of log k compared to the empirical estimator. The estimators achieving these bounds are simple and run in time linear in the number of samples. Our lower bounds provide explicit constructions of distributions with different R´ enyi entropies that are hard to distinguish.
## I. INTRODUCTION
## A. Shannon and R´ enyi entropies
One of the most commonly used measure of randomness of a distribution p over a discrete set X is its Shannon entropy
$$H ( p ) \stackrel { \text{def} } { = } \sum _ { x \in \mathcal { X } } p _ { x } \log \frac { 1 } { p _ { x } }.$$
An initial version of this paper [1] was presented at the ACM Symposium on Discrete Algorithms (SODA), 2015.
The estimation of Shannon entropy has several applications, including measuring genetic diversity [37], quantifying neural activity [32], [29], network anomaly detection [20], and others. It was recently shown that estimating the Shannon entropy of a discrete distribution p over k elements to a given additive accuracy requires Q ( k /log k ) independent samples from p [33], [41]; see [16], [43] for subsequent extensions. This number of samples grows near-linearly with the alphabet size and is only a logarithmic factor smaller than the Q ( k ) samples needed to learn p itself to within a small statistical distance.
glyph[negationslash]
A popular generalization of Shannon entropy is the R´ enyi entropy of order a ≥ 0, defined for a = 1 by
$$H _ { \alpha } ( p ) \stackrel { \text{def} } { = } \frac { 1 } { 1 - \alpha } \log \sum _ { x \in \mathcal { X } } p _ { x } ^ { \alpha }$$
and for a = 1 by
$$H _ { 1 } ( \mathbf p ) \stackrel { \text{def} } { = } \lim _ { \alpha \rightarrow 1 } H _ { \alpha } ( \mathbf p ).$$
It was shown in the seminal paper [36] that R´ enyi entropy of order 1 is Shannon entropy, namely H 1 ( p ) = H ( p , and for all other orders it is the unique extension of Shannon entropy when ) of the four requirements in Shannon entropy's axiomatic definition, continuity, symmetry, and normalization are kept but grouping is restricted to only additivity over independent random variables ( c f . [13]).
R´ enyi entropy too has many applications. It is often used as a bound on Shannon entropy [26], [29], [12], and in many applications it replaces Shannon entropy as a measure of randomness [7], [24], [3]. It is also of interest in its own right, with diverse applications to unsupervised learning [44], [15], source adaptation [22], image registration [21], [28], and password guessability [3], [35], [10] among others. In particular, the R´ enyi entropy of order 2, H 2 ( p , measures the quality ) of random number generators [19], [30], determines the number of unbiased bits that can be extracted from a physical source of randomness [14], [6], helps test graph expansion [8] and closeness of distributions [5], [34], and characterizes the number of reads needed to reconstruct a DNA sequence [27].
Motivated by these and other applications, unbiased and heuristic estimators of R´ enyi entropy have been studied in the physics literature following [9], and asymptotically consistent and normal estimates were proposed in [45], [18]. However, no systematic study of the complexity of estimating R´ enyi entropy is available. For example, it was hitherto unknown if the number of samples needed to estimate the R´ enyi entropy of a given order a differs from that required
for Shannon entropy, or whether it varies with the order a , or how it depends on the alphabet size k .
## B. Definitions and results
glyph[negationslash]
We answer these questions by showing that the number of samples needed to estimate H a ( p ) falls into three different ranges. For a < 1 it grows super-linearly with k , for 1 < a ∈ Z it grows almost linearly with k , and most interestingly, for the popular orders 1 < a ∈ Z it grows as Q ( k 1 -1/ a ) , which is much less than the sample complexity of estimating Shannon entropy.
To state the results more precisely we need a few definitions. A R´ enyi-entropy estimator for distributions over support set X is a function f : X ∗ → R mapping a sequence of samples drawn from a distribution to an estimate of its entropy. The sample complexity of an estimator f for distributions over k elements is defined as
$$S _ { \alpha } ^ { f } ( k, \delta, \epsilon ) \stackrel { \text{def} } { = } \min _ { n } \left \{ n \colon p \left ( \left | H _ { \alpha } ( p ) - f \left ( X ^ { n } \right ) \right | > \delta \right ) < \epsilon, \forall p \text{ with } \| p \| _ { 0 } \leq k \right \},$$
i . e ., the minimum number of samples required by f to estimate H a ( p ) of any k -symbol distribution p to a given additive accuracy d with probability greater than 1 -e . The sample complexity of estimating H a ( p ) is then
$$S _ { \alpha } ( k, \delta, \epsilon ) \stackrel { \text{def} } { = } \min _ { f } S _ { \alpha } ^ { f } ( k, \delta, \epsilon ),$$
the least number of samples any estimator needs to estimate H a ( p ) for all k -symbol distributions p, to an additive accuracy d and with probability greater than 1 -e . This is a min-max definition where the goal is to obtain the best estimator for the worst distribution.
The desired accuracy d and confidence 1 -e are typically fixed. We are therefore most interested 1 in the dependence of S a ( k , d e , ) on the alphabet size k and omit the dependence of S a ( k , d e , ) on d and e to write S a ( k ) . In particular, we are interested in the large alphabet regime and focus on the essential growth rate of S a ( k ) as a function of k for large k . Using the standard asymptotic notations, let S a ( k ) = O k ( b ) indicate that for some constant c which may depend on a , d , and e , for all sufficiently large k , S a ( k , d e , ) ≤ c · k b . Similarly, S a ( k ) = Q ( k b ) adds the corresponding W ( k b ) lower bound for S a ( k , d e , ) , for all sufficiently small d and e . Finally, extending the ˜ W notation 2 , we let S a ( k ) = ∼ ∼ W ( k b ) indicate that for every sufficiently
1 Whenever a more refined result indicating the dependence of sample complexity on both k and d is available, we shall use the more elaborate S a ( k , d e , ) notation.
2 The notations ˜ , O ˜ W , and ˜ Q hide poly-logarithmic factors.
small e and arbitrary h > 0, there exist c and d depending on h such that for all k sufficiently large S a ( k , d e , ) > ck b -h , namely S a ( k ) grows polynomially in k with exponent not less than b -h for d ≤ dh .
We show that S a ( k ) behaves differently in three ranges of a . For 0 ≤ a < 1,
$$\widetilde { \Omega } \left ( k ^ { 1 / \alpha } \right ) \leq S _ { \alpha } ( k ) \leq O \left ( \frac { k ^ { 1 / \alpha } } { \log k } \right ),$$
namely the sample complexity grows super-linearly in k and estimating the R´ enyi entropy of these orders is even more difficult than estimating the Shannon entropy. In fact, the upper bound follows from a corresponding result on estimation of power sums considered in [16] (see Section III-C for further discussion). For completeness, we show in Theorem 10 that the empirical estimator requires O k ( 1/ a ) samples and in Theorem 14 prove the improvement by a factor of log k . The lower bound is proved in Theorem 22.
For 1 < a / ∈ N ,
$$\stackrel { \approx } { \Omega } \left ( k \right ) \leq S _ { \alpha } ( k ) \leq O \left ( \frac { k } { \log k } \right ),$$
namely as with Shannon entropy, the sample complexity grows roughly linearly in the alphabet size. The lower bound is proved in Theorem 21. In a conference version of this paper [1], a weaker O k ( ) upper bound was established using the empirical-frequency estimator. For the sake of completeness, we include this result as Theorem 9. The tighter upper bound reported here uses the best polynomial approximation based estimator of [16], [43] and is proved in Theorem 13. In fact, in the Appendix we show that the empirical estimator can't attain this log k improvement and requires W ( k / d ) and W (( k / d ) 1/ a ) samples for a > 1 and a < 1, respectively.
For 1 < a ∈ N and d and e sufficiently small,
$$S _ { \alpha } ( k, \delta, \epsilon ) = \Theta \left ( \frac { k ^ { 1 - 1 / \alpha } } { \delta ^ { 2 } } \right ),$$
and in particular, the sample complexity is strictly sublinear in the alphabet size. The upper and lower bounds are shown in Theorems 12 and 16, respectively. Figure ?? illustrates our results for different ranges of a .
Of the three ranges, the most frequently used, and coincidentally the one for which the results are most surprising, is the last with a = 2, 3, . . .. Some elaboration is in order.
First, for all integral a > 1, H a ( p ) can be estimated with a sublinear number of samples. The most commonly used R´ enyi entropy, H 2 ( p , can be estimated within ) d using just Q ( √ k / d 2 )
a
Fig. 1: Exponent of k in S a ( k ) as a function of a .
samples, and hence R´ enyi entropy can be estimated much more efficiently than Shannon Entropy, a useful property for large-alphabet applications such as language processing genetic analysis.
Also, note that R´ enyi entropy is continuous in the order a . Yet the sample complexity is discontinuous at integer orders. While this makes the estimation of the popular integer-order entropies easier, it may seem contradictory. For instance, to approximate H 2.001 ( p ) one could approximate H 2 ( p ) using significantly fewer samples. The reason for this is that the R´ enyi entropy, while continuous in a , is not uniformly continuous. In fact, as shown in Example 2, the difference between say H 2 ( p ) and H 2.001 ( p ) may increase to infinity when the alphabet-size increases.
It should also be noted that the estimators achieving the upper bounds are simple and run in time linear in the number of samples. Furthermore, the estimators are universal in that they do not require the knowledge of k . On the other hand, the lower bounds on S a ( k ) hold even if the estimator knows k .
## C. The estimators
The power sum of order a of a distribution p over X is
$$P _ { \alpha } ( \mathbf p ) \stackrel { \text{def} } { = } \sum _ { x \in \mathcal { X } } p _ { x } ^ { \alpha },$$
glyph[negationslash]
and is related to the R´ enyi entropy for a = 1 via
$$H _ { \alpha } ( p ) = \frac { 1 } { 1 - \alpha } \log P _ { \alpha } ( p ).$$
Hence estimating H a ( p ) to an additive accuracy of ± d is equivalent to estimating P a ( p ) to a multiplicative accuracy of 2 ± · d ( 1 -a ) . Furthermore, if d a ( -1 ) ≤ 1/2 then estimating P a ( p ) to multiplicative accuracy of 1 ± d ( 1 -a ) /2 ensures a ± d additive accurate estimate of H a ( p . )
We construct estimators for the power-sums of distributions with a multiplicative-accuracy of ( 1 ± d ) and hence obtain an additive-accuracy of Q ( d ) for R´ enyi entropy estimation. We consider the following three different estimators for different ranges of a and with different performance guarantees.
- a) Empirical estimator: The empirical , or plug-in , estimator of P a ( p ) is given by
$$\widehat { P } _ { \alpha } ^ { \text{e} } \overset { \text{def} } { = } \sum _ { x } \left ( \frac { N _ { x } } { n } \right ) ^ { \alpha }.$$
glyph[negationslash]
For a = 1, ̂ P e a is a not an unbiased estimator of P a ( p . However, we show in Theorem 10 that ) for a < 1 the sample complexity of the empirical estimator is O k ( 1/ a ) and in Theorem 9 that for a > 1 it is O k ( ) . In the appendix, we show matching lower bounds thereby characterizing the k -dependence of the sample complexity of empirical estimator.
- b) Bias-corrected estimator: For integral a > 1, the bias-corrected estimator for P a ( p ) is
$$\widehat { P } _ { \alpha } ^ { u } \stackrel { \text{def} } { = } \sum _ { x } \frac { N _ { x } ^ { \underline { \alpha } } } { n ^ { \alpha } },$$
where for integers N and r > 0, N r def = N N ( -1 ) . . . ( N -r + ) 1 . A variation of this estimator was proposed first in [4] for estimating moments of frequencies in a sequence using random samples drawn from it. Theorem 12 shows that for 1 < a ∈ Z , ̂ P u a estimates P a ( p ) within a factor of 1 ± d using O k ( 1 -1/ a / d 2 ) samples, and Theorem 16 shows that this number is optimal up to a constant factor.
c) Polynomial approximation estimator: To obtain a logarithmic improvement in S a ( k ) , we consider the polynomial approximation estimator proposed in [43], [16] for different problems, concurrently to a conference version [1] of this paper. The polynomial approximation estimator first considers the best polynomial approximation of degree d to y a for the interval y ∈ [ 0, 1 ] [39]. Suppose this polynomial is given by a 0 + a 1 y + a 2 y 2 + . . . + a d y d . We roughly divide the samples into two parts. Let Nx ′ and Nx be the multiplicities of x in the first and second parts respectively. The polynomial approximation estimator uses the empirical estimate of p a x for large Nx ′ , but
TABLE I: Performance of estimators and lower bounds for estimating R´ enyi entropy .
| Range of a | Empirical | Bias-corrected | Polynomial | Lower bounds |
|----------------|--------------------------------------|----------------------|---------------------------|---------------------------------|
| a < 1 | O ( k 1/ a d max ( 4,2/ a ) ) | | O ( k 1/ a d 1/ a log k ) | for all h > 0, W ( k 1/ a - h ) |
| a > 1, a / ∈ N | O ( k min ( d 1/ ( a - 1 ) , d 2 ) ) | | O ( k d 1/ a log k ) | for all h > 0, W ( k 1 - h ) |
| a > 1, a ∈ N | O ( k d 2 ) | O ( k 1 - 1/ a d 2 ) | | W ( k 1 - 1/ a d 2 ) |
estimates a polynomial approximation of p a x for a small Nx ′ ; the integer powers of p x in the latter in turn is estimated using the bias-corrected estimator.
The estimator is roughly of the form
$$\widehat { P } _ { \alpha } ^ { d, \tau } \overset { \text{def} } { = } \sum _ { x \colon N _ { x } ^ { \prime } \leq \tau } \left ( \sum _ { m = 0 } ^ { d } \frac { a _ { m } ( 2 \tau ) ^ { \alpha - m } N _ { x } ^ { \underline { m } } } { n ^ { \alpha } } \right ) + \sum _ { x \colon N _ { x } ^ { \prime } > \tau } \left ( \frac { N _ { x } } { n } \right ) ^ { \alpha },$$
where d and t are both O ( log n ) and chosen appropriately.
Theorem 13 and Theorem 14 show that for a > 1 and a < 1, respectively, the sample complexity of ̂ P d , t a is O k ( /log k ) and O k ( 1 a /log k ) , resulting in a reduction in sample complexity of O ( log k ) over the empirical estimator.
Table I summarizes the performance of these estimators in terms of their sample complexity. The last column denote the lower bounds from Section V.
Our goal in this work was to identify the exponent of k in S a ( k ) . In the process, we were able to characterize the sample complexity S a ( k , d e , ) for 1 < a ∈ ◆ . However, we only obtain partial results towards characterizing the sample complexity S a ( k , d e , ) for a general a . Specifically, while we show that the empirical estimator attains the aforementioned exponent for every 0 < a / ∈ ◆ , we note that the polynomial approximation estimator has a lower sample complexity than the empirical estimator. The exact characterization of S a ( k , d e , ) for a general a remains open.
## D. Organization
The rest of the paper is organized as follows. Section II presents basic properties of power sums of distributions and moments of Poisson random variables, which may be of independent interest. The estimation algorithms are analyzed in Section III, in Section III-A we show results for the empirical or plug-in estimate, in Section III-B we provide optimal results for integral a
and finally we provide an improved estimator for non-integral a > 1. Examples and simulation of the proposed estimators are given in Section IV. Section V contains our lower bounds for the sample complexity of estimating R´ enyi entropy. Furthermore, in the Appendix we analyze the performance of the empirical estimator for power-sum estimation with an additive-accuracy and also derive lower bounds for its sample complexity.
## II. TECHNICAL PRELIMINARIES
## A. Bounds on power sums
Consider a distribution p over [ k ] = { 1, . . . , k } . Since R´ enyi entropy is a measure of randomness (see [36] for a detailed discussion), it is maximized by the uniform distribution and the following inequalities hold:
glyph[negationslash]
$$0 \leq H _ { \alpha } ( { \mathbf p } ) \leq \log k, \ \alpha \neq 1,$$
or equivalently
$$1 \leq P _ { \alpha } ( p ) \leq k ^ { 1 - \alpha }, \ \alpha < 1 \quad \text{and} \quad k ^ { 1 - \alpha } \leq P _ { \alpha } ( p ) \leq 1, \quad \alpha > 1.$$
Furthermore, for a > 1, P a + b ( p ) and P a -b ( p ) can be bounded in terms of P a ( p , ) using the monotonicity of norms and of H¨ older means (see, for instance, [11]).
Lemma 1. For every 0 ≤ a ,
$$P _ { 2 \alpha } ( p ) \leq P _ { \alpha } ( p ) ^ { 2 }$$
Further, for a > 1 and 0 ≤ b ≤ a ,
$$P _ { \alpha + \beta } ( \mathbf p ) \leq k ^ { ( \alpha - 1 ) ( \alpha - \beta ) / \alpha } \, P _ { \alpha } ( \mathbf p ) ^ { 2 },$$
and
$$P _ { \alpha - \beta } ( \mathbf p ) \leq k ^ { \beta } \, P _ { \alpha } ( \mathbf p ).$$
Proof. By the monotonicity of norms,
$$P _ { \alpha + \beta } ( \mathbf p ) \leq P _ { \alpha } ( \mathbf p ) ^ { \frac { \alpha + \beta } { \alpha } },$$
which gives
$$\frac { P _ { \alpha + \beta } ( \mathbf p ) } { P _ { \alpha } ( \mathbf p ) ^ { 2 } } \leq P _ { \alpha } ( \mathbf p ) ^ { \frac { \beta } { \alpha } - 1 }.$$
The first inequality follows upon choosing b = a . For 1 < a and 0 ≤ b ≤ a , we get the second by (4). Note that by the monotonicity of H¨ older means, we have
$$\left ( \frac { 1 } { k } \sum _ { x } p _ { x } ^ { \alpha - \beta } \right ) ^ { \frac { 1 } { \alpha - \beta } } \leq \left ( \frac { 1 } { k } \sum _ { x } p _ { x } ^ { \alpha } \right ) ^ { \frac { 1 } { \alpha } }.$$
The final inequality follows upon rearranging the terms and using (4). glyph[squaresolid]
## B. Bounds on moments of a Poisson random variable
Let Poi ( l ) be the Poisson distribution with parameter l . We consider Poisson sampling where N ∼ Poi ( n ) samples are drawn from the distribution p and the multiplicities used in the estimation are based on the sequence X N = X 1 , ..., XN instead of X n . Under Poisson sampling, the multiplicities Nx are distributed as Poi ( n p x ) and are all independent, leading to simpler analysis. To facilitate our analysis under Poisson sampling, we note a few properties of the moments of a Poisson random variable.
We start with the expected value and the variance of falling powers of a Poisson random variable.
Lemma 2. Let X ∼ Poi ( l ) . Then, for all r ∈ N
$$\mathbb { E } [ X ^ { r } ] = \lambda ^ { r }$$
and
Proof. The expectation is
$$\mathbb { V } \text{ar} [ X ^ { \underline { r } } ] \leq \lambda ^ { r } \left ( ( \lambda + r ) ^ { r } - \lambda ^ { r } \right ).$$
$$\mathbb { E } [ X ^ { r } ] & = \sum _ { i = 0 } ^ { \infty } \text{Poi} ( \lambda, i ) \cdot i ^ { r } \\ & = \sum _ { i = r } ^ { \infty } e ^ { - \lambda } \cdot \frac { \lambda ^ { i } } { i! } \cdot \frac { i! } { ( i - r )! } \\ & = \lambda ^ { r } \sum _ { i = 0 } ^ { \infty } e ^ { - \lambda } \cdot \frac { \lambda ^ { i } } { i! } \\ & = \lambda ^ { r }.$$
The variance satisfies
$$\mathbb { E } \left [ ( X ^ { r } ) ^ { 2 } \right ] & = \sum _ { i = 0 } ^ { \infty } \text{Poi} ( \lambda, i ) \cdot ( i ^ { r } ) ^ { 2 } \\ & = \sum _ { i = r } ^ { \infty } e ^ { - \lambda } \cdot \frac { \lambda ^ { i } } { i! } \left ( \frac { i! ^ { 2 } } { ( i - r )! ^ { 2 } } \\ & = \lambda ^ { r } \sum _ { i = 0 } ^ { \infty } e ^ { - \lambda } \cdot \frac { \lambda ^ { i } } { i! } \cdot ( i + r ) ^ { r } \\ & = \lambda ^ { r } \cdot \mathbb { E } [ ( X + r ) ^ { r } ] \\ & \leq \lambda ^ { r } \cdot \mathbb { E } \left [ \sum _ { j = 0 } ^ { r } \binom { r } { j } X ^ { j } \cdot r ^ { r - j } \right ] \\ & = \lambda ^ { r } \cdot \sum _ { j = 0 } ^ { r } \binom { r } { j } \cdot \lambda ^ { j } \cdot r ^ { r - j } \\ & = \lambda ^ { r } ( \lambda + r ) ^ { r }, \\ \\ \text{allows from}$$
where the inequality follows from
$$( X + r ) ^ { \underline { r } } = \prod _ { j = 1 } ^ { r } \left [ ( X + 1 - j ) + r \right ] \leq \sum _ { j = 0 } ^ { r } \binom { r } { j } \cdot X ^ { j } \cdot r ^ { r - j }.$$
Therefore,
$$\text{Var} [ X ^ { \tau } ] = \mathbb { E } \left [ ( X ^ { \tau } ) ^ { 2 } \, \right ] - \left [ \, \mathbb { E } \, X ^ { \tau } \right ] ^ { 2 } \leq \lambda ^ { r } \cdot ( ( \lambda + r ) ^ { r } - \lambda ^ { r } ) \,.$$
The next result establishes a bound on the moments of a Poisson random variable.
Lemma 3. Let X ∼ Poi ( l ) and let b be a positive real number. Then,
$$\mathbb { E } \left \lceil X ^ { \beta } \, \right \rceil \leq 2 ^ { \beta + 2 } \max \{ \lambda, \lambda ^ { \beta } \}.$$
Proof. Let Z = max { l 1/ b , l } .
$$= \max \{ \lambda ^ { 1 / \beta }, \lambda \}. \\ \mathbb { E } \left [ \frac { X \beta } { Z ^ { \beta } } \right ] & \leq \mathbb { E } \left [ \left ( \frac { X } { Z } \right ) ^ { \lceil \beta \rceil } + \left ( \frac { X } { Z } \right ) ^ { \lfloor \beta \rfloor } \right ] \\ & = \sum _ { i = 1 } ^ { \lceil \beta \rceil } \left ( \frac { \lambda } { Z } \right ) ^ { \lceil \beta \rceil } \left ( \lceil \beta \rceil \right ) + \sum _ { i = 1 } ^ { \lfloor \beta \rfloor } \left ( \frac { \lambda } { Z } \right ) ^ { \lfloor \beta \rfloor } \left ( \lceil \beta \rceil \right ) \\ & \leq \sum _ { i = 1 } ^ { \lceil \beta \rceil } \left ( \lceil \beta \rceil \right ) + \sum _ { i = 1 } ^ { \lfloor \beta \rfloor } \left ( \lceil \beta \rceil \right ) \\ & \leq 2 ^ { \beta + 2 }.$$
The first inequality follows from the fact that either X Z / > 1 or ≤ 1. The equality follows from the fact that the integer moments of Poisson distribution are Touchard polynomials in l . The second inequality uses the property that l / Z ≤ 1. Multiplying both sides by Z b results in the lemma. glyph[squaresolid]
We close this section with a bound for | E [ X a ] -l a | , which will be used in the next section and is also of independent interest.
Lemma 4. For X ∼ Poi ( l ) ,
$$| \mathbb { E } [ X ^ { \alpha } ] - \lambda ^ { \alpha } | & \leq \begin{cases} \alpha \left ( 2 ^ { \alpha } \lambda + ( 2 ^ { \alpha } + 1 ) \lambda ^ { \alpha - 1 / 2 } \right ) & \alpha > 1 \\ \min \left \{ \lambda ^ { \alpha }, \lambda ^ { \alpha - 1 } \right \} & \alpha \leq 1. \end{cases}$$
Proof. For a ≤ 1, ( 1 + ) y a ≥ 1 + a y -y 2 for all y ∈ -[ 1, ¥ ] , hence,
$$X ^ { \alpha } & = \lambda ^ { \alpha } \left ( 1 + \left ( \frac { X } { \lambda } - 1 \right ) \right ) ^ { \alpha } \\ & \geq \lambda ^ { \alpha } \left ( 1 + \alpha \left ( \frac { X } { \lambda } - 1 \right ) - \left ( \frac { X } { \lambda } - 1 \right ) ^ { 2 } \right ).$$
Taking expectations on both sides,
$$\mathbb { E } [ X ^ { \alpha } ] & \geq \lambda ^ { \alpha } \left ( 1 + \alpha \mathbb { E } \left [ \frac { X } { \lambda } - 1 \right ] - \mathbb { E } \left [ \left ( \frac { X } { \lambda } - 1 \right ) ^ { 2 } \right ] \right ) \\ & = \lambda ^ { \alpha } \left ( 2 - \frac { 1 } { \lambda } \right ).$$
Since x a is a concave function and X is nonnegative, the previous bound yields
$$| \mathbb { E } [ X ^ { \alpha } ] - \lambda ^ { \alpha } | & = \lambda ^ { \alpha } - \mathbb { E } [ X ^ { \alpha } ] \\ & \leq \min \left \{ \lambda ^ { \alpha }, \lambda ^ { \alpha - 1 } \right \}.$$
For a > 1,
$$| x ^ { \alpha } - y ^ { \alpha } | \leq \alpha | x - y | \left ( x ^ { \alpha - 1 } + y ^ { \alpha - 1 } \right ),$$
hence by the Cauchy-Schwarz Inequality,
$$\i c \, \text{cauy-cuda-nivequ}, \\ \mathbb { E } [ | X ^ { \alpha } - \lambda ^ { \alpha } | ] & \leq \alpha \mathbb { E } \left [ | X - \lambda | \left ( X ^ { \alpha - 1 } + \lambda ^ { \alpha - 1 } \right ) \right ] \\ & \leq \alpha \sqrt { \mathbb { E } [ ( X - \lambda ) ^ { 2 } ] } \sqrt { \mathbb { E } [ ( X ^ { 2 \alpha - 2 } + \lambda ^ { 2 \alpha - 2 } ) } \right ] \\ & \leq \alpha \sqrt { \lambda } \sqrt { \mathbb { E } [ ( X ^ { 2 \alpha - 2 } + \lambda ^ { 2 \alpha - 2 } ) } \right ] \\ & \leq \alpha \sqrt { 2 ^ { 2 \alpha } \max \{ \lambda ^ { 2 }, \lambda ^ { 2 \alpha - 1 } \} + \lambda ^ { 2 \alpha - 1 } } \\ & \leq \alpha \left ( 2 ^ { \alpha } \max \{ \lambda, \lambda ^ { \alpha - 1 / 2 } \} + \lambda ^ { \alpha - 1 / 2 } \right ), \\ \. \quad. \quad. \quad.. \quad.. \quad.$$
where the last-but-one inequality is by Lemma 3.
glyph[squaresolid]
## C. Polynomial approximation of x a
In this section, we review a bound on the error in approximating x a by a d -degree polynomial over a bounded interval. Let P d denote the set of all polynomials of degree less than or equal to d over R . For a continuous function f ( x ) and l > 0, let
$$E _ { d } ( f, [ 0, \lambda ] ) \stackrel { \text{def} } { = } \inf _ { q \in \mathcal { P } _ { d } } \max _ { x \in [ 0, \lambda ] } | q ( x ) - f ( x ) |.$$
Lemma 5 ([39]) . There is a constant c ′ a such that for any d > 0 ,
$$E _ { d } ( x ^ { \alpha }, [ 0, 1 ] ) \leq \frac { c _ { \alpha } ^ { \prime } } { d ^ { 2 \alpha } }.$$
To obtain an estimator which does not require a knowledge of the support size k , we seek a polynomial approximation q a ( x ) of x a with q a ( 0 ) = 0. Such a polynomial q a ( x ) can be obtained by a minor modification of the polynomial q ′ a ( x ) = GLYPH<229> d j = 0 qj x j satisfying the error bound in Lemma 5. Specifically, we use the polynomial q a ( x ) = q ′ a ( x ) -q 0 for which the approximation error is bounded as
$$\text{oua} \varepsilon a \text{ as} \\ \max _ { x \in [ 0, 1 ] } | q _ { \alpha } ( x ) - x ^ { \alpha } | & \leq | q _ { 0 } | + \max _ { x \in [ 0, 1 ] } | q _ { \alpha } ^ { \prime } ( x ) - x ^ { \alpha } | \\ & = | q _ { \alpha } ^ { \prime } ( 0 ) - 0 ^ { \alpha } | + \max _ { x \in [ 0, 1 ] } | q _ { \alpha } ^ { \prime } ( x ) - x ^ { \alpha } | \\ & \leq 2 \max _ { x \in [ 0, 1 ] } | q _ { \alpha } ^ { \prime } ( x ) - x ^ { \alpha } | \\ & = \frac { 2 c _ { \alpha } ^ { \prime } } { d ^ { 2 \alpha } } \\ & \text{def} \quad & \frac { c _ { \alpha } } { d ^ { 2 \alpha } }.$$
To bound the variance of the proposed polynomial approximation estimator, we require a bound on the absolute values of the coefficients of q a ( x ) . The following inequality due to Markov serves this purpose.
Lemma 6 ([23]) . Let p ( x ) = GLYPH<229> d j = 0 c j x j be a degree-d polynomial so that | p x ( ) | ≤ 1 for all x ∈ -[ 1, 1 ] . Then for all j = 0, . . . , m
$$\max _ { j } | c _ { j } | \leq ( \sqrt { 2 } + 1 ) ^ { d }.$$
Since | x a | ≤ 1 for x ∈ [ 0, 1 , ] the approximation bound (5) implies | q a ( x ) | < 1 + c a d 2 a for all x ∈ [ 0, 1 . ] It follows from Lemma 6 that
$$\max _ { m } | a _ { m } | < \left ( 1 + \frac { c _ { \alpha } } { d ^ { 2 \alpha } } \right ) ( \sqrt { 2 } + 1 ) ^ { d }.$$
## III. UPPER BOUNDS ON SAMPLE COMPLEXITY
In this section, we analyze the performances of the estimators we proposed in Section I-C. Our proofs are based on bounding the bias and the variance of the estimators under Poisson sampling. We first describe our general recipe and then analyze the performance of each estimator separately.
Let X 1 , ..., Xn be n independent samples drawn from a distribution p over k symbols. Consider an estimate f a ( X n ) = 1 1 -a log P ̂ a ( n X n , ) of H a ( p ) which depends on X n only through the multiplicities and the sample size. Here P ̂ a ( n X n , ) is the corresponding estimate of P a ( p ) -as discussed in Section I, small additive error in the estimate f a ( X n ) of H a ( p ) is equivalent to small multiplicative error in the estimate P ̂ a ( n X n , ) of P a ( p . ) For simplicity, we analyze a randomized estimator ˜ f a described as follows: For N ∼ Poi ( n /2 , let )
$$\tilde { f } _ { \alpha } \left ( X ^ { n } \right ) = \begin{cases} & \text{constant}, \quad N > n, \\ & \frac { 1 } { 1 - \alpha } \log \widehat { P } _ { \alpha } ( n / 2, X ^ { N } ), \quad N \leq n. \end{cases}$$
The following reduction to Poisson sampling is well-known.
Lemma 7. (Poisson approximation 1) For n ≥ 8 log ( 2/ e ) and N ∼ Poi ( n /2 ) ,
$$\mathbb { P } \left ( | H _ { \alpha } ( \mathbf p ) - \tilde { f } _ { \alpha } \left ( X ^ { n } \right ) | > \delta \right ) \leq \mathbb { P } \left ( | H _ { \alpha } ( \mathbf p ) - \frac { 1 } { 1 - \alpha } \log \widehat { \mathbf P } _ { \alpha } ( n / 2, X ^ { N } ) | > \delta \right ) + \frac { \epsilon } { 2 }.$$
It remains to bound the probability on the right-side above, which can be done provided the bias and the variance of the estimator are bounded.
Lemma 8. For N ∼ Poi ( n , ) let the power sum estimator ̂ P a = ̂ P a ( n X N , ) have bias and variance satisfying
$$\left | \mathbb { E } \left [ \widehat { P } _ { \alpha } \right ] - P _ { \alpha } ( \mathbf p ) \right | & \leq \frac { \delta } { 2 } P _ { \alpha } ( \mathbf p ), \\ \text{Var} \left [ \widehat { P } _ { \alpha } \right ] & \leq \frac { \delta ^ { 2 } } { 1 2 } P _ { \alpha } ( \mathbf p ) ^ { 2 }.$$
Then, there exists an estimator ̂ P ′ a that uses 18 n log 1/ ( e ) samples and ensures
$$\mathbb { P } \left ( \left | \widehat { \mathbb { P } } _ { \alpha } ^ { \prime } - P _ { \alpha } ( \mathbf p ) \right | > \delta \, P _ { \alpha } ( \mathbf p ) \right ) \leq \epsilon.$$
Proof. By Chebyshev's Inequality
$$\mathbb { P } \left ( \left | \widehat { P } _ { \alpha } - P _ { \alpha } ( p ) \right | > \delta \, P _ { \alpha } ( p ) \right ) \leq \mathbb { P } \left ( \left | \widehat { P } _ { \alpha } - \mathbb { E } \left [ \widehat { P } _ { \alpha } \right ] \right | > \frac { \delta } { 2 } \, P _ { \alpha } ( p ) \right ) \leq \frac { 1 } { 3 }.$$
To reduce the probability of error to e , we use the estimate P ̂ a repeatedly for O ( log 1/ ( e )) independent samples X N and take the estimate P ̂ ′ a to be the sample median of the resulting estimates 3 . Specifically, let ̂ P1, ..., P ̂ t denote t -estimates of P a ( p ) obtained by applying P ̂ a to independent sequences X N , and let 1 E i be the indicator function of the event E i = {| ̂ P i -P a ( p ) | > d P a ( p ) } . By the analysis above we have E [ 1 E i ] ≤ 1/3 and hence by Hoeffding's inequality
$$\mathbb { P } \left ( \sum _ { i = 1 } ^ { t } \mathbb { 1 } _ { \mathcal { E } _ { i } } > \frac { t } { 2 } \right ) \leq \exp ( - t / 1 8 ).$$
The claimed bound follows on choosing t = 18 log ( 1/ e ) and noting that if more than half of ̂ P1, ..., P ̂ t satisfy | ̂ P i -P a ( p ) | ≤ d P a ( p , then their median must also satisfy the same condition. )
glyph[squaresolid]
In the remainder of the section, we bound the bias and the variance for our estimators when the number of samples n are of the appropriate order. Denote by f e a , f u , a and f d , t a , respectively, the empirical estimator 1 1 -a log ̂ P e a , the bias-corrected estimator 1 1 -a log ̂ P u , a and the polynomial approximation estimator 1 1 -a log ̂ P d , t a . We begin by analyzing the performances of f e a and f u a and build-up on these steps to analyze f d , t a .
3 This technique is often referred to as the median trick .
## A. Performance of empirical estimator
The empirical estimator was presented in (1). Using the Poisson sampling recipe given above, we derive upper bound for the sample complexity of the empirical estimator by bounding its bias and variance. The resulting bound for a > 1 is given in Theorem 9 and for a < 1 in Theorem 10.
Theorem 9. For a > 1 , 0 < d < 1/2 , and 0 < e < 1 , the estimator f e a satisfies
$$S _ { \alpha } ^ { f _ { \alpha } ^ { e } } ( k, \delta, \epsilon ) \leq O _ { \alpha } \left ( \frac { k } { \min ( \delta ^ { 1 / ( \alpha - 1 ) }, \delta ^ { 2 } ) } \log \frac { 1 } { \epsilon } \right ),$$
for all k sufficiently large.
Proof. Denote l x def = n p . For x a > 1, we bound the bias of the power sum estimator as follows:
$$\text{of. Denote } \lambda _ { x } \stackrel { \text{def} } { = } n p _ { x }. \text{ For } \alpha > 1, \, \text{we bound the bias of the power sum estimator as follows: } \\ \left | \mathbb { E } \left [ \frac { \sum _ { x } N ^ { a } _ { x } } { n ^ { a } } \right ] - P _ { \alpha } ( p ) \right | & \leq \frac { ( a ) } { n ^ { a } } \sum _ { x } | \mathbb { E } [ N ^ { a } _ { x } ] - \lambda ^ { a } _ { x } | \\ & \stackrel { ( b ) } { \leq } \frac { \alpha } { n ^ { a } } \sum _ { x } \left ( 2 ^ { a } \lambda _ { x } + ( 2 ^ { a } + 1 ) \lambda ^ { a - 1 / 2 } _ { x } \right ) \\ & = \frac { \alpha 2 ^ { a } } { n ^ { a - 1 } } + \frac { \alpha ( 2 ^ { a } + 1 ) } { \sqrt { n } } P _ { \alpha - 1 / 2 } ( p ) \\ & \stackrel { ( c ) } { \leq } \alpha \left ( 2 ^ { a } \left ( \frac { k } { n } \right ) ^ { \alpha - 1 } + ( 2 ^ { a } + 1 ) \sqrt { \frac { k } { n } } \right ) P _ { \alpha } ( p ) \\ & \leq 2 \alpha 2 ^ { a } \left [ \left ( \frac { k } { n } \right ) ^ { \alpha - 1 } + \left ( \frac { k } { n } \right ) ^ { 1 / 2 } \right ] P _ { \alpha } ( p ), \\ \text{ere } ( a ) \, \text{is from the triangle inequality, } ( b ) \, \text{from Lemma } 4, \, \text{and } ( c ) \, \text{follows from Lemma } 1$$
where ( a ) is from the triangle inequality, ( b ) from Lemma 4, and ( c ) follows from Lemma 1 and (4). Thus, the bias of the estimator is less than d a ( -1 ) P a ( p /2 when )
$$n \geq k \cdot \left ( \frac { 8 \alpha 2 ^ { \alpha } } { \delta ( \alpha - 1 ) } \right ) ^ { \max ( 2, 1 / ( \alpha - 1 ) ) }.$$
Similarly, to bound the variance, using independence of multiplicities:
$$\text{Similarly, to bound the variance, using independence of multiplicities:} \\ \text{Var} \left [ \sum _ { x } \frac { N _ { x } ^ { \alpha } } { n ^ { \alpha } } \right ] & = \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x } \text{Var} [ N _ { x } ^ { \alpha } ] \\ & = \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x } \text{E} \left [ N _ { x } ^ { 2 \alpha } \right ] - \left [ \text{E} N _ { x } ^ { \alpha } \right ] ^ { 2 } \\ & \leq \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x } \text{E} \left [ N _ { x } ^ { 2 \alpha } \right ] - \lambda _ { x } ^ { 2 \alpha } \\ & \leq \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x } \left | \text{E} \left [ N _ { x } ^ { 2 \alpha } \right ] - \lambda _ { x } ^ { 2 \alpha } \right | \\ & \leq \frac { 2 \alpha } { n ^ { 2 \alpha } } \sum _ { x } \left ( 2 ^ { 2 \alpha } \lambda _ { x } + ( 2 ^ { 2 \alpha } + 1 ) \lambda _ { x } ^ { 2 \alpha - 1 / 2 } \right ) \\ & = \frac { 2 \alpha \alpha ^ { 2 \alpha } } { n ^ { 2 \alpha - 1 } } + \frac { 2 \alpha ( 2 ^ { 2 \alpha } + 1 ) } { \sqrt { n } } P _ { 2 \alpha - 1 / 2 } ( p ) \\ & \leq 2 \alpha 2 ^ { 2 \alpha } \left ( \frac { k } { n } \right ) ^ { 2 \alpha - 1 } P _ { \alpha } ( p ) ^ { 2 } + 2 \alpha ( 2 ^ { 2 \alpha } + 1 ) \left ( \frac { k ^ { \alpha - \frac { 1 } { n } } } { n } \right ) ^ { 1 / 2 } P _ { \alpha } ( p ) ^ { 2 } \\ ( a ) \text{ is from Jensen's inequality since } z ^ { 4 } \text{ is convex and } \text{E} [ N _ { x } ] = \lambda _ { x }, ( c ) \text{ follows from Lemma 1.} \\ \text{Thus, the variance is less than } \delta ^ { 2 } ( \alpha - 1 ) ^ { 2 } P _ { \left ( n \right ) ^ { 2 } / 1 2 } \text{ when}$$
( a ) is from Jensen's inequality since z a is convex and E [ Nx ] = l x , ( c ) follows from Lemma 1. Thus, the variance is less than d 2 ( a -1 ) 2 P a ( p ) 2 /12 when
$$n \geq k \cdot \max \Big \{ \frac { 4 8 \alpha 2 ^ { 2 \alpha } } { \delta ^ { 2 } ( \alpha - 1 ) ^ { 2 } } \Big ) ^ { 1 / ( 2 \alpha - 1 ) }, \Big ( \frac { 9 6 \alpha 2 ^ { 2 \alpha } } { k ^ { 1 / 2 \alpha } \delta ^ { 2 } ( \alpha - 1 ) ^ { 2 } } \Big ) ^ { 2 } \Big \} = k \cdot \Big ( \frac { 4 8 \alpha 2 ^ { 2 \alpha } } { \delta ^ { 2 } ( \alpha - 1 ) ^ { 2 } } \Big ) ^ { 1 / ( 2 \alpha - 1 ) },$$
where the equality holds for k sufficiently large. The theorem follows by using Lemma 8. glyph[squaresolid]
Theorem 10. For a < 1 , d > 0 , and 0 < e < 1 , the estimator f e a satisfies
$$S _ { \alpha } ^ { f _ { \alpha } ^ { e } } ( k, \delta, \epsilon ) \leq O \left ( \frac { k ^ { 1 / \alpha } } { \delta ^ { \max \{ 4, 2 / \alpha \} } } \log \frac { 1 } { \epsilon } \right ).$$
Proof. For a < 1, once again we take a recourse to Lemma 4 to bound the bias as follows:
$$1, \, \text{ice again we take a recourse to Lemina + toDouglas me dias } \begin{array} { c } \\ \left | \mathbb { E } \left [ \frac { \sum _ { \alpha } N _ { \alpha } ^ { \alpha } } { n ^ { \alpha } } \right ] - P _ { \alpha } ( p ) \right | & \leq \frac { 1 } { n ^ { \alpha } } \sum _ { x } | \mathbb { E } [ N _ { x } ^ { \alpha } ] - \lambda _ { x } ^ { \alpha } | \\ & \leq \frac { 1 } { n ^ { \alpha } } \sum _ { x } \min \left ( \lambda _ { x } ^ { \alpha }, \lambda _ { x } ^ { \alpha - 1 } \right ) \\ & \leq \frac { 1 } { n ^ { \alpha } } \left [ \sum _ { x \notin A } \lambda _ { x } ^ { \alpha } + \sum _ { x \in A } \lambda _ { x } ^ { \alpha - 1 } \right ], \end{array}$$
for every subset A ⊂ [ k ] . Upon choosing A = { x : l x ≥ } 1 , we get
$$\partial _ { \ A } \times [ \kappa ] \cdot \cup \text{powlin} \, \ A = \{ \lambda \colon \lambda _ { x } \leq 1 \}, \, \text{we get} \\ \left | \mathbb { E } \left [ \frac { \sum _ { x } N _ { x } ^ { \alpha } } { n ^ { \alpha } } \right ] - P _ { \alpha } ( p ) \right | & \leq 2 \frac { k } { n ^ { \alpha } } \\ & = 2 \left ( \frac { k ^ { 1 / \alpha } } { n } \right ) ^ { \alpha } \\ & \leq 2 P _ { \alpha } ( p ) \left ( \frac { k ^ { 1 / \alpha } } { n } \right ) ^ { \alpha }, \\ \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \ dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots.$$
where the last inequality uses (4). For bounding the variance, note that
$$\text{e} u e a n e q u a n y \text{ uses } ( \nu, \, \mathcal { U } \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \text{V} \, \begin{array} { c } \text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text$$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\font{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\font$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text $\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text
$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text#$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text}$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text'$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text.$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text`$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text1$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text \,
}$$
Consider the first term on the right-side. For a ≤ 1/2, it is bounded above by 0 since z 2 a is concave in z , and for a > 1/2 the bound in (8) and Lemma 1 applies to give
$$\frac { 1 } { n ^ { 2 \alpha } } \sum _ { x } \mathbb { E } \left [ N _ { x } ^ { 2 \alpha } \right ] - \lambda _ { x } ^ { 2 \alpha } \leq 2 \alpha \left ( \frac { c } { n ^ { 2 \alpha - 1 } } + ( c + 1 ) \sqrt { \frac { k } { n } } \right ) P _ { \alpha } ( \mathfrak { p } ) ^ { 2 }.$$
For the second term, we have
$$\sum _ { x } \lambda _ { x } ^ { 2 \alpha } - [ \mathbb { E } N _ { x } ^ { \alpha } ] ^ { 2 } & = \sum _ { x } \left ( \lambda _ { x } ^ { \alpha } - \mathbb { E } [ N _ { x } ^ { \alpha } ] \right ) \left ( \lambda _ { x } ^ { \alpha } + \mathbb { E } [ N _ { x } ^ { \alpha } ] \right ) \\ & \stackrel { ( a ) } { \leq } 2 n ^ { \alpha } P _ { \alpha } ( p ) \left ( \frac { k ^ { 1 / \alpha } } { n } \right ) ^ { \alpha } \sum _ { x } \left ( \lambda _ { x } ^ { \alpha } + \mathbb { E } [ N _ { x } ^ { \alpha } ] \right ) \\ & \stackrel { ( b ) } { \leq } 4 n ^ { 2 \alpha } P _ { \alpha } ( p ) ^ { 2 } \left ( \frac { k ^ { 1 / \alpha } } { n } \right ) ^ { \alpha }, \\. \ e _ { 1 } \ e _ { 2 } \ e _ { 3 } \ e _ { 4 } \ e _ { 5 } \ e _ { 6 } \ e _ { 7 } \ e _ { 8 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 } \ e _ { 9 }$$
where ( a ) is from (9) and ( b ) from the concavity of z a in z . The proof is completed by combining the two bounds above and using Lemma 8. glyph[squaresolid]
In fact, we show in the appendix that the dependence on k implied by the previous two results are optimal.
Theorem 11. Given a sufficiently small d , the sample complexity S f e a a ( k , d e , ) of the empirical estimator
f e a is bounded below as
$$S _ { \alpha } ^ { f ^ { e } _ { \alpha } } ( k, \delta, 0. 9 ) = \begin{cases} \Omega \left ( \frac { k } { \delta } \right ), & \alpha > 1, \\ \Omega \left ( \frac { k ^ { 1 / \alpha } } { \delta ^ { 1 / \alpha } } \right ), & \alpha < 1. \end{cases}$$
While the performance of the empirical estimator is limited by these bounds, below we exhibit estimators that beat these bounds and thus outperform the empirical estimator.
## B. Performance of bias-corrected estimator for integral a
To reduce the sample complexity for integer orders a > 1 to below k , we follow the development of Shannon entropy estimators. Shannon entropy was first estimated via an empirical estimator, analyzed in, for instance, [2]. However, with o ( k ) samples, the bias of the empirical estimator remains high [33]. This bias is reduced by the Miller-Madow correction [25], [33], but even then, O k ( ) samples are needed for a reliable Shannon-entropy estimation [33].
Similarly, we reduce the bias for R´ enyi entropy estimators using unbiased estimators for p a x for integral a . We first describe our estimator, and in Theorem 12 we show that for 1 < a ∈ Z , ̂ P u a estimates P a ( p ) using O k ( 1 -1/ a / d 2 ) samples. Theorem 16 in Section V shows that this number is optimal up to constant factors.
Consider the unbiased estimator for P a ( p ) given by
$$\widehat { P } _ { \alpha } ^ { \mathbf u } \stackrel { \text{def} } { = } \sum _ { x } \frac { N _ { x } ^ { \underline { \alpha } } } { n ^ { \alpha } },$$
which is unbiased since by Lemma 2,
$$\mathbb { E } \left [ \widehat { P } _ { \alpha } ^ { \mathfrak { u } } \right ] = \sum _ { x } \mathbb { E } \left [ \frac { N _ { x } ^ { \underline { \alpha } } } { n ^ { \alpha } } \right ] = \sum _ { x } p _ { x } ^ { \alpha } = P _ { \alpha } ( \mathfrak { p } ).$$
Our bias-corrected estimator for H a ( p ) is
$$\hat { H } _ { \alpha } = \frac { 1 } { 1 - \alpha } \log \widehat { P } _ { \alpha } ^ { u }.$$
The next result provides a bound for the number of samples needed for the bias-corrected estimator.
Theorem 12. For an integer a > 1 , any d > 0 , and 0 < e < 1 , the estimator f u a satisfies
$$S _ { \alpha } ^ { f _ { \alpha } ^ { u } } ( k, \delta, \epsilon ) \leq O \left ( \frac { k ^ { ( \alpha - 1 ) / \alpha } } { \delta ^ { 2 } } \log \frac { 1 } { \epsilon } \right ).$$
Proof. Since the bias is 0, we only need to bound the variance to use Lemma 8. To that end, we have
$$\text{Var} \left [ \frac { \sum _ { x } N _ { x } ^ { \alpha } } { n ^ { \alpha } } \right ] & = \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x } \text{Var} \left [ N _ { x } ^ { \alpha } \right ] \\ & \leq \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x } \left ( \lambda _ { x } ^ { \alpha } ( \lambda _ { x } + \alpha ) ^ { \alpha } - \lambda _ { x } ^ { 2 \alpha } \right ) \\ & = \frac { 1 } { n ^ { 2 \alpha } } \sum _ { r = 0 } ^ { \alpha - 1 } \sum _ { x } \binom { \alpha } { r } \alpha ^ { \alpha - r } \lambda _ { x } ^ { \alpha + r } \\ & = \frac { 1 } { n ^ { 2 \alpha } } \sum _ { r = 0 } ^ { \alpha - 1 } n ^ { \alpha + r } \binom { \alpha } { r } \alpha ^ { \alpha - r } P _ { \alpha + r } ( p ), \\ \text{equality uses Lemma 2. It follows from Lemma 1 that}$$
where the inequality uses Lemma 2. It follows from Lemma 1 that
$$\text{lally uses Lemil} \, e. \, \text{ in follows in Lemil} \, i \, u \, \text{a} \\ \frac { 1 } { n ^ { 2 \alpha } } \frac { \text{Var} \left [ \sum _ { \alpha } N _ { \alpha } ^ { \underline { \alpha } } \right ] } { P _ { \alpha } ( p ) ^ { 2 } } & \leq \frac { 1 } { n ^ { 2 \alpha } } \sum _ { r = 0 } ^ { \alpha - 1 } n ^ { \alpha + r } \binom { \alpha } { r } \alpha ^ { \alpha - r } \frac { P _ { \alpha + r } ( p ) } { P _ { \alpha } ( p ) ^ { 2 } } \\ & \leq \sum _ { r = 0 } ^ { \alpha - 1 } n ^ { r - \alpha } \binom { \alpha } { r } \alpha ^ { \alpha - r } k ^ { ( \alpha - 1 ) ( \alpha - r ) / \alpha } \\ & \leq \sum _ { r = 0 } ^ { \alpha - 1 } \left ( \frac { \alpha ^ { 2 } k ^ { ( \alpha - 1 ) / \alpha } } { n } \right ) ^ { \alpha - r }, \\ \text{van} \, \& ^ { 2 } / 1 2 \, \text{if} \, \alpha ^ { 2 } k ^ { 1 - 1 / \alpha } / n \text{ for all} \, \& \text{efficient} \, \text{v} \, \text{email} \, \text{ Annlvl} \, \alpha \, \text{ Annlvl} \, \alpha \, \text{ nm}$$
which is less than d 2 /12 if a 2 k 1 -1/ a / , for all n d sufficiently small. Applying Lemma 8 completes the proof. glyph[squaresolid]
## C. The polynomial approximation estimator
Concurrently with a conference version of this paper [1], a polynomial approximation based approach was proposed in [16] and [43] for estimating additive functions of the form GLYPH<229> x f ( p x ) . As seen in Theorem 12, polynomials of probabilities have succinct unbiased estimators. Motivated by this observation, instead of estimating f , these papers consider estimating a polynomial that is a good approximation to f . The underlying heuristic for this approach is that the difficulty in estimation arises from small probability symbols since empirical estimation is nearly optimal for symbols with large probabilities. On the other hand, there is no loss in estimating a polynomial approximation of the function of interest for symbols with small probabilities.
In particular, [16] considered the problem of estimating power sums P a ( p ) up to additive accuracy and showed that O ( k 1/ a /log k ) samples suffice for a < 1. Since P a ( p ) ≥ 1 for a < 1, this in turn implies a similar sample complexity for estimating H a ( p ) for a < 1. On the other hand, a > 1, the power sum P a ( p ) ≤ 1 and can be small ( . e g ., it is k 1 -a for the uniform
distribution). In fact, we show in the Appendix that additive-accuracy estimation of power sum is easy for a > 1 and has a constant sample complexity. Therefore, additive guarantees for estimating the power sums are insufficient to estimate the R´ enyi entropy . Nevertheless, our analysis of the polynomial estimator below shows that it attains the O ( log k ) improvement in sample complexity over the empirical estimator even for the case a > 1.
We first give a brief description of the polynomial estimator of [43] and then in Theorem 13 prove that for a > 1 the sample complexity of ̂ P d , t a is O k ( /log k ) . For completeness, we also include a proof for the case a < 1, which is slightly different from the one in [16].
Let N 1 , N 2 be independent Poi ( n ) random variables. We consider Poisson sampling with two set of samples drawn from p, first of size N 1 and the second N 2. Note that the total number of samples N = N 1 + N 2 ∼ Poi 2 ( n ) . The polynomial approximation estimator uses different estimators for different estimated values of symbol probability p x . We use the first N 1 samples for comparing the symbol probabilities p x with t / n and the second is used for estimating p a x . Specifically, denote by Nx and Nx ′ the number of appearances of x in the N 1 and N 2 samples, respectively. Note that both Nx and Nx ′ have the same distribution Poi ( n p x ) . Let t be a threshold, and d be the degree chosen later. Given a threshold t , the polynomial approximation estimator is defined as follows:
Nx ′ > t : For all such symbols, estimate p a x using the empirical estimate ( Nx / n ) a . Nx ′ ≤ t : Suppose q ( x ) = GLYPH<229> d m = 0 amx m is the polynomial satisfying Lemma 5. Since we expect p x to be less than 2 t / n in this case, we estimate p a x using an unbiased estimate of 4 ( 2 t / n ) a q n ( p /2 x t ) , namely
$$\left ( \sum _ { m = 0 } ^ { d } \frac { a _ { m } ( 2 \tau ) ^ { \alpha - m } N _ { x } ^ { \underline { m } } } { n ^ { \alpha } } \right ).$$
Therefore, for a given t and d the combined estimator ̂ P d , t a is
$$\widehat { P } _ { \alpha } ^ { d, \tau } \stackrel { \text{def} } { = } \sum _ { x \colon N _ { x } ^ { \prime } \leq \tau } \left ( \sum _ { m = 0 } ^ { d } \frac { a _ { m } ( 2 \tau ) ^ { \alpha - m } N _ { x } ^ { \underline { m } } } { n ^ { \alpha } } \right ) + \sum _ { x \colon N _ { x } ^ { \prime } > \tau } \left ( \frac { N _ { x } } { n } \right ) ^ { \alpha }.$$
Denoting by ˆ p x the estimated probability of the symbol x , note that the polynomial approximation estimator relies on the empirical estimator when p ˆ x > t / n and uses the the bias-corrected estimator for estimating each term in the polynomial approximation of p a x when ˆ p x ≤ t / . n
4 Note that if | q ( x ) -x a | < e for all x ∈ [ 0, 1 , ] then | h a q ( x / eta ) -x a | < h a e for all x ∈ [ 0, h ] .
We derive upper bounds for the sample complexity of the polynomial approximation estimator.
Theorem 13. For a > 1 , d > 0 , 0 < e < 1 , there exist constants c 1 and c 2 such that the estimator ̂ P d , t a with t = c 1 log n and d = c 2 log n satisfies
$$S _ { \alpha } ^ { \widehat { P } _ { \alpha } ^ { d, \tau } } ( k, \delta, \epsilon ) \leq O \left ( \frac { k } { \log k } \frac { \log ( 1 / \epsilon ) } { \delta ^ { 1 / \alpha } } \right ).$$
Proof. We follow the approach in [43] closely. Choose t = c ∗ log n such that with probability at least 1 -e the events Nx ′ > t and Nx ′ ≤ t do not occur for all symbols x satisfying p x ≤ t / 2 ( n ) and p x > 2 t / , respectively. Or equivalently, with probability at least 1 n -e all symbols x such that Nx ′ > t satisfy p x > t / 2 ( n ) and all symbols such that Nx ′ ≤ t satisfy p x ≤ 2 t / . We n condition on this event throughout the proof. For concreteness, we choose c ∗ = 4, which is a valid choice for n > 20 log ( 1/ e ) by the Poisson tail bound and the union bound.
Let q ( x ) = GLYPH<229> d m = 0 amx m satisfy the polynomial approximation error bound guaranteed by Lemma 5, i . e .,
$$\max _ { x \in ( 0, 1 ) } | q ( x ) - x ^ { \alpha } | < c _ { \alpha } / d ^ { 2 \alpha }$$
To bound the bias of ̂ P d , t a , note first that for Nx ′ < t (assuming p x ≤ 2 t / nsmp and estimating ( 2 t / n ) a q n ( p /2 x t ) )
$$( 2 \tau / n ) ^ { \alpha } q ( n p _ { x } / 2 \tau ) ) \\ \left | \mathbb { E } \left [ \sum _ { m = 0 } ^ { d } \frac { a _ { m } ( 2 \tau ) ^ { \alpha - m } N _ { x } ^ { m } } { n ^ { \alpha } } \right ] - p _ { x } ^ { \alpha } \right | & = \left | \sum _ { m = 0 } ^ { d } a _ { m } \left ( \frac { 2 \tau } { n } \right ) ^ { \alpha - m } p _ { x } ^ { m } - p _ { x } ^ { \alpha } \right | \\ & = \frac { ( 2 \tau ) ^ { \alpha } } { n ^ { \alpha } } \left | \sum _ { m = 0 } ^ { d } a _ { m } \left ( \frac { n p _ { x } } { 2 \tau } \right ) ^ { m } - \left ( \frac { n p _ { x } } { 2 \tau } \right ) ^ { \alpha } \right | \\ & = \frac { ( 2 \tau ) ^ { \alpha } } { n ^ { \alpha } } \left | q \left ( \frac { n p _ { x } } { 2 \tau } \right ) - \left ( \frac { n p _ { x } } { 2 \tau } \right ) ^ { \alpha } \right | \\ & < \frac { ( 2 \tau ) ^ { \alpha } c _ { \alpha } } { ( n d ^ { 2 } ) ^ { \alpha } }, \\ \text{where the last inequality uses (13) and } n p _ { x } / ( 2 \tau ) \leq 1.$$
where the last inequality uses (13) and n p / 2 x ( t ) ≤ 1.
For Nx ′ > t , the bias of empirical part of the power sum is bounded as
$$\begin{array} { c } \cdot & \cdot \\ \left | \mathbb { E } \left [ \left ( \frac { N _ { x } } { n } \right ) ^ { \alpha } \right ] - p _ { x } ^ { \alpha } \right | & \stackrel { ( a ) } { \leq } \alpha c \frac { p _ { x } } { n ^ { \alpha - 1 } } + \alpha ( c + 1 ) \frac { p _ { x } ^ { \alpha - \frac { 1 } { 2 } } } { \sqrt { n } } \\ & \stackrel { ( b ) } { \leq } \alpha c \frac { p _ { x } ^ { \alpha } } { ( \tau / 2 ) ^ { \alpha - 1 } } + \alpha ( c + 1 ) \frac { p _ { x } ^ { \alpha } } { \sqrt { \tau / 2 } }, \end{array}$$
and ( a ) is from Lemma 4 and ( b ) from p x > t / 2 ( n ) , which holds when Nx ′ > t . Thus, by using the triangle inequality and applying the bounds above to each term, we obtain the following bound on the bias of ̂ P d , t a :
$$\begin{array} { c } \ddot { \left | \mathbb { E } \left [ \widehat { \mathbb { P } } _ { \alpha } \right ] - P _ { \alpha } ( \mathbf p ) \right | } & \leq \frac { k ( 2 \tau ) ^ { \alpha } c _ { \alpha } } { ( n d ^ { 2 } ) ^ { \alpha } } + \alpha P _ { \alpha } ( \mathbf p ) \left [ \frac { c } { ( \tau / 2 ) ^ { \alpha - 1 } } + \frac { c + 1 } { \sqrt { \tau / 2 } } \right ] \\ & \leq P _ { \alpha } ( \mathbf p ) \left [ c _ { \alpha } \left ( \frac { k \cdot 2 \tau } { n d ^ { 2 } } \right ) ^ { \alpha } + \frac { \alpha c } { ( \tau / 2 ) ^ { \alpha - 1 } } + \frac { \alpha ( c + 1 ) } { \sqrt { \tau / 2 } } \right ], \\ \cdot & \quad. & \quad. & \quad. \quad \dots \end{array}$$
where the last inequality uses (4).
For variance, independence of multiplicities under Poisson sampling gives
$$\mathbb { V }ar \left [ \widehat { P } _ { \alpha } \right ] = \sum _ { x \colon N _ { x } ^ { \prime } \leq \tau } \mathbb { V }ar \left ( \sum _ { m = 0 } ^ { d } \frac { a _ { m } ( 2 \tau ) ^ { \alpha - m } N _ { x } ^ { m } } { n ^ { \alpha } } \right ) \ + \sum _ { x \colon N _ { x } > \tau } \mathbb { V }ar \left ( \frac { N _ { x } } { n } \right ) ^ { \alpha }.$$
Let a = max m | am | . By Lemma 2, for any x with p x ≤ 2 t / , n
$$\text{Lex u} - \text{illau} _ { m } | u _ { m } | \cdot \text{By Leillula } 4, \text{for only $A$ will } p _ { x } \geq 4 \, \iota / n, \\ \text{Var} \left ( \sum _ { m = 0 } ^ { d } \frac { a _ { m } ( 2 \tau ) ^ { \alpha - m } _ { X } } { n ^ { \alpha } } \right ) & \leq a ^ { 2 } d ^ { 2 } \max _ { 1 \leq m \leq d } \left \{ \frac { ( 2 \tau ) ^ { 2 \alpha - 2 m } } { n ^ { 2 \alpha } } \text{Var} N _ { X } ^ { m } \right \} \\ & \stackrel { ( a ) } { \leq } a ^ { 2 } d ^ { 2 } \max _ { 1 \leq m \leq d } \left \{ \frac { ( 2 \tau ) ^ { 2 \alpha - 2 m } } { n ^ { 2 \alpha } } ( n p _ { x } ) ^ { m } ( ( n p _ { x } + m ) ^ { m } - n p _ { x } ^ { m } ) \right \} \\ & \stackrel { ( b ) } { \leq } \frac { a ^ { 2 } d ^ { 2 } ( 2 \tau + d ) ^ { 2 \alpha } } { n ^ { 2 \alpha } }, \\ \text{who} \, / \alpha \, \text{is from I am} \, \varphi \, \text{ and } ( h ) \, \text{from miocin} \, \varphi \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \,\text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in} \, \varphi \, \text{in}$$
where ( a ) is from Lemma 2, and ( b ) from plugging n p x ≤ 2 t . Furthermore, using similar steps as (8) together with Lemma 4, for x with p x > t / 2 ( n ) we get
$$\text{Var} \left [ \left ( \frac { N _ { x } } { n } \right ) ^ { \alpha } \right ] \leq 2 \alpha c \frac { p _ { x } ^ { 2 \alpha } } { ( \tau / 2 ) ^ { 2 \alpha - 1 } } + 2 \alpha ( c + 1 ) \frac { p _ { x } ^ { 2 \alpha } } { \sqrt { \tau / 2 } }.$$
The two bounds above along with Lemma 1 and (4) yield
$$\text{Var} \left [ \widehat { P } _ { \alpha } \right ] \leq P _ { \alpha } ( p ) ^ { 2 } \left [ \frac { a ^ { 2 } d ^ { 2 } ( 2 \tau + d ) ^ { 2 \alpha } } { n } \left ( \frac { k } { n } \right ) ^ { 2 \alpha - 1 } + \frac { 2 \alpha c } { ( \tau / 2 ) ^ { 2 \alpha - 1 } } + \frac { 2 \alpha ( c + 1 ) } { \sqrt { \tau / 2 } } \right ].$$
For d = t /8 = 1 2 log n , the last terms in (15) are o ( 1 ) which gives
$$\left | \mathbb { E } \left [ \widehat { \mathbf P } _ { \alpha } \right ] - P _ { \alpha } ( \mathbf p ) \right | = P _ { \alpha } ( \mathbf p ) \left ( c _ { \alpha } \left ( \frac { 3 2 k } { ( n \log n ) } \right ) ^ { \alpha } + o ( 1 ) \right ).$$
Recall from (6) that a < ( 1 + c a / d 2 a )( √ 2 + ) 1 d , and therefore, a 2 = O (( √ 2 + ) 1 log n ) = n c 0 for
some c 0 < 1. Using (18) we get
$$\mathbb { V } a r \left [ \widehat { \mathbf P } _ { \alpha } \, \right ] = O \left ( P _ { \alpha } ( \mathbf p ) ^ { 2 } \frac { n ^ { c _ { 0 } } \log ^ { 2 \alpha + 2 } n } { n } \left ( \frac { k } { n } \right ) ^ { 2 \alpha - 1 } \right ).$$
Therefore, the result follows from Lemma 8 for k sufficiently large.
glyph[squaresolid]
We now prove an analogous result for a < 1.
Theorem 14. For a < 1 , d > 0 , 0 < e < 1 , there exist constants c 1 and c 2 such that the estimator ̂ P d , t a with t = c 1 log n and d = c 2 log n satisfies
$$S _ { \alpha } ^ { \widehat { P } _ { \alpha } ^ { d, \tau } } ( k, \delta, \epsilon ) \leq O \left ( \frac { k ^ { 1 / \alpha } } { \log k } \frac { \log ( 1 / \epsilon ) } { \alpha ^ { 2 } \delta ^ { 1 / \alpha } } \right ).$$
Proof. We proceed as in the previous proof and set t to be 4 log n . The contribution to the bias of the estimator for a symbol x with Nx ′ < t remains bounded as in (14). For a symbol x with Nx ′ > t , the bias contribution of the empirical estimator is bounded as
$$\left | \mathbb { E } \left [ \left ( \frac { N _ { x } } { n } \right ) ^ { \alpha } \right ] - p _ { x } ^ { \alpha } \right | & \stackrel { ( a ) } { \leq } \frac { p _ { x } ^ { \alpha - 1 } } { n } \\ & \stackrel { ( b ) } { \leq } \frac { 2 p _ { x } ^ { \alpha } } { \tau },$$
where ( a ) is by Lemma 4 and ( b ) uses p x > t / 2 ( n ) , which holds if Nx ′ > t . Thus, we obtain the following bound on the bias of ̂ P d , t a :
$$\left | \mathbb { E } \left [ \widehat { P } _ { \alpha } \right ] - P _ { \alpha } ( \mathbf p ) \right | \leq & \frac { k ( 2 \tau ) ^ { \alpha } c _ { \alpha } } { ( n d ^ { 2 } ) ^ { \alpha } } + \frac { 2 } { \tau } P _ { \alpha } ( \mathbf p ) \\ \leq & P _ { \alpha } ( \mathbf p ) \left [ c _ { \alpha } \left ( \frac { k ^ { 1 / \alpha } \cdot 2 \tau } { n d ^ { 2 } } \right ) ^ { \alpha } + \frac { 2 } { \tau } \right ],$$
where the last inequaliy is by (4).
To bound the variance, first note that bound (17) still holds for p x ≤ 2 t / . To bound the n contribution to the variance from the terms with n p x > t /2, we borrow steps from the proof of Theorem 10. In particular, (10) gives
$$\text{Var} \left [ \sum _ { x \colon N _ { x } ^ { \prime } > \tau } \frac { N _ { x } ^ { \alpha } } { n ^ { \alpha } } \right ] \leq \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x \colon N _ { x } ^ { \prime } > \tau } \mathbb { E } \left [ N _ { x } ^ { 2 \alpha } \right ] - \lambda _ { x } ^ { 2 \alpha } + \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x \colon N _ { x } ^ { \prime } > \tau } \lambda _ { x } ^ { 2 \alpha } - \left [ \mathbb { E } N _ { x } ^ { \alpha } \right ] ^ { 2 }.$$
The first term can be bounded in the manner of (11) as
$$\frac { 1 } { n ^ { 2 \alpha } } \sum _ { x \colon N _ { x } ^ { \prime } > \tau } \mathbb { E } \left [ N _ { x } ^ { 2 \alpha } \right ] - \lambda _ { x } ^ { 2 \alpha } \leq 2 \alpha \left ( \frac { c } { n ^ { 2 \alpha - 1 } } + ( c + 1 ) \frac { 1 } { \sqrt { \tau / 2 } } \right ) P _ { \alpha } ( \mathbf p ) ^ { 2 },$$
For the second term, we have
$$& \text{the second term, we have} \\ & \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x \colon N _ { x } ^ { \prime } > \tau } \lambda _ { x ^ { 2 \alpha } } ^ { 2 \alpha } - [ \mathbb { E } N _ { x } ^ { \alpha } ] ^ { 2 } = \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x \colon N _ { x } ^ { \prime } > \tau } \left ( \lambda _ { x } ^ { \alpha } - \mathbb { E } [ N _ { x } ^ { \alpha } ] \right ) ( \lambda _ { x } ^ { \alpha } + \mathbb { E } [ N _ { x } ^ { \alpha } ] ) \\ & \stackrel { ( a ) } { \leq } \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x \colon N _ { x } ^ { \prime } > \tau } \left ( \lambda _ { x } ^ { \alpha - 1 } \right ) ( 2 \lambda _ { x } ^ { \alpha } ) \\ & \quad = & 2 \sum _ { x \colon N _ { x } ^ { \prime } > \tau } \frac { p _ { x } ^ { 2 \alpha } } { n p _ { x } } \\ & \stackrel { ( b ) } { \leq } \frac { 4 } { \tau } P _ { \alpha } ( p ) ^ { 2 }, \\ & \text{re} \left ( a \right ) follows from Lemma 4 and concavity of $z^{\alpha}$ in $z$ and $(b)$ from $np_{x}>\tau/2$ and Lemn}$$
where ( a ) follows from Lemma 4 and concavity of z a in z and ( b ) from n p x > t /2 and Lemma 1. Thus, the contribution of the terms corresponding to Nx ′ > t in the bias and the variance are P a ( p ) · o ( 1 ) and P a ( p ) 2 · o ( 1 , ) respectively, and can be ignored. Choosing d = a 2 log n and combining the observations above, we get the following bound for the bias:
$$\left | \mathbb { E } \left [ \widehat { P } _ { \alpha } \right ] - P _ { \alpha } ( \mathbf p ) \right | = P _ { \alpha } ( \mathbf p ) \left ( c _ { \alpha } \left ( \frac { 3 2 k ^ { 1 / \alpha } } { n \log n \alpha ^ { 2 } } \right ) ^ { \alpha } + o ( 1 ) \right ),$$
and, using (17), the following bound for the variance:
$$\text{Var} \left [ \widehat { P } _ { \alpha } \right ] & \leq k \frac { a ^ { 2 } d ^ { 2 } ( 2 \tau + d ) ^ { 2 \alpha } } { n ^ { 2 \alpha } } + P _ { \alpha } ( p ) ^ { 2 } \cdot o ( 1 ) \\ & \leq P _ { \alpha } ( p ) ^ { 2 } \left [ \left ( \frac { a ^ { 2 } } { n ^ { \alpha } } \right ) ( 9 \log n ) ^ { 2 \alpha + 2 } \left ( \frac { k ^ { 1 / \alpha } } { n } \right ) ^ { \alpha } + o ( 1 ) \right ]$$
Here a 2 is the largest squared coefficient of the approximating polynomial and, by (6), is O ( 2 2 c 0 d ) = O n c ( 0 a ) for some c 0 < 1. Thus, a 2 = ( o n a ) and the proof follows by Lemma 8. glyph[squaresolid]
## IV. EXAMPLES AND EXPERIMENTS
We begin by computing R´ enyi entropy for uniform and Zipf distributions; the latter example illustrates the lack of uniform continuity of H a ( p ) in a .
Example 1 . The uniform distribution U k over [ k ] = { 1, . . . , k } is given by
$$p _ { i } = \frac { 1 } { k } \text{ \ for } i \in [ k ].$$
glyph[negationslash]
Its R´ enyi entropy for every order 1 = a ≥ 0, and hence for all a ≥ 0, is
$$H _ { \alpha } ( U _ { k } ) = \frac { 1 } { 1 - \alpha } \log \sum _ { i = 1 } ^ { k } \frac { 1 } { k ^ { \alpha } } = \frac { 1 } { 1 - \alpha } \log k ^ { 1 - \alpha } = \log k.$$
Example 2 . The Zipf distribution Z b , k for b > 0 and k ∈ [ k ] is given by
$$\mathbf p _ { i } = \frac { i ^ { - \beta } } { \sum _ { j = 1 } ^ { k } j ^ { - \beta } } \quad \text{for } i \in [ k ].$$
glyph[negationslash]
Its R´ enyi entropy of order a = 1 is
$$H _ { \alpha } ( Z _ { \beta, k } ) = \frac { 1 } { 1 - \alpha } \log \sum _ { i = 1 } ^ { k } i ^ { - \alpha \beta } - \frac { \alpha } { 1 - \alpha } \log \sum _ { i = 1 } ^ { k } i ^ { - \beta }.$$
Table II summarizes the leading term g k ( ) in the approximation 5 H a ( Z b , k ) ∼ g k ( ) .
TABLE II: The leading terms g k ( ) in the approximations H a ( Z b , k ) ∼ g k ( ) for different values of ab and b . The case ab = 1 and b = 1 corresponds to the Shannon entropy of Z 1, k .
| | b < 1 | b = 1 | b > 1 |
|--------|--------------------|--------------------|--------------------|
| ab < 1 | log k | 1 - ab 1 - a log k | 1 - ab 1 - a log k |
| ab = 1 | a - ab a - 1 log k | 1 2 log k | 1 1 - a log log k |
| ab > 1 | a - ab a - 1 log k | a a - 1 log log k | constant |
In particular, for a > 1
$$H _ { \alpha } ( Z _ { 1, k } ) = \frac { \alpha } { 1 - \alpha } \log \log k + \Theta \left ( \frac { 1 } { k ^ { \alpha - 1 } } \right ) + c ( \alpha ),$$
and the difference | H 2 ( p ) -H 2 + ( e p ) | is O ( e log log k ) . Therefore, even for very small e this difference is unbounded and approaches infinity in the limit as k goes to infinity.
We now illustrate the performance of the proposed estimators for various distributions for a = 2 in Figures 2 and a = 1.5 in Figures 3. For a = 2, we compare the performance of biascorrected and empirical estimators. For a = 1.5, we compare the performance of the polynomialapproximation and the empirical estimator. For the polynomial-approximation estimator, the
5 We say f ( n ) ∼ g n ( ) to denote lim n → ¥ f ( n ) / g n ( ) = 1.
threshold t is chosen as t = ln ( n ) and the approximating polynomial degree is chosen as d = glyph[ceilingleft] 1.5 t glyph[ceilingright] .
We test the performance of these estimators over six different distributions: the uniform distribution, a step distribution with half of the symbols having probability 1/ ( 2 k ) and the other half have probability 3/ 2 ( k ) , Zipf distribution with parameter 3/4 ( pi GLYPH<181> i -3/4 ), Zipf distribution with parameter 1/2 ( pi GLYPH<181> i -1/2 ), a randomly generated distribution using the uniform prior on the probability simplex, and another one generated using the Dirichlet-1/2 prior.
In both the figures the true value is shown in black and the estimated values are color-coded, with the solid line representing their mean estimate and the shaded area corresponding to one standard deviation. As expected, bias-corrected estimators outperform empirical estimators for a = 2 and polynomial-approximation estimators perform better than empirical estimators for a = 1.5.
(d) Zipf with parameter 1/2 (e) Uniform prior (Dirichlet 1)
(c) Zipf with parameter 3/4
(f) Dirichlet 1/2 prior
Fig. 2: R´ enyi entropy estimates for order 2 for support 10000, number of samples ranging from 1000 to 10000, averaged over 100 trials.
(c) Zipf with parameter 3/4
(d) Zipf with parameter 1/2
(e) Uniform prior (Dirichlet 1)
(f) Dirichlet 1/2 prior
Fig. 3: R´nyi entropy estimates for order 1.5 for support 10000, number of samples ranging from e 1000 to 10000, averaged over 100 trials.
## V. LOWER BOUNDS ON SAMPLE COMPLEXITY
glyph[negationslash]
We now establish lower bounds on S a ( k , d e , ) . The proof is based on exhibiting two distributions p and q with H a ( p ) = H a ( q ) such that the set of Nx 's have very similar distribution from p and q, if fewer samples than the claimed lower bound are available. This method is often referred to as Le Cam's two-point method (see, for instance, [46]). The key idea is summarized in the following result which is easy to derive.
Lemma 15. If for two distributions p and q on X and n ∈ ◆ the total variation distance ‖ p n -q n ‖ < e , then one of the following holds for every function ˆ f :
$$p \left ( | H _ { \alpha } ( \mathbf p ) - \hat { f } ( X ^ { n } ) | & \geq \frac { | H _ { \alpha } ( \mathbf p ) - H _ { \alpha } ( \mathbf q ) | } { 2 } \right ) \geq \frac { 1 - \epsilon } { 2 }, \\ o r \ q \left ( | H _ { \alpha } ( \mathbf q ) - \hat { f } ( X ^ { n } ) | & \geq \frac { | H _ { \alpha } ( \mathbf p ) - H _ { \alpha } ( \mathbf q ) | } { 2 } \right ) \geq \frac { 1 - \epsilon } { 2 }.$$
We first prove the lower bound for integers a > 1, which matches the upper bound in Theorem 12 up to a constant factor.
Theorem 16. Given an 1 < a ∈ ◆ and 1 < e < 1 , for every sufficienly small d > 0
$$S _ { \alpha } ( k, \delta, \epsilon ) = \Omega \left ( \frac { k ^ { ( \alpha - 1 ) / \alpha } } { \delta ^ { 2 } } \right ),$$
where the constant implied by W may depend on e .
Proof. We rely on Lemma 15 and exhibit two distributions p and q with appropriate properties. Specifically, consider the following distributions p and q over [ k ] : p1 = 1/ k 1 -1/ a , and for x = 2, . . . , k , p x = ( 1 -p1 ) / ( k -1 ; q1 ) = ( 1 + ) d / k 1 -1/ a , and for x = 2, . . . , k , q x = ( 1 -q1 ) / ( k -1 . )
Then, we have
$$P _ { \alpha } ( \mathbf p ) & = \frac { 1 } { k ^ { \alpha - 1 } } + ( k - 1 ) \cdot \left ( \frac { 1 - \frac { 1 } { k ^ { 1 - 1 / \alpha } } } { k - 1 } \right ) ^ { \alpha } \\ & = \frac { 1 } { k ^ { \alpha - 1 } } + \frac { 1 } { ( k - 1 ) ^ { \alpha - 1 } } \cdot \left ( 1 - \frac { 1 } { k ^ { 1 - 1 / \alpha } } \right ) ^ { \alpha } \\ & = \frac { ( 2 + o _ { k } ( 1 ) ) } { k ^ { \alpha - 1 } }.$$
Similarly,
$$P _ { \alpha } ( \mathbf q ) = \frac { ( 1 + \delta ) ^ { \alpha } } { k ^ { \alpha - 1 } } + ( k - 1 ) \cdot \left ( \frac { 1 - \frac { ( 1 + \delta ) } { k ^ { 1 - 1 / \alpha } } } { k - 1 } \right ) ^ { \alpha }$$
$$= \frac { ( 2 + \alpha \delta + o _ { k } ( 1 ) ) } { k ^ { \alpha - 1 } }.$$
$$\frac { \mu \omega \top \upsilon _ { k } ( 1 ) ) } { k ^ { \alpha - 1 } }.$$
Therefore, | H a ( p ) -H a ( q ) | = W ( d ) . To complete the proof, we show that there exists a constant C e such that ‖ p n -q n ‖ ≤ e if n ≤ C k e 1 -1/ a / d 2 . To that end, we bound the squared Hellinger distance between p n and q n given by
$$h ^ { 2 } ( p, q ) = 2 - 2 \sum _ { x } \sqrt { p _ { x } q _ { x } } = \sum _ { x } ( \sqrt { p _ { x } } - \sqrt { q _ { x } } ) ^ { 2 }.$$
Since for small values of d we have ( 1 + ) d 1/2 < 1 + d ,
$$h ^ { 2 } ( \mathbf p, \mathbf q ) & = \left ( \sqrt { \frac { 1 + \delta } { k ^ { 1 - 1 / \alpha } } } - \sqrt { \frac { 1 } { k ^ { 1 - 1 / \alpha } } } \right ) ^ { 2 } + \left ( \sqrt { 1 - \frac { 1 + \delta } { k ^ { 1 - 1 / \alpha } } } - \sqrt { 1 - \frac { 1 } { k ^ { 1 - 1 / \alpha } } } \right ) ^ { 2 } \\ & = O \left ( \frac { \delta ^ { 2 } } { k ^ { 1 - 1 / \alpha } } \right ). \\ \quad \ \. \quad \.$$
The required bound for ‖ p n -q n ‖ follows using the following standard steps ( c f . [46])
$$\Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Philon _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \quad \Phi _ { \Phi } \\ \quad \quad \quad$$
glyph[squaresolid]
Next, we lower bound S a ( k ) for noninteger a > 1 and show that it must be almost linear in k . While we still rely on Lemma 15 for our lower bound, we take recourse to Poisson sampling to simplify our calculations.
Lemma 17. (Poisson approximation 2) Suppose there exist d e , > 0 such that, with N ∼ Poi 2 ( n , ) for all estimators ˆ f we have
$$\max _ { p \in \mathcal { P } } \mathbb { P } \left ( | H _ { \alpha } ( p ) - \hat { f } _ { \alpha } ( X ^ { N } ) | > \delta \right ) > \epsilon,$$
where P is a fixed family of distributions. Then, for all fixed length estimators ˜ f
$$\max _ { p \in \mathcal { P } } \mathbb { P } \left ( | H _ { \alpha } ( \mathbf p ) - \tilde { f } _ { \alpha } ( X ^ { n } ) | > \delta \right ) > \frac { \epsilon } { 2 },$$
when n > 4 log ( 2/ e ) .
Also, it will be convenient to replace the observations X N with its profile F = F ( X N ) [31], i.e., F = ( F 1 , F 2, . . . ) where F l is the number of elements x that appear l times in the sequence X N . The following well-known result says that for estimating H a ( p , it suffices to consider only ) the functions of the profile.
Lemma 18. (Sufficiency of profiles). Consider an estimator ˆ f such that
$$\mathbb { P } \left ( | H _ { \alpha } ( \mathbf p ) - \hat { f } ( X ^ { N } ) | > \delta \right ) \leq \epsilon, \ \text{ for all } \mathbf p.$$
Then, there exists an estimator ˜ f ( X N ) = ˜ f ( F ) such that
$$\mathbb { P } \left ( | H _ { \alpha } ( \mathbf p ) - \tilde { f } ( \Phi ) | > \delta \right ) \leq \epsilon, \ \text{ for all } \mathbf p.$$
Thus, lower bounds on the sample complexity will follow upon showing a contradiction for the second inequality above when the number of samples n is sufficiently small. We obtain the required contradiction by using Lemma 15 upon showing there are distributions p and q of support-size k such that the following hold:
- (i) There exists d > 0 such that
$$| H _ { \alpha } ( { \mathbf p } ) - H _ { \alpha } ( { \mathbf q } ) | > \delta ;$$
- (ii) denoting by p F and q F , respectively, the distributions on the profiles under Poisson sampling corresponding to underlying distributions p and q, there exist e > 0 such that
$$\| p _ { \Phi } - q _ { \Phi } \| < \epsilon,$$
if n < k c ( a ) .
Therefore, it suffices to find two distributions p and q with different R´ enyi entropies and with small total variation distance between the distributions of their profiles, when n is sufficiently small. For the latter requirement, we recall a result of [42] that allows us to bound the total variation distance in (21) in terms of the differences of power sums | Pa ( p ) -Pa ( q ) | .
Theorem 19. [42] Given distributions p and q such that
$$\max _ { x } \max \{ p _ { x } ; q _ { x } \} \leq \frac { \epsilon } { 4 0 n },$$
for Poisson sampling with N ∼ Poi ( n , it holds that )
$$\| p _ { \Phi } - q _ { \Phi } \| \leq \frac { \epsilon } { 2 } + 5 \sum _ { a } n ^ { a } | P _ { a } ( \mathbf p ) - P _ { a } ( \mathbf q ) |.$$
It remains to construct the required distributions p and q, satisfying (20) and (21) above. By Theorem 19, the total variation distance ‖ p F -q F ‖ can be made small by ensuring that the power sums of distributions p and q are matched, that is, we need distributions p and q with different R´ enyi entropies and identical power sums for as large an order as possible. To that end, for every positive integer d and every vector x = ( x 1 , ..., xd ) ∈ ❘ d , associate with x a distribution p x of support-size dk such that
$$\mathbf p _ { i j } ^ { \mathbf x } = \frac { | x _ { i } | } { k \| \mathbf x \| _ { 1 } }, \quad 1 \leq i \leq d, \, 1 \leq j \leq k.$$
Note that and for all a
$$P _ { a } \left ( \mathbf p ^ { \mathbf x } \right ) = \frac { 1 } { k ^ { a - 1 } } \left ( \frac { \| \mathbf x \| _ { a } } { \| \mathbf x \| _ { 1 } } \right ) ^ { a }.$$
We choose the required distributions p and q, respectively, as p x and p , where the vectors y x and y are given by the next result.
Lemma 20. For every d ∈ ◆ and a not integer, there exist positive vectors x y , ∈ ❘ d such that glyph[negationslash]
glyph[negationslash]
$$\| \mathbf x \| _ { r } & = \| \mathbf y \| _ { r }, \quad 1 \leq r \leq d - 1, \\ \| \mathbf x \| _ { d } & \neq \| \mathbf y \| _ { d }, \\ \| \mathbf x \| _ { \alpha } & \neq \| \mathbf y \| _ { \alpha }.$$
Proof. Let x = ( 1, ..., d )) . Consider the polynomial
$$p ( z ) = ( z - x _ { 1 } ) \dots ( z - x _ { d } ),$$
and q z ( ) = p z ( ) -D , where D is chosen small enough so that q z ( ) has d positive roots. Let
$$H _ { \alpha } ( p ^ { \mathbf x } ) = \log k + \frac { \alpha } { \alpha - 1 } \log \frac { \| \mathbf x \| _ { 1 } } { \| \mathbf x \| _ { \alpha } },$$
y 1 , ..., yd be the roots of the polynomial q z ( ) . By Newton-Girard identities, while the sum of d th power of roots of a polynomial does depend on the constant term, the sum of first d -1 powers of roots of a polynomial do not depend on it. Since p z ( ) and q z ( ) differ only by a constant, it holds that
$$\sum _ { i = 1 } ^ { d } x _ { i } ^ { r } = \sum _ { i = 1 } ^ { d } y _ { i } ^ { r }, \ \ 1 \leq r \leq d - 1,$$
and that glyph[negationslash]
$$\sum _ { i = 1 } ^ { d } x _ { i } ^ { d } \neq \sum _ { i = 1 } ^ { d } y _ { i } ^ { d }.$$
Furthermore, using a first order Taylor approximation, we have
$$y _ { i } - x _ { i } = \frac { \Delta } { p ^ { \prime } ( x _ { i } ) } + o ( \Delta ),$$
and for any differentiable function g ,
$$g ( y _ { i } ) - g ( x _ { i } ) = g ^ { \prime } ( x _ { i } ) ( y _ { i } - x _ { i } ) + o ( | y _ { i } - x _ { i } | ).$$
It follows that
$$\sum _ { i = 1 } ^ { d } g ( y _ { i } ) - g ( x _ { i } ) = \sum _ { i = 1 } ^ { d } \frac { g ^ { \prime } ( x _ { i } ) } { p ^ { \prime } ( x _ { i } ) } \Delta + o ( \Delta ),$$
and so, the left side above is nonzero for all D sufficiently small provided glyph[negationslash]
$$\sum _ { i = 1 } ^ { d } \frac { g ^ { \prime } ( x _ { i } ) } { p ^ { \prime } ( x _ { i } ) } \neq 0.$$
Upon choosing g x ( ) = x a , we get
$$\sum _ { i = 1 } ^ { d } \frac { g ^ { \prime } ( x _ { i } ) } { p ^ { \prime } ( x _ { i } ) } = \frac { \alpha } { d! } \sum _ { i = 1 } ^ { d } \binom { d } { i } \, ( - 1 ) ^ { d - i } \, i ^ { \alpha }.$$
glyph[negationslash]
glyph[negationslash]
Denoting the right side above by h ( a ) , note that h i ( ) = 0 for i = 1, ..., d -1. Since h ( a ) is a linear combination of d exponentials, it cannot have more than d -1 zeros (see, for instance, [40]). Therefore, h ( a ) = 0 for all a / ∈ { 1, ..., d - } 1 ; in particular, ‖ x ‖ a = ‖ y ‖ a for all D sufficiently small. glyph[squaresolid]
We are now in a position to prove our converse results.
Theorem 21. Given a nonintegral a > 1 , for any fixed 0 < e < 1/2 , we have
$$S _ { \alpha } ( k, \delta, \epsilon ) = & \overset { \approx } { \Omega } \, ( k ).$$
Proof. For a fixed d , let distributions p and q be as in the previous proof. Then, as in the proof of Theorem 21, inequality (20) holds by Lemma 20 and (21) holds by Theorem 19 if n < C k 2 ( d -1 / ) d . The theorem follows since d can be arbitrary large. glyph[squaresolid]
Finally, we show that S a ( k ) must be super-linear in k for a < 1.
Theorem 22. Given a < 1 , for every 0 < e < 1/2 , we have
$$S _ { \alpha } ( k, \delta, \epsilon ) = & \overset { \approx } { \Omega } \left ( k ^ { 1 / \alpha } \right ).$$
Proof. Consider distributions p and q on an alphabet of size kd + 1, where
$$p _ { i j } = \frac { p _ { i j } ^ { x } } { k ^ { \beta } } \text{ and } q _ { i j } = \frac { p _ { i j } ^ { x } } { k ^ { \beta } }, \quad 1 \leq i \leq d, \, 1 \leq j \leq k,$$
where the vectors x and y are given by Lemma 20 and b satisfies a ( 1 + b ) < 1, and
$$p _ { 0 } = q _ { 0 } = 1 - \frac { 1 } { k ^ { \beta } }.$$
For this choice of p and q, we have
$$P _ { a } \left ( \mathbf p \right ) = \left ( 1 - \frac { 1 } { k ^ { \beta } } \right ) ^ { a } + \frac { 1 } { k ^ { a \left ( 1 + \beta \right ) - 1 } } \left ( \frac { \| \mathbf x \| _ { a } } { \| \mathbf x \| _ { 1 } } \right ) ^ { a },$$
$$H _ { \alpha } ( { \mathbf p } ) = \frac { 1 - \alpha ( 1 + \beta ) } { 1 - \alpha } \log k + \frac { \alpha } { 1 - \alpha } \log \frac { \| { \mathbf x } \| _ { \alpha } } { \| { \mathbf x } \| _ { 1 } } + O ( k ^ { a ( 1 + \beta ) - 1 } ),$$
$$& \left ( 1 - \frac { 1 } { k ^ { \beta } } \right ) ^ { a } + \frac { 1 } { k ^ { a ( 1 + \beta ) - 1 } } \left ( \frac { \| \mathbf x \| _ { a } } { \| \mathbf x \| _ { 1 } } \right ) ^ { a }, \\ & \frac { 1 - \alpha ( 1 + \beta ) } { 1 - \alpha } \log k + \frac { \alpha } { 1 - \alpha } \log \frac { \| \mathbf x \| _ { \alpha } } { \| \mathbf x \| _ { 1 } } + O ( k ^ { a ( 1 + \beta ) - 1 } ),$$
and similarly for q, which further yields
$$| H _ { \alpha } ( { \mathbf p } ) - H _ { \alpha } ( { \mathbf q } ) | = \frac { \alpha } { 1 - \alpha } \left | \log \frac { \| { \mathbf x } \| _ { \alpha } } { \| { \mathbf y } \| _ { \alpha } } \right | + O ( k ^ { a ( 1 + \beta ) - 1 } ).$$
Therefore, for sufficiently large k , (20) holds by Lemma 20 since a ( 1 + b ) < 1, and for n < C k 2 ( 1 + b -1/ d ) we get (21) by Theorem 19 as
$$\| p _ { \Phi } - q _ { \Phi } \| \leq \frac { \epsilon } { 2 } + 5 \sum _ { a \geq d } \left ( \frac { n } { k ^ { 1 + \beta - 1 / a } } \right ) ^ { a } \leq \epsilon.$$
The theorem follows since d and b < 1/ a -1 are arbitrary.
glyph[squaresolid]
## ACKNOWLEDGEMENTS
The authors thank Chinmay Hegde and Piotr Indyk for helpful discussions and suggestions.
## REFERENCES
- [1] J. Acharya, A. Orlitsky, A. T. Suresh, and H. Tyagi, 'The complexity of estimating r´ enyi entropy,' in Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015 , 2015, pp. 1855-1869.
- [2] A. Antos and I. Kontoyiannis, 'Convergence properties of functional estimates for discrete distributions,' Random Struct. Algorithms , vol. 19, no. 3-4, pp. 163-193, Oct. 2001.
- [3] E. Arikan, 'An inequality on guessing and its application to sequential decoding,' IEEE Transactions on Information Theory , vol. 42, no. 1, pp. 99-105, 1996.
- [4] Z. Bar-Yossef, R. Kumar, and D. Sivakumar, 'Sampling algorithms: lower bounds and applications,' in Proceedings on 33rd Annual ACM Symposium on Theory of Computing, July 6-8, 2001, Heraklion, Crete, Greece , 2001, pp. 266-275.
- [5] T. Batu, L. Fortnow, R. Rubinfeld, W. D. Smith, and P. White, 'Testing closeness of discrete distributions,' J. ACM , vol. 60, no. 1, p. 4, 2013.
- [6] C. Bennett, G. Brassard, C. Crepeau, and U. Maurer, 'Generalized privacy amplification,' IEEE Transactions on Information Theory , vol. 41, no. 6, Nov 1995.
- [7] I. Csisz´ ar, 'Generalized cutoff rates and Renyi's information measures,' IEEE Transactions on Information Theory , vol. 41, no. 1, pp. 26-34, Jan. 1995.
- [8] O. Goldreich and D. Ron, 'On testing expansion in bounded-degree graphs,' Electronic Colloquium on Computational Complexity (ECCC) , vol. 7, no. 20, 2000.
- [9] P. Grassberger, 'Finite sample corrections to entropy and dimension estimates,' Physics Letters A , vol. 128, no. 6, pp. 369-373, 1988.
- [10] M. K. Hanawal and R. Sundaresan, 'Guessing revisited: A large deviations approach,' IEEE Transactions on Information Theory , vol. 57, no. 1, pp. 70-78, 2011.
- [11] G. Hardy, J. E. Littlewood, and G. P´ olya, Inequalities. 2nd edition . Cambridge University Press, 1952.
- [12] N. J. A. Harvey, J. Nelson, and K. Onak, 'Sketching and streaming entropy via approximation theory,' in 49th Annual IEEE Symposium on Foundations of Computer Science, FOCS 2008, October 25-28, 2008, Philadelphia, PA, USA , 2008, pp. 489-498.
- [13] V. M. Ilic and M. S. Stankovic, 'A unified characterization of generalized information and certainty measures,' CoRR , vol. abs/1310.4896, 2013. [Online]. Available: http://arxiv.org/abs/1310.4896
- [14] R. Impagliazzo and D. Zuckerman, 'How to recycle random bits,' in FOCS , 1989.
- [15] R. Jenssen, K. Hild, D. Erdogmus, J. Principe, and T. Eltoft, 'Clustering using Renyi's entropy,' in Proceedings of the International Joint Conference on Neural Networks . IEEE, 2003.
- [16] J. Jiao, K. Venkat, Y. Han, and T. Weissman, 'Minimax estimation of functionals of discrete distributions,' IEEE Transactions on Information Theory , vol. 61, no. 5, pp. 2835-2885, May 2015.
- [17] J. Jiao, K. Venkat, and T. Weissman, 'Maximum likelihood estimation of functionals of discrete distributions,' CoRR , vol. abs/1406.6959, 2014.
- [18] D. K¨ allberg, N. Leonenko, and O. Seleznjev, 'Statistical inference for r´ enyi entropy functionals,' CoRR , vol. abs/1103.4977, 2011.
- [19] D. E. Knuth, The Art of Computer Programming, Volume III: Sorting and Searching . Addison-Wesley, 1973.
- [20] A. Lall, V. Sekar, M. Ogihara, J. Xu, and H. Zhang, 'Data streaming algorithms for estimating entropy of network traffic,' SIGMETRICS Perform. Eval. Rev. , vol. 34, no. 1, pp. 145-156, Jun. 2006.
- [21] B. Ma, A. O. H. III, J. D. Gorman, and O. J. J. Michel, 'Image registration with minimum spanning tree algorithm,' in ICIP , 2000, pp. 481-484.
- [22] Y. Mansour, M. Mohri, and A. Rostamizadeh, 'Multiple source adaptation and the Renyi divergence,' CoRR , vol. abs/1205.2628, 2012.
- [23] V. Markov, 'On functions deviating least from zero in a given interval,' Izdat. Imp. Akad. Nauk, St. Petersburg , pp. 218-258, 1892.
- [24] J. Massey, 'Guessing and entropy,' in Information Theory, 1994. Proceedings., 1994 IEEE International Symposium on , Jun 1994, pp. 204-.
- [25] G. A. Miller, 'Note on the bias of information estimates,' Information theory in psychology: Problems and methods , vol. 2, pp. 95-100, 1955.
- [26] A. Mokkadem, 'Estimation of the entropy and information of absolutely continuous random variables,' IEEE Transactions on Information Theory , vol. 35, no. 1, pp. 193-196, 1989.
- [27] A. Motahari, G. Bresler, and D. Tse, 'Information theory of dna shotgun sequencing,' Information Theory, IEEE Transactions on , vol. 59, no. 10, pp. 6273-6289, Oct 2013.
- [28] H. Neemuchwala, A. O. Hero, S. Z., and P. L. Carson, 'Image registration methods in high-dimensional space,' Int. J. Imaging Systems and Technology , vol. 16, no. 5, pp. 130-145, 2006.
- [29] I. Nemenman, W. Bialek, and R. R. de Ruyter van Steveninck, 'Entropy and information in neural spike trains: Progress on the sampling problem,' Physical Review E , vol. 69, pp. 056 111-056 111, 2004.
- [30] P. C. V. Oorschot and M. J. Wiener, 'Parallel collision search with cryptanalytic applications,' Journal of Cryptology , vol. 12, pp. 1-28, 1999.
- [31] A. Orlitsky, N. P. Santhanam, K. Viswanathan, and J. Zhang, 'On modeling profiles instead of values,' 2004.
- [32] L. Paninski, 'Estimation of entropy and mutual information,' Neural Computation , vol. 15, no. 6, pp. 1191-1253, 2003.
- [33] --, 'Estimating entropy on m bins given fewer than m samples,' IEEE Transactions on Information Theory , vol. 50, no. 9, pp. 2200-2203, 2004.
- [34] --, 'A coincidence-based test for uniformity given very sparsely sampled discrete data,' vol. 54, no. 10, pp. 4750-4755, 2008.
- [35] C.-E. Pfister and W. Sullivan, 'Renyi entropy, guesswork moments, and large deviations,' IEEE Transactions on Information Theory , vol. 50, no. 11, pp. 2794-2800, Nov 2004.
- [36] A. R´ enyi, 'On measures of entropy and information,' in Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Contributions to the Theory of Statistics , 1961, pp. 547-561.
- [37] P. S. Shenkin, B. Erman, and L. D. Mastrandrea, 'Information-theoretical entropy as a measure of sequence variability,' Proteins , vol. 11, no. 4, pp. 297-313, 1991.
- [38] E. V. Slud, 'Distribution Inequalities for the Binomial Law,' The Annals of Probability , vol. 5, no. 3, pp. 404-412, 1977.
- [39] A. F. Timan, Theory of Approximation of Functions of a Real Variable . Pergamon Press, 1963.
- [40] T. Tossavainen, 'On the zeros of finite sums of exponential functions,' Australian Mathematical Society Gazette , vol. 33, no. 1, pp. 47-50, 2006.
- [41] G. Valiant and P. Valiant, 'Estimating the unseen: an n/log(n)-sample estimator for entropy and support size, shown optimal via new clts,' 2011.
- [42] P. Valiant, 'Testing symmetric properties of distributions,' in STOC , 2008.
- [43] Y. Wu and P. Yang, 'Minimax rates of entropy estimation on large alphabets via best polynomial approximation,' CoRR , vol. abs/1407.0381v1, 2014.
- [44] D. Xu, 'Energy, entropy and information potential for neural computation,' Ph.D. dissertation, University of Florida, 1998.
- [45] D. Xu and D. Erdogmuns, 'Renyi's entropy, divergence and their nonparametric estimators,' in Information Theoretic Learning , ser. Information Science and Statistics. Springer New York, 2010, pp. 47-102.
- [46] B. Yu, 'Assouad, Fano, and Le Cam,' in Festschrift for Lucien Le Cam . Springer New York, 1997, pp. 423-435.
## APPENDIX A: ESTIMATING POWER SUMS
The broader problem of estimating smooth functionals of distributions was considered in [41]. Independently and concurrently with this work, [16] considered estimating more general functionals and applied their technique to estimating the power sums of a distribution to a given additive accuracy. Letting S P + ( a k ) denote the number of samples needed to estimate P a ( p ) to a given additive accuracy, [16] showed that for a < 1,
$$S _ { \alpha } ^ { P + } ( k ) = \Theta \left ( \frac { k ^ { 1 / \alpha } } { \log k } \right ),$$
and [17] showed that for 1 < a < 2,
$$S _ { \alpha } ^ { P + } ( k ) \leq O \left ( k ^ { 2 / \alpha - 1 } \right ).$$
In fact, using techniques similar to multiplicative guarantees on P a ( p ) we show that for S P + ( a k ) is a constant independent of k for all k > 1.
Since P a ( p ) > 1 for a < 1, power sum estimation to a fixed additive accuracy implies also a fixed multiplicative accuracy, and therefore
$$S _ { \alpha } ( k ) = \Theta ( S _ { \alpha } ^ { P \times } ( k ) ) \leq O ( S _ { \alpha } ^ { P + } ( k ) ),$$
namely for estimation to an additive accuracy, R´ enyi entropy requires fewer samples than power sums. Similarly, P a ( p ) < 1 for a > 1, and therefore
$$S _ { \alpha } ( k ) = \Theta ( S _ { \alpha } ^ { P \times } ( k ) ) \geq \Omega ( S _ { \alpha } ^ { P + } ( k ) ),$$
namely for an additive accuracy in this range, R´ enyi entropy requires more samples than power sums.
It follows that the power sum estimation results in [16], [17] and the R´ enyi-entropy estimation results in this paper complement each other in several ways. For example, for a < 1,
$$\widetilde { \Omega } \left ( k ^ { 1 / \alpha } \right ) \leq S _ { \alpha } ( k ) = \Theta ( S _ { \alpha } ^ { P \times } ( k ) ) \leq O ( S _ { \alpha } ^ { P + } ( k ) ) \leq O \left ( \frac { k ^ { 1 / \alpha } } { \log k } \right ),$$
where the first inequality follows from Theorem 22 and the last follows from the upper-bound (22) derived in [16] using a polynomial approximation estimator . Hence, for a < 1, estimating power sums to additive and multiplicative accuracy require a comparable number of samples.
On the other hand, for a > 1, Theorems 9 and 21 imply that for non integer a , ∼ ∼ W ( k ) ≤ S P × a ( k ) ≤ O k ( ) , while in the Appendix we show that for 1 < a , S P + ( a k ) is a constant. Hence in this range, power sum estimation to a multiplicative accuracy requires considerably more samples than estimation to an additive accuracy.
We now show that the empirical estimator requires a constant number of samples to estimate P a ( p ) independent of k , i . e ., S P + ( a k ) = O ( 1 . ) In view of Lemma 8, it suffices to bound the bias and variance of the empirical estimator. Concurrently with this work, similar results were obtained in an updated version of [16].
As before, we comsider Poisson sampling with N ∼ Poi ( n ) samples. The empirical or plug-in estimator of P a ( p ) is
$$\widehat { P } _ { \alpha } ^ { \text{e} } \stackrel { \text{def} } { = } \sum _ { x } \left ( \frac { N _ { x } } { n } \right ) ^ { \alpha }.$$
The next result shows that the bias and the variance of the empirical estimator are o ( 1 . )
Lemma 23. For an appropriately chosen constant c > 0 , the bias and the variance of the empirical estimator are bounded above as
$$\left | \widehat { P } _ { \alpha } ^ { e } - P _ { \alpha } ( p ) \right | & \leq 2 c \max \{ n ^ { - ( \alpha - 1 ) }, n ^ { - 1 / 2 } \}, \\ \text{Var} [ \widehat { P } _ { \alpha } ] & \leq 2 c \max \{ n ^ { - ( 2 \alpha - 1 ) }, n ^ { - 1 / 2 } \},$$
for all n ≥ 1 .
Proof. Denoting l x = n p , we get the following bound on the bias for an appropriately chosen x
constant c :
$$\left | \widehat { P } _ { \alpha } ^ { \text{e} } - P _ { \alpha } ( \text{p} ) \right | & \leq \frac { 1 } { n ^ { \alpha } } \sum _ { \lambda _ { x } \leq 1 } | \mathbb { E } [ N _ { x } ^ { \alpha } ] - \lambda _ { x } | + \frac { 1 } { n ^ { \alpha } } \sum _ { \lambda _ { x } > 1 } | \mathbb { E } [ N _ { x } ^ { \alpha } ] - \lambda _ { x } | \\ & \leq \frac { c } { n ^ { \alpha } } \sum _ { \lambda _ { x } \leq 1 } \lambda _ { x } + \frac { c } { n ^ { \alpha } } \sum _ { \lambda _ { x } > 1 } \left ( \lambda _ { x } + \lambda _ { x } ^ { \alpha - 1 / 2 } \right ), \\. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad.$$
where the last inequality holds by Lemma 4 and Lemma 2since x a is convex in x . Noting GLYPH<229> i l x = n , we get
$$\left | \widehat { P } _ { \alpha } ^ { \text{e} } - P _ { \alpha } ( \mathbf p ) \right | \leq \frac { c } { n ^ { \alpha - 1 } } + \frac { c } { n ^ { \alpha } } \sum _ { \lambda _ { x } > 1 } \lambda _ { x } ^ { \alpha - 1 / 2 }.$$
Similarly, proceeding as in the proof of Theorem 9, the variance of the empirical estimator is bounded as
$$\text{Var} [ \widehat { P } _ { \alpha } ] & = \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x \in \mathcal { X } } \mathbb { E } \left [ N _ { x } ^ { 2 \alpha } \right ] - \mathbb { E } [ N _ { x } ^ { \alpha } ] ^ { 2 } \\ & \leq \frac { 1 } { n ^ { 2 \alpha } } \sum _ { x \in \mathcal { X } } \left | \mathbb { E } \left [ N _ { x } ^ { 2 \alpha } \right ] - \lambda _ { x } ^ { 2 \alpha } \right | \\ & \leq \frac { c } { n ^ { 2 \alpha - 1 } } + \frac { c } { n ^ { 2 \alpha } } \sum _ { \lambda _ { x } > 1 } \lambda _ { x } ^ { 2 \alpha - 1 / 2 }. \\ \text{non show\in that}$$
The proof is completed upon showing that
$$\sum _ { \lambda _ { x } > 1 } \lambda _ { x } ^ { \alpha - 1 / 2 } \leq \max \{ n, n ^ { \alpha - 1 / 2 } \}, \quad \alpha > 1.$$
To that end, note that for a < 3/2
$$\sum _ { \lambda _ { x } > 1 } \lambda _ { x } ^ { \alpha - 1 / 2 } \leq \sum _ { \lambda _ { x } > 1 } \lambda _ { x } \leq n, \quad \alpha < 3 / 2.$$
Further, since x a -1/2 is convex for a ≥ 3/2, the summation above is maximized when one of the l x 's is n and the remaining equal 0 which yields
$$\sum _ { \lambda _ { x } > 1 } \lambda _ { x } ^ { \alpha - 1 / 2 } \leq n ^ { \alpha - 1 / 2 }, \ \alpha \geq 3 / 2,$$
and completes the proof.
glyph[squaresolid]
APPENDIX B: LOWER BOUND FOR SAMPLE COMPLEXITY OF EMPIRICAL ESTIMATOR
We now derive lower bounds for the sample complexity of the empirical estimator of H a ( p . )
Lemma 24. Given a < 1 and d < c a for a constant c a depending only on a , the sample complexity S f e a a ( k , d e , ) of the empirical estimator f e a is bounded below as
$$S _ { \alpha } ^ { f _ { \alpha } ^ { e } } ( k, \delta, 0. 9 ) = \Omega \left ( \frac { k } { \delta } \right ).$$
Proof. We prove the lower bound for the uniform distributon over k symbols in two steps. We first show that for any constant c 1 > 1 if n < k / 1 c then the additive approximation error is at least d with probability one, for every d < log c 1 . Then, assuming that n ≥ k / 1, we show that c the additve approximation error is at least d with probability greater than 0.9 if n < k / . d
For the first claim, we assume without loss of generality that n ≤ k , since otherwise the proof is complete. Note that for a > 1 the function ( p i -y ) a +( p j + ) y a is decreasing in y for all y such that ( p i -y ) > ( p j + ) y . Thus, the minimum value of GLYPH<229> x ( Nx n ) a is attained when each Nx is either 0 or 1. It follows that
$$\widehat { P } _ { \alpha } ^ { \text{e} } = \sum _ { x } \left ( \frac { N _ { x } } { n } \right ) ^ { \alpha } \geq \frac { 1 } { n ^ { \alpha - 1 } },$$
which is the same as
$$H _ { \alpha } ( { \mathbf p } ) - \frac { 1 } { \alpha - 1 } \log \frac { 1 } { \widehat { P } _ { \alpha } ^ { e } } \geq \log \frac { k } { n }.$$
Hence, for any c 1 > 1 and n < k / 1 c and any 0 ≤ d ≤ log c 1 , the additive approximation error is more than d with probability one.
Moving to the second claim, suppose now n > k / 1. We first show that with high probability, c the multiplicities of a linear fraction of k symbols should be at least a factor of standard deviaton higher than the mean. Specifically, let
$$A = \sum _ { x } 1 \left ( N _ { x } \geq \frac { n } { k } + c _ { 2 } \sqrt { \frac { n } { k } \left ( 1 - \frac { 1 } { k } \right ) } \right ).$$
Then,
$$\mathbb { E } [ A ] & = \sum _ { x } \mathbb { E } \left [ 1 \left ( N _ { x } \geq \frac { n } { \bar { k } } + c _ { 2 } \sqrt { \frac { n } { \bar { k } } \left ( 1 - \frac { 1 } { \bar { k } } \right ) } \right ) \right ] \\ & = k \cdot p \left ( N _ { x } \geq \frac { n } { \bar { k } } + c _ { 2 } \sqrt { \frac { n } { \bar { k } } \left ( 1 - \frac { 1 } { \bar { k } } \right ) } \right ) \\ & \geq k Q ( c _ { 2 } ),$$
where Q denotes the Q -function, i . e ., the tail of the standard normal random variable, and the final inequality uses Slud's inequality [38, Theorem 2.1].
Note that A is a function of n i.i.d. random variables X 1 , X 2, . . . , Xn , and changing any one Xi changes A by at most 2. Hence, by McDiarmid's inequality,
$$\Pr ( A \geq \mathbb { E } [ A ] - \sqrt { 8 n } ) \geq 1 - e ^ { - 4 } \geq 0. 9.$$
Therefore, for all k sufficiently large (depending on d ) and denoting c = Q c ( 2 ) /2, at least ck symbols occur more than n k + c 2 √ n k times with probability greater than 0.9. Using the fact that ( p i -y ) a +( p j + ) y a is decreasing if ( p i -y ) > ( p j + ) y once more, we get
$$\Pi \i o w s { \ e c u a n o r e u a n } \bar { k } ^ { + } + \varepsilon _ { 2 } \sqrt { k } \ u c o w s { \ e c u a n o r e u a n } \varsigma _ { i } - y ) ^ { \alpha } + ( p _ { j } + y ) ^ { \alpha } & \text{is decreasing if } ( p _ { i } - y ) > ( p _ { j } + y ) \text{ once more, we get} \\ \sum _ { x \in \lambda } \frac { N _ { x } ^ { \alpha } } { n ^ { \alpha } } & = \sum _ { x \in \lambda _ { x } ^ { \alpha } } \frac { N _ { x } ^ { \alpha } } { n ^ { \alpha } } + \sum _ { x \in \lambda _ { x } ^ { \alpha } } \frac { N _ { x } ^ { \alpha } } { n ^ { \alpha } } \\ & \geq c k \left ( \frac { 1 } { k } + c _ { 2 } \sqrt { \frac { 1 } { n k } } \right ) ^ { \alpha } + ( 1 - c ) k \left ( \frac { 1 } { k } - \frac { c c _ { 2 } } { 1 - c } \sqrt { \frac { 1 } { n k } } \right ) ^ { \alpha } \\ & = \frac { 1 } { k ^ { \alpha - 1 } } \left [ c \left ( 1 + c _ { 2 } \sqrt { \frac { k } { n } } \right ) ^ { \alpha } + ( 1 - c ) \left ( 1 - \frac { c c _ { 2 } } { 1 - c } \sqrt { \frac { k } { n } } \right ) ^ { \alpha } \right ] \\ & \geq \frac { 1 } { k ^ { \alpha - 1 } } \left [ c \left ( 1 + c _ { 2 } \sqrt { \frac { k } { n } } \right ) ^ { \alpha } + ( 1 - c ) \left ( 1 - \frac { \alpha c c _ { 2 } } { 1 - c } \sqrt { \frac { k } { n } } \right ) \right ] \\ & \geq \frac { 1 } { k ^ { \alpha - 1 } } \left [ c \left ( 1 + \alpha c _ { 2 } \sqrt { \frac { k } { n } } + c _ { 4 } \frac { k } { n } \right ) + ( 1 - c ) \left ( 1 - \frac { \alpha c c _ { 2 } } { 1 - c } \sqrt { \frac { k } { n } } \right ) \right ] \\ & = \frac { 1 } { k ^ { \alpha - 1 } } \left ( 1 + c c _ { 4 } \frac { k } { n } \right ) \\ \text{ here the second inequality is by Bernoulli} ^ { \prime } s \text{ inequality and the third inequality holds for even} \\ < \alpha ( \alpha - 1 ) ( c _ { x }. \sqrt { c _ { x } } ^ { 2 } / \text{ Therefore with non-habilitiv} > 0 9$$
where the second inequality is by Bernoulli's inequality and the third inequality holds for every c 4 ≤ a a ( -1 )( c 2 √ c 1 ) a -2 /2. Therefore, with probability ≥ 0.9,
$$H _ { \alpha } ( p ) - \frac { 1 } { \alpha - 1 } \log \frac { 1 } { \widehat { P } _ { \alpha } ^ { \text{e} } } \geq \frac { 1 } { \alpha - 1 } \log \left ( 1 + c c _ { 4 } \frac { k } { n } \right ),$$
which yields the desired bound.
glyph[squaresolid]
Lemma 25. Given a < 1 and d < c a for a constant c a depending only on a , the sample complexity S f e a a ( k , d e , ) of the empirical estimator f e a is bounded as
$$S _ { \alpha } ^ { f _ { \alpha } ^ { e } } ( k, \delta, 0. 9 ) = \Omega \left ( \frac { k ^ { 1 / \alpha } } { \delta ^ { 1 / \alpha } } \right ).$$
Proof. We proceed as in the proof of the previous lemma. However, instead of using the uniform distribution, we use a distribution which has one 'heavy element' and is uniform conditioned on the occurance of the remainder. The key observation is that there will be roughly n a occurances of the 'light elements'. Thus, when we account for the error in the estimation of the
contribution of light elements to the power sum, we can replace n with n 1/ a in our analysis of the previous lemma, which yields the required bound for sample complexity.
Specifically, consider a distribution with one heavy element 0 such that
$$\mathbf p _ { 0 } = 1 - \frac { \delta } { n ^ { 1 - \alpha } }, \text{ and } \mathbf p _ { i } = \frac { \delta } { k n ^ { 1 - \alpha } }, \text{ } \ 1 \leq i \leq k.$$
Thus,
$$P _ { \alpha } ( \mathbf p ) = \left ( 1 - \frac { \delta } { n ^ { 1 - \alpha } } \right ) ^ { \alpha } + \delta ^ { \alpha } \left ( \frac { k } { n ^ { \alpha } } \right ) ^ { 1 - \alpha }.$$
We begin by analyzing the estimate of the second term in power sum, namely
$$\sum _ { i \in [ k ] } \left ( \frac { N _ { i } } { n } \right ) ^ { \alpha }.$$
Let R = GLYPH<229> i ∈ [ k ] Ni be the total number of occurances of light elements. Since R is a binomial ( n , d n a -1 ) random variable, for every constant c > 0
$$\mathbb { P } \left ( 1 - c < \frac { R } { \delta n ^ { \alpha } } < 1 + c \right ) \geq 1 - \frac { 1 } { c ^ { 2 } n }.$$
In the remainder of the proof, we shall assume that this large probability event holds.
As in the proof of the previous lemma, we first prove a d independent lower bound for sample complexity. To that end, we fix d = 1 in the definition of p. Assuming ( 1 + ) c n a ≤ k , which implies R ≤ k , and using the fact that ( p i -y ) a -( p j + ) y a is increasing in y if ( p i -y ) > ( p j + ) y , we get
$$\widehat { P } _ { \alpha } ^ { \text{e} } & \leq 1 + \left ( \frac { R } { n } \right ) ^ { \alpha } \sum _ { i \in [ k ] } \left ( \frac { N _ { i } } { R } \right ) ^ { \alpha } \\ & \leq 1 + \frac { ( 1 + c ) ^ { \alpha } } { n ^ { \alpha } ( 1 - \alpha ) } \sum _ { i \in [ k ] } \left ( \frac { N _ { i } } { R } \right ) ^ { \alpha } \\ & \leq 1 + \frac { ( 1 + c ) ^ { \alpha } } { n ^ { \alpha } ( 1 - \alpha ) } R ^ { 1 - \alpha } \\ & \leq 3, \\ \psi \quad \, \psi \quad \, \cdots \quad \, \psi \quad \, \psi \quad \, \quad \,.$$
where the last inequality uses R ≤ ( 1 + ) c n a ≤ 2 n a . Thus, the empirical estimate is at most 3 with probability close to 1 when k (and therefore n ) large. It follows from (23) that
$$H _ { \alpha } ( p ) - \frac { 1 } { 1 - \alpha } \log \widehat { P } _ { \alpha } ^ { e } \geq \log \frac { k } { 3 n ^ { \alpha } }.$$
Therefore, for all c 1 > 1, d < log 3 c 1 and k sufficiently large, at least ( k / 1 c ) 1/ a samples are needed to get a d -additive approximation of H a ( p ) with probability of error less than 1 -1/ ( c 2 n ) . Note that we only needed to assume R ≤ ( 10/9 ) n a , an event with probability greater than 0.9, to get the contradiction above. Thus, we may assume that n ≥ ( k / 1 c ) 1/ a . Under this assumption, for k sufficiently large, n is sufficiently large so that ( 1 -c ) n a ≤ R ≤ ( 1 + ) c n a holds with probability arbitrarily close to 1.
Next, assuming that n ≥ ( k / 1 c ) 1/ a , we obtain a d -dependent lower bound for sample complexity of the empirical estimator. We use the p mentioned above with a general d and assume that the large probability event
$$( 1 - c ) \leq \frac { R } { \delta n ^ { \alpha } } \leq ( 1 + c )$$
holds. Note that conditioned on each value of R , the random variables ( Ni , i ∈ [ k ]) have a multinomial distribution with uniform probabilities, i . e ., these random variables behave as if we drew R i.i.d. samples from a uniform distribution on [ k ] elements. Thus, we can follow the proof of the previous lemma mutatis mutandis . We now define A as
$$A = \sum _ { x } 1 \left ( N _ { x } \leq \frac { n } { k } - c _ { 2 } \sqrt { \frac { n } { k } \left ( 1 - \frac { 1 } { k } \right ) } \right ).$$
and satisfies Then,
$$\mathbb { E } [ A ] & = \sum _ { x } \mathbb { E } \left [ 1 \left ( N _ { x } \leq \frac { n } { k } - c _ { 2 } \sqrt { \frac { n } { k } \left ( 1 - \frac { 1 } { k } \right ) } \right ) \right ] \\ & = k \cdot p \left ( N _ { x } \leq \frac { n } { k } - c _ { 2 } \sqrt { \frac { n } { k } \left ( 1 - \frac { 1 } { k } \right ) } \right ).$$
To lower bound p ( Nx ≤ n k -c 2 √ ( n k 1 -1 k ) ) Slud's inequality is no longer available (since it may not hold for Bin ( n p , ) with p > 1/2 and that is the regime of interest for the lower tail probability bounds needed here). Instead we take recourse to a combination of Bohman's inequality and Anderson-Samuel inequality, as suggested in [38, Eqns. (i) and (ii)]. It can be verified that the condition for [38, Eqns. (ii)] holds, and therefore,
$$p \left ( N _ { x } \leq \frac { n } { k } - c _ { 2 } \sqrt { \frac { n } { k } \left ( 1 - \frac { 1 } { k } \right ) } \right ) \geq Q ( c _ { 2 } ).$$
Continuing as in the proof of the previous lemma, we get that the following holds with
conditional probability greater than 0.9 given each value of R satisfying (24):
$$\sum _ { i \in [ k ] } \Big { ( } \frac { N _ { x } } { R } \Big { ) } ^ { \alpha } & \leq k ^ { 1 - \alpha } \left ( 1 - c _ { 3 } \frac { k } { R } \right ) \\ & \leq k ^ { 1 - \alpha } \left ( 1 - c _ { 4 } \frac { k } { \delta n ^ { \alpha } } \right ),$$
where c 3 is a sufficiently small constant such that ( 1 + ) x a ≤ 1 + a x -c 3 x 2 for all x ≥ 0 and c 4 = c 3/ 1 ( + ) c . Thus,
$$\i u s, \\ \widehat { P } _ { \alpha } ^ { \text{e} } & \leq 1 + \left ( \frac { R } { n } \right ) ^ { \alpha } \sum _ { i \in [ k ] } \left ( \frac { N _ { i } } { R } \right ) ^ { \alpha } \\ & \leq 1 + \left ( \frac { R } { n } \right ) ^ { \alpha } k ^ { 1 - \alpha } \left ( 1 - c _ { 4 } \frac { k } { \delta n ^ { \alpha } } \right ) \\ & \leq 1 + ( 1 + c ) ^ { \alpha } \, \delta ^ { \alpha } \left ( \frac { k } { n ^ { \alpha } } \right ) ^ { 1 - \alpha } \left ( 1 - c _ { 4 } \frac { k } { \delta n ^ { \alpha } } \right ). \\ \i u ^ { \alpha } \cup \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cup \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cap \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cap \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cap \cap \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cap \cup \cap \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \cup \vend$$
Denoting y = ( k / n a ) and choosing c 1 and c small enough such that ̂ P e a ≤ 2, for all sufficiently large n we get from (23) that
$$\Pi \cap \varpi \ u a \, \\ \frac { P _ { \alpha } ( p ) } { \widehat { P } _ { \alpha } ^ { e } } & \geq \frac { 1 - \delta + y ^ { 1 - \alpha } } { 1 + \delta ^ { \alpha } ( 1 + c ) ^ { \alpha } y ^ { 1 - \alpha } - ( 1 + c ) ^ { \alpha } c _ { 4 } \delta ^ { \alpha - 1 } y ^ { 2 - \alpha } } \\ & \geq \frac { 1 - \delta + y ^ { 1 - \alpha } } { 1 + y ^ { 1 - \alpha } - \delta ^ { \alpha - 1 } y ^ { 2 - \alpha } } \\ & \geq 1 - \frac { \delta } { 2 } + \frac { \delta ^ { \alpha - 1 } y ^ { 2 - \alpha } } { 2 }, \\ \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \colon$$
where the second inequality uses the fact that d a ( 1 + ) c a y 1 -a -( 1 + ) c a c 4 d a -1 y 2 -a is negative, c 4 > 1 and d < 1. Therefore, P a ( p ) ̂ P e a ≥ 1 + d if y 2 -a ≥ 3 d 2 -a , which completes the proof. glyph[squaresolid]
| null |
[
"Jayadev Acharya",
"Alon Orlitsky",
"Ananda Theertha Suresh",
"Himanshu Tyagi"
] |
2014-08-02T18:52:52+00:00
|
2016-03-10T08:35:51+00:00
|
[
"cs.IT",
"cs.DS",
"cs.LG",
"math.IT"
] |
Estimating Renyi Entropy of Discrete Distributions
|
It was recently shown that estimating the Shannon entropy $H({\rm p})$ of a
discrete $k$-symbol distribution ${\rm p}$ requires $\Theta(k/\log k)$ samples,
a number that grows near-linearly in the support size. In many applications
$H({\rm p})$ can be replaced by the more general R\'enyi entropy of order
$\alpha$, $H_\alpha({\rm p})$. We determine the number of samples needed to
estimate $H_\alpha({\rm p})$ for all $\alpha$, showing that $\alpha < 1$
requires a super-linear, roughly $k^{1/\alpha}$ samples, noninteger $\alpha>1$
requires a near-linear $k$ samples, but, perhaps surprisingly, integer
$\alpha>1$ requires only $\Theta(k^{1-1/\alpha})$ samples. Furthermore,
developing on a recently established connection between polynomial
approximation and estimation of additive functions of the form $\sum_{x} f({\rm
p}_x)$, we reduce the sample complexity for noninteger values of $\alpha$ by a
factor of $\log k$ compared to the empirical estimator. The estimators
achieving these bounds are simple and run in time linear in the number of
samples. Our lower bounds provide explicit constructions of distributions with
different R\'enyi entropies that are hard to distinguish.
|
1408.1001v3
|
## Predictability of threshold exceedances in dynamical systems
Tam´ as B´dai o ∗
Meteorological Institute, University of Hamburg, Grindelberg 5, 20144 Hamburg, Germany
## Abstract
In a low-order model of the general circulation of the atmosphere we examine the predictability of threshold exceedance events of certain observables. The likelihood of such binary events - the cornerstone also for the categoric (as opposed to probabilistic) prediction of threshold exceedences - is established from long time series of one or more observables of the same system. The prediction skill is measured by a summary index of the ROC curve that relates the hit- and false alarm rates. Our results for the examined systems suggest that exceedances of higher thresholds are more predictable ; or in other words: rare large magnitude, i.e., extreme, events are more predictable than frequent typical events. We find this to hold provided that the bin size for binning time series data is optimized, but not necessarily otherwise. This can be viewed as a confirmation of a counterintuitive (and seemingly contrafactual) statement that was previously formulated for more simple autoregressive stochastic processes. However, we argue that for dynamical systems in general it may be typical only, but not universally true. We argue that when there is a sufficient amount of data depending on the precision of observation, the skill of a class of data-driven categoric predictions of threshold exceedences approximates the skill of the analogous model-driven prediction, assuming strictly no model errors. Therefore, stronger extremes in terms of higher threshold levels are more predictable both in case of data- and model-driven prediction. Furthermore, we show that a quantity commonly regarded as a measure of predictability, the finite-time maximal Lyapunov exponent, does not correspond directly to the ROC-based measure of prediction skill when they are viewed as functions of the prediction lead time and the threshold level. This points to the fact that even if the Lyapunov exponent as an intrinsic property of the system, measuring the instability of trajectories, determines predictability, it does that in a nontrivial manner.
Keywords: Extreme event, Data-driven prediction, Precursory structure, Prediction skill, ROC curve, Finite-time Lyapunov exponent
## 1. Introduction
Extreme events have fundamental importance to life, as they are often associated with survival and losses. - Extreme events to do
∗ [email protected], +49 40 42838 9205
with gains or amusement receive far less attention in general, dissociated from individual events. Rare and large magnitude events of interest arise in physical, technological, social, and other systems [1]. The classical theory of extremes in uncorrelated sequences (or, in sequences in which the auto-correlation is
decaying sufficiently fast) [2, 3, 4] has a statistical orientation; it is notand cannot be concerned with prediction or with uncovering mechanisms that can produce extremes; but it is rather concerned with e.g. expected return times, which can be useful in designing structures of a certain required life time, such as sea walls [5].
Since Newton revolutionized science, it has become a paradigm that predictions should be based on validated models. These models describing fluctuating phenomena often take the form of a system of differential equations, also referred to as a dynamical system . Since the work of Lorenz it has become clear that even though some phenomena can be modeled quite accurately, they can be inherently unpredictable because of their extreme sensitivity to initial conditions [6]. Such systems are called chaotic, characterized by positive Lyapunov exponents. This imposes a time horizon on predictions; beyond that only statistical properties can be robustly estimated, which is what classical extreme value theory is concerned with. In contrast with that, in our analysis we consider prediction lead times shorter than the decorrelation time in a time series.
In the context of model-based or modeldriven predictions (MDP) equivalent with initial value problems for deterministic differential equations, like e.g. a weather forecast, one can often read that extremes are much harder to predict. Unfortunately a systematic study of the dependence of some appropriate prediction skill score - or a measure of predictability in a more general sense - of any model on the magnitude of events is still lacking. Inaccuracy of the model may be an important factor leading to such a dependence of its prediction skill on the event magnitude, beside details of its chaotic nature. In contrast, in pure datadriven prediction (DDP), model errors are not present, as the basis of the prediction (of any kind) does not involve a model in the form of equations or an algorithm, only observational/measurement data. Instead, beyond errors in measuring the present conditions (as with initial conditions for MDP), the prediction is compromised by the finite size of the data set. That is, the said virtue of DDP can be exploited - when employing it in its pure form - only if enough and good quality data (with a high precision of observation and high signal-to-noise ratio) is available [7].
One might expect that the slogan that 'extremes in comparison with more moderate events are harder to predict' extends to DDP. In fact, just the opposite has been reported by Hallerberg and Kantz [8] for simple autoregressive processes at least, indifferently to whether the probability distribution is exponentially decaying or according to a powerlaw, and also for some observational data [9]: stronger events are easier to predict . This counter-intuitive statement is based on a measure of prediction skill that derives from the so-called receiver operating characteristic (ROC) curve [10, 11] that takes into account the true positives - meaning that an event is correctly predicted to happen - as well as the false negatives. Concerning rare events, such a measure of prediction skill is regarded [12] more meaningful than other proper [13] socalled skill scores for probabilistic predictions like the Brier or Ignorance scores. This is so, because the ROC statistics has been viewed to not depend on the relative frequency of events (only that the accurate evaluation of the statistics requires a sufficient number of events). The latter characteristic is thought to allow for the comparison of the ROC-predictability of events between two situations where the events have different frequency [12].
Whether the above statement [8] can be maintained in case of more complex processes has been an open question so far - addressed
but not settled with a consensus. Recently two studies [14, 15] have been published concerning the predictability of extreme events in dynamical systems with seemingly contradictory results as to whether stronger events are more predictable. Franzke [14] applied the method set out in [10] to predict extreme threshold exceedances in a systematically derived stochastic dynamical system representing climate variability by the resolved (slow) variable(s) and weather variability by noise in place of the unresolved (fast) variable(s) [16]. He maintained the earlier statement [8] in this case, measuring the prediction skill by the ROC statistics, but on the basis of considering only two high threshold values. On the other hand, Sterk et al. [15] considered a number of dynamical systems of various complexity, and various physical observables. They evaluated finite-time maximal Lyapunov exponents (FTMLE) of trajectories that lead to extremes, and concluded that no generally applicable statements can be made, but the predictability of extremes depends on the system (and so the attractor geometry) and on the observable in question, as well as the prediction lead time. We emphasize that in their study the authors did not take model errors into account.
To summarize the essence of the above review, we can list three different views encountered in the literature regarding the predictability of extremes:
- (1) Stronger extremes are better predictable.
- (2) Stronger extremes are less predictable.
- (3) Stronger extremes can be better or less predictable depending on various factors.
Without giving details, e.g. assumptions of these statements, they seem to be contradictory to each other. On this basis we set out the following objectives for the present paper:
- (i) Keeping to the assumption of (1), we evaluate the predictability of peak-overthreshold events measured by a ROCbased quantity, using time series data of finite length produced by the Lorenz84 model [17]. With an attention to (3), we evaluate (i.a) the dependence of predictability itself on various factors, and also what is more relevant to the question: (i.b) the magnitudedependence of predictability - whether increasing/decreasing or nonmonotonic depending on some of those same factors.
- (ii) We argue for an analogy between a certain class of DDPs and MDP, and that the latter is usually understood as something that below we will refer to as an on-demand MDP, in which case any input data belongs to a single time instant. We believe this is an assumption of (2). This objective (ii) is to reconcile (1) and (2), suggesting that (2) can be true when model errors are present, even if the predictability is measured by the same ROCbased quantity as that assumed by (1).
- (iii) For on-demand MDP and the analogous DDP where the time of input is arbitrary, we will be able to carry out an assessment of the lead time-dependence of the predictability of what we will call threshold-exceedance-in-an-interval events in a straightforward manner. This will turn out to have a bearing on the magnitude-dependence of predictability. This objective (iii) together with (i.b) are to revisit point (1), possibly extending that point from stochastic processes to dynamical systems.
- (iv) However, to show that (3) does not necessarily contradict (1), we recall that (3)
was stated on the basis of measuring predictability by FTMLEs. This also assumes an arbitrary input time on-demand MDP or analogous DDP. Accordingly, in the autonomous L84 we calculate the FTMLEs of trajectories that lead to extremes, and compare their average to the ROC-based measure of prediction skill looking for any qualitative mismatch.
To motivate our top objective (i) and (iii) we remark that DDP is gaining increasing prominence nowadays given that data is relatively much more easily accessible than models. This is certainly the case with geophysical phenomena that we are primarily interested in, such as meteorology. Furthermore, performing predictions based on data can be far less costly than those based on simulating complex models, while the skill may not be much worse [18].
Next we recapitulate the methodology of the applied prediction scheme and the used ROCbased measure of prediction skill. Lorenz's 1984 model of global atmospheric circulation, simulated to produce time series data for the purpose of assessing the predictability of threshold exeedance events, is also briefly described. Subsequently, in Sec. 3, we present our results on the dependence of predictability on several factors, such as: the makeup of the so-called precursory structure - made use for a prediction - in terms of the observables involved, the prediction lead time, and the magnitude of extreme events. These results pertaining to objectives (i) and (iii) are summarized in a compact table format in Sec. 4 and discussed subsequently regarding objectives (ii) and (iv). To close our presentation we pose a few open questions for future research into practical aspects of the prediction of extremes, which might potentially have theoretical ramifications. We provide in Appendices A
and B, respectively, the description of an algorithm for finding the maximum of a function of one variable and a definition of the finite-time Lyapunov exponents.
## 2. Methodology
## 2.1. Prediction of threshold exceedances by precursors
Our aim is to predict large excursions of some (scalar) physical observable x , exceeding a chosen threshold level x ∗ , before that exceedance happens. Figure 1 pictures the situation as the observable evolves continuously in time, x = x t ( ) , t ∈ R , occasionally exceeding the threshold. We intend to examine situations when x t ( ) is generated by a process that can be described by an ordinary or stochastic differential equation, examples of which for our case study will be briefly described in Sec. 2.2. When one does not have a validated model, only the time series x t ( ) , one can still perform DDP; when the opposite is the case, one can do MDP.
## 2.1.1. Formal setting
The following methodological description regarding the prediction task closely follows [10]. It applies to DDP as well as MDP, on which we comment in the end of this section. We introduce a discrete-time binary event variable:
$$\chi _ { n } = \begin{cases} \ 1, & x ( t _ { n } ) > x _ { * } \\ 0, & x ( t _ { n } ) < x _ { * } \end{cases} \quad \ ( 1 )$$
where the times t n , n ∈ Z , belong to consecutive apexes, that is, local maxima or peaks, of the continuous evolution of x for consecutive values of n , and in general they are not equally 'spaced'. Thereby the continuous-time evolution is discretized . (This can be viewed as a restriction on generality, and it will be relaxed further below.) This approach is often
Figure 1: Illustration of the prediction problem: given a continuous-time chaotic time series x t ( ) featuring apexes that every so often overshoot a relatively high threshold x ∗ (straight horizontal line), we want to predict these overshoots, i.e., extreme events, at the time, say, of the immediately preceding apex. The time series shown was in fact generated by simulating the Lorenz 84 model (9) with F = 8 , and we plotted the first component of the solution x . The time between two subsequent apexes above the threshold is called the return time t r of extremes.
referred to as the peak-over-threshold (POT) approach. The prediction is based on a likewise discrete-time precursory structure x n ∈ R M of size M , whose different members, observables desirably related to x , may belong to different times, e.g. t n -d m , preceding the current time t n , specified by delays d m ∈ Z , m = 1 , . . . , M . We call t n -t n -min( d m ) > 0 the prediction lead time. Our binary prediction for χ n at t n -min( d m ) is defined as:
$$\hat { \chi } _ { n } = \begin{cases} \ 1, & \mathcal { L } ( \mathbf x _ { n } ) > \mathcal { L } _ { * } \\ 0, & \mathcal { L } ( \mathbf x _ { n } ) < \mathcal { L } _ { * } \end{cases} \quad ( 2 ) \quad \text{$\mathfrak $0$}$$
for an integral formulation of e.g. P ( x ) which applies the Heaviside step function as a filter. Note that Eq. (3) expresses Bayes' theorem relating the conditional probabilities: the likelihood and the posterior probability. Our prediction ˆ χ n is controlled by a threshold L ∗ ∈ [min( L ) , max( L )] of stringency on L . Note that an actual choice is meant to be made as to the applied value of L ∗ in practice, for which reason this kind of prediction is not probabilistic, but we call it a categoric prediction.
based on the likelihood function:
$$\mathcal { L } ( { \mathbf x } ) = \mathbb { P } _ { \chi | { \mathbf x } } ( \chi = 1, { \mathbf x } ) = \mathcal { P } ( { \mathbf x } ) / p ( { \mathbf x } ). \ \ ( 3 ) \quad \ \ I.$$
In the above P ( x ) = p x | χ ( x , χ = 1) P χ ( χ = 1) is the posterior probability density function (PDF) of x , and p ( x ) is the 'process' PDF, i.e., the basic PDF generated by the considered process . 1 Refer to the appendix of [19]
1 The probability density p x | χ ( x , χ = 1) of x conditioned on some realized value of χ is usually denoted
Depending on L ∗ , the rate of true positives, or the hit rate , i.e., the frequency of making correct predictions, yields as follows:
$$& \mathfrak { / } ) \quad H ( \mathcal { L } _ { * } ) = \frac { \int _ { \mathbb { R } ^ { M } } d V _ { x } \mathcal { P } ( { \mathbf x } ) \mathcal { H } ( \mathcal { L } ( { \mathbf x } ) - \mathcal { L } _ { * } ) } { \int _ { \mathbb { R } ^ { M } } d V _ { \mathbf x } \mathcal { P } ( { \mathbf x } ) }. \ ( 4 )$$
In the above dV x is a volume element in the precursory space, and H · ( ) is the Heaviside
more simply as p ( x | χ = 1) , but we want to emphasize that we consider a function of two variables. Also, it would create ambiguity if the symbol p without a subscript was to be reused to denote another function, and we prefer to reserve p for the process PDF.
step function. Another measure of the overall goodness or skill of prediction is the false alarm rate :
$$F ( \mathcal { L } _ { * } ) = \frac { \int _ { \mathbb { R } ^ { M } } d V _ { \mathbf x } [ p ( \mathbf x ) - \mathcal { P } ( \mathbf x ) ] \mathcal { H } ( \mathcal { L } ( \mathbf x ) - \mathcal { L } _ { * } ) } { \int _ { \mathbb { R } ^ { M } } d V _ { \mathbf x } [ p ( \mathbf x ) - \mathcal { P } ( \mathbf x ) ] }. \quad \stackrel { \infty } { \Xi } } _ { 0. 4 } \Big | \\ \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \text{$c$} \quad \end{cases}$$
Clearly, one can achieve a very good hit rate by reducing the stringency, but in fact [10] always at the price of an increased false alarm rate. Figure 2 shows an example of how the two measures of skill depend on the stringency in terms of a parametric plot or curve { ( F ( L ∗ ) , H ( L ∗ )) } , which is referred to as the receiver operating characteristic (ROC) curve . With the extremal choices, L ∗ = 0 and 1, we have ( F = 1 , H = 1) and ( F = 0 , H = 0) , respectively, i.e., the ROC curve stretches from corner to corner. It is a diagonal straight line with no prediction skill at all (over random predictions ˆ χ with P (ˆ = 1) = χ L ∗ ), and situated above the diagonal with any skill. In the same diagram another ROC curve is also shown, to be referred to as P -ROC curve, obtained by writing within the scope of the Heaviside function in the definitions (4), (5), and also in (2), P instead of L and P ∗ instead of L ∗ . Note that P ∗ ∈ [min( P ) , max( P )] . This is based on the intuitive strategy, expressed by the conditional probability p x | χ ( x , χ = 1) , that one looks at what happens before extreme events. From Eq. (3) one can see that following this strategy the posterior PDF is just the likelihood that such states lead to an extreme event weighed by the relative frequency of those states, whereby the 'predictive potential of relatively infrequent states is suppressed'. It can be shown [10, 9] that as a result of this the P -ROC curve will be always wholly underneath the L -ROC curve, making this intuitive strategy inferior. Furthermore, the L -ROC curve is always con-
Figure 2: Example L - (above, blue) and P -ROC curves (below, green) corresponding to the scenario of the autonomous L84 in Fig. 1 when x n = x n -1 and x ∗ = 1 8 . . The distributions that these ROC curves derive from can be seen in Fig. 3 drawn by the thin black line.
cave [11], while the P -ROC curve is not necessarily concave. Besides, in accordance with the above statement on the trade-off situation, F L ( L ∗ ) , H L ( L ∗ ) , F P ( P ∗ ) , H P ( P ∗ ) are all monotonic functions, and, therefore, so are e.g. H L ( F L ) and H P ( F P ) .
The ideal situation when extreme events ( χ = 1 ) and nonevents ( χ = 0 ) can be predicted with certainty ( ˆ χ = χ ) is represented by the ( F = 0 , H = 1) corner in the ROC diagram. In this case no choice has to be made on the applied stringency L ∗ . Clearly this is possible only in case of the deterministic but not the stochastic version of a model, and there are further factors - to be demonstrated in Sec. 3 - that can compromise the prediction skill. In the nonideal situation an optimal L ∗ is to be chosen. A unique optimum exists only in terms of a single-objective optimization problem, defined by a scalar-valued cost function. However, in our case the minimization of the false alarm rate and the maximization of the hit rate are both 'valid' objectives. It takes
a specific application to be possibly able to define a scalar-valued cost function C F,H ( ) . For our general assessment of predictability we choose to consider the intuitive measure:
$$D = \min _ { \mathcal { L } } ( \sqrt { F ^ { 2 } + ( H - 1 ) ^ { 2 } } ), \quad ( 6 )$$
the distance of the ROC curve from the ideal corner. With no prediction skill at all: D = √ 2 2 / . Other summary statistics for the ROC curve have also been defined, such as the area under the curve [10], or the slope H ′ F ( F = 0) [8]. Unlike these two, the distance D can be associated to actual predictions specified by an actual choice for L ∗ . We note that it is not trivial to interpret what the comparison of D with a proper skill score of probabilistic prediction means.
## 2.1.2. Numerical issues
Perhaps the most obvious factor that compromises the prediction skill in the data-driven framework is the finite size N of the data set: { x , n x n } , n = 1 , . . . , N . The distributions p ( x ) , P ( x ) , L ( x ) will be approximated in our study by histograms { p b } , {P } b , {L } b , b = 1 , . . . , B , of a certain uniform bin size ∆ x ∈ R M ; different values of b can be assigned to the different bins by a sensible algorithm. Let us denote by ∆ x the unique linear bin size applied in all M dimensions after suitable nondimensionalization. Note that in the coarse-grained situation {L } b derives from { p b } and {P } b much the same way as with the continuous functions according to Eq. (3). With the discrete formulation of Eqs. (4) and (5), accordingly, the ROC curve turns into (the graph of) a staircase (function), given by a set of discrete data points: { ( H , F b b ) } , b = 1 , . . . , B , belonging to stringency levels {L ∗ ,b } = {L } b :
$$H _ { b } \ = \ \frac { \sum _ { b ^ { \prime } = 1 } ^ { B } \mathcal { P } _ { b ^ { \prime } } \mathcal { H } ( \mathcal { L } _ { b ^ { \prime } } - \mathcal { L } _ { b } ) } { \sum _ { b ^ { \prime } = 1 } ^ { B } \mathcal { P } _ { b ^ { \prime } } } \quad \ \ ( 7 )$$
$$F _ { b } \ = \ \frac { \sum _ { b ^ { \prime } = 1 } ^ { B } \stackrel { \nu } { ( p _ { b ^ { \prime } } - \mathcal { P } _ { b ^ { \prime } } ) } \mathcal { H } ( \mathcal { L } _ { b ^ { \prime } } - \mathcal { L } _ { b } ) } { N - \sum _ { b ^ { \prime } = 1 } ^ { B } \mathcal { P } _ { b ^ { \prime } } } ( 8 )$$
Note that in the above the histograms { p b } and {P } b do not need to be normalized; and e.g. {L } b and {L } b ′ , b, b ′ = 1 , . . . , B , denote the same set. Note also that if L b does not-, then neither do H b and F b change monotonically with increasing b .
The above estimation of the measures of skill is not conservative , 2 however, which is to do with small histogram counts and associated statistical errors. An approach to fix this problem is the following. The available data is divided equally into 'training' and 'evaluation' data sets. Then, the conservative estimates are defined again by Eqs. (7) and (8), but the different terms appearing in them are associated with different data sets: {L } b is derived from the training data set, and { p b } and {P } b are derived from the evaluation data set. Note that the latter requires the use of the same grid forming the bins in case of the training and evaluation data sets.
A further issue to do with small bin sizes when many bins contain a single data point is that the 'ROC staircase' can have an excessively large last step. This is so, because bins that contain single data points tend to have empty counterparts mutually between the 'training' and 'evaluation' data sets. This way D > √ 2 / 2 can even be realized.
2 By 'conservative estimation of the skill' we mean that an overestimation of skill (e.g. D estimated to be smaller than the true value) is excessively unlikely. This entails an appropriate sign of the bias of the estimator, and a standard deviation of the estimator much smaller than the modulus of the bias.
Too large bin sizes would of course also deteriorate the prediction skill. Therefore, there should be an optimal bin size yielding (locally) minimal D . Our numerical experience shows that there is always, for any given prediction lead time or threshold level x ∗ , a unique (globally) optimal uniform bin size defining the regular grid.
## 2.1.3. Relationship of data- and model-driven predictability
We note finally that the above description of evaluating predictability applies clearly to the case of data-driven prediction. However, evaluating the model-driven predictability of binary exceedance events measured by the same ROC-based measure of prediction skill can be done in exactly the same way: by simulating long trajectories of the model and by binning the resulting time series data. This is more obviously true when MDP is thought of in the sense that the complete PDFs, p ( x ) , P ( x ) , L ( x ) , or rather the histograms, { p b } , {P } b , {L } b , are established preliminary to making any predictions. We will refer to such an MDP as an 'archival' MDP. Note that in this sense any DDP is archival. The archival MDP and the corresponding DDP differ only in that in principle unlimited data is available to establish the PDFs for MDP. But MDP can be thought of also in the sense that the model is only simulated 'on-demand', to produce an ensemble forecast from which the likelihood of a threshold exceedance can be established. The case of on-demand MDP corresponds to imposing a constraint on archival MDP or DDP, namely that the components of the precursory structure x of the archival MDP or DDP belong to the same time instant ( d m = d m ′ , m,m ′ ∈ [1 , ..., M ] ), x thereby representing the initial conditions for the on-demand MDP. Note that it is allowed that x excludes some variables that determine the considered phe- nomenon ( M < d d , being the dimension of the phase space, as detailed in Appendix B); the excluded variables can be initialized arbitrarily, or possibly as a random sample from a probability distribution. We refer here to the practice of stochastic parametrization of unresolved processes in weather forecast models.
As for the definition of the binary event in case of on-demand MDP the following can also be taken: in an event a chosen observable exceeds a set threshold in a chosen future time interval, defined by a leading window of width ∆ T at a lead time T ahead of the present time. Let us call this an event of threshold exceedance in an interval (TEI event in short). In contrast to considering the threshold exceedance of apexes of the time evolution, i.e., POT events, here T can be set arbitrarily, not restricted to discrete values; accordingly, χ t ( ) and x ( ) t are defined in continuoustime. We will take this approach when assessing predictability with respect to the prediction lead time in Sec. 3.2.1.
The bin size ∆ x for a DDP would in this case correspond roughly to the precision δx of measuring initial conditions for a MDP. (The difference is that ∆ x corresponds to averaging of the distributions, e.g. L ( x ) , over predefined fixed bins, whereas δx corresponds to a smoothing of the distribution: ˜ (˜) L x = ∫ d x L ( x ) p e (˜ x -x ) , where δx can be thought of as the standard deviation of the measurement error distribution p e (˜ x -x ) .) In order to well-approximate the skill of MDP for a given measurement precision, the time series has to be long enough so that the likelihood for all bins, even those which cover a relatively small portion of the invariant measure of the attractor, is well-approximated on the first place. That is, the ratio of the prediction skill of DDP and that of MDP (with no model errors), which is always smaller than unity, depends on N ∆ x .
## 2.2. The model climate
We carry out the assessment of the predictability of extremes in a model of geophysical relevance. It constitutes a nonlinear dynamical system featuring complex chaotic deterministic dynamics. Yet, it is simple enough - involving just three scalar prognostic variables -to yield time series data relatively inexpensively, and to allow for a more tangible demonstration of some aspects of predictability. Our choice of a model, Lorenz's model of global atmospheric circulation (L84) with standard parameter settings, reads as follows [17]:
$$\dot { x } & = - y ^ { 2 } - z ^ { 2 } \, - x / 4 \, + F / 4, \quad \begin{pmatrix} \Re \\ p \end{pmatrix} \\ \dot { y } & = x y - 4 x z \, - y + 1, \quad \begin{pmatrix} 0 \\ 0 \\ \tilde { z } & = x z + 4 x y \, - z. \end{pmatrix}$$
The model describes - in a very coarse manner [20] - the meridional heat transport via eddies, represented by principal mode amplitudes y and z , given rise by the baroclininc instability of the midlatitude jet, represented by its average speed x . The instability occurs for appropriate conditions defined by the large scale meridional temperature gradient, represented in the model by F , due to differential heating between the equator and the poles. The equations are nondimensionalized with respect to time by the average damping time of eddies, being about 5 days. This model enjoys popularity in teaching [21, 6] as well as theoretically oriented weather and climate research [22, 23, 24, 25].
We will examine the autonomous dynamics in perpetual winter conditions realized by, say, F = 8 , since this gives rise to chaotic dynamics, which is nontrivial from the point of view of predictability. Let us label the autonomous dynamics/model by M1, as a synonym of (9), when the forcing takes the above indicated form. In order to assess the de- pendence of predictability on intrinsic system properties, according to objective (i.a), we will consider also nonautonomous dynamics. It is achieved by introducing some driving or time-dependent forcing to the L84 system in the form of F t ( ) = F 0 + Ax t ˜( ) , where the fluctuating process (in a mathematical sense) ˜( ) x t can be seen as an unresolved, i.e., physically not modeled, (physical) process. However, from the point of view of data-driven predictability, as described in Sec. 2.1, autonomous and nonautonomous systems are not distinguished - the driving mimics additional degrees of freedom of the system. We can represent additional degrees of freedom of comparable time scales to that of the resolved dynamics, τ L 84 ≈ 4 , by a continuous smooth chaotic process, such as the first component of the classical Lorenz equations (L63): { ˙ ˜ = x τ -1 σ y (˜ -˜) x , y ˙ ˜ = τ -1 ( ρx ˜ -˜ y -˜˜) xz , z ˙ ˜ = τ -1 ( -βz ˜ + ˜˜) xy } , with an appropriate choice for the time-scale-tuning parameter τ as follows. With the common choice for a chaotic solution of the original ( τ = 1 ) equations: σ = 10 , ρ = 28 , β = 8 3 / , the time scale of L63 is τ L 63 = 0 7 . . The latter is defined by the crossover frequency in the power spectrum. Setting some other value for the timescale-tuning parameter we can achieve a new time scale: ττ L 63 . Therefore, a comparable time scale can be achieved by τ = τ L 84 /τ L 63 . However, it is not ττ L 63 that we want to set equal to τ L 84 . But rather, we regard the time scale of the driving comparable or approximately equal to the intrinsic time scale when L63 exerts maximal response of the driven L84 in terms of extremal behavior, measured e.g. by either the kurtosis or a high quantile of the distribution of x , as found in [24]. We can call this a kind of resonance, and define a time scale τ ′ such that it is unity in resonance. Based on our finding reported in [24] τ ′ ≈ 0 4 . τ . Note that τ ′ ≈ 2 3 . > 1 when
ττ L 63 = τ L 84 . Beside the resonant τ ′ = 1 scenario (M2) we will also consider one when the driving is much faster than the model climate: τ ′ = 1 4 / (M3). Assuming an even larger time scale separation between the resolved and unresolved processes one can apply an uncorrelated white noise (WN) driving: ˜( ) x t = ξ , ∫ t -∞ dtξ = W t , where W t is a Wiener process. We label the resulting model by M4. We use a coupling strength A = 0 025 . in case of the L63-driving, and an appropriate choice of A in case of the WN-driving that gives the same variance of the driving.
## 3. Results
In this paper we focus primarily on the predictability of threshold exceedances of the first component x of the various L84 models M1-4, whose symbol happens to coincide with that of our generic observable x . We will consider one other observable, and if it is not explicitly said, we mean to speak about the main observable x . Also, unless explicitly otherwise said, figures for D are based on the L -ROC curve.
## 3.1. Predictability of peak-over-threshold events
Histograms in this paper are constructed from sets of about 5 × 10 6 discrete data points each, resulting from appropriately long simulations. For the following results we simulate the autonomous and L63-driven L84 (M13) using Matlab's ode45 , which integrator chooses the time step size h adaptively. This is to make use of the event-handling capability of ode45 for the purpose of locating smooth apexes of x t ( ) . We employ the explicit order 1.5 strong scheme described in [24] to integrate M4 the WN-driven L84 with fixed h = 0 01 . .
## 3.1.1. Dependence on the makeup of the precursory structure and on intrinsic properties the model
The most simple case of a discrete-time precursory structure is that of the previous peak value of the observable whose threshold exceedances are to be predicted: x n = x n -1 . The two panels of Fig. 3 show the posterior PDF and the likelihood function, respectively, represented by adequate histograms. The cases of the autonomous, i.e., undriven, L84 (M1), and the L63-driven L84 (M3) are shown in one diagram side-by-side. As expected, the driving smooths out features of both distributions seen in Fig. 3, however, contrary to expectations: e.g. the likelihood (b) can be even enhanced by driving (see for example 1 2 . < x n -1 < 1 4 . ). Furthermore we point out that the relationship given by Eq. (3) is manifested in the more broad structures of the distribution of the likelihood as compared to that of the posterior probability density. This is so because P ( x ) tends to be peaked where p ( x ) is peaked. As already mentioned in Sec. 2.1, this broadening ought to be reflected in the relative positions of the respective ROC curves. In fact the pair of ROC curves in Fig. 2 belong to the present prediction scenario considering M1.
It is an intuitive expectation that the predictability can be improved by relying on more information by means of extending the precursory structure. The most simple step in this direction - in the realistic context when only one variable is available or practical to observe is that beside the previous data point we monitor also the one just before that, i.e., x n = ( x n -1 , x n -2 ) . The distributions for the same scenarios as considered before in Fig. 3 are displayed in Fig. 4. The bivariate distributions are visualized by color plots; and beside the distributions, on the left we also display scatter plots of data points (but for a better visibility of features we plot fewer points than those
Figure 3: Distributions of the likelihood L and posterior probability density P when x n = x n -1 and x ∗ = 1 8 . . The solid thin black line and thick gray line correspond, respectively, to models M1 and M3 defined in Sec. 2.2. Corresponding ROC curves in case of M1 are shown in Fig. 2.
that the histograms are constructed from). In the case of the autonomous L84 (M1) the scatter plot [Fig. 4 (a)] reveals a fractal pattern with a distinctive filamentary structure. This could lead one to think that there is a one-toone or unique relationship between subsequent pairs of ( x n -1 , x n -2 ) , which is also called a mapping or map [6]. This can be confirmed by looking at the distribution of the likelihood [Fig. 4 (b)], which takes on the maximum value of unity wherever the scatter plot exhibits fractality. - Because of the uniqueness, we have a deterministic rule to predict the next value, and so we can tell with certainty whether it will exceed the threshold. In regions where a lack of clear fractality is observed, e.g. around ( x n -1 , x n -2 ) = (1 4 1 4) . , . , L < 1 consistently. The exhibited pattern of the scatter plot can give the intuition that the lack of uniqueness is a result of 'looking at' a curved surface living in 3D 'from a poor angle' so that the 2D view of some parts of the surface is obstructed by other parts of it. In other words, the surface looks folded. In fact, Takens' embedding theorem [26] states that an attractor of Hausdorff dimension D 0 can (always) be embedded by M > 2 D 0 number of delay variables. For us this means an unfolded appearance. In our case D 0 ≈ 1 6 . [24], and so for uniqueness we need maximum M = 4 . This does not 'encourage' us that we can have an unobstructed 2D view, although neither does it say that we cannot have. In fact, in our case we can have such a view, to be described next.
Let us bear in mind that the discrete x n data belong to apexes of the continuous x t ( ) . In these points ˙ x = 0 . Wecan use this fact in conjunction with the first component of the equations of L84 (9) to determine that the apexes are situated on the surface: y 2 + z 2 = -x/ 4 + aF . This can be viewed as a Poincar´ e surface of intersection that defines a slice of the attractor - the Poincar´ e section [6]. For any fixed x we recognize the equation of a circle. That is, the surface itself is locally conical, which approximation applies well to the chaotic at-
tractor with F = 8 extending between about [-0.5,2.5] wrt. x , as seen in Fig. 1. Such a surface can be rectified on the plane spanned by the azimuthal angle
$$\phi = \arctan ( y / z ), \text{ \quad \ \ } ( 1 0 )$$
periodic in e.g. [0 , 2 π ] , and x . Therefore, there exists a unique mapping between subsequent pairs of ( x , φ n n ) . This allows for an unfolded view of the Poincar´ e section, and so for the prediction of the next apex with certainty. Accordingly, as seen in Fig. 4 (g), the scatter plot exhibits fractality everywhere, and L = 1 (or 0) also everywhere [Fig. 4 (h)]. This certainty is compromised in the numerics only by the effect of coarse-graining, when {L } b may be less than unity due to the finite data set size.
By introducing a driving as defined in Sec. 2.2, the dimensionality of the problem increases. Therefore, the same precursory structure of only two variables is inevitably insufficient for predictions with certainty. The example of M3 shows that the scatter plot becomes area-filling [Fig. 4 (d) and (j)], and we will have distributions of the likelihood that take on all values between 0 and 1 [Fig. 4 (e) and (k)]. However, as long as the forcing strength is moderate, the new features tend to develop through a smearing of the original ones. Evidently, this applies partially in our case.
ROC curves that derive from the distributions seen in Figs. 3 and 4 are shown in Fig. 5. L -ROC curves are always above the P -ROC curves, and the corresponding ROC curves for undriven and driven versions of L84 have the same relation consistently. In particular, for the scenario of the undriven L84 when x n = ( x n -1 , φ n -1 ) , the ROC curves approach very near the ideal corner of certainty.
For the various scenarios considered the distances D (6) from the corner, as a summary measure, is provided in Table 1. Besides the
Figure 5: L -and P -ROC curves for scenarios specified in Table 1 marked with the corresponding numbers framed in boxes.
scenarios treated in Fig. 5, data is provided in the table also for several other scenarios as follows. First off, we used white noise driving too. Unexpectedly, the predictability with a single delay variable, x n = x n -1 , is better for this driven case than the undriven one. This is unchanged even if we apply a smoothing to the time series. Note that x of the WN-driven L84 (M4) is a red noise-like nonsmooth process. An improvement of predictability by smoothing is achieved only with larger precursory structures (compare the values between round brackets with the numbers just above each in the said table). The effect of improved predictability having introduced a driving could be due to a stabilization of the trajectory by noise. However, instead of x , considering an-
Figure 4: Scatter plots (left) and distributions of the likelihood L (middle) and posterior probability density P (right). In each diagram the color scale ranges from dark blue for 0 to dark red for the maximal density value, which is unity for L for the presented scenarios, but various different values for P . We used x ∗ = 1 8 . . The first and third (second and fourth) row concern M1 (M3). The following correspondence between the labeled panels above and scenarios labeled by boxed numbers in Table 1 and Fig. 5 stand, respectively: (b) 7 (c) 8 (e) 5 (f) 6 (h) 11 (i) 12 (k) 9 (l) 10.
other observable, namely, the total cyclonic activity in the model,
$$r = \sqrt { y ^ { 2 } + z ^ { 2 } },$$
we do not observe the same effect; see Table 2, and note that the process of r , contrary to that of x , is smooth. Therefore, the more likely cause of the unexpected effect is that the geometry of the attractor is altered by noise in a favorable manner, when the Poincar´ e section 'looks' less folded 'in view of x ' (but not r ).
Beside a fast ( τ ′ = 1 4 / ) L63-driving (M3) we also considered a slower one with matching time scales of driving and model climate, i.e., τ ′ = 1 (M2). The rationale for this is that we expect that the delay variables, with delay times determined dominantly by the main system, would be able to pick up more information on the driving of a longer decorrelation time. However, while this mechanism should be at work, an improvement of predictability is not registered, but on the contrary. This is so also when using one more delay variable: x n = ( x n -1 , x n -2 , x n -3 ) , or when considering observable r . The likely cause of this is that, as a counter-effect, the trajectory is destabilized by the driving more so with τ ′ = 1 than 1/4.
Beside - but not independent of - the issue of the choice of the observables to make up the precursory structure, we can make an interesting observation to do with the size of the precursory structure too. Extending the two delay variables with a third one (all of the same type) did improve the predictability. However, it was still not as good as with the shorter precursory structure involving the azimuthal angle φ . This is consistent with Takens' embedding theorem, as mentioned earlier.
## 3.1.2. Dependence on the event-magnitude
We evaluate now the dependence of predictability on the event-magnitude to achieve our objective (i.b). That is, we construct D x ( ∗ ) , with computations for an array of sample points of x ∗ . The numerical value of D x ( ∗ ) does vary with the bin size ∆ x . This variation has a single minimum in all cases checked (and so we assume that this is always the case), but these are different ∆ x opt values for different values of x ∗ . We intend to construct the D x ( ∗ ) which is optimized for all values of x ∗ , and this is what we regard as the 'dependence of predictability on the event-magnitude'. This result is achieved by constructing the unoptimized D x ( ∗ ) for a range of ∆ x (or B ) values, and plot these curves in a single diagram. The range of ∆ x values should include the optimal values belonging to all x ∗ 's. Then, the lower envelope of these curves will represent the optimized D x ( ∗ ) . In Fig. 6 this construction is shown for M1, M3, M4, sideby-side, considering the precursory structures x n = ( x n -1 , x n -2 ) and x n = ( x n -1 , φ n -1 ) , φ being defined by (10). For each model we see a decreasing trend of D x ( ∗ ) , but only for the largest values of x ∗ (from about 1.8), and only on coarse scales of x ∗ , i.e., on smaller scales of x ∗ the variation of D x ( ∗ ) can be nonmonotonic . 3 In effect we have assessed the monotonicity of the magnitude-dependence D x ( ∗ ) depending on other factors, the precursory structure and also the model, and we have found a rather robust behavior.
Next we carry out a thought experiment raising continuously a horizontal line in the scatter plots in the left column of Fig. 4. We monitor for what x n -1 levels do new features enter into- or features present exit from what is above the line. The measure of points above
3 A blowup approaching the largest value of x , and so x ∗ , is due to the undersampling of the probability distributions, and so it is to be disregarded. The undersampling does not show up for M1 because the tail of the process PDF p x ( x ∗ ) , to be defined shortly below, does not decay slowly like in the other models.
Table 1: Summary measure for the ROC-statistics: the 'distance' D (6) from the ideal case of all events correctly predicted without making any false alarms. The models M1-4 assessed are defined in Sec. 2.2. The threshold is the same in all cases, that is, x ∗ = 1 8 . . Histograms were constructed with B 1 / 2 = 200 bins in all dimensions in the respective ranges where data points are present; the 'resolution' that the B 1 / 2 = 200 bins and the N ≈ 5 × 10 6 data points give is depicted in the middle and right columns of Fig. 4. The number of significant digits were determined based on only two independent realizations. The figures in round brackets were obtained by a smoothing of the nonsmooth white noise-driven time series over a moving window of width of a nondimensional time unit. The numbers framed by boxes indicate the correspondence with results shown in Fig. 5.
| Precursory structure | M1 | M1 | M2 | M2 | M3 | M3 | M4 | M4 |
|---------------------------------|-----------|-----------|--------|--------|-----------|----------|----------------|----------------|
| Precursory structure | P | L | P | L | P | L | P | L |
| x n - 1 | 0.414 4 | 0.359 3 | 0.480 | 0.409 | 0.490 2 | 0.410 1 | 0.390 (0.374) | 0.357 (0.358) |
| ( x n - 1 , x n - 2 ) | 0.062 8 | 0.027 7 | 0.278 | 0.217 | 0.265 6 | 0.197 5 | 0.527 (0.312) | 0.320 (0.226) |
| ( x n - 1 , φ n - 1 ) | 0.0108 12 | 0.0054 11 | 0.1803 | 0.1116 | 0.1840 10 | 0.1114 9 | 0.2930 (0.206) | 0.2587 (0.156) |
| ( x n - 1 , x n - 2 , x n - 3 ) | 0.0177 | 0.0087 | 0.2456 | 0.1434 | 0.2257 | 0.1350 | 0.4917 (0.323) | 0.2828 (0.189) |
Table 2: Same as in Table 1 but with observable r of cyclonic activity, and r ∗ = 1 8 . . The number of significant digits is taken to be the same as in case of observable x , i.e., not based on a number of independent realizations.
| Precursory structure | M1 | M1 | M2 | M2 | M3 | M3 | M4 | M4 |
|---------------------------------|--------|--------|--------|--------|--------|--------|--------|--------|
| Precursory structure | P | L | P | L | P | L | P | L |
| r n - 1 | 0.475 | 0.329 | 0.513 | 0.368 | 0.503 | 0.354 | 0.599 | 0.420 |
| ( r n - 1 , r n - 2 ) | 0.054 | 0.025 | 0.192 | 0.123 | 0.171 | 0.098 | 0.288 | 0.181 |
| ( r n - 1 , r n - 2 , r n - 3 ) | 0.0150 | 0.0069 | 0.1667 | 0.0802 | 0.1433 | 0.0660 | 0.2829 | 0.1440 |
Figure 6: Event magnitude-dependence of predictability of POT events. The thin black and gray lines depict the dependence using the precursory structure ( x n -1 , x n -2 ) and ( x n -1 , φ n -1 ) , respectively, for models (a) M1, (b) M3 (result qualitatively representative of those for M2), (c) M4. In each scenario a bundle of curves belong to B 1 / 2 = 50 j , j = 1 , . . . , 8 number of bins in one dimension. The lower envelope of the bundle approximates the optimized D x ( ∗ ) . In panel (a) the optimal B 1 / 2 is larger than 400 (and checked to be larger than even 1000), but we believe that the same qualitative behavior persist up to the optimum. Dot markers on the curves mark out sample values of x ∗ . In panel (b) the (rescaled) PDF of x is included for reference, depicted by the lowermost thick black line.
the line is in fact given by the denominator in (4) as the integral of the posterior probability density, which is nothing but P χ ( χ = 1) , because ∫ R M dV p x x | χ ( x , χ = 1) = 1 . Considering the meaning of the event variable (1), P χ ( χ = 1) as a function of the threshold x ∗ only is equivalent with the complementary distribution function belonging to the process PDF p x ( x ∗ ) of one variable x ∗ pertaining to the observable of concern x . Therefore, what we monitor in effect is the points or levels x ∗ of discontinuities or fast changes of p x ( x ∗ ) . The levels of the major discontinuities are discernible in Fig. 4. Note that in the unperturbed L84 because of the quadratic tangency of the filaments to horizontal lines, for increasing x ∗ , before a discontinuous drop of the density p x ( x ∗ ) , it actually increases. In the driven L84 there are no discontinuities, they are 'washed out', but, as the lowermost curve in panel (b) of Fig. 6 shows, the density p x ( x ∗ ) features strong nonmonotonicities at the levels of interest. We observe that these levels coincide well with the locations of 'humps' of D x ( ∗ ) in all cases displayed in Fig. 6 (a) and (b). This suggests that the behavior of D x ( ∗ ) is controlled at least in part by p x ( x ∗ ) . This raises the question whether the decay of p x ( x ∗ ) is responsible for the decreasing nature of D x ( ∗ ) , at least in the present situation. We will revisit this question in Sec. 4.
## 3.2. Predictability of threshold-exceedancein-an-interval events
Here we attend to our objective (iii).
## 3.2.1. Dependence on the prediction lead time
Let us emphasize that when the precursory space can embed the attractor - defining an ideal precursory structure - the effect of the destabilization of trajectories, mentioned in Sec. 3.1.1, can influence data-driven predictability only in the practical sense of having a finite trajectory length, i.e., finite data set size. Because of the latter, a coarsegraining is inevitable in constructing the histograms. When establishing a correspondence
of DDP of POT events with on-demand MDP in terms of the Poincar´ e mapping (not the original flow), the coarse-graining can be translated into terms of errors in measuring initial conditions/precursory observables 4 of size bounded from above by the histogram bin size. In case of a chaotic trajectory the error in tracing the trajectory forward in time grows exponentially fast (at least while the error is still small). In MDP this manifests in the spreading out of the ensemble. Fixing the ensemble size, the evaluation of the likelihood of an event will have a larger statistical error the more spreadout the ensemble is. The latter would typically correspond to a longer prediction lead time. The larger errors in estimating the likelihood should clearly precipitate in a deterioration of the overall prediction skill. Because of the correspondence, this deterioration carries over to DDP. Note that when the attractor cannot be embedded in the precursory space, the instabilities of trajectories take effect also if N →∞ , because some initial conditions are randomly initialized.
In the situation with an ideal precursory structure as discussed in Sec 3.1.1, the likelihood was evaluated to be nil or unity, or that with a very good approximation, because the prediction lead time was in fact limited by the typical time scale of the system, given that from one apex we intended to predict the next one. We could evaluate the dependence of predictability on the lead time by looking further than the next apex to predict. However, instead of this exercise we prefer to map out the predictability as a continuous function of the
4 For this point it does not matter whether the Poincar´ e mapping can be constructed analytically to facilitate the on-demand MDP of POT events. In fact, it is not possible in general even for the most simple chaotic flows. In that case only archival MDP of POT events is possible, whose skill, nevertheless, should be the same as that of the hypothetical on-demand MDP.
prediction lead time instead of its discrete advances. Our preference is partly due to the fact that the discrete advances are not known 'apriori', i.e., before integrating the system. That is, next we examine the predictability of not POT but TEI events.
In practice T , ∆ T , and t can take values that are integer multiples of the trajectory sampling time, which latter is chosen small anyway in order to secure good accuracy of tracing out trajectories by numerical integration of (9). In fact, for this exercise we use the classical Runge-Kutta algorithm/stochastic integrator mentioned above in case of M1-3/M4 with fixed h = 0 01 . , and we save the state in every 5th time step (to make sure that there will never be two trajectory points subsequent in time in one bin); furthermore we apply ∆ T = 2 × 5 × h . As for the precursory structure we will take the triplet of the system state variables ( x, y, z ) . Therefore, the threshold x ∗ , the interval length ∆ T , and the dynamics itself, determine an event volume in phase space. We then generate data points for the tri-variate histogram {P } b by identifying trajectory sample points in the event volume and trace them backward in time by T . We do this not onthe-fly during the simulation, but as a postprocessing of the pregenerated long time series data produced by numerical integration. For now we choose a bin size rather arbitrarily, and by 'predictability' now we mean predictability conditional to the fixed bin size, not the best possible predictability - given a fixed data set - as a result of some optimization.
The results of evaluating the predictability in terms of the distance D for M1 and M4 are displayed in Fig. 7. We can make a number of observations. First, the driven system is less predictable, as expected. Second, the predictability is declining in both cases with increasing prediction lead time, as it should, given the chaotic dynamics and the finite data
set size. We have checked that beyond the range of T shown the curves approach the √ 2 2 / asymptote, belonging to a straight diagonal ROC curve, meaning no prediction skill at all. This is also expected. Third, on shorter scales D T ( ) is not monotonic in either case, unlike in the well-known case of an autoregressive AR(1) process of order one, studied regarding data-driven predictability by Hallerberg and Kantz [8]. In the case of L84 the deterministic or autonomous part of its equations results in a chaotic dynamics which is much more complex than the linear deterministic term of the AR(1). In particular, the deterministic term of AR(1) need to have a stabilizing effect on the trajectories in order to have a bounded dynamics, while, although on the compact chaotic attractor of L84 trajectories are bounded, they are unstable in a long-term average sense measured by a positive Lyapunov exponent. This instability can also deteriorate predictability, beside a stochastic part if any. On short-terms, however, the deterministic trajectory can experience stable periods, which periods are associated with the return of skill admitted by the plateaus, or negative slopes even, of D T ( ) .
## 3.3. Dependence on the event-magnitude
In comparison with the AR(1) process, a further matter of interest is the dependence of predictability on the threshold level. The counterintuitive finding in case of AR(1) was reported by Hallerberg and Kantz [8], namely, that stronger extremes - indifferently to the distribution that the process realizes - are more predictable. The obvious question to ask is, then, whether the latter holds also in case of processes with a more complex deterministic dynamics subjected (or not) to stochastic forcing.
Figure 8 (a) and (b) show the predictability as a function of the prediction lead time as
Figure 7: Predictability as a function of the prediction lead time. The curve on top (blue) is obtained for M4, and the other one (green) for M1. Histograms were constructed with B 1 / 3 = 40 bins in all dimensions in the respective ranges where data points are present.
well as the threshold level for M1 and M4, respectively. Firstly, the nonmonotonic nature of D T ( ) is prevalent on any fixed threshold level. Secondly, we observe that while for some fixed prediction lead times stronger events are more predictable, i.e., D x ( ∗ ) is a decreasing function, it is just the opposite for some other T 's. That is, the above statement for AR(1) does not seem to hold in general for more complex dynamical systems.
However, Fig. 8 (b) of the noisy L84 admits values of D > √ 2 / 2 , which should be erroneous. In fact the reason for this error is that the corresponding ROC curves or staircases (not shown) do not extend to the corner (1,1), or more precisely, they feature an excessively large last step - for the reason stated in the end of Sec. 2.1.2. This is so because the back-traced data points from the event volume are spread out in relatively large domains of the phase space due to the relatively large prediction lead times and strong instabilities (on average) of the trajectories. As also mentioned already, there is a unique optimal (uniform) bin size yielding a minimal D . Wedetermined numerically this optimum for each sam-
Figure 8: Predictability in terms of D as a function of the prediction lead time T and threshold level x ∗ . Results are shown for models M1 and M4 in panels (a) and (b), respectively. Histograms were constructed with B 1 / 3 = 100 bins in all dimensions in the respective ranges where data points are present. Notice the different ranges of D shown.
ple combination of ( T, x ∗ ) separately. This was done using a simple algorithm of maximum finding detailed in Appendix A, which is suitable for treating nonsmooth functions of one variable. The result of this optimization for M1 and M4 can be seen in Fig. 9 (a) and (c), respectively. The surprising outcome with the bin size optimization is that stronger events are generally better predictable, reinstating the rule found by Hallerberg and Kantz [8] for the simple stochastic process of AR(1). Only in case of M4 do we see an anomaly for very high thresholds, which could well be the same undersampling effect what was seen in Fig. 6. A further observation is that the nonmonotonic T -dependence is suppressed/gone almost completely for M1/M4.
The optimal number of bins are shown in Fig. 10. Comparing these diagrams with the corresponding ones in Fig. 8 one can notice that larger values of the unoptimized D correspond to fewer and so larger optimal bins. The reason for this is that in these situations the trajectories are more unstable and therefore they scatter in a larger volume, which 'asks for' increasing the bin size in order to have a better estimate of the likelihood in those bins.
Finally, we note that MDP cannot involve such optimization; the bin size is determined by the precision of observation only. Nevertheless, the unoptimized result in Fig. 8 does not represent model-driven predictability either, because in many bins there is an insufficient number of points for the evaluation of the likelihood. For a given data set size the likelihood can be well-approximated in most bins with a bin size larger than the optimal one for DDP. For the data set size in our analysis we evaluate the model-driven predictability as a function of the prediction lead time T and the threshold level x ∗ for a 'hypothetical' observational precision that derives from approximately the largest optimal bin size in the considered ranges of T and x ∗ , taken to be B 1 3 / = 25 bins in all three dimensions. The interesting result is that also the model-driven predictability is the better the stronger the extremes.
## 4. Summary and discussion
We examined the predictability of threshold exceedance 'extreme' events in a simple but chaotic continuous-time dynamical system or 'flow'. Given the nature of the problem, namely, that extremes are rare, we chose an
Table 3: Summary of how different factors/choices influence predictability measured by D (6). Where appropriate, the mechanism is specified. Wherever not explicitly specified, the point applies to both POT and TEI events. Whether a point applies to data- or model-driven prediction, or both, should be clear from the context. A positive/arbitrary/negative effect on predictability is denoted by the symbol ↑ / glyph[arrowbothv] / ↓ . We denote the ceiling function by glyph[ceilingleft]·glyph[ceilingright] .
| Factor/choice | Effect/mechanism | Support (proof/example) |
|-----------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| 0. L -ROC vs P -ROC | ↑ L -ROC always better | [10, 9]; Fig. 4 (b) vs (c), (e) vs (f), (h) vs (i), (k) vs (l); Fig. 5 |
| 1. Makeup of the pre- cursory structure (PS) (assuming other factors fixed) | 1.1 Size of PS M Extending PS with additional observable, M ′ = M +1 : ↑ Always better | Embedding theorem [26] Compare figures in Tables 1 and 2 (T1, T2), in each column separately, row-wise [disregard row 3 (r3) of T1]. |
| 1. Makeup of the pre- cursory structure (PS) (assuming other factors fixed) | 1.2 Choice of observables | |
| 1. Makeup of the pre- cursory structure (PS) (assuming other factors fixed) | 1.2.1 A PS of size M ≥ 1 can outperform another PS of size M ′ , 2 glyph[ceilingleft] D 0 glyph[ceilingright] ≥ M ′ ≥ 1 no matter M ′ > M | Compare r3 to r4 in T1, wrt. each column separately. |
| 1. Makeup of the pre- cursory structure (PS) (assuming other factors fixed) | 1.2.2 Fixing the size M ≤ 2 glyph[ceilingleft] D 0 glyph[ceilingright] , there are an infinity of PSs with varying performance, and there should be an optimal, best-performing, one. Ideal case: precursory space can fully embed attractor. | Compare Fig. 4 (a) to (g) and (b) to (h); and r3 of T1 to other rows; and cases 11, 12 of Fig. 5 to other cases in it. |
| 2. Intrinsic properties of the system, i.e., the sys- tem itself. Therefore, the listed properties change 'in tandem' in general. | Overall effect: glyph[arrowbothv] Arbitrary, since at least one of the intrinsic properties have an arbitrary contribution, and they change in tandem. | In general the effects cannot be demonstrated in isolation (see examples for 2.1.1, 2.2), only if one property can be changed independently (see example for 2.1.2). |
| 2. Intrinsic properties of the system, i.e., the sys- tem itself. Therefore, the listed properties change 'in tandem' in general. | 2.1 Attractor | |
| 2. Intrinsic properties of the system, i.e., the sys- tem itself. Therefore, the listed properties change 'in tandem' in general. | 2.1.1 Larger dimension D 0 : ↓ The attractor would look more folded Note that 2.1.1 is not a converse of 1.1, because 2.1.2-3, 2.2 changes along with 2.1.1. | Compare r1 to r2 and r3 to r4 in Fig. 4; and figures in T1, T2 under M1 to M2-3 in each row separately. Exception: M1 to M4 in r1 of T1, dominated by other intrinsic factors |
| 2. Intrinsic properties of the system, i.e., the sys- tem itself. Therefore, the listed properties change 'in tandem' in general. | 2.1.2 Geometry: glyph[arrowbothv] Can enhance or suppress foldedness 2.1.3 Natural measure (probability distribution) glyph[arrowbothv] Can enhance or suppress the significance of foldedness | No such exception in T2 as in T1 wrt. 2.1.1, and note that the changes in 2.2 and 2.1.1 between M1 and M4 are the same. Note that it has no significance that also the PSs for M1 and M4 are not the same. |
| 2. Intrinsic properties of the system, i.e., the sys- tem itself. Therefore, the listed properties change 'in tandem' in general. | 2.2 Stronger instability of trajectories: ↓ Faster spread of ensemble of trajectories Note: no effect only in case of N →∞ and when PS fully embeds attractor in the same time | Compare M2 to M3 in T1 or T2. MLE of L63 increases for decreasing τ ′ , but the impact on L84 wrt. either 2.1.2-3 or 2.2 clearly counters and outweighs that (except for r1 in T1). |
| 3. Data set size N (other factors being arbitrarily fixed) | Smaller N : ↓ Larger statistical errors in estimating L | Intuitive statement not checked empirically |
| 4. Precision of measurement δx | Larger δx : ↓ Smoothing or coarser-graining in estimating L | Intuitive statement not checked empirically |
| 5. Bin size ∆ x ( N being finite and fixed) | glyph[arrowbothv] An optimum ∆ x opt exists, which might always be unique. | Our experience in all cases examined in numerics |
| 6. Prediction lead time T (only for TEI events) | Increasing T : glyph[arrowbothv] Possibly nonmonotonic D ( T ) | Fig. 8 |
| 6. Prediction lead time T (only for TEI events) | ↓ Increasing D ( T ) for ∆ x opt | Fig. 9 (a,c) |
| 7. Event magnitude x ∗ | Increasing x ∗ : POT events: ↑ glyph[arrowbothv] D ( x ∗ ) decreasing for highest thresholds on coarse scales, but could be nonmonotonic on 'fine' scales even for ∆ x opt 20 | Fig. 6 |
| 7. Event magnitude x ∗ | TEI events: glyph[arrowbothv] Possibly nonmonotonic D ( x ∗ ) | Fig. 8 |
| 7. Event magnitude x ∗ | ↑ Decreasing D ( x ∗ ) for ∆ x opt | Fig. 9 (a,c) |
(a)
0.2
0.15
0.1
0.05
0
10
0.4
0.3
0.2
0.1
0
10
D
(c)
D
5
T
5
2
2
2
5
T
1.5
0
x
∗
1.5
0
x
∗
T
Figure 9: Predictability in terms of D as a function of the prediction lead time T and threshold level x ∗ . Results are shown for models M1 and M4 in panels (a)-(b) and (c)-(d), respectively. The pair of panels in each column correspond to panels (a) and (b) of Fig. 8 but (a), (c) having the histogram bin size [uniform wrt. one histogram but different for each combination of ( T, x ∗ )] optimized, and (b), (d) with B 1 / 3 = 25 bins in all dimensions. Notice the different ranges of D shown.
arguably [12] appropriate measure of prediction skill for assessing predictability: the ROC statistics, more specifically, a distance measure D from the ideal situation of having all events successfully predicted without any false alarms. According to our top objective (i) and (iii) set out in the Introduction, we examined the dependence of predictability on various factors. Our conclusions are collected in a systematic form in Table 3 (T3); in the column on the right we refer to our results presented in this paper, and previous results reported by others, to support the statements in the middle column.
Among the factors we did not list the choice of observable for x whose extremal values are concerned. The reason for this is that we cannot make a statement of general interest with respect to this choice. The predictability of extremes concerning different observables x and x ′ can be compared only if we name conditions that have to be satisfied in both cases. Such a condition on the threshold levels x ∗ and x ′ ∗ can hardly be given objectively; one possibility is that the threshold level should belong to the same quantile of the respective process distributions p x ( ) and p x ( ′ ) . However, this is still a subjective condition. And since the pre-
0
1.5
x
∗
2.5
2.5
(d)
0.4
0.3
0.2
0.1
0
10
D
2.5
Figure 10: Optimal number of bins ( B 1 / 3 opt ). Panels (a) and (b) belong to (a) and (c) of Fig. 9. The colorbar applies to both diagrams.
dictability strongly depends on the threshold, a comparison is hardly possible . 5
Instead of merely the distance D , we can compare the monotonicity or trend (increasing or decreasing) of D x ( ∗ ) and D x ( ′ ∗ ) . This is the difference between objectives (i.a) and (i.b). The interesting finding in this regard is that D x ( ∗ ) is monotonically decreasing concerning TEI events - although only if the bin size of
5 This argument can be applied also to the intrinsic properties, i.e., the choice of the system in general (point 2 in T3). However, it can be of general interest to compare the predictability with respect to the same physical observable while only slightly changing the system, e.g. by changing a parameter, or by considering different types of perturbations as illustrated by our model choices M1-4.
histograms is optimized (see point/conclusion 7 (c7) in T3). Concerning POT events the situation is somewhat more intricate, but on coarse scales of the threshold level we observe the same behavior. It appears to be robust, being the same qualitatively for two different precursory structures and for all of M14. Since we found this effect in an arbitrarily chosen dynamical system, and with respect to an arbitrarily chosen observable of it, we suggest that it might be a rather typical behavior. This would be a nontrivial generalization of the same statement made by Hallerberg and Kantz [8] concerning autoregressive processes. A theoretical argument why this should or rather should not be always true is yet to be provided, however.
We point out that the above observation was made in the special case, among other cases, when the precursory space is identical to the phase space. There was one case concerning POT events (to do with the gray lines in Fig. 6), and another one concerning TEI events (see e.g. Fig. 9). Therefore, the effect of folding referred to under points 2.1.1-3 of T3 does not take place. However, we have already seen evidence that even in this case this is not the stability of trajectories (point 2.2 of T3) alone, or not that measured by the average finite-time maximal Lyapunov exponent 〈 λ ( T ) 〉 (defined in Appendix B), that determines D : As seen in Figs. 8 (a) and 9 (b), the bin size alone can change the monotonicity or trend of D x ( ∗ ) , while 〈 λ ( T ) ( x ∗ ) 〉 is obviously unchanged. The latter is shown in Fig. 11. This mismatch is not a finite data set size numerical effect, as lim ∆ x → 0 lim N →∞ D = 1 with no ( T, x ∗ ) -dependence that could match that of 〈 λ ( T ) 〉 . We might say rather that (assuming N → ∞ for DDP, or that we concern on-demand MDP) ∆ x (or δx ) controls the 'filtering' of intrinsic properties in determining D . We note that, provided that ∆ x > 0 (or δx > 0 ), a similar
Figure 11: Predictability in terms of the average finitetime maximal Lyapunov exponent 〈 λ ( T ) 〉 as a function of the prediction lead time T and threshold level x ∗ , corresponding to Fig. 8 (a). The exact correspondence is established by averaging the FTMLEs over the event volume of TEI events. The latter is defined in Sec. 3.2.1, and the corresponding average is formulated by Eq. (14).
filtering role can be played also by x ∗ . However, some other role seems to be played by the process PDF p x ( x ∗ ) too (see Fig. 6 (b)). With this, we believe to have reached our objective (iv).
Advanced algorithms to bin data will be the objective of our future research. In this regard it is envisaged that a theory linking formally some suitable measure of the instability of trajectories etc. with the prediction skill in terms of D could indicate the optimal grid in a straightforward manner, rather than having to find this grid by conducting a costly general iterative optimization procedure. Furthermore, it is often the case that the data set is very limited for pure data-driven prediction, while some model, even if inaccurate at the current stage of its development, is known. In this case it would be most beneficial if a combined data- and (archival or on-demand) model-driven prediction technique could exploit fully the assets of data and model at hand. We will concentrate efforts to develop such techniques.
Finally we remark that point (2) of the Introduction does seem to contradict our conclusion (c7). This is so given that we argued in Sec. 2.1.1 - according to our objective (ii) - that (c7) should apply also to model-driven predictions. However, it might be crucial to assume no model errors in order to reach (c7). This assumption certainly never holds in practice, and it might indeed lead to point (2) to hold. Model errors can be easily modeled, say, by taking M3 as the truth, and M4 as the model. Another reason for the apparent contradiction between (2) and (c7) may also be that they assume different measures of prediction skill. In order to possibly reach completely our objective (ii), we should- and we wish to examine these issues in a separate work.
## Acknowledgments
Much of this work was carried out at the Max Planck Institute for the Physics of Complex Systems, Dresden, Germany; the scholarship awarded by the Max Planck Society is gratefully acknowledged. The work was supported also by the NAMASTE project owned by Valerio Lucarini (under the ERC grant No. 257106). The author is indebted for useful discussions with Holger Kantz, Tam´ as T´ el, and Christian Franzke. Useful comments on the manuscript by Frank Lunkeit are gratefully acknowledged. The author is thankful to two anonymous reviewers for many helpful suggestions to improve this work.
## Appendices
## A. Algorithm for finding the approximate global maximum of a nonsmooth function of one variable
The ROC-based measure of prediction skill D is a discontinuous function of the linear bin size ∆ x due to the finite data set size N . Conversely, if there was infinite data available, it would be a continuous function. We will assume here that with finite N D , features a single minimum. If a function is discontinuous or nonsmooth, the Newton-Raphson algorithm that relies on the derivative cannot be applied to find a global minimum.
Instead of the bin size ∆ x , we will specify the number B of bins (or B 1 /n along a single dimension in the n -dimensional precursory space) in the domain where data points are found. Our experience is that, given the data available, as specified in Sec. 2.2, and the ranges of T and x ∗ desired to be explored, the optimal number B 1 3 / opt of bins is between, say, 6 and 200. One could evaluate D for all intermediate integers to find the one that gives the smallest D . However, one can do better than that. The following algorithm is applicable to smooth functions f ( x ) possessing a single maximum, but also to discontinuous/nonsmooth or discrete approximations of such functions, provided that the rootmean-square error of approximation is relatively small (loosely speaking: smaller than the 'elevation of the maximum'). To start with, we define five equally spaced values of the independent variable x ∈ R determined by the choice for the smallest and largest values: x i,j , i = 1 , . . . , 5 , x 2 3 4 / / ,j = ( x 1 1 3 / / ,j + x 3 5 5 / / ,j ) / 2 , j being the iteration variable. (A rounding can be applied if integer values of x i,j are accepted only.) Initially we set x 1 0 , = 6 and x 5 0 , = 200 . Then in each iteration, j = 1 2 , , . . . , we check the following cases:
$$\mathfrak { e } \quad & \text{ Case 1} \ \max _ { i } [ f ( x _ { i, j } ) ] = f ( x _ { 3, j } ) \to \\ \vdots \quad & x _ { 1, j + 1 } = x _ { 2, j }, \ x _ { 5, j + 1 } = x _ { 4, j }$$
$$\text{kil} \\ \text{in} \quad \begin{array} { c c c } \text{Case 2} \ f ( x _ { 3, j } ) & > & f ( x _ { 2, j } ) & \& \ f ( x _ { 4, j } ) & > \\ f ( x _ { 3, j } ) \to \\ x _ { 1, j + 1 } = x _ { 3, j }, & x _ { 3, j + 1 } = x _ { 4, j } \end{array}$$
$$\xi, \, \text{it} \quad \text{ Case 3} \ f ( x _ { 3, j } ) \ \ < \ \ f ( x _ { 2, j } ) \ \ & \ f ( x _ { 4, j } ) \ \ < \\ \text{in-} \quad \ \text{$x_{5,j+1}=x_{3,j}$, $x_{3,j+1}=x_{2,j}$}$$
$$\text{Case 4 otherwise} \rightarrow$$
$$\max [ f ( x ) ] \approx x _ { 3, j }$$
Case 4 is never encountered in case of a smooth function featuring a single maximum, and the iteration would go on indefinitely without a stopping condition. Considering discontinuous/nonsmooth or discrete approximants of such functions the iteration is terminated in finite time ( j ).
## B. Finite-time Lyapunov exponent
A well-known measure of predictability is the positive maximal Lyapunov exponent (MLE), which approximates the average rate of the exponential separation of very close trajectories on a chaotic attractor [6]. Sterk et al. [15] evaluated the finite-time version of this measure to assess the predictability, with some lead time T , of extreme events. We consider here 'apriori' known nonautonomous dynamical systems ˙ y = f ( y, t ) in a d -dimensional phase space with generic initial condition y 0 = y y , t ( 0 = t 0 , t 0 ) ∈ R d , where y ( · , · , · ) denotes the two-time evolution operator.
The spectrum of finite-time T Lyapunov exponents (FTLE) λ ( T ) i can be defined in a pullback sense [27] as follows:
$$& \lambda _ { i } ^ { ( T ) } = \lim _ { t _ { 0 } \to - \infty } \frac { \ln ( s _ { i } ^ { 1 / 2 } ( t, t _ { 0 } ) ) - \ln ( s _ { i } ^ { 1 / 2 } ( t - T, t _ { 0 } ) ) } { T }, \\ & \colon$$
i = 1 , . . . , d , where s i ( t, t 0 ) : det( Y ( t, t 0 ) · Y T ( t, t 0 ) -sI ) = 0 are the singular values of the deformation gradient, Y = ∂y/∂y 0 , governed by the variational equation:
$$\dot { Y } = \frac { \partial f } { \partial y } \Big | _ { y ( y _ { 0 }, t, t _ { 0 } ) } \cdot Y. \quad \quad ( 1 2 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
The initial condition is not arbitrary but implied as Y ( y , t 0 0 , t 0 ) = I . Note that the LEs are recovered as: λ i = lim T →∞ λ ( T ) i . We will omit the index i to denote the MLE simply by λ or λ ( T ) . Clearly, by λ ( T ) ( ) t the predictability from the present time t -T of a trajectory at the future time t is defined. Note that by the inversion t = ( t y, t 0 , y 0 ) we have λ ( T ) ( y, t y , 0 ) .
A summary statistics for this measure of predictability can be defined, generalizing the proposal of Sterk et al. [15], by an ensemble average over parts of the pullback or snapshot attractor [24] that realize extreme events in terms of some physical observable x y ( ) :
$$\langle \lambda ^ { ( T ) } ( t ) \rangle = & \int \mu ( y, t ) d V _ { y } \lambda ^ { ( T ) } ( y, t ) \times \quad ( 1 3 ) \quad \quad \quad [ 3 ] { 3 } \\ & \mathcal { H } ( x ( y ) - x _ { * } ), \quad \quad \quad$$
where µ y, t ( ) is the natural measure supported by the snapshot attractor [25]. Alternatively, the average can be taken over parts of the snapshot attractor collecting the ensemble of trajectories that would cross the threshold in a leading window of time of width ∆ T at time t (TEI events):
$$\frac { t _ { 0 } ) ) } { }, \quad \langle \lambda ^ { ( T ) } ( t ) \rangle = & \int \mu ( y, t ) d V _ { y } \lambda ^ { ( T ) } ( y, t ) \times \\ \dots & \quad \mathcal { H } ( x ( y ) - x _ { * } ) \times \\ \dots$$
where y 0 ( y, t, t 0 ) is obtained by the inversion of the two-time evolution operator. The average FTLE or FTMLE 〈 λ ( T ) 〉 is dissimilar to D in that it is not calculated from predicted data, but rather it expresses an intrinsic property of the system that determines predictability. Nevertheless, we will compare figures obtained for 〈 λ ( T ) 〉 by (14) and D , at least in case of the autonomous dynamics when the attractor is timeinvariant.
## References
- [1] M. Ghil, P. Yiou, S. Hallegatte, B. D. Malamud, P. Naveau, A. Soloviev, P. Friederichs, V. Keilis-Borok, D. Kondrashov, V. Kossobokov, O. Mestre, C. Nicolis, H. W. Rust, P. Shebalin, M. Vrac, A. Witt, I. Zaliapin, Extreme events: dynamics, statistics and prediction, Nonlinear Processes in Geophysics 18 (3) (2011) 295-350. doi:10.5194/npg-18-295-2011. URL http://www. nonlin-processes-geophys.net/ 18/295/2011/
- [2] R. A. Fisher, L. H. C. Tippett, Limiting forms of the frequency distribution of the largest and smallest member of a sample, Proc. Cambridge Phil. Soc. 24 (1928) 180-190.
- [3] B. V. Gnedenko, Sur la distribution limite du terme maximum d'une serie aleatoire, Annals of Mathematics 44 (1943) 423-453.
- [4] M. R. Leadbetter, G. Lindgren, H. Rootz´ en, Extremes and Related Properties of Random Sequences and Processes, Springer-Verlag, 1983.
- [5] S. Coles, An Introduction to Statistical Modeling of Extreme Values, Springer, 2001.
- [6] T. T´ el, M. Gruiz, Chaotic Dynamics, Cambridge University Press, Cambridge, 2006.
- [7] M. I. Bogachev, A. Bunde, On the predictability of extreme events in records with linear and nonlinear long-range memory: Efficiency and noise
robustness, Physica A: Statistical Mechanics and its Applications 390 (12) (2011) 2240 - 2250. doi:http://dx.doi.org/10.1016/j.physa.2011.02.024. URL http://www.sciencedirect. com/science/article/pii/
S0378437111001506
- [8] S. Hallerberg, H. Kantz, How does the quality of a prediction depend on the magnitude of the events under study?, Nonlinear Processes in Geophysics 15 (2) (2008) 321-331. doi:10.5194/npg-15-321-
- 2008. URL http://www. nonlin-processes-geophys.net/ 15/321/2008/
- [9] J. M. Miotto, E. G. Altmann, , PLoS ONE 9 (11) (2014) e111506. doi:10.1371/journal.pone.0111506.
- URL http://dx.doi.org/10.1371% 2Fjournal.pone.0111506
- [10] S. Hallerberg, J. Br¨ ocker, H. Kantz, Prediction of extreme events, in: R. V. Donner, S. M. Barbosa (Eds.), Nonlinear Time Series Analysis in the Geosciences, Vol. 112, Springer, 2008.
- [11] J. P. Egan, Signal Detection Theory and ROC Analysis, Academic Press, 1975.
- [12] H. Kantz, Dynamics and statistic of extreme events, in: E. Estrada, M. Fox, D. Higham, G.-L. Oppo (Eds.), Network Science, Springer-Verlag, London, 2010.
- [13] J. Br¨cker, L. A. Smith, Scoring probabilistic foreo casts: the importance of being proper, Wea. Forecasting 22 (2007) 382-388.
- [14] C. Franzke, Predictability of extreme events in a nonlinear stochastic-dynamical model, Phys. Rev. E 85 (2012) 031134. doi:10.1103/PhysRevE.85.031134.
- URL http://link.aps.org/doi/10. 1103/PhysRevE.85.031134
- [15] A. E. Sterk, M. P. Holland, P. Rabassa, H. W. Broer, R. Vitolo, Predictability of extreme values in geophysical models, Nonlinear Processes in Geophysics 19 (5) (2012) 529-539. doi:10.5194/npg-19-529-2012.
- URL http://www. nonlin-processes-geophys.net/ 19/529/2012/
- [16] A. J. Majda, C. Franzke, D. Crommelin, Normal forms for reduced stochastic climate models, Proceedings of the National Academy of Sciences 106 (10) (2009) 3649-3653.
doi:10.1073/pnas.0900173106.
URL http://www.pnas.org/content/ 106/10/3649.abstract
- [17] E. N. Lorenz, Irregularity: a fundamental property of the atmosphere, Tellus 36A (1984) 98-110.
- [18] S. Siegert, J. Br¨ ocker, H. Kantz, Skill of data based predictions versus dynamical models -case study on extreme temperature anomalies, arXiv:1312.4323.
- [19] S. Hallerberg, H. Kantz, Influence of the event magnitude on the predictability of an extreme event, Phys. Rev. E 77 (2008) 011108. doi:10.1103/PhysRevE.77.011108.
URL http://link.aps.org/doi/10. 1103/PhysRevE.77.011108
- [20] L. van Veen, Baroclinic flow and the lorenz84 model, International Journal of Bifurcation and Chaos 13 (08) (2003) 2117-2139. doi:10.1142/S0218127403007904.
- [21]
- A. Provenzale, N. J. Balmford, Chaos and structures in geophysics and astrophysics (Woods Hole lecture notes) (1999). URL http://www.whoi.edu/ fileserver.do?id=21476&pt=10&p= 17353
- [22] C. Masoller, A. S. Schifino, L. Romanelli, Characterization of strange attractors of lorenz model of general circulation of the atmosphere, Chaos, Solitons & Fractals 6 (0) (1995) 357 -366, complex Systems in Computational Physics. doi:http://dx.doi.org/10.1016/09600779(95)80041-E.
- URL http://www.sciencedirect. com/science/article/pii/ 096007799580041E
- [23] C. Nicolis, S. Vannitsem, J.-F. Royer, Short-range predictability of the atmosphere: Mechanisms for superexponential error growth, Quarterly Journal of the Royal Meteorological Society 121 (523) (1995) 705-722. doi:10.1002/qj.49712152312. URL http://dx.doi.org/10.1002/qj. 49712152312
- [24] T. B´ odai, G. K´ arolyi, T. T´ el, Driving a conceptual model climate by different processes: Snapshot attractors and extreme events, Phys. Rev. E 87 (2013) 022822. doi:10.1103/PhysRevE.87.022822.
URL http://link.aps.org/doi/10. 1103/PhysRevE.87.022822
[25]
G. Dr´ otos, T. B´dai, T. T´l, Probabilistic concepts
o
e
arXiv:http://www.pnas.org/content/106/10/3649.full.pdf+html, n a changing climate:
i
a snapshot attractor pic-
ture, Accepted for publication in the Journal of Climate.
[26] F. Takens, Detecting strange attractors in turbulence, in: A. Rand, L.-S. Young (Eds.), Dynamical Systems and Turbulence, Vol. 898, SpringerVerlag, 2008, pp. 366-381.
[27] T. B´ odai, E. G. Altmann, A. Endler, Stochastic perturbations in open chaotic systems: Random versus noisy maps, Phys. Rev. E 87 (2013) 042902. doi:10.1103/PhysRevE.87.042902.
URL http://link.aps.org/doi/10. 1103/PhysRevE.87.042902
|
10.1016/j.physd.2015.08.007
|
[
"Tamas Bodai"
] |
2014-08-05T15:24:12+00:00
|
2015-09-07T12:52:56+00:00
|
[
"nlin.CD",
"physics.ao-ph"
] |
Predictability of threshold exceedances in dynamical systems
|
In a low-order model of the general circulation of the atmosphere we examine
the predictability of threshold exceedance events of certain observables. The
likelihood of such binary events -- the cornerstone also for the categoric (as
opposed to probabilistic) prediction of threshold exceedences -- is established
from long time series of one or more observables of the same system. The
prediction skill is measured by a summary index of the ROC curve that relates
the hit- and false alarm rates. Our results for the examined systems suggest
that exceedances of higher thresholds are more predictable; or in other words:
rare large magnitude, i.e., extreme, events are more predictable than frequent
typical events. We find this to hold provided that the bin size for binning
time series data is optimized, but not necessarily otherwise. This can be
viewed as a confirmation of a counterintuitive (and seemingly contrafactual)
statement that was previously formulated for more simple autoregressive
stochastic processes. However, we argue that for dynamical systems in general
it may be typical only, but not universally true. We argue that when there is a
sufficient amount of data depending on the precision of observation, the skill
of a class of data-driven categoric predictions of threshold exceedences
approximates the skill of the analogous model-driven prediction, assuming
strictly no model errors. Furthermore, we show that a quantity commonly
regarded as a measure of predictability, the finite-time maximal Lyapunov
exponent, does not correspond directly to the ROC-based measure of prediction
skill when they are viewed as functions of the prediction lead time and the
threshold level. This points to the fact that even if the Lyapunov exponent as
an intrinsic property of the system, measuring the instability of trajectories,
determines predictability, it does that in a nontrivial manner.
|
1408.1002v1
|
## Signalling entropy: a novel network-theoretical framework for systems analysis and interpretation of functional omic data
Andrew E. Teschendorff a,b, ∗ , Peter Sollich , Reimer Kuehn c c a CAS-MPG Partner Institute for Computational Biology, Chinese Academy of Sciences, Shanghai Institute for Biological Sciences, 320 Yue Yang Road, Shanghai 200031, China. b Statistical Cancer Genomics, Paul O'Gorman Building, UCL Cancer Institute, University College London, London WC1E 6BT, UK. c Department of Mathematics, King's College London, London WC2R 2LS, UK.
## Abstract
A key challenge in systems biology is the elucidation of the underlying principles, or fundamental laws, which determine the cellular phenotype. Understanding how these fundamental principles are altered in diseases like cancer is important for translating basic scientific knowledge into clinical advances. While significant progress is being made, with the identification of novel drug targets and treatments by means of systems biological methods, our fundamental systems level understanding of why certain treatments succeed and others fail is still lacking. We here advocate a novel methodological framework for systems analysis and interpretation of molecular omic data, which is based on statistical mechanical principles. Specifically, we propose the notion of cellular signalling entropy (or uncertainty), as a novel means of analysing and interpreting omic data, and more fundamentally, as a means of elucidating systems-level principles underlying basic biology and disease. We describe the power of signalling entropy to discriminate cells according to differentiation potential and cancer status. We further argue the case for an empirical cellular entropy-robustness correlation theorem and demonstrate its existence in cancer cell line drug sensitivity data. Specifically, we find that high signalling entropy correlates with drug resistance and further describe how entropy could be used to identify the achilles heels of cancer cells. In summary, signalling entropy is a deep and powerful concept, based on
∗ [email protected]
August 6, 2014
rigorous statistical mechanical principles, which, with improved data quality and coverage, will allow a much deeper understanding of the systems biological principles underlying normal and disease physiology.
Keywords: entropy; network; signalling; genomics; drug resistance; cancer; differentiation; stem cell
## 1. Introduction
Recent advances in biotechnology are allowing us to measure cellular properties at an unprecedented level of detail [1]. For instance, it is now possible to routinely measure various molecular entities (e.g. DNA methylation, mRNA and protein expression, SNPs) genome-wide in hundreds if not thousands of cellular specimens [2]. In addition, other molecular data detailing interactions between proteins or between transcription factors and regulatory DNA elements are growing at a rapid pace [1]. All these types of data are now widely referred to collectively as 'omic' data. The complexity and high-dimensional nature of this omic data presents a daunting challenge to those wishing to analyse and interpret the data [1]. The difficulty of analysing omic data is further compounded by the inherent complexity of cellular systems. Cells are prime examples of organized complex systems, capable of highly stable and predictable behaviour, yet an understanding of how this deterministic behaviour emerges from what is a highly complex and probabilistic pattern of dynamic interactions between numerous intra and extracellular components, still eludes us [1]. Thus, elucidating the systemsbiological laws or principles dictating cellular phenotypes is also key for an improved analysis and interpretation of omic data. Furthermore, important biological phenomena such as cellular differentiation are fundamentally altered in diseases like cancer [3]. Hence, an attempt to understand how cellular properties emerge at a systems level from the properties seen at the individual gene level is not only a key endeavour for the systems biology community, but also for those wanting to translate basic insights into effective medical advances [4, 5].
It is now well accepted that at a fundamental level most biological systems are best modeled in terms of spatial interactions between specific entities (e.g. neurons in the case of the brain), which may or may not be dynamically changing in time [6, 7]. It therefore seems natural to also use the mathematical and physical framework of networks to help us analyse and in-
terpret omic data at a systems level [8, 9]. Indeed, the cellular phenotype is determined to a large extent by the precise pattern of molecular interactions taking place in the cell, i.e. a molecular interaction network [10]. Although this network has spatial and dynamic dimensions which at present remain largely unexplored due to technological or logistical limitations, there is already a growing number of examples where network-based analysis strategies have been instrumental [11, 12, 13]. For instance, a deeper understanding of why sustained treatment with EGFR inhibitors can lead to dramatic sensitization of cancer cell lines to cytotoxic agents was possible thanks to a systems approach [13]. Another study used reverse engineering network approaches to identify and validate drug targets in glioblastoma multiforme, to be further tested in clinical trials [14]. What is key to appreciate here is that these successes have been achieved in spite of noisy and incomplete data, suggesting that there are simple yet deep systems biological principles underlying cellular biology that we can already probe and exploit with current technology and data. Thus, with future improvements in data quality and coverage, network-based analysis frameworks will play an ever increasing and important role in systems biology, specially at the level of systems analysis and interpretation [4]. Therefore, it is also imperative to develop novel, more powerful, network-theoretical methods for the systems analysis of omic data. In adopting a network's perspective for the analysis and interpretation of omic data, there are in principle two different (but not mutually exclusive) approaches one can take. One possibility is to infer (i.e. reverse engineer) the networks from genome-wide data [15]. Most of these applications have done this in the context of gene expression data, with the earliest approaches using clustering or co-expression to postulate gene interdependencies [16]. Partial correlations and Graphical Gaussian Models have proved useful as a means of further refining correlation networks by allowing one to infer the more likely direct interactions while simultaneously also filtering out those which are more likely to be indirect [17]. These methods remain popular and continue to be studied and improved upon [18, 19]. Other methods have drawn on advanced concepts from information theory, for instance ARACNe ('Algorithm for the Reconstruction of Accurate Cellular Networks') has been shown to be successful in infering regulatory networks in B-cells [15].
In stark contrast to reverse engineering methods, another class of algorithms have used structural biological networks from the outset, using these as scaffolds to integrate with omic data. Specifically, by using a structural network one can sparsify the correlation networks inferred from reverse-engineering
approaches, thus providing another means of filtering out correlations that are more likely to be indirect [9]. Besides, integration with a structural network automatically provides an improved framework for biological interpretation [10, 20, 21, 22]. The structural networks themselves are typically derived from large databases, which detail literature curated experimentally verified interactions, including interactions derived from Yeast 2 Hybrid screens (Y2H) [23]. The main example is that of protein protein interaction (PPI) maps, which have been generated using a number of different complementary experimental and in-silico approaches, and merging these maps from these different sources together has been shown to be an effective approach in generating more comprehensive high-confidence interaction networks [24, 25]. PPI networks have been used mainly as a means of integrating and analysing gene expression data (see e.g. [20, 26, 27]). More recently, this approach has also been successfully applied in the DNA methylation context, for instance it has been shown that epigenetic changes associated with age often target specific gene modules and signalling pathways [28].
Another class of methods that have used structural networks, PPIs in particular, have integrated them with gene expression data to define an approximation to the underlying signaling dynamics on the network, thus allowing more in-depth exploration of the interplay between network topology and gene expression. Typically, these studies have used the notion of random walks on weighted graphs where the weights are constructed from differential expression statistics, and where the aim is to identify nodes (genes) in the network which may be important in dictating the signaling flows within the pathological state. For instance, among these random walk methods is NetRank, a modification of the Google PageRank algorithm, which was able to identify novel, robust, network based biomarkers for survival time in various cancers [29, 30]. Other random walk based approaches, aimed at identifying causal drivers of specific phenotypes (e.g. expression or cancer), have modeled signal transduction between genes in the causal and phenotypic layers as flows in an electric circuit diagram, an elegant formulation capable of not only identifying the likely causal genes but also of tracing the key pathways of information flow or dysregulation [31, 32]. Random walk theory has also been employed in the development of differential network methodologies. An example is NetWalk [33], which is similar to NetRank but allows differential signaling fluxes to be inferred. This approach was successful in identifying and validating the glucose metabolic pathway as a key determinant of lapatinib resistance in ERBB2 positive breast cancer patients [34].
Another important concept to have emerged recently is that of network rewiring [35, 36, 37]. This refers to the changes in interaction patterns that accompany changes in the cellular phenotype. Network rewiring embodies the concept that it is the changes in the interaction patterns, and not just the changes in absolute gene expression or protein activity, that are the main determinants of the cellular phenotype. That network rewiring may be key to understanding cellular phenotypes was most convincingly demonstrated in a differential epistasis mapping study conducted in yeast cells exposed to a DNA damaging agent [35]. Specifically, what this study demonstrated is that responses to perturbations or cellular stresses are best understood in terms of the specific rewiring of protein complexes and functional modules. Thus, this conceptual framework of network rewiring may apply equally well to the genetic perturbations and cellular stresses underlying disease pathologies like cancer.
In this article we advocate a network-theoretical framework based on statistical mechanical principles and more specifically on the notion of signalling entropy [8, 9]. This theoretical framework integrates gene expression (but in principle also other functional data) with a PPI network, merging existing concepts such as signaling dynamics (i.e. random walks) and network rewiring with that of signalling entropy. In previous work we have shown how signalling entropy (i) provides a proxy to the elevation in Waddington's epigenetic landscape, correlating with a cell's differentiation potential [38], (ii) how it can be used to identify signaling pathways and nodes important in differentiation and cancer [38, 8, 9], and (iii) how it predicts two cancer system-omic hallmarks: (a) cancer is characterised by an increase in signalling entropy and (b) local signaling entropy changes anti-correlate with differential gene expression [8, 9]. Here, we present and unify the different signaling entropy measures used previously and further explore a novel application of signalling entropy to understanding drug sensitivity profiles in cancer cell lines. Specifically, we first use simulated data to justify the existence of an entropy-robustness theorem, and subsequently provide empirical evidence for this theorem by demonstrating that increases in local signalling entropy correlate with drug resistance (robustness). We further show the importance of network topology in dictating the signalling entropy changes underlying drug response. In addition, we provide R-functions implementing the entropy rate calculation and ranking of genes according to differential entropy, all freely available from sourceforge.net/projects/signalentropy/files/
.
## 2. Materials and Methods
## 2.1. Basic rationale and motivation for Signalling Entropy: understanding systems biology through uncertainty
Loosely defined, entropy of a system, refers to a measure of the disorder, randomness or uncertainty of processes underlying the system's state. In the context of a single cell, signalling entropy will reflect the amount of overall disorder, randomness or uncertainty in how information, i.e. signaling, is passed on in the molecular interaction network. At a fundamental level, all signaling in a cell is probabilistic, determined in part by the relative cellular concentrations of the interacting molecules. Hence, in discussing cellular signalling entropy, it is useful to picture a molecular interaction network in which edges represent possible interactions and with the edge weights reflecting the relative probabilities of interaction ( Fig.1A ). Thus, an interaction network in a high-entropy state is characterised by signaling interaction probabilities that are all fairly similar in value, whereas a low-entropy state will be characterised by specific signalling interactions possessing much higher weights ( Fig.1A ).
Why would this type of signalling entropy, loosely defined as the amount of uncertainty in the signaling interaction patterns, be useful to systems biology? One way to think of signalling entropy is as representing signaling promiscuity, which has been proposed as a key systems feature underlying the pluripotent or multipotent capacity of cells ( Fig.1B ) [39, 40, 41, 42]. Indeed, it has been suggested that pluripotency is an intrinsic statistical mechanical property, best defined at the cellular population level [40]. Specifically, it has been demonstrated that pluripotent stem cells exhibit remarkable heterogeneity in gene expression levels, including well known stem cell markers such as NANOG [40]. It is also well known that a large number of genes, many encoding transcription factors, exhibit low-levels of expression in stem cells, yet simultaneously are being kept in a poised chromatin state, allowing immediate activation if this were required [43]. Thus, in a pluripotent stem cell like state, signal transduction is in a highly egalitarian and, thus, promiscuous state, i.e. a state of high signalling entropy. Conversely, differentiation leads, by necessity, to activation of specific transcription factors and pathways and thus to a lowering in the uncertainty of signaling patterns, and thus to a lowering of entropy. We recently demonstrated, using gene expression profiles of over 800 samples, comprising cells of all major stages of differentiation, including human embryonic stem cells (hESCs), induced pluripotent
stem cells (iPSCs), multipotent cell types (e.g. hematopoietic stem cells (HSCs)), and terminally differentiated cells within these respective lineages, that signalling entropy not only correlates with differentiation potential but that it provides a highly quantitative measure of potency [38]. Indeed, we showed that signalling entropy provides a reasonably good approximation to the energy potential in Waddington's epigenetic landscape [38].
Here we decided to explore the concept of signalling entropy in relation to cellular robustness and specifically to drug resistance in cancer. That signalling entropy may be informative of a cell's robustness is a proposal that stems from a general (but unproven) theorem, first proposed by Manke and Demetrius [44, 45, 46]: namely, that a system's entropy and robustness are correlated. Mathematically, this can be expressed as ∆ S ∆ R > 0, which states that a positive change in a system's entropy (i.e. ∆ S > 0) must lead to an increase in robustness (∆ R > 0). Now, cells are fairly robust entities, having evolved the capacity to buffer the intra-and-extracellular stresses and noise which they are constantly exposed to [47, 39]. Much of this overall stability and robustness likely stems from the topological features of the underlying signaling and regulatory networks, for instance features such as scale-freeness and hierarchical modularity, which are thought to have emerged through natural evolutionary processes such as gene duplication [10, 48, 49]. However, another key feature which contributes to cellular robustness is cross-talk and signalling pathway redundancy [47]. Pathway redundancy refers to a situation where a cell has the choice of transmitting signals via two or more possible routes. In the language of statistical mechanics, this corresponds to a state of high uncertainty (or entropy) in signaling. High signalling entropy could thus underpin a cell's robustness to perturbations, suggesting that a cell's entropy and robustness may indeed be correlated ( Fig.2A ). Consistent with this, pathway redundancy is also well recognized to be a key feature underlying drug resistance of cancer cells [50].
Further supporting the notion that entropy and robustness may be correlated, we previously showed that (i) cancer is characterised by a global increase in signalling entropy compared to its respective normal tissue [9], in line with the observation that cancer cells are specially robust to generic perturbations, and (ii) that a gene's differential entropy between normal and cancer tissue is generally speaking anticorrelated with its differential expression [9], consistent with the view that cancer cells are specially sensitive to drug interventions that target overexpressed oncogenes, a phenomenon known as oncogene addiction ( Fig.2B ) [3]. Interestingly, the observations
that differential entropy and differential expression are anti-correlated and that cancer is characterised globally by an increase in entropy [9], are also consistent with the prevalent view that most driver mutations are indeed inactivating, targeting tumor suppressor genes [51, 52, 53]. Hence, based on all of these observations and insights, we posited that signalling entropy could also prove useful as a means of predicting drug resistance in cancer cells.
## 2.2. The Signalling Entropy Method: definitions and construction
Briefly, we review the definitions and construction of signalling entropy as used in our previous studies [8, 9, 38]. The construction relies on a comprehensive and high-confidence PPI network which is integrated with gene expression data [24, 9] (see Appendix A ). Briefly, the PPI is used as a scaffold, and edge weights are constructed from the gene expression data to approximate the interaction or signaling probabilities between the corresponding proteins in the PPI. Thus, the gene expression data provides the biological context in which to modulate the PPI interaction probabilities. To compute signalling entropy requires the estimation of a stochastic matrix, reflecting these interaction probablities over the network.
The construction of the stochastic matrix can proceed in two different ways. In the earliest studies we used a construction which was done at the level of phenotypes [8, 9]. Under this model, edge weights w ij between proteins i and j were constructed from the correlation between the expression levels of the corresponding genes i and j , as assessed over independent samples all representing the same phenotype. Estimating the correlations over independent samples, all within the same phenotype, can be viewed as representing a multifactorial perturbation experiment, with e.g. genetic differences between individuals mimicking specific perturbations, and thus allowing influences to be inferred. Thus, this approach hinges on the assumption that highly correlated genes, whose coding proteins interact, are more likely to be interacting in the given phenotype than two genes which do not correlate. The use of a PPI as a scaffold is important to filter out significant correlations which only result from indirect influences. The correlations themselves can be defined in many different ways, for instance, using simple Pearson correlations or nonlinear measures such as Mutual Information. For example, one way to define the weights is as w ij = 1 2 (1 + c ij ) with c ij describing the Pearson correlation coefficient between genes i and j . This definition guarantees positivity, but also treats positive and negative correlations differently, which makes biological sense because activating and inhibitory interactions normally have
completely distinct consequences on downstream signalling. Thus, the above weight scheme treats zero or insignificant correlations as intermediate, consistent with the view that an absent interaction is neither inhibitory nor activating. However, other choices of weights are possible: e.g. w ij = | c ij | , which treats negative and positive correlations on an equal footing. Once edge weights are defined as above, these are then normalised to define the stochastic matrix p ij over the network,
$$p _ { i j } = \frac { w _ { i j } } { \sum _ { k \in \mathcal { N } _ { i } } w _ { i k } },$$
glyph[negationslash]
with N i denoting the PPI neighbors of gene i . Thus, p ij is the probability of interaction between genes i and j , and as required, ∑ j p ij = 1. However, there is no requirement for p ij to be doubly stochastic, i.e. P is in general not a symmetric matrix. Hence, edges are bi-directional with different weights labeling the probability of signal transduction from i to j , and that from j to i ( p ij = p ji ).
An alternative to the above construction of the stochastic matrix is to invoke the mass action principle, i.e. one now assumes that the probability of interaction in a given sample is proportional to the product of expression values of the corresponding genes in that sample [38]. Thus, the PPI is again used as a scaffold to only allow interactions supported by the PPI network, but the weights are defined using the mass action principle, as
$$w _ { i j } \in E _ { i } E _ { j }$$
where E i denotes the normalised expression intensity value (or normalised RNA-Seq read count) of gene i . An important advantage of this construction is that the stochastic matrix is now sample specific, as the expression values are unique to each sample.
Given a stochastic matrix, p ij , constructed using one of the two methods above, one can now define a local Shannon entropy for each gene i as
$$\tilde { S } _ { i } = - \frac { 1 } { \log k _ { i } } \sum _ { k \in \mathcal { N } _ { i } } p _ { i k } \log p _ { i k },$$
where k i denotes the degree of gene i in the PPI network. The normalisation is optional but ensures that this local Shannon entropy is normalised between 0 and 1. Clearly, if only one weight is non-zero, then the entropy
attains its minimal value (0), representing a state of determinism or lowest uncertainty. Conversely, if all edges emanating from i carry the same weight, the entropy is maximal (1), representing the case of highly promiscuous signaling. In principle, local Shannon entropies can thus be compared between phenotypes to identify genes where there are changes in the uncertainty of signaling. In the case where entropies are estimated at the phenotype level, jackknife approaches can be used to derive corresponding differential entropy statistics [9]. Deriving statistics is important because node degree has a dramatic influence on the entropy variance, with high degree nodes exhibiting significantly lower variability in absolute terms, which nevertheless could be highly significant [9]. In the case where entropies are estimated at the level of individual samples, ordinary statistical tests (e.g. rank sum tests) can be used to derive sensible P-values, assuming of course that enough samples exist within the phenotypes being compared.
In addition to the local entropy, it is also of interest to consider statistical properties of the distribution of local entropies, for instance their average. Comparing the average of local entropies between two phenotypes would correspond to a comparison of non-equilibrium entropy rates. To see this, consider the formal definition of the entropy rate SR [54, 55], i.e.
$$S R = \sum _ { i = 1 } ^ { n } \pi _ { i } S _ { i },$$
where π i is the stationary distribution (equivalently the left eigenvector with unit eigenvalue) of P (i.e. πP = π ), and where now
$$S _ { i } = - \sum _ { k \in \mathcal { N } _ { i } } p _ { i k } \log p _ { i k }.$$
Note that the entropy rate SR is an equilibrium entropy since it involves the stationary distribution of the random walker. As such, the entropy rate also depends on the global topology of the network. Thus, the entropy rate is a weighted average of the local unnormalized entropies, with the weights specified by the stationary distribution. It follows that comparing the unweighted averages of local entropies reflects a comparison of a non-equilibrium entropy rate since the stationary distribution is never used.
## 2.3. The importance of the integrated weighted network in estimating signalling entropy
The entropy rate constructed using the mass action principle is sample specific. We previously demonstrated that this entropy rate was highly discriminative of the differentiation potential of samples within a developmental lineage, as well as being highly discriminative of normal and cancer tissue [38] ( Fig.3A ). Since the entropy rate takes as input a PPI network and a sample's genome-wide expression profile, the significance of the resulting entropy values, as well as that of the difference between entropy rates, also needs to be assessed relative to a null distribution in which the putative information content between network and gene expression values is non-existent. In fact, since the weights in the network determine the entropy rate and these weights are dependent on both the specific network nodes and their respective gene expression profiles, it is natural to assess the significance of the entropy rate by 'destroying' the mutual information between the network nodes and their gene expression profiles, for instance by randomising (i.e. permuting) the gene expression profiles over the network. Thus, under this randomisation, the topological properties of the network remain fixed, but the weights are redefined. Application of this randomisation procedure to the normal/cancer expression set considered previously [38] ( Fig.3A ) shows that the discriminatory potential is significantly reduced upon permuting the gene expression values over the network ( Fig.3B-C ). Importantly, we observe that the entropy rate is much higher in the normal and cancer states compared to the rates obtained upon randomisation of the gene expression profiles ( Fig.3B ), indicating that both normal and cancer states are characterised by a higher level of signaling promiscuity compared to a network with random weights. That the discrimination between normal and cancer is significantly reduced in the randomly weighted network further demonstrates that there is substantial mutual information between the PPI network and the gene expression profiles, thus justifying the signalling entropy approach.
## 2.4. Signalling entropy R-package: input, output and code availability
A vignette/manual and user-friendly R-scripts that allow computation of the entropy rate is available at the following url: sourceforge.net/projects/signalentropy/files/ . Here we briefly describe the salient aspects of this package:
The input: The main R-script provided ( CompSR ) takes as input a userspecified PPI network, and a genome-wide expression vector representing
the gene expression profile of a sample. It is assumed that this has been generated using either Affy, Illumina or RNA-Sequencing. In principle one ought to use as gene expression value the normalised unlogged intensity (Affy/Illu) or RNA-seq count, since this is what should be proportional to the number of RNA transcripts in the sample. However, in practice we advise taking the log-transformed normalised value since the log-transformation provides a better compromise between proportionality and regularisation, i.e. some regularisation is advisable since the underlying kinetic reactions are also regular.
The output: The R-functions provided in the package then allow the user to estimate the global entropy rate for a given sample, as well as the local normalised entropies for each gene in the integrated network. If a phenotype is specified then genes can be ranked according to the correlation strength of their local entropy to the phenotype. Thus, the signalling entropy method allows us to assess (i) if the overall levels of signalling promiscuity is different between phenotypes, which could be important, for instance, to compare the pluripotent capacity of iPSCs generated via different protocols or to assess the stem cell nature of putative cancer stem cells [38], and (ii) to rank genes according to differential entropy between phenotypes, allowing key signalling genes associated with differentiation, metastasis or cancer to be identified [8, 9, 38].
## 3. Results
## 3.1. Signalling entropy and cellular robustness
Our previous observation that signalling entropy is increased in cancer [9], and that one of the key characteristics of cancer cells is their robustness to intervention and environmental stresses, suggested to us that high cellular signalling entropy may be a defining feature of a cell that is robust to general perturbations. In addition, cancer cells often exhibit the phenomenon of oncogene addiction, whereby they become overly reliant on the activation of a specific oncogenic pathway, rendering them less robust to targeted intervention [3]. Since oncogenes are normally overexpressed in cancer, it follows by the second cancer system-omic hallmark [9], that their lower signalling entropy may underpin their increased sensitivity to targeted drugs ( Fig.2 ). Based on these insights, we posited that cellular signalling entropy may be correlated to the overall cellular system's robustness to perturbations.
In order to explore the relation between signalling entropy and robustness
in a general context, it is useful to consider another network property of the stochastic matrix, namely the global mixing rate. This mixing rate is defined formally as the inverse of the number of timesteps for which one has to evolve a random walk on the graph so that its probability distribution is close to the stationary distribution, independently of the starting node ( Appendix B ). This is a notion that is closer to that of robustness or resilience as defined by Demetrius and Manke [44, 45, 46], allowing the mixing rate to be viewed as a crude proxy of a system's overall robustness to generic perturbations. Furthermore, the global mixing rate can be easily estimated as
$$\mu _ { R } = - \log S L E M$$
where SLEM is the second largest (right) eigenvalue modulus of the stochastic matrix ( Appendix B ). Thus, we first asked how this global mixing rate is related to the signalling entropy rate.
For a regular network of degree d it can be easily shown that the entropy rate SR = log d , whilst results on graph theory also show that for sufficiently large regular graphs, µ R ∝ log d [56]. Hence, at least for regular networks a direct correlation exists. It follows that for non-regular graphs with tight degree distributions, e.g. Erd¨ os-Renyi (ER) graphs, the entropy and mixing rates should also be approximately correlated. Indeed, using an empirical computational approach to evaluate the entropy and mixing rates for ER graphs with variable entropy rates, we were able to confirm this correlation ( Appendix C, Fig.4A ). Next, we wanted to investigate if this relationship also holds for networks with more realistic topologies than ER graphs. Hence, we generated connected networks on the order of 500 nodes by random subsampling 1000 nodes from our large protein interaction network ( ∼ 8000 nodes) followed by extraction of the maximally connected component ( Appendix D ). We verified that these networks possessed approximate scale-free topologies with clustering coefficients which were also significantly higher than for ER graphs. As before, for each generated network, stochastic matrices of variable entropy rates were defined. Signalling entropy and mixing rates were then estimated for each of these networks, and subsequently averaged over an ensemble of such graphs. As with the random Poisson (ER) graphs, average mixing and entropy rates were highly correlated ( Fig.4B ).
Having demonstrated a direct correlation between these two measures on simulated data, we next asked if such a correlation could also be present in
the full protein interaction networks and with the stochastic matrices derived from real expression data. Thus, we computed the global entropy and mixing rates for 488 cancer cell lines from the Cancer Cell Line Encylopedia (CCLE) ( Appendix E ) [57]. Remarkably, despite the dynamic ranges of both entropy and mixing rates being much smaller ( Fig.4C ) compared to those of the artificially induced stochastic matrices (c.f Fig.4A-B ), we were still able to detect a statistically significant correlation between entropy and mixing rates, as estimated across the 488 cell lines ( Fig.4C ). Thus, all these results support the view that global entropy and mixing rates are correlated, albeit only in an average/ensemble sense.
## 3.2. Local signalling entropy predicts drug sensitivity
In the case of realistic expression and PPI data, the observed correlation between entropy and mixing rates was statistically significant but weak ( Fig.4C ). This could be easily attributed to the fact that in real biological networks, the global mixing rate is a very poor measure of cellular robustness. In fact, it is well known that cellular robustness is highly sensitive to which genes undergo the perturbation. For instance, in mice some genes are embryonically lethal, whereas others are not [10]. Robustness itself also admits many different definitions. Because of this, we decided to investigate signalling entropy in relation to other more objective measures of cellular robustness. One such measure is drug sensitivity, for instance, IC50 values, which measure the amount of drug dose required to inhibit cell proliferation by 50%. According to this measure, a cell that is insensitive to drug treatment is highly robust to that particular treatment. Since most drugs target specific genes, we decided to explore the relation, if any, between the local signalling entropy of drug targets and their associated drug sensitivity measures. Specifically, we hypothesized that since local entropy provides a proxy for local pathway redundancy, that it would correlate with drug resistance. To test this, we first computed for each of the 8038 genes in the PPI network its local signalling entropy in each of the 488 CCLE cancer cell-lines. To identify associations between the 24 drug sensitivity profiles and the 8038 local network entropies, we computed non-linear rank (Spearman) correlation coefficients across the 488 cancer cell-lines, resulting in 24 × 8038 correlations and associated P-values. We observed that there were many more significant associations than could be accounted for by random chance ( Fig.5A ), with the overall strength of association very similar to that seen between gene
expression and drug sensitivity ( Fig.5B ).
One would expect the targets of specific drugs to be highly informative of the sensitivity of the corresponding drugs. We used the CancerResource [58] to identify targets of the 24 drugs considered here, and found a total of 154 drug-target pairs. For 134 of these pairs we could compute a P-value of association between local entropy and drug sensitivity with 76 pairs (i.e. 57%) exhibiting a significant association ( Fig.5C ). This was similar to the proportion (54%) of observed significant associations between gene expression and drug sensitivity ( Fig.5D ). However, interestingly, only 42 of these significant drug-target pairs were in common between the 76 obtained using signalling entropy and the 72 obtained using gene expression. Importantly, while the significant associations between gene expression and drug sensitivity involved preferentially positive correlations, in the case of signalling entropy most of the significant correlations were negative ( Fig.5C-D ), exactly in line with our hypothesis that high entropy implies drug resistance. Thus, as expected, cell-lines with highly expressed drug targets were more sensitive to treatment by the corresponding drug, but equally, drug targets exhibiting a relatively low signalling entropy were also predictive of drug sensitivity.
To formally demonstrate that local signalling entropy adds predictive power over gene expression, we considered bi-variate regression models including both the target's gene expression as well as the target's signalling entropy. Using such bivariate models and likelihood ratio tests we found that in the majority of cases where signalling entropy was significantly associated with drug sensitivity that it did so independently of gene expression, adding predictive value ( Fig.5E ). Top ranked drug-target pairs where signalling entropy added most predictive value over gene expression included Topotecan/ TP53 and Paclitaxel/ MYC ( Fig.5F ).
To further demonstrate that the observed associations between local signalling entropy and drug sensitivity are statistically meaningful, we conducted a control test, in which we replaced in the multivariate model the signalling entropy of the target with a non-local entropy computed by randomly replacing the PPI neighbours of the target with other 'far-away' genes in the network. For a considerable fraction (41%) of drug-target pairs, the original multivariate models including the local entropy constituted better predictive models than those with the non-local entropy (false discovery rate < 0 25), indicating that the observed associations are driven, at least partly, . by the network structure.
## 3.3. High signalling entropy of intra-cellular signaling hubs is a hallmark of drug resistance
Among the drug-target pairs for which signalling entropy was a significant predictor of drug sensitivity, we observed a striking non-linear association with the topological degree of the targets in the network ( Fig.6A ). In fact, for hubs in the network, most of which encode nodes located in the intracellular signaling hierarchy, high signalling entropy was exclusively associated with drug resistance (negative SCC). Examples included well-known intracellular signalling hubs like HDAC1, HDAC2, AKT1, TP53, STAT3, MYC ). Some intracellular non-hubs (e.g. CASP9, BCL2L1, BIRC3 ) also exhibited negative correlations between signalling entropy and drug sensitivity. Among targets for which high signalling entropy predicted drug sensitivity, we observed several membrane receptors (e.g ERBB2, ERBB3, EGFR, MET ) and growth factors (e.g HBEGF, EGF, TGFA ). Given that the correlation coefficients were estimated across samples (cell-lines) and that the underlying network topology is unchanged between samples, the observed non-linear relation between the directionality of the correlation and node degree is a highly non-trivial finding. We also observed a clear dependence on the main signaling domain of the target, with intracellular hubs showing preferential anticorrelations, in contrast to growth factors and membrane receptors which exhibited positive and negative correlations in equal proportion ( Fig.6B ). Thus, we can conclude from these analyses that cancer associated changes to the interaction patterns of intra-cellular hubs are key determinants of drug resistance. In particular, in cancers where the local entropy at these hubs is increased, as a result of increased promiscuous signaling, drugs targeting these hubs are less likely to be effective.
## 4. Discussion and Conclusions
Here we have advocated a fairly novel methodological framework, based on the notion of signalling entropy, to help analyze and interpret functional omic data sets. The method uses a structural network, i.e. a PPI network, from the outset, and integrates this with gene expression data, using local and global signalling entropy measures to estimate the amount of uncertainty in the network signaling patterns. We made the case as to why uncertainty or entropy might be a key concept to consider when analysing biological data. In previous work [9, 38], we showed how signalling entropy can be used to estimate the differentiation potential of a cellular sample in the context of
normal differentiation processes, as well as demonstrating that signalling entropy also provides a highly accurate discriminator of cancer phenotypes. In this study we focused on a novel application of signalling entropy to understanding cellular robustness in the context of cancer drug sensitivity screens. Our main findings are that (i) local signalling entropy measures add predictive value over models that only use gene expression, (ii) that the local entropy of drug targets generally correlates positively with drug resistance, and (iii) that increased local entropy of intra-cellular hubs in cancer cells is a key hallmark of drug resistance.
These results are consistent and suggestive of an underlying entropy-robustness correlation theorem, as envisaged by previous authors [44]. Here, we provided additional empirical justification for such a theorem, using both simulated as well as real data, and using drug sensitivity measures as proxies for local robustness measures. A more detailed theoretical analysis of local mixing and entropy rates and incorporation of additional information (e.g. phosphorylation states of kinases, protein expression,..etc) when estimating entropies on real data, will undoubtedly lead to further improvements in our systemslevel understanding of how entropy/uncertainty dictates cellular phenotypes. From a practical perspective, we have already shown in previous work [38] how local network entropies could be used to identify key signaling pathways in differentiation. It will therefore be interesting in future to apply the signalling entropy framework in more detail to identify specific signaling nodes/pathways underlying drug sensitivity/resistance.
Our results have also confirmed the importance of network topology (e.g. hubness and therefore scale-freeness) in dictating drug resistance patterns. Thus, it will be interesting to continue to explore the deep relation between topological features such as scale-freeness and hierarchical modularity in relation to the gene expression patterns seen in normal and disease physiology. It is entirely plausible that, although our current data and network models are only mere caricatures of the real biological networks, that underlying fundamental systems biology principles characterising cellular phenotypes can still be gleaned from this data. Indeed, our discovery that increased signalling entropy correlates with drug resistance demonstrates that such fundamental principles can already be inferred from existing data resources.
It will also be of interest to consider other potential applications of the signaling entropy method. For instance, one application could be to the identification of functional driver aberrations in cancer. This would first use epigenomic (e.g. DNA methylation and histone modification profiles) and
genomic information (SNPs, CNVs) together with matched gene or protein expression data to identify functional epigenetic/genetic aberrations in cancer. Signalling entropy would subsequently be used as a means of identifying those aberrations which also cause a fundamental rewiring of the network. With multi-dimensional matched omic data readily available from the TCGA/ICGC, this represents another potentially important application of the signalling entropy method. Another important future application of the signaling entropy method would be to single-cell data, which is poised to become ever more important [59]. So far, all signaling entropy analyses have been performed on cell populations, yet single-cell analysis will be necessary to disentangle the entropies at the single-cell and population-cell levels.
In summary, we envisage that signalling entropy will become a key concept in future analyses and interpretation of biological data.
## Acknowledgements
AET is supported by the Chinese Academy of Sciences, Shanghai Institute for Biological Sciences and the Max-Planck Gesellshaft. RK and PS acknowledge funding under FP7/2007-2013/grant agreement nr. 290038.
## Appendix A. The protein protein interaction (PPI) network
We downloaded the complete human protein interaction network from Pathway Commons ( www.pathwaycommons.org ) (Jun.2012) [24], which brings together protein interactions from several distinct sources. We built a protein protein interaction (PPI) network from integrating the following sources: the Human Protein Reference Database (HPRD) [23], the National Cancer Institute Nature Pathway Interaction Database (NCI-PID) ( pid.nci.nih.gov ), the Interactome (Intact) http://www.ebi.ac.uk/intact/ and the Molecular Interaction Database (MINT) http://mint.bio.uniroma2.it/mint/ . Protein interactions in this network include physical stable interactions such as those defining protein complexes, as well as transient interactions such as posttranslational modifications and enzymatic reactions found in signal transduction pathways, including 20 highly curated immune and cancer signaling pathways from NetPath ( www.netpath.org ) [60]. We focused on nonredundant interactions, only included nodes with an Entrez gene ID annotation and focused on the maximally conntected component, resulting in a connected network of 10,720 nodes (unique Entrez IDs) and 152,889 documented interactions.
## Appendix B. Mixing rate results for a general random walk on a connected graph
Suppose that we have a connected graph with an ergodic Markov chain defined on it, given by a stochastic matrix P with stationary distribution π (i.e. πP = π ). We further assume that the detailed balance equation, π p i ij = π p j ji holds. Defining the diagonal matrix Π ≡ diag( π , ..., π 1 N ), the detailed balance equation can be rewritten as Π P = P T Π. That the matrix P is stochastic means that each row of P sums to 1. Equivalently, the vector with all unit entries, 1, is a right eigenvector of P , i.e. P 1 = 1. Note also that the stationary distribution π is a left eigenvector of P with eigenvalue 1. The Perron-Frobenius theorem further implies that all other eigenvalues are less than 1 in magnitude. If detailed balance holds, all eigenvalues are also real.
The global mixing rate can be defined by considering the rate at which the node visitation probabilities of a random walker approaches that of the stationary distribution, independently of the starting position. Formally, if we let Q t i ( ) denote the probability that at time t we find the walker at node i , then the mixing rate, µ R , is defined by [61]
$$\mu _ { R } = \lim _ { t \rightarrow \infty } \sup \max _ { i } | Q _ { i } ( t ) - \pi _ { i } | ^ { 1 / t }.$$
Denoting by Q t ( ) the column vector with elements Q t i ( ), one can write
$$Q ( t ) = ( P ^ { t } ) ^ { T } Q ( 0 ).$$
To determine the elements, Q t i ( ), of this vector, it is convenient to first introduce the matrix M = Π 1 2 P Π -1 2 . This is because M t = Π 1 2 P t Π -1 2 , and so P t can be rewritten in terms of M , but also because M satisfies the following lemma:
Lemma 1. M has the same eigenvalues as P and if u a is an eigenvector of M , then r a = Π -1 2 u a and l a = Π 1 2 u a are right and left eigenvectors of P .
Proof of Lemma 1. Suppose that Mu a = λ u a a . It then follows that
$$P ( \Pi ^ { - \frac { 1 } { 2 } } u _ { a } ) = \lambda _ { a } ( \Pi ^ { - \frac { 1 } { 2 } } u _ { a } ).$$
In the case of the left-eigenvector, multiply Mu a = λ u a a from the left with Π . Then, 1 2
$$\Pi P \Pi ^ { - 1 } ( \Pi ^ { \frac { 1 } { 2 } } u _ { a } ) = \lambda _ { a } ( \Pi ^ { \frac { 1 } { 2 } } u _ { a } ).$$
Detailed balance implies that Π P = P T Π, so P T l a = λ l a a . Taking the transpose of this implies that l a is indeed a left-eigenvector of P .
The significance of the above lemma becomes clear in light of the detailed balance equation, which implies that M = M T , and so M and M t can be orthogonally diagonalized. In particular, we can express Q t i ( ) in terms of the eigenvalue decomposition of M t , as
$$Q _ { i } ( t ) = \sum _ { a } q _ { a } | \lambda _ { a } | ^ { t } \tilde { u } _ { a i } \pi _ { i } ^ { 1 / 2 }$$
, where q a = ∑ j ˜ u aj π -1 2 / j Q j (0) and where ˜ u a is the a 'th unit norm eigenvector of M . Since Π 1 is the top unit norm eigenvector of 1 2 M (using previous lemma), which has an eigenvalue of 1, it follows that q 1 = 1 and hence that
$$Q _ { i } ( t ) = \pi _ { i } + \sum _ { a \geq 2 } q _ { a } | \lambda _ { a } | ^ { t } \tilde { u } _ { a i } \pi _ { i } ^ { 1 / 2 }.$$
It follows that
$$| Q _ { i } ( t ) - \pi _ { i } | = \sum _ { a \geq 2 } q _ { a } | \lambda _ { a } | ^ { t } \tilde { u } _ { a i } \pi _ { i } ^ { 1 / 2 }.$$
Since 1 = | λ 1 | ≥ | λ 2 | ≥ | λ 3 | . . . , we can conclude that as t →∞ , the rate at which Q i approaches the stationary value π i is determined by the modulus of the second largest eigenvalue (the Second Largest Eigenvalue ModulusSLEM). The global mixing rate µ R can thus be estimated as
$$\mu _ { R } \approx - \log | \lambda _ { 2 } | = - \log S L E M$$
.
## Appendix C. Entropy and mixing rates in simulated weighted Erd¨so Renyi graphs
For large regular graphs of degree d , the mixing rate is proportional to log d [56] and thus directly proportional to the entropy rate ( SR = log d for a regular graph of degree d ). By extrapolation, we can thus reasonably assume that for any sufficiently large graph with a tight degree distribution, such as random Erdos-Renyi (ER) graphs, that the entropy and mixing rates will also be correlated, albeit perhaps only in an average ensemble sense. The
glyph[negationslash]
analytical demonstration of this is beyond the scope of this work. Hence, we adopt a computational empirical approach to see if the entropy and mixing rates may indeed be correlated. In detail (and without loss of generality concerning the end result) we considered ER graphs of size 100 nodes and average degree 10 (other values for these parameters gave similar answers). We built an ensemble of 100 distinct ER graphs, and for each of these we constructed a family of weighted networks, parameterised by a parameter glyph[epsilon1] which controls the level of signalling entropy. Specifically, we gradually shifted the weight distribution around each node i (with degree d i ), in such a way that p ij = glyph[epsilon1]/d i for j = k and p ij = 1 -glyph[epsilon1] d i ( d i -1) for j = k , with 0 ≤ glyph[epsilon1] ≤ 1 and with k labeling a randomly chosen neighbor of node i . Thus, glyph[epsilon1] is a parameter that directly controls the uncertainty in the information flow, i.e. the entropy rate, since for glyph[epsilon1] = 1 we have that p ij = A /d ij i ( A ij is the symmetric adjacency matrix of the graph), whilst for glyph[epsilon1] = 0, p ij = δ ik , i.e. the information flow from node i can only proceed along one node (node k ). Thus, one would expect the entropy rate to decrease as glyph[epsilon1] is decreased to zero. For each value of glyph[epsilon1] we thus have an ensemble of 100 ER-graphs, for each of which the entropy and mixing rates can be computed. Finally, at each value of glyph[epsilon1] the entropy and mixing rates are averaged over the ensemble.
## Appendix D. Entropy and mixing rates in simulated weighted subgraphs of a PPI network
The analysis described above was performed also for maximally connected subnetworks generated from the underlying realistic PPI network described earlier. Specifically, we randomly subsampled 1000 nodes from the 8038 node PPI network, and extracted the maximally connected subnetwork, which resulted (on average) in a subnetwork of approximately 500 nodes. A family of stochastic matrices of variable entropy rates were constructed as explained above and for each resulting weighted network we estimated the entropy and mixing rates. Finally, ensemble averages over 100 different realisations were computed.
## Appendix E. The Cancer Cell Line Encyclopedia (CCLE) data
We used the gene expression data and drug sensitivity profiles as provided in the previous publication [57]. Briefly, integration of the gene expression data with our PPI network resulted in a maximally connected component
consisting of 8038 genes/proteins. There were 488 cell-lines with gene expression and drug sensitivity profiles for 24 drugs. As a measure of drug response we used the Activity Area [57] since this measure gave optimal results when correlating drug response to gene expression levels of well established drug targets.
## References
- [1] L. Hood, L. Rowen, The human genome project: big science transforms biology and medicine, Genome Med 5 (9) (2013) 79.
- [2] N. TCGA, Integrated genomic analyses of ovarian carcinoma, Nature 474 (7353) (2011) 609-615.
- [3] D. Hanahan, R. A. Weinberg, Hallmarks of cancer: the next generation, Cell 144 (5) (2011) 646-674.
- [4] A. L. Barabasi, N. Gulbahce, J. Loscalzo, Network medicine: a networkbased approach to human disease, Nat Rev Genet 12 (1) (2011) 56-68.
- [5] P. Creixell, E. M. Schoof, J. T. Erler, R. Linding, Navigating cancer network attractors for tumor-specific therapy, Nat. Biotechnol. 30 (9) (2012) 842-848.
- [6] G. Weng, U. S. Bhalla, R. Iyengar, Complexity in biological signaling systems, Science 284 (5411) (1999) 92-96.
- [7] A. L. Barabasi, Network science, Philos Trans A Math Phys Eng Sci 371 (1987) (2013) 20120375.
- [8] A. E. Teschendorff, S. Severini, Increased entropy of signal transduction in the cancer metastasis phenotype, BMC Syst Biol 4 (2010) 104.
- [9] J. West, G. Bianconi, S. Severini, A. E. Teschendorff, Differential network entropy reveals cancer system hallmarks, Sci Rep 2 (2012) 802.
- [10] A. L. Barabasi, Z. N. Oltvai, Network biology: understanding the cell's functional organization, Nat Rev Genet 5 (2) (2004) 101-113.
- [11] X. Robin, P. Creixell, O. Radetskaya, C. Costa Santini, J. Longden, R. Linding, Personalized network-based treatments in oncology, Clin. Pharmacol. Ther.
| [12] | J. T. Erler, R. Linding, Network medicine strikes a blow against breast cancer, Cell 149 (4) (2012) 731-733. |
|--------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [13] | M. J. Lee, A. S. Ye, A. K. Gardino, A. M. Heijink, P. K. Sorger, G. MacBeath, M. B. Yaffe, Sequential application of anticancer drugs en- hances cell death by rewiring apoptotic signaling networks, Cell 149 (4) (2012) 780-794. |
| [14] | M. S. Carro, W. K. Lim, M. J. Alvarez, R. J. Bollo, X. Zhao, E. Y. Snyder, E. P. Sulman, S. L. Anne, F. Doetsch, H. Colman, A. Lasorella, K. Aldape, A. Califano, A. Iavarone, The transcriptional network for mesenchymal transformation of brain tumours, Nature 463 (7279) (2010) 318-325. |
| [15] | C. Lefebvre, G. Rieckhof, A. Califano, Reverse engineering human reg- ulatory networks, Wiley Interdiscip Rev Syst Biol Med 4 (4) (2012) 311-25. |
| [16] | E. Segal, N. Friedman, D. Koller, A. Regev, A module map showing conditional activity of expression modules in cancer, Nat Genet 36 (10) (2004) 1090-1098. |
| [17] | R. Opgen-Rhein, K. Strimmer, From correlation to causation networks: a simple approximate learning algorithm, BMC Syst Biol 1 (2007) 37. |
| [18] | B. Barzel, A. L. Barabasi, Network link prediction by global silencing of indirect correlations, Nat Biotechnol 31 (8) (2013) 720-5. |
| [19] | S. Feizi, D. Marbach, M. Medard, M. Kellis, Network deconvolution as a general method to distinguish direct dependencies in, Nat Biotechnol 31 (8) (2013) 726-33. |
| [20] | H. Y. Chuang, E. Lee, Y. T. Liu, D. Lee, T. Ideker, Network-based classification of breast cancer metastasis, Mol Syst Biol 3 (2007) 140. |
| [21] | J. Dutkowski, T. Ideker, Protein networks as logic functions in develop- ment and cancer, PLoS Comput Biol 7 (9) (2011) e1002180. |
| [22] | K. Mitra, A. R. Carvunis, S. K. Ramesh, T. Ideker, Integrative ap- proaches for finding modular structure in biological networks, Nat Rev Genet 14 (10) (2013) 719-32. |
| [23] | T. S. Prasad, K. Kandasamy, A. Pandey, Human protein reference database and human proteinpedia as discovery tools for systems biology, Methods Mol Biol 577 (2009) 67-79. |
|--------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [24] | E. G. Cerami, B. E. Gross, E. Demir, I. Rodchenkov, O. Babur, N. An- war, N. Schultz, G. D. Bader, C. Sander, Pathway commons, a web resource for biological pathway data, Nucleic Acids Res 39 (Database) (2011) D685-D690. |
| [25] | M. Vidal, M. E. Cusick, A. L. Barabasi, Interactome networks and hu- man disease, Cell 144 (6) (2011) 986-998. |
| [26] | I. Ulitsky, R. Shamir, Identification of functional modules using network topology and high-throughput data, BMC Syst Biol 1 (2007) 8. |
| [27] | I. W. Taylor, R. Linding, D. Warde-Farley, Y. Liu, C. Pesquita, D. Faria, S. Bull, T. Pawson, Q. Morris, J. L. Wrana, Dynamic modularity in pro- tein interaction networks predicts breast cancer outcome, Nat Biotechnol 27 (2) (2009) 199-204. |
| [28] | J. West, S. Beck, X. Wang, A. E. Teschendorff, An integrative network algorithm identifies age-associated differential methylation interactome hotspots targeting stem-cell differentiation pathways, Sci Rep 3 (2013) 1630. |
| [29] | J. Roy, C. Winter, Z. Isik, M. Schroeder, Network information improves cancer outcome prediction, Brief Bioinform. |
| [30] | C. Winter, G. Kristiansen, S. Kersting, J. Roy, D. Aust, T. Knosel, P. Rummele, B. Jahnke, V. Hentrich, F. Ruckert, M. Niedergethmann, W. Weichert, M. Bahra, H. J. Schlitt, U. Settmacher, H. Friess, M. Buch- ler, H. D. Saeger, M. Schroeder, C. Pilarsky, R. Grutzmann, Google goes cancer: improving outcome prediction for cancer patients by, PLoS Comput Biol 8 (5) (2012) e1002511. |
| [31] | S. Suthram, A. Beyer, R. M. Karp, Y. Eldar, T. Ideker, eqed: an efficient method for interpreting eqtl associations using protein, Mol Syst Biol 4 (2008) 162. |
| [32] | Y. A. Kim, S. Wuchty, T. M. Przytycka, Identifying causal genes and dysregulated pathways in complex diseases, PLoS Comput Biol 7 (3) (2011) e1001095. |
|--------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [33] | K. Komurov, M. A. White, P. T. Ram, Use of data-biased random walks on graphs for the retrieval of context-specific networks from genomic data, PLoS Comput Biol 6 (8). |
| [34] | K. Komurov, J. T. Tseng, M. Muller, E. G. Seviour, T. J. Moss, L. Yang, D. Nagrath, P. T. Ram, The glucose-deprivation network counteracts lapatinib-induced toxicity in, Mol Syst Biol 8 (2012) 596. |
| [35] | S. Bandyopadhyay, M. Mehta, D. Kuo, M. K. Sung, R. Chuang, E. J. Jaehnig, B. Bodenmiller, K. Licon, W. Copeland, M. Shales, D. Fiedler, J. Dutkowski, A. Gunol, H. van Attikum, K. M. Shokat, R. D. Kolodner, W. K. Huh, R. Aebersold, M. C. Keogh, N. J. Krogan, T. Ideker, Rewiring of genetic networks in response to dna damage, Science 330 (6009) (2010) 1385-1389. |
| [36] | A. Califano, Rewiring makes the difference, Mol Syst Biol 7 (2011) 463. |
| [37] | T. Ideker, N. J. Krogan, Differential network biology, Mol Syst Biol 8 (2012) 565. |
| [38] | C. R. Banerji, D. Miranda-Saavedra, S. Severini, M. Widschwendter, T. Enver, J. X. Zhou, A. E. Teschendorff, Cellular network entropy as the energy potential in waddington's differentiation landscape, Sci Rep 3 (2013) 3039. |
| [39] | H. H. Chang, M. Hemberg, M. Barahona, D. E. Ingber, S. Huang, Transcriptome-wide noise controls lineage choice in mammalian progen- itor cells, Nature 453 (7194) (2008) 544-547. |
| [40] | B. D. Macarthur, I. R. Lemischka, Statistical mechanics of pluripotency, Cell 154 (3) (2013) 484-489. |
| [41] | C. Furusawa, K. Kaneko, Chaotic expression dynamics implies pluripo- tency: when theory and experiment meet, Biol Direct. 15 (4) (2009) 17. |
| [42] | C. Furusawa, K. Kaneko, A dynamical systems view of stem cell biology, Science 338 (6104) (2012) 215-7. |
|--------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [43] | T. I. Lee, R. G. Jenner, L. A. Boyer, M. G. Guenther, S. S. Levine, R. M. Kumar, B. Chevalier, S. E. Johnstone, M. F. Cole, K. Isono, H. Koseki, T. Fuchikami, K. Abe, H. L. Murray, J. P. Zucker, B. Yuan, G. W. Bell, E. Herbolsheimer, N. M. Hannett, K. Sun, D. T. Odom, A. P. Otte, T. L. Volkert, D. P. Bartel, D. A. Melton, D. K. Gifford, R. Jaenisch, R. A. Young, Control of developmental regulators by polycomb in human embryonic stem cells, Cell 125 (2) (2006) 301-313. |
| [44] | L. Demetrius, V. M. Grundlach, G. Ochs, Complexity and demographic stability in population models, Theo Pop Biol 65 (2004) 211-225. |
| [45] | L. Demetrius, T. Manke, Robustness and network evolution-an entropic principle, Physica A 346 (2005) 682-696. |
| [46] | T. Manke, L. Demetrius, M. Vingron, An entropic characterization of protein interaction networks and cellular robustness, J R Soc Interface 3 (11) (2006) 843-850. |
| [47] | J. Stelling, U. Sauer, Z. Szallasi, F. J. Doyle, J. Doyle, Robustness of cellular functions, Cell 118 (6) (2004) 675-685. |
| [48] | J. D. Han, N. Bertin, T. Hao, D. S. Goldberg, G. F. Berriz, L. V. Zhang, D. Dupuy, A. J. Walhout, M. E. Cusick, F. P. Roth, M. Vidal, Evidence for dynamically organized modularity in the yeast protein- protein interaction network, Nature 430 (6995) (2004) 88-93. |
| [49] | J. Li, Z. Yuan, Z. Zhang, The cellular robustness by gene redundancy in budding yeast, PLoS Genet 6 (11) (2010) e1001187. |
| [50] | J. A. Engelman, L. Chen, X. Tan, K. Crosby, A. R. Guimaraes, R. Upad- hyay, M. Maira, K. McNamara, S. A. Perera, Y. Song, L. R. Chirieac, R. Kaur, A. Lightbown, J. Simendinger, T. Li, R. F. Padera, C. Garcia- Echeverria, R. Weissleder, U. Mahmood, L. C. Cantley, K. K. Wong, Effective use of PI3K and MEK inhibitors to treat mutant Kras G12D and PIK3CA H1047R murine lung cancers, Nat. Med. 14 (12) (2008) 1351-1356. |
- [51] T. Sjoblom, S. Jones, L. D. Wood, D. W. Parsons, J. Lin, T. D. Barber, D. Mandelker, R. J. Leary, J. Ptak, N. Silliman, S. Szabo, P. Buckhaults, C. Farrell, P. Meeh, S. D. Markowitz, J. Willis, D. Dawson, J. K. Willson, A. F. Gazdar, J. Hartigan, L. Wu, C. Liu, G. Parmigiani, B. H. Park, K. E. Bachman, N. Papadopoulos, B. Vogelstein, K. W. Kinzler, V. E. Velculescu, The consensus coding sequences of human breast and colorectal cancers, Science 314 (5797) (2006) 268-274.
- [52] L. D. Wood, D. W. Parsons, S. Jones, J. Lin, T. Sjblom, R. J. Leary, D. Shen, S. M. Boca, T. Barber, J. Ptak, N. Silliman, S. Szabo, Z. Dezso, V. Ustyanksky, T. Nikolskaya, Y. Nikolsky, R. Karchin, P. A. Wilson, J. S. Kaminker, Z. Zhang, R. Croshaw, J. Willis, D. Dawson, M. Shipitsin, J. K. Willson, S. Sukumar, K. Polyak, B. H. Park, C. L. Pethiyagoda, P. V. Pant, D. G. Ballinger, A. B. Sparks, J. Hartigan, D. R. Smith, E. Suh, N. Papadopoulos, P. Buckhaults, S. D. Markowitz, G. Parmigiani, K. W. Kinzler, V. E. Velculescu, B. Vogelstein, The genomic landscapes of human breast and colorectal cancers, Science 318 (5853) (2007) 1108-1113.
- [53] B. Vogelstein, N. Papadopoulos, V. E. Velculescu, S. Zhou, L. A. Diaz, K. W. Kinzler, Cancer genome landscapes, Science 339 (6127) (2013) 1546-58.
- [54] V. Latora, M. Baranger, Kolmogorov-sinai entropy rate versus physical entropy, Phys Rev Lett 82 (3) (1999) 520-524.
- [55] J. Gomez-Gardenes, V. Latora, Entropy rate of diffusion processes on complex networks, Phys Rev E Stat Nonlin Soft Matter Phys 78 (6) (2008) 065102.
- [56] J. Friedman, On the second eigenvalue and random walks in random d-regular graphs, Combinatorica 11 (1991) 331-262.
- [57] J. Barretina, G. Caponigro, N. Stransky, K. Venkatesan, A. A. Margolin, S. Kim, C. J. Wilson, J. Lehr, G. V. Kryukov, D. Sonkin, A. Reddy, M. Liu, L. Murray, M. F. Berger, J. E. Monahan, P. Morais, J. Meltzer, A. Korejwa, J. Jan-Valbuena, F. A. Mapa, J. Thibault, E. Bric-Furlong, P. Raman, A. Shipway, I. H. Engels, J. Cheng, G. K. Yu, J. Yu, P. Aspesi Jr, M. de Silva, K. Jagtap, M. D. Jones, L. Wang, C. Hatton, E. Palescandolo, S. Gupta, S. Mahan, C. Sougnez, R. C. Onofrio,
- T. Liefeld, L. MacConaill, W. Winckler, M. Reich, N. Li, J. P. Mesirov, S. B. Gabriel, G. Getz, K. Ardlie, V. Chan, V. E. Myer, B. L. Weber, J. Porter, M. Warmuth, P. Finan, J. L. Harris, M. Meyerson, T. R. Golub, M. P. Morrissey, W. R. Sellers, R. Schlegel, L. A. Garraway, The cancer cell line encyclopedia enables predictive modelling of anticancer drug sensitivity, Nature 483 (7391) (2012) 603-607.
- [58] J. Ahmed, T. Meinel, M. Dunkel, M. S. Murgueitio, R. Adams, C. Blasse, A. Eckert, S. Preissner, R. Preissner, Cancerresource: a comprehensive database of cancer-relevant proteins and compound interactions supported by experimental knowledge, Nucleic Acids Res 39 (Database) (2011) D960-D967.
- [59] M. R. Speicher, Single-cell analysis: toward the clinic, Genome Med 5 (8) (2013) 74.
- [60] K. Kandasamy, S. S. Mohan, R. Raju, S. Keerthikumar, G. S. Kumar, A. K. Venugopal, D. Telikicherla, J. D. Navarro, S. Mathivanan, C. Pecquet, S. K. Gollapudi, S. G. Tattikota, S. Mohan, H. Padhukasahasram, Y. Subbannayya, R. Goel, H. K. Jacob, J. Zhong, R. Sekhar, V. Nanjappa, L. Balakrishnan, R. Subbaiah, Y. L. Ramachandra, B. A. Rahiman, T. S. Prasad, J. X. Lin, J. C. Houtman, S. Desiderio, J. C. Renauld, S. N. Constantinescu, O. Ohara, T. Hirano, M. Kubo, S. Singh, P. Khatri, S. Draghici, G. D. Bader, C. Sander, W. J. Leonard, A. Pandey, Netpath: a public resource of curated signal transduction pathways, Genome Biol 11 (1) (2010) R3.
- [61] L. Lovasz, Random walks on graphs: A survey, in: Combinatorics, Paul Erdos is Eighty, Vol. 2 of Mathematical Studies, Bolyai Society, Keszthely, Hungary, 1993, pp. 1-46.
## Figures
Figure 1: Signalling Entropy: understanding systems biology through uncertainty. A) A caricature model of a cellular interaction network with edge widths/color indicating the relative probabilities of interaction. On the left and right, we depict states of high and low signalling entropy, respectively. At the cellular population level, this translates into samples of high and low intra-sample heterogeneity, respectively. B) Signalling entropy correlates with pluripotency as demonstrated in our previous work (Banerji et al 2013). The pluripotent state is a highly promiscuous signaling state, generating high intra-sample heterogeneity, and allowing the stem cell population to differentiate into any cell type. In contrast, in a terminally differentiated state, signaling is almost deterministic, reflecting activation of very specific pathways in the majority of cells, leading to a highly homogeneous and differentiated cell population. Thus, signalling entropy defines the height in Waddington's differentiation landscape.
Figure 2: Signalling entropy and cellular robustness: A) Signalling entropy ought to correlate with cellular robustness. The inequality encapsulates this information by stating that a decrease in signalling entropy (i.e. if ∆ S < 0), then the system's robustness R must also decrease, i.e. ∆ R < 0, so that the product ∆ S ∆ R > 0. Observe how in the low entropy state, random removal of edges through e.g. inactivating mutations, can lead to deactivation of a key signaling pathway connecting a given start and end nodes (shown in orange). In the high entropy state, the same perturbations do not prevent signal transduction between the two orange nodes. B) Depicted are the effects of two major forms of cancer perturbation. In the upper panel, inactivation (typically of tumour suppressors), leads to underexpression and a corresponding loss of correlations/interactions with neighbors in the PPI network. This is tantamount to a state of increased entropy and drug intervention is unlikely to be effective. In the lower panel, we depict the case of an oncogene, which is overexpressed in cancer. This overexpression leads to activation of a specific oncogenic pathway which results in oncogene addiction and increased sensitivity to targeted drug intervention. Thus, local signalling entropy and robustness (as determined by response to a drug), may also correlate locally.
Figure 3: Entropy rate in normal and cancer tissue: A) Boxplots of sample specific entropy rates comparing normal liver and liver cancer samples. Expression data set is the one used in [38]. B) As A), but also shown are the sample specific entropy rates obtained by randomly permuting the gene expression values over the network. Note how the entropy rates for the normal and cancer states are significantly reduced upon permutation and are no longer highly discriminative between normal and cancer. C) ROC curves and associated AUC normal-cancer discriminatory values for the unpermuted and permutated cases depicted in A) and B).
Figure 4: Correlation between global entropy and mixing rates. A) Plotted is the entropy rate against the mixing rate for Erdos-Renyi graphs of 100 nodes and average degree 10. The light blue lines show the results over 100 different network realisations, with the dark blue line representing the ensemble average. B) As A) but for connected subnetworks of average size 500 nodes, generated by random subsampling of 1000 nodes from the full PPI network of 8038 nodes. As in A), a range of edge weight distributions were considered reflecting variable entropy rates. The light blue lines show the results over 100 different realisations, with the dark blue line representing the ensemble average. C) Scatterplot of the global entropy versus mixing rates for the 488 Cancer Cell-Line Encyclopedia (CCLE) samples. Spearman rank correlation coefficient (SCC) and associated P-value are given (left panel). Wilcoxon rank sum test P-value between high and low mixing rate groups (as defined by tertiles) (right panel).
Figure 5: Anti-correlation between local signalling entropy and drug sensitivity. A) Histogram of Spearman rank correlation P-values between drug sensitivities (n=24) and local signalling entropies (n=8038 genes), as computed over the 488 CCLE cell-lines. B) As A) but for gene expression instead of signalling entropy. C) Scatterplot of Spearman rank Correlation Coefficient (SCC) between local signalling entropy (sigS) and drug sensitivity (DS) against -log 10 P-value for each of 134 drug gene target pairs. D) As C) but for gene expression instead of local entropy. In C) & D) , we give the distribution of significant positive and negative correlations and the associated Binomial test P-value. E) Drug target gene pairs ranked according to negative SCC (cyan color) between signalling entropy and drug sensitivity. Only pairs where at least one of entropy or gene expression were significantly associated are shown. Upper panels show the SCC (cyan=strong negative SCC, white=zero or non-significant SCC, magenta=strong positive SCC), while lower panels show the corresponding P-values with the last row showing the P-value from the Likelihood Ratio Test (LRT) assessing the added predictive value of signalling entropy over gene expression. The darker the tones of red the more significant the P-values are, whilst white indicates non-significance. F) A subset of E), with pairs now ranked according to the LRT P-value.
Figure 6: High signalling entropy of intra-cellular hubs confers drug resistance:
Upper panel plots the topological degree of drug targets (x-axis) against the Spearman rank Correlation Coefficient (SCC) between its local signalling entropy and drug sensitivity, as assessed over the CCLE samples. EC=target annotated as extra-cellular, MR=target annotated as membrane receptor, IC=target annotated as intra-cellular. Left lower panel shows the difference in SCC values between the IC and EC+MR targets. Wilcoxon rank sum test P-value given. Right lower panel shows the difference in SCC values between the IC and EC+MR targets, where now the SCC value were computed between gene expression and drug sensitivity. Wilcoxon rank sum test P-value given.
|
10.1016/j.ymeth.2014.03.013
|
[
"Andrew Teschendorff",
"Peter Sollich",
"Reimer Kuehn"
] |
2014-08-05T15:25:02+00:00
|
2014-08-05T15:25:02+00:00
|
[
"q-bio.MN",
"q-bio.GN"
] |
Signalling Entropy: a novel network-theoretical framework for systems analysis and interpretation of functional omic data
|
A key challenge in systems biology is the elucidation of the underlying
principles, or fundamental laws, which determine the cellular phenotype.
Understanding how these fundamental principles are altered in diseases like
cancer is important for translating basic scientific knowledge into clinical
advances. While significant progress is being made, with the identification of
novel drug targets and treatments by means of systems biological methods, our
fundamental systems level understanding of why certain treatments succeed and
others fail is still lacking. We here advocate a novel methodological framework
for systems analysis and interpretation of molecular omic data, which is based
on statistical mechanical principles. Specifically, we propose the notion of
cellular signalling entropy (or uncertainty), as a novel means of analysing and
interpreting omic data, and more fundamentally, as a means of elucidating
systems-level principles underlying basic biology and disease. We describe the
power of signalling entropy to discriminate cells according to differentiation
potential and cancer status. We further argue the case for an empirical
cellular entropy-robustness correlation theorem and demonstrate its existence
in cancer cell line drug sensitivity data. Specifically, we find that high
signalling entropy correlates with drug resistance and further describe how
entropy could be used to identify the achilles heels of cancer cells. In
summary, signalling entropy is a deep and powerful concept, based on rigorous
statistical mechanical principles, which, with improved data quality and
coverage, will allow a much deeper understanding of the systems biological
principles underlying normal and disease physiology.
|
1408.1003v2
|
## Self-consistent model of spin accumulation magnetoresistance in ferromagnet-insulator-semiconductor tunnel junctions
Ian Appelbaum, ∗ Holly N. Tinkey, and Pengke Li
Department of Physics and Center for Nanophysics and Advanced Materials, U. Maryland, College Park, MD 20742
Spin accumulation in a paramagnetic semiconductor due to voltage-biased current tunneling from a polarized ferromagnet is experimentally manifest as a small additional spin-dependent resistance. We describe a rigorous model incorporating the necessary self-consistency between electrochemical potential splitting, spin-dependent injection current, and applied voltage that can be used to simulate this so-called '3T' signal as a function of temperature, doping, ferromagnet bulk spin polarization, tunnel barrier features and conduction nonlinearity, and junction voltage bias.
IntroductionOver the past decade or so, substantial progress has been made in understanding the conditions required for achieving spin-polarized electron transport in otherwise nonmagnetic semiconductors. 1 However, the specific constraints imposed 2-5 empirically necessitate the use of unconventional fabrication techniques 6,7 , precise control over material growth, and often elaborate high-resolution lithographic and deposition procedures 8,9 . When it was suggested 10 that these complications could be circumvented simply by analyzing the magnetic-field dependence of local magnetoresistance due to spin precession 11,12 in large-area single ferromagnet (FM)-insulator-semiconductor tunnel junctions, interest in the experimental spintronics research community was noticeably raised. A significant report in 2009 asserted 13 that evidence for spin 'accumulation' in bulk Si persisted through room-temperature in this type of '3T' device (so-named because of 3-terminal configurations intended to eliminate series ohmic voltage drops through the semiconductor electrode). As a result of these fantastic claims, many in the field quickly and uncritically accepted the new approach as a genuine breakthrough allowing easy access to spin transport properties such as lifetime and diffusion coefficient. 14-18
oto group 31 have characteristics (smaller magnitude and dephasing fields, strong temperature dependence, etc.) consistent with true spin accumulation, many specious conclusions on spin transport properties of several important materials have been drawn by others due to a fundamental misunderstanding of the underlying physics behind the 3T technique. The present paper aims at constructing a rigorous scheme to model such expected genuine results and provide a quantitative means to compare experiment to theory. In particular, we incorporate the essential self-consistency between electrochemical potential splitting, the spin-dependent injection current that induces it, and applied external voltage to simulate magnetoresistance measurements of FM-insulatorsemiconductor tunnel junctions as a function of all extrinsic parameters e.g. temperature, doping, ferromagnet bulk spin polarization, tunnel barrier features and conduction nonlinearity, and junction voltage bias.
Others, however, were more cautious and exposed inconsistencies that became evident after a more skeptical analysis of the relevant experimental parameters. For instance, spin lifetimes were found to be largely insensitive to doping polarity or concentration, temperature, tunnel barrier material 19 , or semiconductor 20 , all contrary to expectations set by electron spin resonance measurements. Even when normal metals replaced the semiconductor, devices produced the same signals - independent of the spin-orbit interaction strength which drives relaxation rates! 21 Furthermore, voltage signals are often several orders of magnitude larger than what is possible if due to injection-driven spin accumulation in the bulk. It now appears clear that instead of spin accumulation, it is rather the inelastic transport pathways provided by defect-localized electronic states in or near the tunnel barrier that can play a dominant role in device behavior. 22-26
Although the original report by the Minnesota group 10 on devices using epitaxial Fe/GaAs, and recent work from the Kyushu group on clean CoFe/Si 27-30 and Ky-
BackgroundThe steady-state solution to the coupled transport equations for up/down spin electron density n ↑ / ↓ for boundary conditions corresponding to spin injection at z = 0 into a homogeneous semi-infinite conductor z ≥ 0 with fixed diffusion coefficient D , drift velocity v , and spin flip rate 1 / 2 τ is n ↑ -n ↓ = Pne -z/L , where P is the electron spin polarization at the injection site, n = n ↑ + n ↓ , and L = vτ 2 + √ ( vτ 2 ) 2 + Dτ is the 'downstream' drift-diffusion transport lengthscale. 32 In steady-state, we must supply to this region enough polarized electrons as are lost to spin flips (relaxation-time approximation), resulting in
$$\frac { ( n _ { \uparrow } - n _ { \downarrow } ) L } { \tau } = \frac { \beta J } { q }, \quad \quad \quad ( 1 )$$
where J = qnv is injected charge current density, β is current spin polarization from the FM, and q is the fundamental electron charge.
Asymptotic formsWe will now attempt to estimate the consequences of Eq. 1 for the case of injection into a semiconductor of arbitrary n , by making use of simple asymptotic forms of maximum density imbalance n ↑ -n ↓ in the statistically degenerate and nondegenerate regime. For the former case, thermal energy k B T /lessmuch µ -E C , where µ is chemical potential and E C is conduction band
minimum. Then we have n ↑ -n ↓ /similarequal ∆ µ 2 · D µ ( -E C ), where the density of states at the chemical potential is D µ ( -E C ) ≈ 3 n 2( µ -E C ) , allowing us to calculate the chemical potential splitting
$$\Delta \mu = \mu _ { \uparrow } - \mu _ { \downarrow } \simeq \frac { 2 ( \mu - E _ { C } ) } { 3 } \left [ \frac { 2 \beta J \tau } { q n L } \right ] \,. \, ( \text{degenerate} ) \ ( 2 ) \quad \text{poton} \\ - \ \. \ \. \ \. \ \. \ \. \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
For the latter, nondegenerate regime when k B T /greatermuch µ -E C , we can use the classical Boltzmann distribution and effective conduction band electron density N C = 2 -1 / 2 ( m ∗ SC k B T π /planckover2pi1 2 ) 3 / 2 in n ↑ / ↓ = n 2 ± βJτ 2 qL ≈ N C 2 e ( µ ↑ / ↓ -E C ) /k B T , giving
$$\mu _ { \uparrow / \downarrow } - E _ { C } & \approx k _ { B } T \ln \left ( \frac { n } { N _ { C } } \pm \frac { \beta J \tau } { q L N _ { C } } \right ) \quad \begin{array} { c } 1 \\ 1 \\ \end{array} \\ & = k _ { B } T \left [ \ln \left ( \frac { n } { N _ { C } } \right ) + \ln \left ( 1 \pm \frac { \beta J \tau } { q L n } \right ) \right ]. \quad \begin{array} { c } 1 \\ \end{array} \\ \text{For dilute sin} \text{ densities from weak} \text{ sin} \text{ iniation or} \quad \begin{array} { c } 1 \end{array}$$
For dilute spin densities from weak spin injection or strong relaxation, we can expand the second term to first order, yielding
$$\Delta \mu \simeq k _ { B } T \left [ \frac { 2 \beta J \tau } { q L n } \right ]. \quad ( n o n d e g e n e r a t e ) \quad ( 3 ) \quad \text{the } \frac { 1 c } { \text{and} } \\ \quad \text{--} \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \........................................................................................................................................................................................................ \..................................................................................................................................................................................................... \. \................................................................................................................................................................................................... \... \................................................................................................................................................................................................... ....................................................................................................................................................................................................... ... ................................................................................................................................................................................................... ... \................................................................................................................................................................................................... \. ..................................................................................................................................................................................................... . ..................................................................................................................................................................................................... \. . ................................................................................................................................................................................................... . . ................................................................................................................................................................................................... \... ................................................................................................................................................................................................... \. \. \................................................................................................................................................................................................. \..... \................................................................................................................................................................................................. \... \. \................................................................................................................................................................................................. \. \... \................................................................................................................................................................................................. \. \. \. \............................................................................................................................................................................................... \....... \..............................................................................................................................................................................................$$
Notice that Eqs. 2 and 3 nominally differ only by the energy scale prefactor. However, nondegenerate systems at low temperatures typically have much smaller density n , so that (all other parameters being equal) the unitless quantity in square brackets is much larger in magnitude. Nevertheless, since this quantity is simply the ratio of steady-state spin imbalance to the equilibrium electron density, it is bounded by unity in the regime where the relaxation-time approximation is valid.
This chemical potential splitting ∆ µ resulting from spin injection is not measured directly in an experiment. In the most na¨ ıve approach, we can treat injection and detection as separate events. In open-circuit detection appropriate for four-terminal non-local devices with ferromagnetic contacts 12 , injection creates a spin splitting ∆ , and due to FM conductance asymmetry (bulk spin µ polarization) β = σ ↑ -σ ↓ σ ↑ + σ ↓ a voltage
$$\Delta V = \beta \frac { \Delta \mu } { 2 q }$$
develops to maintain zero net carrier flow across the interface. Dephasing the spins via precession in a perpendicular field will suppress this voltage, allowing for a direct experimental measurement of ∆ V , the additional voltage necessary to drive a fixed current due to spin accumulation magnetoresistance.
Using Eqs. 2, 3, and 4, we can estimate voltage signals of order ∆ V /apprle β Jτ 2 q 2 Ln E , where the energy scale E is given by the maximum of the thermal or Fermi energy. While useful to provide gross predictions of an upper bound on expected signals, this approach is however not rigorously correct since we have in fact a biased junction used for both injection and detection that cannot be considered separate processes. The spin-dependent chemical potential imbalance necessary for the spin accumulation itself changes the injection rates of spin up and down nonlinearly, so a self-consistent solution is needed. Importantly, this method, freed from reliance on asymptotic expressions used to obtain Eqs. 2 and 3, will be inherently able to address both i. ) the intermediate regime between degenerate and nondegenerate conditions where most experiments were performed; and ii. ) conditions of high spin injection rates, long spin lifetimes, or the deep nondegenerate regime where βJτ qLn ∼ 1 (unobtainable in metals) and first-order expansion of the logarithm function used to obtain Eq. 3 fails.
Self-consistent modelConsider the FM-semiconductor tunnel junction illustrated in Fig. 1. Since our focus here is on a purely elastic tunneling model, we ignore Fermi level pinning from interface states resulting in thermionic emission-dominated transport into the depletion region, which would otherwise potentially complicate detection 33 .
To begin the self-consistent calculation, we initialize the unknown P = n ↑ -n ↓ n ↑ + n ↓ = 0 so that n ↑ = n ↓ = n/ 2 and µ ↑ = µ ↓ = µ 0 in the conduction band. We can then calculate the current density J = J 0 = J ↑ + J ↓ which flows under voltage V = V 0 , using an extension of the 1-dimensional transport model 34 which sums over FM cathode states:
$$\ u d e. \\ \L o \text{ of } \quad J _ { \uparrow / \downarrow } = & q \frac { 1 \pm \beta } { 2 } \frac { 2 m _ { F M } ^ { * } E _ { F } } { h ^ { 3 } } \times \\ \L t \text{ the } \quad \int _ { 0 } ^ { \infty } \left [ f _ { T } ( E - E _ { F } ) - f _ { T } ( E - \mu _ { \uparrow / \downarrow } ) \right ] \mathfrak { I } ( E, V ) d E, \\ \text{rom } \quad \int _ { 0 } ^ { \infty } \left [ f _ { T } \right ] \left ( \Lambda \right ) \right ] \Lambda \right ) \Lambda \right ) \Lambda \right ) \Lambda \right ) \Lambda \right ) \Lambda \right ) \Lambda \right ) \Lambda \right ) \Lambda \right ) \Lambda \right ) \Lambda \right ) \Lambda \right ) \Lambda \right )$$
where f T is the Fermi-Dirac occupation function at temperature T , and we employ the usual semiclassical approximation for incoherent tunneling transmission coefficient T , 5 correct to lowest order in the ratios of E and qV to the total tunnel barrier height E F +Φ, where Φ is the tunnel barrier internal work function:
$$\mathfrak { T } ( E, V ) = \frac { 4 k _ { F M } \kappa } { k _ { F M } ^ { 2 } + \kappa ^ { 2 } } \frac { 4 k _ { S C } \kappa } { k _ { S C } ^ { 2 } + \kappa ^ { 2 } } e ^ { - 2 \kappa d }, \text{ where } \text{ } ( 6 )$$
$$\kappa & = \sqrt { 2 m _ { I } ^ { * } ( E _ { F } + \Phi - q V / 2 - E ) } / \hbar {, } \quad \ ( 7 ) \\ k _ { F M } & = \sqrt { 2 m _ { F M } ^ { * } E } / \hbar {, } \, \text{and} \quad \ ( 8 )$$
Notice that k SC , and therefore T , vanishes for electrons with energy in the forbidden gap.
$$k _ { F M } & = \overset { \mathbf v } { \sqrt { 2 m _ { F M } ^ { * } E } / \hbar {, } \text{and} } & ( 8 ) \\ k _ { S C } & = \Re \sqrt { 2 m _ { S C } ^ { * } ( E - E _ { C } ) } / \hbar {. } & ( 9 ) \\ \text{Notice that } k _ { S C }, \text{ and therefore } \mathfrak { T }, \text{ vanishes for electrons}$$
µ
FIG. 1. (Color online) (a) Schematic energy band diagram identifying parameters in the model simulating spin accumulation magnetoresistance in the ferromagnet (FM)-insulatorsemiconductor (SC) tunnel junction. Blue arrows show equilibrium electrochemical potential µ 0 splitting due to spinpolarized electron injection. (b) Schematic flow diagram for the self-consistent algorithm used to calculate '3T' magnetoresistance signal ∆ V at constant current density J 0 and electron density n due to spin accumulation and precession in a magnetic field.
Now, we perform the following four steps in a loop to obtain self-consistency between the injection rate J and the spin accumulation ∆ µ by iterative adjustment of the junction bias voltage V (see Fig. 1 (b)):
[1.] Using the present value of accumulated polarization P , find µ ↑ and µ ↓ consistent with n = n ↑ + n ↓ from the sum over semiconductor states
$$& n _ { \uparrow / \downarrow } = \\ & \frac { ( 1 \pm P ) n } { 2 } = g \int _ { 0 } ^ { \infty } f _ { T } ( E - \mu _ { \uparrow / \downarrow } ) D _ { \frac { 1 } { 2 } } ( E - E _ { C } ) d E, \ \ ( 1 0 ) \\ & \quad \ \.$$
where g is the conduction band degeneracy (e.g. g = 6
for Si, = 4 for Ge), and the single-spin density of states 3 / 2 √ inverse
D 1 2 ( E ) = 1 √ 2 π 2 ( m ∗ SC /planckover2pi1 2 ) E . This nonlinear problem is solved using a binomial search algorithm.
[2.] Using the new values for µ ↑ / ↓ , employ Eqs. 5-9 to find the applied voltage V > V 0 necessary to maintain fixed current density J 0 = J ↑ + J ↓ . Again, this is a nonlinear inverse problem, which can be solved using a binomial search algorithm.
[3.] Using the new values for spin-dependent current densities J ↑ ↓ / , update the accumulated polarization from the relaxation-time approximation result (Eq. 1) in the form P = ( J ↑ -J ↓ ) τ qnL . By rewriting the drift-diffusion lengthscale as
$$L = \sqrt { D \tau + \left ( \frac { J _ { 0 } \tau } { 2 n q } \right ) ^ { 2 } } + \frac { J _ { 0 } \tau } { 2 n q }, \quad \quad ( 1 1 )$$
we see that in the high injection current limit, even if β = 1, this implies (to lowest order)
$$P & \lesssim 1 - \frac { q ^ { 2 } n ^ { 2 } D } { \tau J _ { 0 } ^ { 2 } }. \quad \quad \ ( 1 2 ) \\, \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot$$
The spin polarization is thus always strictly less than unity as required.
[4.] Repeat steps 1-3 until convergence of µ ↑ / ↓ . When completed, the self-consistent voltage necessary in constant-current mode is the difference between the spin-independent (average) chemical potential in the SC, ¯ = µ µ ↑ + µ ↓ 2 , and the FM Fermi energy, implying a change in applied voltage in a perpendicular magnetic field necessary to fully dephase the spins of
$$\Delta V = \frac { E _ { F } - \bar { \mu } } { q } - V _ { 0 }. \text{ \quad \ \ } ( 1 3 )$$
Given the structure of expressions above, we qualitatively expect that self consistent iteration is most important under a particular set of circumstances. Of course, this includes the condition that ∆ V is comparable to V , so that changes in ∆ µ affect large changes in the ratios of spin up/down injected current. Usually, this requires highly conductive tunnel barriers and large injected current spin polarization β . However, Eq. 12 shows that the steady-state density polarization P saturates at high J 0 , making self-consistency unnecessary beyond one iteration in this limit. Therefore, assuming a fixed n τ , , and D , a self-consistent approach is most important when P and dP dJ are jointly maximized, which occurs for
$$J _ { 0 } ^ { * } & \approx q n \sqrt { 2 D / \tau }. \quad & ( 1 4 ) \\ \text{recognizable as the current for which the}$$
This value is recognizable as the current for which the contributions from drift and diffusion are approximately equal in Eq. 11.
ResultsNow we discuss the results of our selfconsistent method. In Fig. 2, we see an initial error relative to converged value of several percent in ∆ . µ
FIG. 2. (Color online) Convergence of the self-consistent scheme, showing the exponential reduction of relative error with iteration of calculation steps 1-4 discussed in the text.
However, we also see fast exponential convergence toward negligible error upon successive iteration. Here, we use parameters E F = 5eV, d = 1nm, Φ = 1eV, n = 10 19 cm -3 (just below ambient N C ≈ 3 × 10 19 ), m ∗ FM = m 0 , and m ∗ SC = 0 36 . m 0 . At T = 300K, V 0 = 100mV drives a current density J 0 ≈ 3 × 10 A/cm , 4 2 on the order of J ∗ 0 given in Eq. 14. With a FM polarization of β = 0 4, . D = 10cm /s, and 2 τ = 10ns, we obtain a steady state spin accumulation in the SC of approximately P = 0 17 and ∆ . µ ≈ 0 6mV. Note that, . since these parameters give 2 βJτ qnL ≈ 0 5 . ∼ 1, the results represent the extreme high-injection limit of validity and are used here merely to illustrate the model behavior at current densities close to J ∗ 0 (Eq. 14). Also shown in Fig. 2 is the small relative error of the tunnel current density during the voltage adjustment, illustrating the stability of the method and effectiveness of the binomial search algorithm as applied to Eq. 5.
The current-voltage relationship for the same junction is shown in Fig. 3(a), where the nonlinear behavior expected from tunneling can clearly be seen. In the inset, the corresponding voltage dependence of converged variables shows that increasing the current injection with higher voltage results in saturation of the accumulated polarization ( ∝ ∆ ) and a suppression of spin accumuµ lation signal (∆ V ). This latter phenomenon can be understood by considering the effect of an increasing conductance nonlinearity at high voltage on the constraints of fixed current density while reaching self-consistency.
The calculations presented so far have used parameters intended to approach the current density in Eq. 14. Actual experiments performed to date have, however, far smaller tunnel junction conductances. By adjusting d = 1 5nm and Φ =1.7eV (appropriate for Al O . 2 3 barriers), the current density dramatically decreases to more realistic values, as shown for several SC donor densities in Fig. 3(b). Although the current-voltage relationship is only weakly dependent on doping in this restricted range, the spin accumulation voltage signals shown in the inset clearly demonstrate a stronger sensitivity. Importantly, we see here that the magnitude of ∆ V (tens of nV) is orders of magnitude smaller than in the high injection
FIG. 3. (Color online) (a) Current-voltage relation in the strong injection current regime. Self-consistent convergence data shown in Fig. 2 for point indicated. Inset: Corresponding spin accumulation voltage and electrochemical splitting. (b) I-V in the more realistic weak injection regime, showing little dependence on donor concentration. Inset: Spin accumulation signals over the same voltage range.
limit above; for a comparison to the asymptotic expressions given earlier, Eqs. 3 and 4 predict ∆ V ≈ 8 µ V for n = 5 × 10 18 cm -3 , and using Eq. 2 for n = 5 × 10 19 cm -3 yields ∆ V ≈ 500nV. Both degenerate and nondegenerate asymptotic approaches sharply overestimate the expected accumulation signal for these intermediate electron densities.
Temperature dependence of the spin accumulation voltage signals is shown in Fig. 4 with a reference voltage of V 0 =0.5V. The current density varies only slightly over this range (inset), but ∆ V increases dramatically with temperature over ≈ 100K. This behavior (constant at low temperatures in the quasi-degenerate regime and linearly increasing in the quasi-nondegenerate regime) is consistent with the trends predicted by Eqs. 2 and 3. However, the calculated signal magnitudes are much smaller than indicated by the coarse approximations.
ConclusionTo illustrate this apparent discrepancy, we compare the asymptotic forms at 300K to the calculated self-consistent results for low injection J 0 = 1A/cm 2 as a function of electron density n in Fig. 5. We have highlighted three regions, separated by two vertical lines, where µ -E C < -5 k B T (nondegenerate),
FIG. 4. (Color online) Temperature dependence of spin accumulation voltage signals at a fixed V 0 =0.5V. Inset: current density at this bias.
FIG. 5. (Color online) comparison of asymptotic forms to selfconsistent values for '3T' magnetoresistance voltage ∆ V , for J 0 =1 A/cm 2 at 300K.
µ -E C > +5 k B T (degenerate), and the intermediate
- ∗ [email protected]
- 1 J. Fabian et al. , Acta Phys. Slovaca 57 , 565 (2007).
- 2 M. Johnson and R. Silsbee, Phys. Rev. B 35 , 4959 (1987).
- 3 E. Rashba, Phys. Rev. B 62 , R16267 (2000).
- 4 G. Schmidt et al. , Phys. Rev. B 62 , R4790 (2000).
- 5 J. C. Slonczewski, Phys. Rev. B 39 , 6995 (1989).
- 6 I. Appelbaum et al. , Nature 447 , 295 (2007).
- 7 B. Huang et al. , Phys. Rev. Lett. 99 , 177209 (2007).
- 8 X. Lou et al. , Nat. Phys. 3 , 197 (2007).
- 9 T. Suzuki et al. , Appl. Phys. Express 4 , 3 (2011).
- 10 X. Lou et al. , Phys. Rev. Lett. 96 , 176603 (2006).
- 11 W. Hanle, Z. Phys. 30 , 93 (1924).
- 12 M. Johnson and R. Silsbee, Phys. Rev. Lett. 55 , 1790 (1985).
- 13 S. P. Dash et al. , Nature 462 , 245 (2009).
- 14 C. Li et al. , Nature Comm. 2 , 245 (2011).
- 15 R. Jansen et al. , Phys. Rev. B 82 , 241305(R) (2010).
- 16 K. Jeon et al. , Appl. Phys. Lett. 98 , 3 (2011).
- 17 N. Gray and A. Tiwari,
regime where µ ≈ E C . Clearly, the numerically calculated results trend toward the asymptotic curves in the doping level extremes, with especially good approximation in the degenerate limit. However, the intermediate and weakly nondegenerate regime show two orders of magnitude of disagreement. Even at the solidlynondegenerate density of n = 10 15 cm -3 , at least an order of magnitude separates the results.
The lack of agreement for low electron densities can be understood by considering that as n decreases, junction zero-bias resistance increases and the tunnel bias is substantially larger for the same driven current density J 0 . The junction nonlinearity is then appreciable, and the suppression of ∆ V seen at high bias in the insets to Figs. 3(a) and (b) becomes substantial. This explanation is corroborated by the fact that, all other parameters being equal, deviation from Eqs. 2 and 3 across all densities worsens as J 0 (and hence V 0 ) is chosen to be larger. The problem at lower electron doping makes it clear that the asymptotic expression for the nondegenerate case Eq. 3 is merely an optimistic upper bound, and that realistic signals from devices in this regime (more correctly simulated with the self-consistent scheme described here) will in fact be systematically much lower. Even for the degenerate case, voltage changes from spin-accumulation magnetoresistance induced by precession and dephasing are far lower than those experimental measurements attributed to this mechanism. For example, the original claim of room-temperature spin accumulation in 1.8 × 10 19 cm -3 n-Si with ∆ V ≈ 1 2mV for . J = 4A/cm 2 (see Ref. 13) is so far above physically reasonable values that it (and the results of other room-temperature measurements 14-18 ) is vertically off-scale in Fig. 5.
We gratefully acknowledge support from ONR under contract N000141410317, NSF under contracts ECCS1231855 and DGE-1322106 (H.N.T.), and DTRA under contract HDTRA1-13-1-0013.
Appl. Phys. Lett. 98 , 102112 (2011).
- 18 A. Dankert et al. , Sci. Rep. 3 , 3196 (2013).
- 19 O. van't Erve et al. , J. Appl. Phys. 113 , 17C502 (2013).
- 20 B. Birkner et al. , Phys. Rev. B 87 , 081405 (2013).
- 21 O. Txoperena et al. , Appl. Phys. Lett. 102 , 192406 (2013).
- 22 M. Tran et al. , Phys. Rev. Lett. 102 , 036601 (2009).
- 23 A. Jain et al. , Phys. Rev. Lett. 109 , 106603 (2012).
- 24 R. Jansen et al. , Phys. Rev. B 85 , 134420 (2012).
- 25 H. Tinkey et al. , Appl. Phys. Lett. 104 , 232410 (2014).
- 26 Y. Song and H. Dery, (2014), arXiv:1401.7649.
- 27 Y. Ando et al. , Appl. Phys. Lett. 99 , 132511 (2011).
- 28 Y. Ando et al. , Phys. Rev. B 85 , 035320 (2012).
- 29 M. Ishikawa et al. , Appl. Phys. Lett. 100 , 252404 (2012).
- 30 K. Hamaya et al. , J. Appl. Phys. 113 , 17C501 (2013).
- 31 M. Kameno et al. , Appl. Phys. Lett. 104 , 092409 (2014).
- 32 Z. Yu and M. Flatt´, Phys. Rev. B e 66 , 201202 (2002).
- 33 R. Jansen and B. Min, Phys. Rev. Lett. 99 , 246604 (2007).
- 34 W. A. Harrison, Phys. Rev. 123 , 85 (1961).
|
10.1103/PhysRevB.90.220402
|
[
"Ian Appelbaum",
"Holly N. Tinkey",
"Pengke Li"
] |
2014-08-05T15:25:27+00:00
|
2014-10-01T17:44:37+00:00
|
[
"cond-mat.mes-hall",
"cond-mat.other"
] |
Self-consistent model of spin accumulation magnetoresistance in ferromagnet-insulator-semiconductor tunnel junctions
|
Spin accumulation in a paramagnetic semiconductor due to voltage-biased
current tunneling from a polarized ferromagnet is experimentally manifest as a
small additional spin-dependent resistance. We describe a rigorous model
incorporating the necessary self-consistency between electrochemical potential
splitting, spin-dependent injection current, and applied voltage that can be
used to simulate this so-called "3T" signal as a function of temperature,
doping, ferromagnet bulk spin polarization, tunnel barrier features and
conduction nonlinearity, and junction voltage bias.
|
1408.1004v1
|
## Title: Ventricular torsional relation to ventricular fiber arrangement
Authors: Saeed Ranjbar , Ph.D, Tohid Emami Meybodi , MD, , Mahmood emami Meybodi , MD
## Research institute:
- 1Modarres Hospital, Institute of Cardiovascular Research, Shahid Beheshti University of Medical Science, Tehran, Iran.
- 2Shahid sadoughi university of medical sciences, Yazd, Iran
## *Corresponding Author: Saeed Ranjbar
The full postal address of the corresponding author:
Modarres Hospital, Institute of Cardiovascular Research, Shahid Beheshti University of Medical Science, Tehran, Iran
E-mail: [email protected]
Tel/FAX: +9821 22083106
## Abstract:
Left ventricular torsion from helically oriented myofibers is a key parameter of cardiac performance. Physicians observing heart motion on echocardiograms, during cardiac catheterization, or in the operating room, are impressed by the twisting or rotary motion of the left ventricle during systole. Conceptually, the heart has been treated as a pressure chamber. The rotary or torsional deformation has been poorly understood by basic scientists and has lacked clinical relevance. The aim of this paper attempts to discuss about this question: Is ventricular twisting related to ventricular fiber arrangement? That is dependent to an assumed model of the left ventricular structure.
Key words: Left ventricular torsion, Heart Models, Myocardial fiber arrangement
## Ventricular torsional relation to ventricular fiber arrangement
The twisting motion of the heart is believed to be secondary to the arrangement of the muscle fibers. Pettigrew [1] performed careful dissection of the heart of mammals and man, demonstrating 7 muscle layers. The three outer layers spiral with an increasing angle from the perpendicular, while the fourth layer is horizontal. The three inner layers spiral in the opposite direction, increasing toward the vertical. The layers are arranged in opposition so that 1 opposes 7, 2 opposes 6, and 3 opposes 5, with the fourth layer being a connecting layer. Pettigrew postulated that one triplet of muscles contracts during systole and the other stores energy that is utilized in diastole. In his view, the motion of the heart muscle is like that of a torsional pendulum. A reproduction of one of his anatomic dissections can be seen in Fig.1 [38].
Figure 1: Anatomical dissections
Myocardial fiber orientation was examined by Streeter et al. [2] He reported that there was a well-ordered distribution varying from 60° (from the circumferential axis with positive being toward the base) on the endocardium to approximately -60° on the epicardium of the heart. He found that this fiber angle increased in systole by approximately 7° near the base and 19° near the apex relative to their counterparts in diastole, suggesting a torsion during contraction of die left ventricle.
More recently, Fernandez-Tran and Hurle [3] confirmed that there are three muscle fiber layers in the heart wall based on orientation. They described superficial (subpicardial), middle and deep (subendocardial), muscle layers which are similar for both the left and right ventricle with the exception of the middle layer which is found only in the left ventricle.
The opposing epicardial and endomyocardial muscle bundles might be important functionally. If potential energy is stored in these muscle fibers, the sudden release might promote filling during early diastole or the period of rapid ventricular filling. The proposed muscle bundles would also give a morphological correlate to the twisting or wringing motion of the left ventricle.
Since the discovery of the helical ventricular myocardial band by Francisco Torrent-Guasp >50 years ago [4] and the functional impact of the myocardial band [5, 6], many scientists have debated the validity of this concept. In any case, Torrent-Guasp's helical spiral concept may find an ideal connection with the spiralization of the outflow tracts and great arteries probably being related to the asymmetric intracardiac flow [7, 8] and/or to the spiral pattern detected at cellular/molecular level [9]. If the normal rightward spiralization represents the best hydrodynamic pattern, our first aim should be to try to restore the geometric pattern as close as possible to normality, avoiding, as it sometimes happens, the most abstruse surgical choice. In conclusion, the form and function of the heart are inevitably interdependent and we believe this to be true at each phylogenetic and ontogenetic stage. To treat heart diseases as best we can, we must try to understand the structure, function and deep mechanisms of the normal heart from its 'origin'. Once we have acquired as much knowledge as possible in this enormous field, we will then need to do the most difficult thing: mimic nature! [39]
## Models of heart structure
Eight conceptual models of cardiac structure are summarized in Fig. 2 and Fig. 3. The variety of proposed structures is visually striking this is due, in a large part, to the different levels of cardiac structure described. Model 6 (Fig. 2, part 6) groups the myocardium into regional functional units analogous to skeletal muscles. Models 1-5 (Fig. 2, parts 1-5) are continuum concepts stressing to differing degrees the anisotropic interconnectivity inherent in cardiac structure. Models 1 and 2 (Fig. 2, parts 1 and 2) describe fiber orientation. Model 3 (Fig. 2, part 3) describes changes through layers from epicardium to endocardium. Models 4 and 5 (Fig. 2, parts 4 and 5) examine the myolaminar structure, with less emphasis on fibre orientation. Some descriptions of Model 7 (Fig. 2, part 7), the HVMB, have emphasized discrete structural bundles [10] while other reports have proposed the band within a continuum framework [11]. Model 8 describe a dynamic orientation contraction (through the cardiac cycle) of every individual myocardial fiber could be created by adding together the sequential steps of the multiple fragmented sectors of that fiber. This way we attempted to mechanically illustrate the global LV model Fig. 3. We serve to describe explicitly three models of heart that one has been provided by authors. [40]
Figure 2: Mechanisms of heart in different models
Doughnut and pretzel models - Models 1 and 2 (Fig. 2, parts 1 and 2)
The fiber path models are based on the synthesis of many measurements of angles of fiber orientation through the ventricular wall and differ primarily in their scope (Model 1 is left ventricle (LV) only; Model 2 is LV and right ventricle (RV)). No argument is made for discrete traceable fiber paths - rather a recognizable general fiber trajectory [12]. These models reconstruct the long accepted helical pattern of fiber orientation, which has recently been confirmed by DT-MRI [13, 21-26] (Fig. 3). DT-MRI fiber tracing algorithms (Zhukov and Barr [14] and Kondratieva et al. [15], applied to canine data; Schmid et al. [13], applied to porcine data; Rohmer et al. [16], applied to human data) and automatic computational constructive visualization methods (Chen et al. [17,18]), used to extract meaningful visual information from histologically recorded three-dimensional fiber orientation datasets [19,20] (recorded from the rabbit heart), produce models remarkably similar to that proposed by Streeter (Fig. 2, part 1).
Peskin [27] has carried out an asymptotic analysis of Model 1 and has derived the fiber architecture of the heart from first principles.
These models form a conceptual model of cardiac structure from the one macro- and microscopically observed characteristic of principal fiber direction. As such, the nested doughnuts/pretzel surfaces are abstract concepts no discrete biological equivalents exist. It is possible to create the surfaces by dissection following the observed principal fiber direction, as demonstrated in early studies by Torrent-Guasp [12,28], but in so doing much information is lost. The dissection and histological methods used do not record data of local branching in directions other than the predominant fiber orientation, so by definition no detail of local tissue organization is modelled. Locally branching fibers are smoothed to a single orientation, which is again smoothed to a global ventricular fiber orientation. With reference to Grant's principles, these models represent the highest order schema of whole heart fiber orientation in isolation. It might be assumed that the fiber maps produced by algorithmic analysis and computer visualization may be more than a conceptual model, but the considerations above apply equally. The DT-MRI fiber-tracing algorithms track the principle fiber orientation from the primary eigenvector only [40]. The automatic computational constructive methods were applied to histological fiber orientation datasets which only record principle fiber orientation. Being limited to tracking the principle fiber orientation alone, these methods cannot reconstruct any detail of myolaminar structure, and as such reproduce idealized fiber-tracing dissections. Much evidence points to myolaminae as the central feature of the wall motion mechanism [29-33]. As such the doughnut and pretzel models represent geometric abstractions of cardiac structure. Their primary uses may be in (i) refining more histologically detailed models for the constraint of fiber orientation and (ii) in modeling the spread of the cardiac action potential, the conduction of which is significantly influenced by principal fiber direction [34].
## Model 8 (Fig.3)
This model provides a dynamic orientation contraction (through the cardiac cycle) of every individual myocardial fiber could be created by adding together the sequential steps of the multiple fragmented sectors of that fiber. Our study shows that in normal cases myocardial fiber paths initiate from the posterior-basal region of the heart, continues through the LV free wall, reaches the septum, loops around the apex, ascends, and ends at the superior-anterior edge of LV [35, 41 ].
Figure 3: The rout of a myofiber in the left ventricle
## Philosophical approach for description of a complex structure
How can so many models co-exist? Some must be wrong or is it possible that different models are true representations of the heart when considered from different perspectives? One 1965 review on cardiac structure stands out as a philosophical foundation upon which all future studies should be based. Robert P. Grant, M.D., in his Notes on the Muscular Architecture of the Left Ventricle [36], puts forward the following principles (along with an excellent description of cardiac structure):
- (1) As evidenced from numerous dissections, a given fiber segment has branching connections with other fiber segments in several different directions -as such the myocardial structure is a syncytium-like arrangement [37].
- (2) The cardiac structure problem is therefore a three dimensional network problem,
- (3) Structure understood from study of the network depends on the level on which the structure is approached; whether an attempt is made to describe the predominant behavior in the entire ventricle(s) or at the other extreme, to describe the average branching from a specific location,
- (4) Many models of heart structure can therefore be proposed depending on the conceptual approach,
- (5) Statistical study of the branching must be added to geometry for an accurate picture of myocardial architecture,
- (6) In consideration of such statistical problems one approach is to construct models for different degrees of generalization of the problem,
- (7) No single model gives the whole story but together they provide a schema,
- (8) The existence of a syncytium-like arrangement does not in itself dictate that no separate 'bundles' are present; however, due to the complex structure it is possible to construct by dissection bizarre arrangements which have no underlying anatomical reality.
- (9) Any dissection of the myocardium may represent (a) a valid schema from an infinite set of valid schema, (b) a bizarre and meaningless pathway or (c) a grouping of fiber paths within the syncytium of such general shared fiber direction that it can be considered a physiological or anatomical bundle,
- (10) A unique anatomical bundle is not the same as a unique physiological bundle - a connection between separate anatomical structures may produce one physiological structure,
- (11) A statistical approach is required to demonstrate any independent anatomical entity.
When viewed from this perspective it is not surprising that many models exist, that these are not all mutually compatible and that argument continues. Further evidence for these principles is presented in a dissection study by Fox and Hutchins [37], where principle (1) is re-stated and emphasized: 'The only level of the network of cells that can be referred to accurately as a fiber is a single cell. The 'fiber' is often only one cell in length before it splits and branches'. When considering the structural debate it is logical to return to Grant's principles - an approach adopted in this article. It should be noted that although Grant suspected regional variations in fiber branching within the left ventricle, he was special whether these local prevalence would statistically warrant consideration as separate bundles. One feature of the myocardial structure problem Grant did not consider was that significant structural differences may exist between individuals of the same species. [40]
## Conclusion:
Studies to date indicate that the left ventricle undergoes a torsional deformation, twisting in a counterclockwise direction during systole and then returning in a clockwise direction during diastole. LV myocardial models enable physicians to diagnose and follow-up many cardiac diseases and the torsional, deformation is likely related to the myocardial models. The opposing bands of muscle in the ventricular wall appear to be morphologically correlated with function.
Disclosure:
There is no disclosure.
## References:
- [1]- Pettigrew JB: Design in Nature, Vol. 2, Longman's Green and Co., London, 1908.
- [2]- Streeter DD. Spotnitz HM, Patel DP, Ross J Jr., Sonnenblick EH: Fiber orientation in the canine left ventricle during diastole and systole. Circ Res 24:339-347, 1969.
- [3]- Fernandez-Teran MA, Hurle JM: Myocardial fiber architecture of the human heart ventricles. Anat Rec 204:137-147, 1982.
- [18]- Chen M, Clayton RH, Holden AV, Tucker JV. Constructive volume geometry applied to visualisation of cardiac anatomy and electrophysiology. Int J Bifurcat Chaos 2003;13.
- [19]- Chen M, Clayton RH, Holden AV, Tucker JV. Visualising cardiac anatomy using constructive volume geometry. In: Magnin IE, Montagnat J, Clarysse P, Nenonen J, Katila T, editors. Functional imaging and modelling of the heart. Lyon, France: Springer; 2003; 2:30-8.
- [20]- LeGrice I, Hunter P, Young A, Smaill B. The architecture of the heart: a data-based model. Philos T Roy Soc A 2001;1217-32.
- [21]- Hsu EW, Muzikant AL, Matulevicius SA, Penland RC, Henriquez CS. Magnetic resonance myocardial fiber-orientation mapping with direct histological correlation. Am J Physiol 1998;274:H1627-34.
- [22]- Scollan DF, Holmes A, Winslow R, Forder J. Histological validation of myocardial microstructure obtained from diffusion tensor magnetic resonance imaging. Am J Physiol 1998;275:H2308-1.
- [23]- Scollan DF, Holmes A, Zhang J, Winslow RL. Reconstruction of cardiac ventricular geometry and fiber orientation using magnetic resonance imaging. Ann Biomed Eng 2000;28:934-44.
- [24]- Hsu EW, Henriquez CS. Myocardial fiber orientation mapping using reduced encoding diffusion tensor imaging. J Cardiov Magn Reson 2001;3:339-48.
- [25]- Geerts L, Bovendeerd P, Nicolay K, Arts T. Characterization of the normal cardiac myofiber field in goat measured with MR-diffusion tensor imaging. Am J Physiol Heart Circ Physiol 2002;283:H139-45.
- [26]- Hsu EW. Myocardial fiber orientation mapping via MR diffusion tensor imaging. In: Joint EMBS-BMES conference. Houston, TX: IEEE; 2002; 11:69-70.
- [27]- Peskin CS. Fiber architecture of the left ventricular wall: an asymptotic analysis. Commun Pur Appl Math 1989;42:79-113.
- [28]- Torrent-Guasp F. The cardiac muscle. Madrid: Juan March Foundation Ltd.; 1973.
- [29]- Chen J, Liu W, Zhang H, Lacy L, Yang X, Song SK, Wickline SA, Yu X. Regional ventricular wall thickening reflects changes in cardiac fiber and sheet structure during contraction: quantification with diffusion tensor MRI. Am J Physiol Heart Circ Physiol 2005;289:H1898-907.
- [30]- LeGrice IJ, Takayama Y, Covell JW. Transverse shear along myocardial cleavage planes provides a mechanism for normal systolic wall thickening. Circ Res 1995;77:182-93.
- [31]- Costa KD, Takayama Y, McCulloch AD, Covell JW. Laminar fiber architecture and three-dimensional systolic mechanics in canine ventricular myocardium. Am J Physiol 1999;276:H595-607.
- [32]Arts T, Costa KD, Covell JW, McCulloch AD. Relating myocardial laminar architecture to shear strain and muscle fiber orientation. Am J Physiol Heart Circ Physiol 2001;280:H2222-9.
- [33]- Spotnitz HM, Spotnitz WD, Cottrell TS, Spiro D, Sonnenblick EH. Cellular basis for volume related wall thickness changes in the rat left ventricle. J Mol Cell Cardiol 1974;6:317-31.
- [34]Clayton RH. Computational models of normal and abnormal action potential propagation in cardiac tissue: linking experimental and clinical cardiology. Physiol Meas 2001;22:R15-34.
- [35] Ranjbar S., Karvandi M., Ajzachi M. System and method modeling left ventricle of heart. US Patent. 2013; patent number: 8,414,490.
- [36]- Grant RP. Notes on the muscular architecture of the left ventricle. Circulation 1965;32:301-8.
- [37]- Fox CC, Hutchins GM. The architecture of the human ventricular myocardium.Johns Hopkins Med J 1972;130:289-99.
- [38]- Frank A. Baciewicz Jr., David G. Penney, William A. Marinelli, Ralph Marinelli. Torsional Ventricular Motion and Rotary Blood Flow. RESEARCH IN PROGRESS
- [39]- Antonio Amodeo. Marco Oliverio. Paolo Versacci. Bruno Marino. Spiral shapes in heart and shells: when form and function do matter. European Journal of Cardio-Thoracic Surgery. 2012; 41: 473-475
- [40]- Stephen H. Gilbert , Alan P. Benson, Pan Li, Arun V. Holden. Regional localization of left ventricular sheet structure: integration with current models of cardiac fiber, sheet and band structure. European Journal of Cardio-Thoracic Surgery. 2007; 32: 231-249.
- [41]- Mersedeh Karvandi, Saeed Ranjbar, Seyed Ahmad Hassantash. Mechanical mitral valve modeling: Advancing the field through emerging science. International Journal of Medical Imaging. 2014; 2(2): 24-28.
| [4]- Torrent-Guasp F. Anatomia Funciónal del Corazón. La actividad ventricular diastólica y sistólica. Madrid: Paz Montalvo, 1957. |
|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [5]- Torrent-Guasp F, Kocica MJ, Corno A, Komeda M, Cox J, Flotats A et al. Systolic ventricular filling. Eur J Cardiothorac Surg 2004;25: 376-86. |
| [6]- Kocica MJ, Corno AF, Carreras-Costa F, Ballester-Rodes M, Moghbel MC, Cueva CN et al. The helical ventricular myocardial band: global, threedimensional, functional architecture of the ventricular myocardium. Eur J Cardiothorac Surg 2006;29:S21 40. |
| [7]- Kilner PJ, Yang GZ, Wilkes AJ, Mohiaddin RH, Firmin DN, Yacoub MH. Asymmetric redirection of flow through the heart. Nature 2000;404: 759-61. |
| [8]- Corno AF, Kocica MJ, Torrent-Guasp F. The helical ventricular myocardial band of Torrent-Guasp: potential implications in congenital heart defects. Eur J Cardiothorac Surg 2006;29: S61-8. |
| [9]- Pasipoularides A. Heart's Vortex: Intracardiac Blood Flow Phenomena. Shelton, CT: People's Medical Publishing House, 2010. |
| [10]- Torrent-Guasp F, Ballester M, Buckberg GD, Carreras F, Flotats A, Carrio I, Ferreira A, Samuels LE, Narula J. Spatial orientation of the ventricular muscle band: physiologic contribution and surgical implications. J Thorac Cardiovasc Surg 2001;122:389-92. |
| [11]- Kocica MJ, Corno AF, Carreras-Costa F, Ballester-Rodes M, Moghbel MC, Cueva CNC, Lackovic V, Kanjuh VI, Torrent-Guasp F. The helical ventricular myocardial band: global, three-dimensional, functional architecture of the ventricular myocardium. Eur J Cardiothorac Surg 2006;29:S21-40. |
| [12]- Streeter Jr DD. Gross morphology and fiber geometry of the heart. Baltimore: Williams and Wilkins; 1979; 23: 61-112. |
| [13]- Schmid P, Jaermann T, Boesiger P, Niederer PF, Lunkenheimer PP, Cryer CW, Anderson RH. Ventricular myocardial architecture as visualised in postmortem swine hearts using magnetic resonance diffusion tensor imaging. Eur J Cardiothorac Surg 2005;27:468-72. |
| [14]- Zhukov L, Barr AH. Heart-muscle fiber reconstruction from diffusion tensor MRI. In: Proceedings of the 14th IEEE visualization 2003 (VIS'03). IEEE Computer Society; 2003. |
| [15]- Kondratieva P, Kruger J, Westermann R. The application of GPU particle tracing to diffusion tensor field visualization. In: Proceedings IEEE visualization 2005. |
| [16]- Schmid P, Jaermann T, Boesiger P, Niederer PF, Lunkenheimer PP, Cryer CW, Anderson RH. Ventricular myocardial architecture as visualised in postmortem swine hearts using magnetic resonance diffusion tensor imaging. Eur J Cardiothorac Surg 2005;27:468-72. |
| [17]- Rohmer D, Sitek A, Gullberg GT. Reconstruction and visualization of fiber and sheet structure with regularized tensor diffusion MRI in the human heart. Lawrence Berkeley National Laboratory Publication. LBNL-60277, 2006 |
| null |
[
"Saeed Ranjbar",
"Tohid Emami Meybodi",
"Mahmood Emami Meybodi"
] |
2014-08-02T22:24:59+00:00
|
2014-08-02T22:24:59+00:00
|
[
"physics.med-ph"
] |
Ventricular torsional relation to ventricular fiber arrangement
|
Left ventricular torsion from helically oriented myofibers is a key parameter
of cardiac performance. Physicians observing heart motion on echocardiograms,
during cardiac catheterization, or in the operating room, are impressed by the
twisting or rotary motion of the left ventricle during systole. Conceptually,
the heart has been treated as a pressure chamber. The rotary or torsional
deformation has been poorly understood by basic scientists and has lacked
clinical relevance. The aim of this paper attempts to discuss about this
question: Is ventricular twisting related to ventricular fiber arrangement?
That is dependent to an assumed model of the left ventricular structure.
|
1408.1005v1
|
## LHC prospects for minimal decaying Dark Matter
## Giorgio Arcadi, Laura Covi and Federico Dradi
Institute for Theoretical Physics, Georg-August University Göttingen, Friedrich-Hund-Platz 1, Göttingen, D-37077 Germany
Abstract. We study the possible signals at LHC of the minimal models of decaying dark matter. Those models are characterized by the fact that DM interacts with SM particles through renormalizable coupling with an additional heavier charged state. Such interaction allows to produce a substantial abundance of DM in the early Universe via the decay of the charged heavy state, either in- or out-of-equilibrium. Moreover additional couplings of the charged particle open up decay channels for the DM, which can nevertheless be sufficiently long-lived to be a good DM candidate and within reach of future Indirect Detection observations. We compare the cosmologically favored parameter regions to the LHC discovery reach and discuss the possibility of simultaneous detection of DM decay in Indirect Detection.
| Contents | Contents | Contents |
|------------|-------------------|------------|
| 1 | Introduction | 1 |
| 2 | The model | 2 |
| 3 | Collider analysis | 4 |
| 3.1 | Colored scalar | 6 |
| 3.2 | EW-charged scalar | 14 |
| 4 | Discussion | 20 |
| 5 | Conclusions | 23 |
## 1 Introduction
After the first run of the LHC, no signal of New Physics has been found yet in the traditionally DM-motivated channels containing missing energy [1, 2]. While the next LHC run could still bring a WIMP-like signal, especially from the electroweakly charged sector, we would like here to consider a different type of scenario, connected to the cosmologically well-motivated case of decaying Dark Matter.
In fact, it is perfectly possible that the DM particle is not stable, but just very longlived and in such case immediately two usual DM assumptions are lifted: a) no symmetry is introduced to justify the stability of Dark Matter and therefore the particle can interact also as single state with charged SM and non-SM states; b) to ensure survival until the present age, the DM particle has to interact with suppressed couplings to any sector and therefore can be naturally produced not through the WIMP mechanism, which relies on electroweaksize cross-section, but via the FIMP or SuperWIMP mechanisms, based on much smaller couplings.
A very minimal setting of this type has been recently proposed in [3], featuring a DM Maiorana fermion and a scalar charged under the SM gauge group. It is minimal in the sense that just a minimal particle content has been added to the SM and that the additional couplings are all renormalizable. Such a scheme can be embedded in more complex models, like supersymmetry with R-parity violation, but the main phenomenological characteristics are independent from the particle physics framework, at least for what concerns DM phenomenology. We will thus here concentrate on such minimal constructions as the more conservative case, since in the presence of a larger field content or more couplings, additional signals could be accessible at LHC or in DM indirect detection.
Due to the long lifetime of the DM, requiring at least one of the additional charged particles to be within the LHC reach constrains all couplings to be suppressed and therefore points also to regions of (near)-collider-metastability for these exotic states. It is therefore natural to look for such particles at the LHC in the presence of displaced vertices or metastable tracks (MP). These channels have recently attracted more attention [4-15] and are indispensable in order to cover a wide range of cosmologically interesting scenarios [16-23] like the one presented here.
This paper is organized as follows: we will discuss the type of minimal models we study in section 2 together with the connection to DM indirect detection and the DM production mechanisms. We will then study in section 3 the LHC phenomenology in the two scenarios of scalar field carrying color or only electroweak charge. Section 4 will be then devoted to the discussion of the results while our conclusions will be given in Section 5.
## 2 The model
We consider here the minimal model introduced in [3] (see also [24-26] for simular setups) featuring a Maiorana fermion, singlet with respect to the SM gauge group and dark matter candidate, and a single scalar field multiplet Σ f , non-trivially charged under at least one of the SM gauge groups. We assume that these fields interact among themselves and with the SM only via renormalizable Yukawa-type couplings according to the quantum number chosen for Σ f . In particular the interaction between the DM and Σ f will be of the form:
$$L _ { \text{eff} } = \lambda _ { \psi f } \ \bar$$
where ψ is the DM Majorana fermion, f any chiral SM fermion and Σ f denotes a scalar with quantum numbers equal to f .
No symmetry is imposed to guarantee the stability of the DM, allowing couplings of Σ f with only SM fermions. Depending on the quantum numbers a rather broad variety of operators may arise:
$$\begin{array} {$$
where q, glyph[lscript] denote the SM SU (2) L LH doublets, while u, d, e are RH SU (2) L singlets, and the superscript c indicates the charge-conjugated field, f c = Cf ¯ t while ˜ Σ f ≡ iσ 2 Σ ∗ f . On the right the quantum numbers of the Σ f fields are specified according to the SM gauge groups SU (3) c × SU (2) L × U (1) Y . We are here suppressing flavour indices, even if some couplings like ¯ qq , glyph[lscript]glyph[lscript] c ¯ c must be antisymmetric in flavour and vanish for a single generation, and considering in each line the presence of a single scalar field Σ f . We have then that the new particle sector can just be described by two mass scales m ψ and m Σ and a few Yukawa couplings. The scalar field Σ f is also coupled, according to its assignment of quantum numbers, given above, to the SM group gauge bosons.
The interactions (2.1) and (2.2) induce three-body decays for the Dark Matter with a rate given by, up to kinematical and multiplicity factors:
$$\Gamma _ { \text{DM} } = \frac { c _ { f } | \$$
with c f = counting the number of degrees of freedom of the intermediate Σ f ( c f = 6 3 3 2 1 , , , , for Σ q , Σ u , Σ d , Σ glyph[lscript] , Σ e respectively). The decay channels can be different depending on the
quantum numbers of the intermediate particle Σ f . We distinguish substantially four types of decay channels:
$$\psi & \to \ \bar { u } u \nu, \bar { d } d$$
For simplicity we focus on signals of the type ffν with f being either a quark or a charged lepton of any generation, denoted by l . We will also assume that the decays of DM are flavour conserving. If this is not the case bounds from flavour violation decays of mesons and leptons arise (see e.g. [27] for a complete list). However they are sensitively weaker than the one imposed by cosmology and DM ID. We will also assume that, for a given assignment of the quantum numbers of Σ f , only one of the allowed operators, as reported in (2.2), dominates. We just comment that some of the effective lagrangians (2.2) violate both lepton and baryon number and in case of contemporary presence of lepton and baryon number violating operators, very strong constraints from the stability of the proton arise [28]. We will neglect here this possibility.
All the considered DM decay channels are already severely constrained from Indirect Detection. The hadronic decay channels, namely ddν and uuν are mostly constrained by antiproton searches, which give bounds on the DM lifetime varying between 10 26 -10 28 s for m ψ > 100 GeV [26] and can become as stringent as 10 29 s for values of the DM mass down to 1 GeV [29]. In the case of the leptonic decays, comparably severe constraints, ranging from approximately 10 27 s to 10 29 s, according to the decay channel, for DM masses between 10 GeV and 2 TeV, are obtained by the recent measurements by AMS of the positron flux and positron fraction [30]. In the case of the Σ glyph[lscript] mediator, one-loop induced decay processes in Zγ and νγ may also be important. In particular the latter can originate monochromatic γ -ray lines. Current searches give bounds which can be as strong as 10 29 ÷ 30 seconds for m ψ = 1 -10 GeV [24, 31, 32]. All these bounds can be satisfied only for a very small value of the product of the couplings, namely λ ψf ¯ λ glyph[lessorsimilar] 10 -(16 ÷ 22) , for masses of the scalar field Σ f within the kinematical reach of the LHC.
While the present lower limits on the DM lifetime only constrain the product of the couplings, the DM production mechanisms in the early universe also provide bounds on the ratio of the couplings. In fact, in this kind of models, the correct amount of DM relic density can be generated from the decays of the field Σ f either in thermal equilibrium (freeze-in production [19, 33-36]) or out-of-equilibrium (SuperWIMP production [37-39]) 1 .
Since the two processes take place at quite different cosmological epochs, the DM relic
1 It is as well possible, alternatively, that the DM acquires its relic density through the WIMP paradigm. However this setup corresponds to the simple scenario with λ ψf ∼ O (1) and an irrelevant coupling of Σ f to SM fields. We will not consider this possibility in this work (see instead e.g. [40] for a scenario of this kind with DM coupling to the top quark).
density can expressed, in very good approximation, as the sum of the two contributions [3]:
$$\mathbb { m } _ { \Sigma } \right. \right. \right. \$$
where x = m ψ m Σ , g Σ and g ∗ are, respectively, the internal degrees of freedom of the field Σ f and the relativistic degrees of freedom of the primordial plasma at the time of DM production, τ Σ is the lifetime of the Σ f field defined, up to kinematical factors, as:
$$\tau _ { \Sigma } ^ { - 1 } = \Gamma _ { \Sigma } = \frac { | \lambda _ { \psi \, f } | ^ { 2 } + | \overline { \lambda } | ^ { 2 } } { 8 \pi } \, m _ { \Sigma }$$
while Γ(Σ f → fDM ) and Br (Σ f → fDM ) are the decay rate and branching fraction of the field Σ f into DM. Neglecting the final state masses, we have then:
$$\Gamma \left ( \Sigma _ { f } \rightarrow f D M \right ) = \frac { | \lambda _ { \psi f } | ^ { 2 } } { 8 \pi } \, m _ { \Sigma }, \ \ B r \left ( \Sigma _ { f } \rightarrow f D M \right ) = \frac { | \lambda _ { \psi f } | ^ { 2 } } { | \lambda _ { \psi f } | ^ { 2 } + | \bar { \lambda } | ^ { 2 } } \, \quad \ \ ( 2. 7 )$$
We note that the freeze-in contribution is proportional to the decay rate of Σ f into DM, which is in turn proportional to λ Σ f , independently of the other coupling, while the SuperWIMP contribution depends on the branching fraction of decay of Σ f into DM as well as on its mass and charge, strongly influencing the relic density Ω Σ h 2 . But in general from eq. (2.5) we see that both production mechanisms are inefficient if the branching fraction of Σ f decay into DM becomes too small. On the other hand a too large DM coupling to Σ f can easily cause overproduction. Imposing the cosmological value of the Dark Matter density for ψ fixes some definite ranges of the couplings λ, λ or equivalently τ Σ , Br (Σ f → fDM ) as a function of the mass scales, m ,m ψ Σ .
In the next sections we will investigate the reach of LHC in the detection of the charged field Σ f and investigate which signal can be expected in the parameter regions favored by a successful cosmological DM production and possible DM decay.
## 3 Collider analysis
Contrary to the DM, the scalar field Σ f is charged under SM gauge interactions which may give its efficient production at the LHC, if kinematically allowed. Since the Yukawa couplings with the quarks are much smaller than any of the gauge couplings, the main production channels at a proton-proton collider are gluon fusion into a scalar-antiscalar pair, for colored Σ f , or Drell-Yan production, for the electroweakly or electromagnetically charged case. In either cases the production rate is practically independent on the details of the DM model, and given just by the mass and charge of the field Σ f . We will estimate here the NLO production rates by computing the LO cross-section with the package MadGraph 5 [41] and correcting with a constant NLO k-factor, depending on the channel.
For any given assignment of its quantum numbers the scalar particle features two kind of decay channels after its production (We are implicitly assuming that in the case of the SU (2) doublets Σ q and Σ glyph[lscript] the two components are mass degenerate. If this is not the case, in
addition to the processes described below, the decay of the heavier component of the doublet into a W, either on or off shell, and the lighter one is open. We will better clarify this point later in the text). We have first of all decays into a DM particle and a standard model fermion, with rate proportional to λ 2 ψf , which can be classified as follows:
$$\begin{array} { c c c c } \Phi & & \\ & \Sigma _ { f } \to & u \psi & \text{for} & \Sigma _ { q }, \Sigma _ { u }, \Sigma _ { d } \\ & \Sigma _ { f } \to & d \psi & \text{for} & \Sigma _ { q }, \Sigma _ { d } \\ & \Sigma _ { f } \to & l \psi & \text{for} & \Sigma _ { \ell }, \Sigma _ { e } \\ & \Sigma _ { f } \to & \nu \psi & \text{for} & \Sigma _ { \ell } \end{array}$$
where l is a charged lepton. We see that only in the case of Σ glyph[lscript] the decay can be into an invisible final state νψ , but in that case also the visible channel into a charged lepton is present. So the decays in general give rise to a kink in the observable charged track/jet due to the Σ f decay.
The scalar field can decay as well into two SM fermions, with a rate governed by λ 2 , according the following channels:
$$\text{g.m.s.} \\ \Sigma _ { f } & \to \ q ^ { ^ { \prime } } _ { \ } ^ { \ } = \text{for} \quad \Sigma _ { u }, \Sigma _ { d }, \Sigma _ { \ell } \\ \Sigma _ { f } & \to \ q l \quad \text{for} \quad \Sigma _ { q }, \Sigma _ { d } \\ \Sigma _ { f } & \to \ q \nu \quad \text{for} \quad \Sigma _ { q }, \Sigma _ { d } \\ \Sigma _ { f } & \to \ l \bar { l } \quad \text{for} \quad \Sigma _ { \ell } \\ \Sigma _ { f } & \to \ l \nu \quad \text{for} \quad \Sigma _ { \ell }, \Sigma _ { e } \\ \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \end{array}$$
where l is, again, a charged lepton while q is an up or down-type quark. In view of the dependence of the decay rate of Σ f on the λ ψf , λ -type coupling, a tight relation exists between possible signals at LHC of such decays and the constraints from the DM phenomenology, being governed by the same couplings. As already argued in [3] the constraints from ID and from the cosmological abundance of the DM require very low values of the couplings, namely λ, λ glyph[lessorsimilar] 10 -(7 ÷ 8) , thus implying that the decay vertices result displaced with respect to the production ones and may even lie outside the detector.
In order to determine the LHC capability of detecting this kind of decays we will adopt the method introduced in [42]. We have then identified three possible detection regions, referring for definiteness to the design of the CMS detector (see [43] for a detailed description of the detector), namely the Pixel and the Tracker (inner detector), and an 'outside' region. The inner detector regions are sensitive to the shorter lifetimes and allow for the detection of displaced vertices (see for example [44], as well as [45] for similar searches from the ATLAS collaboration). The 'outside' region is instead suitable for the detection of very long-lived particles , whose corresponding signal is constituted by outgoing tracks, thus related to the scalar field itself rather than its decays. On recent times searches of decay of stopped particles [46] in the detector have been considered as well. This is an very intriguing possibility since this last kind of searches result complementary to the ones of outgoing tracks. We will further comment on this later in the text.
We have generated several samples of pair produced Σ f , at 14 TeV of centre of mass energy, corresponding to different assignments of its quantum numbers and different masses, and determined the spatial distribution of the decay vertices from the kinematic variables of the events and the decay rate Γ Σ as described in [42]. Note that we consider here a straight-line motion of the particle after production, neglecting the magnetic field deflection
and the interactions with the intervening matter, which could even increase the number of decays in the inner part of the detector by bending the trajectory or slowing down the decaying particle 2 . Assuming 100 % detector efficiency, in order to claim the discovery for a given scenario we have required the presence of at least 10 decay events in one of the components in which the detector is schematized, i.e. in the pixel or the tracker, or outside the detector 3 . The optimal scenario, and thus main focus of our analysis, is however a 'double' LHC detection scenario, consisting in the contemporary detection of at least 10 events in one of the components of the inner detector, namely pixel or tracker, and 10 tracks leaving the detector. This indeed would allow for a cross-check in the measurement of the lifetime of the scalar particle as well as a better discrimination of possible backgrounds.
For each of the cases considered we have performed the analysis for three definite luminosities, namely 25 fb -1 , 300 fb -1 and 3000 fb -1 , in order to determine the feasibility of a next future discovery during run II, as well as the maximal discovery reach considering the full LHC data set and, finally, after a high luminosity run.
Since the analysis employed is not sensitive to the particular type of decay products of Σ f , as long as a vertex (or kink) can be observed and happens in the detector, we will from now on refer to a schematic setup described by just four model parameters: the mass m Σ f of the scalar field, the ratio x = m /m ψ Σ f and two couplings λ and λ ′ , referring, respectively, to the decay of Σ f to the DM or to only SM states. We have as well considered an equivalent representation in terms of the Σ f lifetime and DM branching fraction, more directly connected to the phenomenological observables at the LHC, and also helpful to translate the results obtained in different particle physics setups. In either case we will find the region of parameter space where displaced vertices and/or metastable tracks may be seen and compare it to the cosmologically viable parameter space.
As will be discussed later, it would be very important, in order to relate an hypothetical LHC signal to the DM properties, to distinguish both the decay channels of Σ f . A necessary condition for the identification of a particular channel is that the product of the total number of events times the corresponding branching fraction is large enough, in one of the detector regions where it is possible to observe the decay products of the scalar field. A proper determination of the number of events needed would require the full detector simulation, accounting for the capability of reconstruction of the various decay products.
In the following subsections we will investigate separately the scenarios of color and electroweakly charged scalar particle.
## 3.1 Colored scalar
The first case that we are going to consider is when the field Σ f carries color charge. Colored states are expected to be more efficiently produced at the LHC. For definiteness we will consider a Σ d -type field in our analysis. As already mentioned, the Σ f = q,u,d pairs are produced through gluon fusion and thus the production cross section is substantially the same for the three kind of states, apart a possible enhancement in the case of Σ q because of multiplicity. We will indeed assume, for this scenario, that the two components of the doublet Σ q are exactly degenerate in mass. If this is not the case the heaviest state of the doublet could decay into the lightest one and a W boson, if kinematically possible, or two quarks or leptons (through
2 Such effects could be captured only by a full detector simulation, which is beyond the scope of this paper. 3 Notice that this is a rather conservative requirement which prevents the effect of statistical fluctuations and ensures stable numerical results. For a pure Poisson distribution and no background 5 such events correspond to a discovery at the 95% CL.
an off-shell W). For mass splittings above ∼ 1 GeV this decay channel has a branching fraction substantially equal to one and leads to prompt decays of the heavy state in case its production is accessible at the LHC. For sizable enough mass splittings, such that the momentum of decay products can survive analysis cuts (e.g. quality of signal requirements, background discrimination cuts), this decay can be detected and, then, the signals discussed in the following would result accompanied by prompt jets or leptons. On the other hand the required mass splitting would imply a sensitive suppression of the pair production of the heaviest states of the doublet and thus a small number of this kind of events. This might not be the case of the Σ glyph[lscript] -type particles, as will be clarified in the next subsection.
According to the method described above we have generated samples of events of Σ d pair production normalizing the cross-section computed by Madgraph with a k-factor, accounting for NLO effects, which has been taken from the numerical package Prospino [47], given the similarities with the supersymmetric case of a stop squark. From the determination of the production cross-section it is possible to infer a general upper limit on the LHC reach at a given luminosity L , from the relation N ev = σ pp → Σ Σ d ∗ d L ≥ 5 (10 events correspond to 5 pair produced Σ ) where N ev represents the number of produced pairs Σ Σ d ∗ d irrespectively of the position of the decay vertices. For the luminosities considered in our analysis the LHC reach ranges from around 1600 GeV at L = 25 fb -1 to a maximal value of 2200 GeV corresponding to L = 3000 fb -1 . We have then computed the spatial distribution of the decay vertices for several values of the mass of Σ d , namely 800, 1600 and 2200 GeV. Masses below 800 GeV are currently excluded, for the range of lifetimes relevant for our analysis, by current searches of detector stable particles [48, 49].
Figure 1 : LHC reach in the plane of the pure SM λ ′ vs DM coupling λ for (from left to right) L = 300 3000 , fb -1 for m Σ d = 800 GeV. The region below the solid line corresponds to at least 10 metastable tracks, while the regions above the dashed/dash-dotted lines to at least 10 decay events in the tracker or pixel detector respectively.
We report two examples of results of our analysis in fig. (1), for m Σ d = 800 GeV and L = 300 3000 , fb -1 , where we have identified the region corresponding to more than 10 decay events in the pixel, tracker and outside the detector. We see that the contours run piecewise
parallel to the axis, since in most of the parameter space one single coupling dominates the total decay rate. The detection regions for pixel and tracker are very similar, since the difference in volume is practically compensated by the distance from the interaction point.
The searches for displaced vertices and particles escaping from the detector are highly complementary: the first has a maximal reach at low lifetimes, corresponding to high values of λ, λ ′ , while the latter is able to probe efficiently the very long lifetimes, i.e. low values of λ, λ ′ . Combining both search strategies allows to cover practically the whole parameter space.
The strong requirement of detection of both types of signal, i.e. displaced vertices in the pixel/tracker and metastable tracks, is realized only in the narrow regions comprised between the iso-contours in the plane ( λ, λ ′ ) representing the detection of exactly 10 events in the pixel, tracker and outside the detector. The size of these strips is expected to increase with the integrated luminosity and, instead, to shrink once increasing the mass of the scalar, because of the lower number of particles produced. For the highest values of the mass, corresponding to approximately 2200 GeV the whole parameter space in the couplings might only be probed by the high luminosity upgrade of the LHC as long as both types of signal, either displaced vertex or metastable track, are considered.
The capability of LHC detection of a displaced decay of Σ d or its metastable track can be confronted with the requirement of the correct cosmological DM abundance via Σ d decay and, possibly, a detection of decaying DM. In case of a colored scalar the correlation between the DM phenomenology and the LHC predictions is rather straightforward since the DM production is substantially dominated by the first (freeze-in) contribution in eq. (2.5) and the couplings can be analytically determined as function of x , m Σ d and the DM relic density and lifetime as [3]:
$$\lambda \simeq 1. 5 9 \times 1 0 ^ { - 1 2 } x ^ { - 1 / 2 } \left ( \frac { g _ { * } } { 1 0 0 } \right ) ^ { 3 / 4 } \left ( \frac { \Omega _ { \text{CDM} } h ^ { 2 } } { 0. 1 1 } \right ) ^ { 1 / 2 } g _ { \Sigma } ^ { - 1 / 2 }$$
$$\lambda ^ { \prime } \simeq 0. 9 1 \times 1 0 ^ { - 1 2 } \, x ^ { - 2 } \left ( \frac { g _ { * } } { 1 0 0 } \right ) ^ { - 3 / 4 } \left ( \frac { m _ { \Sigma _ { f } } } { 1 T e V } \right ) ^ { - 1 / 2 } g _ { \Sigma } ^ { 1 / 2 } \left ( \frac { \tau _ { \psi } } { 1 0 ^ { 2 7 } s } \right ) ^ { - 1 / 2 } \left ( \frac { \Omega _ { \text{CDM} } h ^ { 2 } } { 0. 1 1 } \right ) ^ { - 1 / 2 } \quad ( 3. 3 )$$
From these relations we can determine the cosmologically preferred parameter space in the plane of the couplings. We show indeed in fig. (2), as solid lines, the isolines of the correct value of the DM relic density for m Σ d = 800 GeV and some values of x ranging from 10 -3 to 0 5 . . They appear as vertical lines since the freeze-in mechanim is independent from λ ′ . This curves can be confronted with the contours of the reach in the three detector regions. The panels of fig. (2) report the LHC reach for the three values of luminosity considered in our analysis. Fig. (3) shows an analogous analysis for the values m Σ d = 1600 2200 , GeV. In this case we have considered only L = 3000 fb -1 since we expect a statistically relevant number of events only for this very high luminosity. In fig. (2) and (3) we have reported as well (dashed lines) the values of the DM lifetime, for the chosen combinations of parameters, near the present bounds. As already stated, we have assumed, for the scenario of colored scalar field, uu dd ν ( ) as the only relevant decay channel for the DM and thus applied the bounds of [26] in the m ψ > 100 GeV region and of [29] 4 at lower masses. For any value of x , the intersection of the corresponding solid and dashed lines corresponds to a DM with the
4 The bounds presented here actually refer to two body decays in fermion pairs and then result conservative since in our setup part of the energy of the products is carried away by the neutrinos.
l ¢
10
10
-
-
10
6
8
-
10
10
10
10
-
12
-
14
-
16
pp fi S
outside pixel
tracker
MP
n
pi
<
n
tr
-
14
<
10
d
S *
d
&
m
S
d
=
800 GeV
excluded today
10
10
0.5
=
=
0.1
x
x
10
2
-
10
=
x
-
12
3
-
10
=
x
l
l ¢
10
-
25
fb
<
10
x
=
-
10
x
=
-
1
3
10
x
=
x
10
10
-
6
10
-
10
8
-
10
10
10
=
2
-
10
0.5
10
-
8
pp fi S
d
S *
outside pixel
tracker
n
pi
<
10
MP
n
excluded today
tr
10
-
14
<
10
=
0.5
1
2
-
-
10
x
=
x
10
10
=
x
-
12
10
800 GeV
&
L
=
3000
fb
-
1
n
out
-
10
10
<
x
=
-
10
3
x
=
10
x
=
-
2
-
10
x
=
0.5
10
-
8
l
Figure 2 : Contours of the correct DM relic density (solid lines) and of the reference value of 10 28 s of its lifetime (dashed lines) color-coded according to the values of x reported in the plot. A combined detection of Σ d at LHC and of the DM candidate via ID can be achieved if the lines of the DM relic density and lifetime, corresponding to a given value of x , intersect within the discovery region between the iso-contours labeled as pixel-tracked and outside. The plot refers to m Σ d = 800 GeV and L = 25 fb -1 (left plot), L = 300 fb -1 (left plot), L = 3000 fb -1 (bottom plot). The yellow shaded region delimited by the yellow thick long-dashed line is already excluded by current searches for metastable particles.
correct relic density and a lifetime approximately coinciding with the current observational bounds for the assumed dominant decay channel; consequently the parameter space above the DM lifetime curves may be already excluded. Notice that the actual ID exclusion region depends strongly on the DM decay channel and DM mass and is affected from astrophysical uncertainties in the propagation modeling as discussed in [26].
The most favorable scenario, consisting in a multiple detection of the DM and Σ d decay, respectively in cosmic rays and at the LHC, with the latter satisfying the double LHC
1
d
&
1
10
-
12
-
14
-
16
&
L
=
n
out
-
m
S
d
=
3
-
10
=
x
Figure 3 : The same as fig. (2) but with m Σ d = 1600 GeV (left plot) and m Σ d = 2200 GeV (right plot) and L = 3000 fb -1 in both cases.
detection requirement, is potentially feasible, for a given value of the pair ( x, m Σ d ) , when the corresponding isolines of the DM relic density and lifetime intersect inside the double detection region. Outside this region the contemporary ID detection of DM and only one type of LHC signal, i.e. displaced vertices (above the white strip) or metastable tracks (below the white strip), is anyway still feasible. We remark, however, that the region below the white strip, corresponding to very long lifetimes, is already constrained, for the lower values of the mass of the scalar, by current searches of detector stable particles. We have reformulated, for the scenario under consideration, these constraints using the procedure described in [42] and reported the excluded region in fig. (2). The region below the 'double' detection strip is nearly ruled out for m Σ d = 800 GeV. The limit from detector stable particles weakens very quickly with increasing mass of the scalar particle and it is substantially irrelevant for masses above 1 TeV.
The double LHC detection region corresponds, for m Σ d = 800 GeV, to a rather definite range of values of x comprised between 10 -2 and 10 -1 . This range is reduced at higher values of m Σ d because of the decreased size of the LHC double detection strip. For the highest possible value m Σ d the combined detection prospects are substantially limited to x glyph[similarequal] 10 -2 .
It is also interesting to reformulate the previous results in a more model independent way in terms of the pair ( Γ -1 Σ ( s ) , Br (Σ d → DM ) ) as done in fig. (4), (5). Using these parameters the LHC detection regions are just delimited by horizontal lines of constant Γ Σ values. The green (violet) shaded regions in the plot represent the regions in which it is possible to detect more than 10 events in the pixel/tracker (outside region). 'Double' signals are accessible instead in the middle white strip. The combined LHC detection of the Σ d , in at least one of the two channels, and DM indirect detection are again achieved whenever the isolines (dashdotted) of the reference DM lifetime and of the correct relic density (solid) for a fixed value of x cross in the LHC 'double' discovery region. Above this strip it is still possible to observe displaced vertices at the LHC and have an ID DM signal for small values of x . At large values of x , instead, only metastable particle signals are compatible with DM ID.
As evident from the plots, simultaneous ID/LHC detection, with the latter in the form
Figure 4 : LHC detection reach compared with the constraints from DM phenomenology in the plane ( Br (Σ d → DM ) , Γ -1 Σ d ) . The shaded green (magenta) region corresponds to more than 10 decay events happening in the pixel or tracker (outside the detector). The double detection region is the white strip comprised between the shaded regions. We have fixed m Σ d = 800 GeV and L = 25 fb -1 (left plot) and L = 300 fb -1 (right plot). The yellow shaded region, below the thick long-dashed yellow line, is excluded by current bounds on metastable particles.
of a double signal, i.e. displaced decay plus metastable track, could be achieved in the next future, namely for luminosities up to 300 fb -1 , for values of Br Σ d → dψ lower than 10 -3 . As a consequence the observation at the LHC of both the DM and only SM decay channels of Σ d appears difficult but it is not a priori excluded provided that there is a high statistics.
In order to investigate this possibility we have performed a more focused study on a benchmark set of parameters, namely m Σ d = 800 GeV, corresponding to the maximal production cross-section, m ψ = 10 GeV and λ and λ ′ fixed to reproduce the correct DM relic density and a DM lifetime of approximately 10 28 seconds. We have reported on table (1) the number of decay events, expected at the three values of the luminosity, namely 25, 300 and 3000 fb -1 , together with an estimate of the number of events in the two different type of decay channels, namely in DM plus SM fermion and two SM fermions.
As evident, no events of decay into DM can be observed inside the detector (namely pixel or tracker), even in a high luminosity run of LHC. This is a consequence of the very suppressed branching ratio of decay of Σ d into DM, ≈ 10 -3 , required to reconcile the correct DM relic density with observable decays of the latter. A potential observation of decay of Σ d inside the detector seems thus limited to the pure SM channel. Possible dedicated searches should be then optimized for the search of displaced vertices with multi-jet/leptons and low amounts of missing energy.
The observation of the DM channel might be feasible considering the possibility of observing decays of stopped particles in the detector. Indeed we have until now assumed that the decays of Σ d can be observed in the pixel/tracker region while beyond those only escaping tracks associated to the particle Σ d itself are detected. In reality colored/electromagnetically charged metastable particles lose energy by interacting with the detector material and a frac-
Figure 5 : The same as fig. (4) but with m Σ d = 800 GeV (upper plot), m Σ d = 1600 GeV (left plot) and m Σ d = 2200 GeV (right plot) and L = 3000 fb -1 in all cases.
tion of them can be stopped inside the detector itself. Their late time decays can be observed in the intervals between the collisions of proton beams [50, 51]. These kind of searches are characterized by rather low efficiencies, in particular because of the low, namely glyph[lessorsimilar] 10% , fractions of stopped particles; as a consequence one should focus on the regions in which the number of decays outside the pixel/tracker regions is maximal, possibly renouncing to a statistically significant number of events inside the detector. So searches for stopped particles are in some sense complementary to the searches discussed in this work. At the same time, as can be noted e.g. in fig. (4), (5), in the regions where the number of particles decaying outside the detector is maximal, the isolines of DM lifetime and relic density cross at values of the branching ratios of order 0.5 and thus a contemporary detection of both decays of Σ d is feasible provided the presence of statistically relevant population of stopped particles.
This possibility is confirmed by the outcome of the analysis reported in tab. (2), where a benchmark with ( m ,m ψ Σ d ) = (500 1000) , GeV has been considered and the couplings set
Table 1 : Number of decay events, total as well as separately in the two kind of decay channels (DM or SM only), which is expected to be observed, in the three detection regions, at the indicated values of the luminosity, for a Σ d production scenario corresponding to the benchmark set of masses m Σ d = 800 GeV and m ψ = 10 GeV. The coupling λ and λ ′ have been chosen in such a way that the DM achieves the correct relic density, through the freeze-in mechanism, and its lifetime is 10 28 s, just beyond present ID limits. The values of λ and λ ′ are, respectively, 1 8 . × 10 -11 and 5 5 . × 10 -10 .
| Part of detector | Total | Σ → DM | Σ → SM only |
|--------------------|---------|----------|---------------|
| L = 25 fb - 1 | | | |
| Pixel | 31 | 0 | 31 |
| Tracker | 63 | 0 | 63 |
| Out | 454 | 1 | 453 |
| L = 300 fb - 1 | | | |
| Pixel | 378 | 0 | 378 |
| Tracker | 752 | 1 | 751 |
| Out | 5445 | 6 | 5439 |
| L = 3000 fb - 1 | | | |
| Pixel | 3785 | 4 | 3781 |
| Tracker | 7522 | 8 | 7514 |
| Out | 54446 | 57 | 54389 |
Table 2 : The same as tab. (1) but for m Σ = 1 TeV, x = 0 5 . and the couplings λ and λ ′ set to, respectively, 1 2 . × 10 -12 and 3 6 . × 10 -12 , in order to achieve the correct DM relic density and a DM lifetime of 10 28 s.
| Part of detector | Total | Σ → DM | Σ → SM only |
|--------------------|---------|----------|---------------|
| L = 25 fb - 1 | | | |
| Pixel | 0 | 0 | 0 |
| Tracker | 0 | 0 | 0 |
| Out | 178 | 19 | 159 |
| L = 300 fb - 1 | | | |
| Pixel | 0 | 0 | 0 |
| Tracker | 0 | 0 | 0 |
| Out | 2133 | 222 | 1911 |
| L = 3000 fb - 1 | | | |
| Pixel | 0 | 0 | 0 |
| Tracker | 0 | 0 | 0 |
| Out | 21334 | 2225 | 19109 |
in order to achieve, besides the correct DM relic density, a lifetime of the DM of 10 28 s and a branching ratio of decay of Σ into DM of approximately 10% . For luminosities above 300 fb -1 , a sizable number of events could be observed in both the decay channels of Σ d , possibly compensating the low efficiency in the detection of stopped particles. In order to
Table 3 : The same as tab. (1) but for m Σ = 1 TeV, x = 10 -6 and the couplings λ and λ ′ set to, respectively, 6 5 . × 10 -10 and 1 5 . × 10 -9 . For this choice of parameters Br (Σ d → DM ) ∼ 0 15 . and the DM relic density is entirely achieved through the freeze-in mechanism. The DM lifetime instead exceeds of many orders of magnitude the sensitivity of present and next future detectors.
| Part of detector | Total | Σ → DM | Σ → SM only |
|--------------------|---------|----------|---------------|
| L = 25 fb - 1 | | | |
| Pixel | 49 | 8 | 41 |
| Tracker | 58 | 9 | 49 |
| Out | 2 | 0 | 2 |
| L = 300 fb - 1 | | | |
| Pixel | 585 | 92 | 493 |
| Tracker | 692 | 109 | 583 |
| Out | 25 | 4 | 21 |
| L = 3000 fb - 1 | | | |
| Pixel | 5848 | 925 | 4923 |
| Tracker | 6923 | 1094 | 5829 |
| Out | 250 | 40 | 210 |
quantitatively explore this scenario an analysis accounting for the typology of decay products of Σ d as well as a simulation of the detector are however needed. This is beyond the scope of this work and will be left to a future study.
The contemporary detection of both decay channels of Σ d can be feasible as well at very low values of x , thus corresponding to very light DM particles. Indeed in such a case it is possible to set the two couplings λ and λ ′ to comparable values without conflicting with ID constraints because of the strong enhancement, namely ∝ x -5 of the DM lifetime. We also notice that higher values of the coupling λ are favored by the DM relic density since it tends to be suppressed as x . In such scenario we thus expect Σ f particles to be relatively short-lived and thus most of the decay events happen in the pixel/tracker region. This is confirmed by the results shown in tab. (3) in which we have considered a benchmark model with the very low value x = 10 -6 . As we expected, most of the decay events lie in the inner detector with statistically relevant populations in both the decay channels. For luminosities above 300 fb -1 it is nonetheless possible to observe more than 10 escaping tracks; the LHC thus provides an optimal reconstruction of the properties of the scalar field. On the other hand for the very low value of x considered the Indirect Detection of DM is not possible since its lifetime largely exceeds present and next future experimental sensitivity. As a consequence, in this kind of setup, LHC is the only probe of the model under consideration.
## 3.2 EW-charged scalar
In this section we present a similar analysis for Σ e,glyph[lscript] -type scalar field. We will first assume, analogously to the previous scenario, that the two components of the SU (2) L doublet are exactly degenerate in mass, such that the analysis can be carried out with the exact same steps as before.
The case in which, instead, a sizable mass splitting is present is also phenomenologically
intriguing. Indeed one of the components of Σ glyph[lscript] is electrically neutral and remains undetectable at LHC in case of decays outside the Pixel/Tracker region, thus behaving like an additional DM component. Additional interesting collider signatures can arise in the case in which both the components are accessible to LHC production and decays in W boson (either on- or off-shell) are allowed. We will comment on this possibility at the end of this subsection.
We show in fig. (6) the production cross-section of the Σ glyph[lscript] and Σ e type fields, compared with the one of Σ d . As we can see the production cross-sections are sensitively lower, with respect to the colored case, with a maximal mass reach, corresponding to the high luminosity upgrade of LHC, of approximately 1400 GeV. At the same time the limits from detector stable
Figure 6 : NLO cross-section for several kinds of Σ f field, namely Σ d , Σ glyph[lscript] and Σ e . We report as well the minimal value of the production cross-section, for two values of the luminosity, namely 300 and 3000 fb -1 , needed to give 5 pairs of particles.
particles are substantially relaxed with respect to the case of colored particles so that we can consider values of the mass as low as 300 -400 GeV.
We also notice that the production cross-section of Σ glyph[lscript] pairs is sensitively larger with respect to the case of a SU (2) L singlet. This enhancement is substantially due to the process pp → Σ Σ ± glyph[lscript] 0 glyph[lscript] , with Σ ± glyph[lscript] and Σ 0 glyph[lscript] being, respectively, the electrically charged and neutral components of the SU (2) doublet. On the other hand the increase in the expected number of events depends on the lifetime of Σ glyph[lscript] . It is indeed maximal at shorter lifetimes, when the scalar field decays prevalently in the inner detector, since displaced vertices can be detected both for electrically charged and neutral mother states, while it is more moderate at the longest lifetimes since only the charged component of the SU (2) L doublet can manifest as metastable tracks while the neutral one escapes detection.
Fig. (7) and (8) show the LHC reach, in the three detection regions, as function of the couplings λ and λ ′ and for some fixed values of x , for two values of m Σ glyph[lscript],e , namely 400 and 800 GeV. In the lower mass scenario it is again present an upper bound on the lifetime of the scalar field coming from current searches of disappearing tracks. Notice that this last limit is stronger in the case of Σ glyph[lscript] type field as consequence of the higher cross-section at a given value of the mass (see [42] for details.).
Contrary to the colored case it is possible to have sizable or even dominant contribution
Figure 7 : LHC reach, in the plane ( λ, λ ′ ) , in the Pixel (red dot-dashed lines), Tracker (red dashed lines) and outside the detector (red solid lines), for an integrated luminosity of 300 fb -1 for a Σ e (left panel) and a Σ glyph[lscript] (right panel) scalar field of mass 400 GeV. The solid lines represent the cosmological value of the DM relic density for x = { 10 -3 , 10 -2 , 0 1 . , 0 5 . } , while the short-dashed lines represent a reference value of 10 28 s, approximately corresponding to the current experimental sensitivity on the DM lifetime for the considered set of values of x . The yellow region below the thick long-dashed yellow line is excluded by searches of metastable particles.
Figure 8 : The same as fig. (7) but for m Σ e,glyph[lscript] = 800 GeV.
from the SuperWIMP mechanism at masses accessible to the LHC production. However, as evident form the plot, this occurs only at very small values of the couplings, such that the decays of the scalar field occur substantially only outside the detector.
We have as well reformulated our results in the plane (BR( Σ f → DM)Γ -1 Σ ). As evident
from fig. (9) and (10) the LHC 'double' detection region defined in the previous sections is extremely narrow, as consequence of the lower number of expected events, due to the lower production cross-sections, and it is already closed, for masses of 800 GeV, in the case of Σ e -type field, which thus features extremely poor detection prospects. We also notice that the crossing of curves of the relic density and the DM lifetime occurs, in the double detection strip, at values of the branching ratio at most of the order of 10 -3 which again makes very difficult (likely impossible) the detection of the decay channel of Σ glyph[lscript],e into DM.
1
Figure 9 : LHC reach for Σ e (left) and Σ glyph[lscript] (right) in the plane ( Br (Σ e,glyph[lscript] → DM ) , Γ -1 Σ ) for integrated luminosity of 300 fb -1 and m Σ e,glyph[lscript] = 400 GeV. The violet, green and blue region represent the regions where less then 10 events are expected, respectively in the 'outside', tracker and pixel regions. The 'double detection' region, defined in the main text, is thus the white strip between the shaded regions. The solid and dashed lines give again the cosmological relic density and the reference DM lifetime for the set x = { 10 -3 , 10 -2 , 0 1 . , 0 5 . } . The yellow region below the thick dashed yellow line is excluded by current searches of metastable particles.
Analogously to the previous section, we have investigated the prospects of detection of Σ glyph[lscript],e from a more quantitative perspective focussing on a benchmark set of parameters. We have indeed considered the pair production of a Σ glyph[lscript] particle with an assignment of (10,400) GeV for ( m ,m ψ Σ glyph[lscript] ) and fixed the couplings to give the correct relic density and a lifetime compatible with experimental limits. As evident from the results reported in table (4) the number of expected events is drastically reduced with respect to the colored case. A statistically relevant number of decay events both in the inner and outer part of the detector appears only at luminosities above 300 fb -1 . An eventual discovery of this scenario will thus require more time with respect to the case of color charged scalar field. We notice moreover that the low production cross section, combined with the low branching ratio ∼ 10 -3 , does not allow for any event with Σ glyph[lscript] decaying into DM, even considering a high luminosity upgrade of the LHC. Since our choice of the mass of Σ glyph[lscript] corresponds to the maximal cross-section available, it appears evident that, in the case of Σ f charged only with respect to electroweak interactions, it will not be possible to identify at LHC the peculiar feature of our model, namely the presence of two (i.e. DM+SM and SM only) decay channels.
Figure 10 : The same as fig. (9) but for m Σ e,glyph[lscript] = 800 GeV. A 'double detection' region is present only in the case of Σ glyph[lscript] -type field.
Table 4 : Number of decay events, total as well as separately in the two kind of decay channels (DM or SM only), expected in the three detection regions at the LHC for the indicated values of luminosity. The benchmark chosen consist in a Σ glyph[lscript] -type field with mass of 400 GeV while the DM mass is set to 10 GeV. The pair ( λ, λ ′ ) has been fixed to ( 5 × 10 -12 , 10 -9 ).
| Part of detector | Total | Σ → DM | Σ → SM only |
|--------------------|---------|----------|---------------|
| L = 25 fb - 1 | | | |
| Pixel | 1 | 0 | 1 |
| Tracker | 2 | 0 | 2 |
| Out | 23 | 0 | 23 |
| L = 300 fb - 1 | | | |
| Pixel | 14 | 0 | 14 |
| Tracker | 26 | 0 | 26 |
| Out | 270 | 0 | 270 |
| L = 3000 fb - 1 | | | |
| Pixel | 135 | 0 | 135 |
| Tracker | 260 | 0 | 260 |
| Out | 2705 | 0 | 2705 |
Analogously to the colored scenario we can consider signals outside the double detection regions. In the high lifetime region only the decay channel into SM particles might be accessible to possible searches of stopped particles since the small number of events, consequence of the low production cross section, does not allow to have a statistically relevant number of events in both the decay channels 5 . Alternatively we can consider very low values of the DM
5 We also notice that there are not, at the moment, searches of only electroweakly interacting stopped particles. On general grounds we expect a lower detection efficiency with respect to the case of color interaction.
Table 5 : Number of decay events, total as well as separately in the two kind of decay channels (DM or SM only), which is expected in the three detection regions at the LHC for the indicated values of luminosity. The benchmark chosen consist in a Σ glyph[lscript] -type field with mass of 800 GeV while the DM mass is set to 8 × 10 -4 GeV. The pair ( λ, λ ′ ) has been fixed to ( 8 × 10 -10 , 7 8 . × 10 -10 ) and corresponds to a lifetime many orders of magnitude above current and next future experimental sensitivity. No ID detection is thus expected in this case.
| Part of detector | Total | Σ → DM | Σ → SM only |
|--------------------|---------|----------|---------------|
| L = 25 fb - 1 | | | |
| Pixel | 1 | 1 | 0 |
| Tracker | 2 | 1 | 1 |
| Out | 8 | 4 | 4 |
| L = 300 fb - 1 | | | |
| Pixel | 12 | 6 | 6 |
| Tracker | 27 | 14 | 13 |
| Out | 94 | 48 | 46 |
| L = 3000 fb - 1 | | | |
| Pixel | 120 | 62 | 58 |
| Tracker | 270 | 139 | 131 |
| Out | 941 | 484 | 457 |
mass in order to evade ID constraints and for x ∼ 10 -6 it might be again possible to observe both the Σ glyph[lscript] decay channels, renouncing however to the possibility of next future detection of DM decays in cosmic rays. A possible outcome in this kind of scenario is reported in tab. (5), where it is shown the number of expected observed decay events in the case of Σ glyph[lscript] -type field with mass of 800 GeV and for a very low value of the DM mass corresponding to x = 10 -6 . The value of the ( λ, λ ′ ) pair, respectively 8 × 10 -10 , 7 8 . × 10 -10 , guarantees the correct DM relic density, through the freeze-in mechanism, and a branching fraction of decay of the scalar field into DM of approximately 50% . Although the choice of the parameters is substantially analogous to the low mass benchmark studied in the case of color charged scalar (the lifetimes of the scalar field differ approximately by a factor 2) the majority of the events now lie in the 'outside' region. This is due to the fact the Σ glyph[lscript],e are more boosted with respect to colored scalars. At the higher luminosities we have anyway a number of events in the inner detector, for both decay channels, satisfying our discovery requirement. We have thus again a good capability for LHC of providing information on the model under consideration, renouncing however at any prospects of indirect detection of the decays of the DM.
We conclude this section describing how the LHC signals may result altered, in the Σ glyph[lscript] scenario, in the presence of a sizable mass splitting in the EW multiplet. As already mentioned in this case the heaviest component of the doublet dominantly decays into the lighter one and a W boson (or into two quarks/leptons) and these decays result prompt for mass-splitting above ∼ 1 GeV. Contrary to the colored case, we can have a sizable production of the heavy component through the process pp → Σ Σ ± l 0 l , being mediated by the W boson, which can account for ∼ 50% of the total production cross-section even for sizable mass splittings. The range of signals results therefore enriched if the detection of the products of the prompt decays
of the heavy component of the doublet is possible.
In the case m Σ ± > m Σ 0 the decays Σ ± → Σ 0 ff ′ produce displaced vertices accompanied by prompt leptons and jets, for short lifetimes of Σ 0 , or a signal consisting of jet/leptons+missing energy, customarily studied in supersymmetric setups [52, 53], in the case that Σ 0 decays outside the inner detector 6 . In the opposite case m Σ ± < m 0 Σ prompt jets/leptons are associated to displaced vertices or metastable tracks.
## 4 Discussion
In this section we summarize the outcome of the previous analysis and discuss in greater detail whether an hypothetical next future LHC signal can unambiguously discriminate the underlying particle physics scenario.
We show in fig. (11) and (12) the LHC reach, in the two scenarios of Σ d -and Σ glyph[lscript] -type field, as function of the mass of the scalar field and of the coupling λ ′ , for the two assignments of the DM mass of 10 and 100 GeV. The luminosity has been set to 300 fb -1 . The coupling λ has been determined as function of the other parameters, using eq. (2.5), (2.6) and (2.7), according the requirement of the correct DM relic density.
The red, blue and green lines in the plots correspond to the observation of 10 events respectively in the Tracker, Pixel and outside the detector region. The green region comprised between this lines is the 'double detection' region, in which the observation of at least 10 events in the inner detector and at least 10 escaping tracks. Above this region only displaced vertices can be observed at a statistically relevant amount while below metastable tracks are the only visible signal. The yellow regions in the plots are already excluded, for the considered values of the DM mass, by constraints from DM indirect detection 7 . As evident the possibility of 'double' LHC detection is already ruled out for DM masses above 100 GeV in the case of colored scalar and for masses above 10 GeV in the case of EW charged scalar. Nevertheless for the colored case, as can be seen in the left panel of fig. (11) as well as in the benchmark in Table 1, the future indirect detection region just below the present bound lies exactly in the double detection corner for a DM mass of 10 GeV and m Σ < 1500 GeV and in that case possibly all the four parameters of the model could be within reach in the next future. The possibility of observing only displaced vertices is disfavored as well by ID for these values of the DM mass, ad exception of the highest value of the scalar field mass, at the boundaries of the LHC reach. The more severe exclusion in this last case is due to the fact that the production of lighter scalar particles is considered, in turn implying higher decay rates for the DM.
We can now investigate in more detail which kind of information can be inferred from an hypothetical future LHC signal. At high DM masses, namely above 100 (10) GeV in the colored (only electroweakly interacting) Σ f scenario, the only signal which can be expected are metastable tracks. In such case it will be possible to give a lower bound on the scalar field lifetime or perhaps to measure it, although maybe with less precision with respect to a double detection scenario, from the decay of stopped Σ f . From this it may be possible to determine the sum λ 2 + λ ′ 2 while, in absence of observation of the decay processes, it is
6 Notice that we could as well consider monojet + missing energy signals in the case of long lived pair produced Σ 0 l , totally analogous to the DM pair production scenarios. However in our setup pair production of Σ 0 l occurs only through a strongly off-shell Z-boson giving a rather weak signal [2, 54].
7 Notice that our bounds from ID detection are rather conservative. In the case of Σ glyph[lscript],e -type field we can obtain, for example, weaker bounds assuming decays only in some flavour states like, e.g., τ leptons.
Figure 11 : Summary of the possible observed signals at LHC, at L = 300 fb -1 , for a Σ d -type field, as function of its mass and the coupling λ ′ . The other coupling, λ , has been fixed in order to reproduce the correct DM relic density while the DM mass has been set to 10 GeV (left plot) and 100 GeV (right plot). The red, blue and green lines correspond to the observation of 10 events respectively in the Tracker, Pixel and outside the detector region. In the 'double' detection region, the green shaded region between this lines, a number ≥ 10 of decay events in the Pixel and the Tracker and more than 10 tracks leave the detector. The yellow shaded region, above the thick yellow dashed line, is excluded by constraints from indirect detection of DM decay.
Figure 12 : The same as fig.(11) but for the case of a Σ glyph[lscript] -type field.
not possible to determine the single values of the two couplings, which could allow to infer the DM production mechanism. This task could be achieved in case of an Indirect Detection
of dark matter decay which would provide, besides the value of the DM mass, information on a different combination of the two couplings, namely their product λλ ′ . The complete identification of the model could then be achieved by verifying that the DM relic density, computed with the reconstructed parameters, matches the cosmological value.
At intermediate DM mass scales the LHC 'double' detection region is instead still viable. This scenario guarantees the optimal reconstruction of the lifetime of the scalar field, as well as its mass which can be inferred by the cross section and by the energy of the decay products. As already pointed out these informations alone do not allow to infer the DM relevant properties like the mass, which it is in any case not accessible at the LHC, and the production mechanism. Indeed the latter would require the knowledge of the single values of the couplings λ and λ ′ while the lifetime of Σ f depends on the sum λ 2 + λ ′ 2 . The two couplings could be singularly inferred in case it is possible to distinguish, through the decay products in the displaced vertices, the two decay channels of the scalar field. However our study based on the relative branching fractions of the two decay channels of the scalar fields, confirmed from a quantitative perspective by the study of the two benchmarks reported in tab. (1) and (4), has shown that only the pure SM decay channel is accessible to LHC detection. In this case it is possible to infer, from the determination of the lifetime of the scalar field, only the value of λ ′ (assuming a negligible branching fraction of decay into DM). A LHC signal only can thus neither provide evidence of the existence of the DM nor information on its production mechanism. A correlation with a DM ID signal is again mandatory for determining the remaining parameters. In both the scenarios proposed above the capability of full determination of the model under consideration is hence limited to the regions of the parameters space which lie in proximity of the current experimental sensitivity to DM Indirect Detection.
For very low DM masses, LHC is instead the only probe of the model under consideration while ID is not achievable (as a consequence, bounds like the ones shown in fig. (11) and (12), are completely evaded). On the contrary the two branching fractions of decay into the scalar field, namely the one into DM+SM and only SM, can be comparable within the double detection region. This statement is again confirmed at the quantitative level by the study of two benchmarks, one for colored and one for only EW interacting scalar, reported in tab. (3) and (5), which show the presence of statistically relevant number of events, for both the two decay channels, in the inner detector as well as in the 'outside' region. In case of identification of the two decay channels it is possible to infer the values of λ and λ ′ as well as the m Σ f . The DM mass is instead not directly accessible from observations but might be determined from the requirement of the correct relic density, given its proportionality to x = m /m ψ Σ f .
We remark anyway that the capability of disentangling the decay channels of the scalar field is actually model dependent since, according to which of the operators in (2.1) and (2.2) determine the decay processes of Σ f , we might have scenarios with very different decay products in the two channels (for example in the case of Σ glyph[lscript] field we can have a decay into DM and a charged lepton opposed to a decay into two jets triggered by the coupling λ 2 3 , glyph[lscript] ) as well as rather similar signals, e.g. in the case in which the scalar field decays into a neutrino and another SM fermion. We also remark that the various kinds of final states, e.g. jets, leptons etc..., possibly emerging from displaced vertices have not the same capability of reconstruction, at a given lifetime (see e.g. [13] for a discussion). In order to properly address this issue a refinement of our analysis, including a simulation of the detector, is required.
The statements discussed until now substantially hold for all the assignments of the SM charge of the Σ f particle. In absence of a detector analysis, sensitive to the decay products,
the main difference is between the cases of colored or only electroweakly charged Σ f with the latter requiring higher luminosities for a discovery and featuring a more limited mass reach, in view of the lower production cross-section. In the case of colored mediator one could consider complementary searches, with respect to the one discussed in this paper, relying on the possibility of observing very late decays of Σ f particles stopped in the detector. At these very long lifetimes the observation of both decay channels of Σ f is compatible with constraints from DM phenomenology. Our study reported in tab. (2) shows that this possibility is potentially feasible even once the actual low efficiencies in this kind of searches are accounted for. On the other hand we remark again that in order to properly determine whether the different decay channels can be discriminated a simulation of the detector is mandatory and that a definite statement is left to a future study. In case of possible identifications of the two decay channels it is again possible to determine their relative branching fractions and then directly the couplings λ and λ ′ .
## 5 Conclusions
We have considered the LHC detection prospects of a very simple and economical extension of the Standard Model featuring a decaying Maiorana fermion as DM candidate and a metastable scalar field with non-trivial quantum numbers under the SM gauge group. The potential LHC signal is constituted by the observation of the displaced decays of the scalar field, either into the DM and a SM fermion or into two SM fermions, or of the scalar field itself as disappearing track escaping the detector. We have investigated the possibility of detecting this kind of signals in the regions of the parameters space favored by the requirement of the correct DM relic density and a future Indirect Detection of the DM decays.
In the case of color charged scalar field a statistically relevant population of both displaced vertices and metastable tracks might be observed in next future 14 TeV LHC run up to masses of the order of 1500 GeV and up to approximately 2200 considering a high luminosity run; however bounds from DM Indirect Detection already exclude the possibility of "double" detection for DM masses above 100 GeV, leaving open only a possible detection of metastable tracks. In the parameter region corresponding to the 'double' LHC detection unfortunately only the decay channel into just SM fermions is accessible to observation, since the branching ratio of the scalar field into DM is of the order ≤ 10 -2 -10 -3 . The two decay channels of the Σ f field might be instead contemporarily observed in the case of very long Σ f lifetimes, with negligible amount of events in the inner detector, considering the possibility that a fraction of Σ f can be stopped in the detector material and decay at later time. This possibility should be investigated through a dedicated study. The contemporary detection of the two decay channels of Σ f can occur as well at very low, i.e. glyph[lessmuch] 1 GeV DM masses. In this case there would be the observation of mostly displaced vertices in the inner detector, while DM decays in Indirect Detection are unfortunately beyond present and future observational capabilities because of the extremely long DM lifetimes.
In the case in which, instead, Σ f features only electroweak interactions the maximal mass reach is limited to approximately 1400 TeV and the high luminosity upgrade is most likely needed to probe the parameters space of model. Because of the lower masses of Σ f involved in the analysis the bounds from DM ID are more severe, with respect to the case of a colored scalar field, such that under the most conservative assumptions the 'double' LHC detection is not possible for DM masses above 10 GeV. There are as well poorer prospects of detection of stopped particles because of the lower number of events and of stopping probability. The
scenario of contemporary detection of the two decay channels at very low DM masses is instead feasible although kinematic effects imply, contrary to the colored case, a greater number of decays in the outer detector region. The range of possible signals is enriched in the case the scalar field is a SU (2) L doublet with a mass splitting between the components. In particular, in the case of Σ glyph[lscript] type field, with lighter neutral component, the signals presented above can be accompanied as well by conventional missing energy signatures and prompt tracks.
In most of the parameter space of the scenario presented, LHC experiments and Indirect Detection observations provide complementary information on the model and both signals are necessary in order to confirm the identity of the Dark Matter and verify the production mechanism. In particular if the mass of the Dark Matter is in the 1-10 GeV range and the Σ f scalar within the kinematical reach of the next LHC run, i.e. below 1600 GeV for the colored case or 1100 GeV for the EW case, there is a clear chance to obtain multiple signals in the very near future.
## Acknowledgments
The authors thank Marco Nardecchia for enlighting discussions.
The authors acknowledge partial support from the European Union FP7 ITN-INVISIBLES (Marie Curie Actions, PITN-GA-2011-289442).
## References
- [1] ATLAS Collaboration Collaboration, Search for New Phenomena in Monojet plus Missing Transverse Momentum Final States using 10fb-1 of pp Collisions at sqrts=8 TeV with the ATLAS detector at the LHC , Tech. Rep. ATLAS-CONF-2012-147, CERN, Geneva, Nov, 2012.
- [2] CMS Collaboration Collaboration, Search for new physics in monojet events in pp collisions at sqrt(s)= 8 TeV , Tech. Rep. CMS-PAS-EXO-12-048, CERN, Geneva, 2013.
- [3] G. Arcadi and L. Covi, Minimal Decaying Dark Matter and the LHC , JCAP 1308 (2013) 005, [ arXiv:1305.6587 ].
- [4] T. Han, Z. Si, K. M. Zurek, and M. J. Strassler, Phenomenology of hidden valleys at hadron colliders , JHEP 0807 (2008) 008, [ arXiv:0712.2041 ].
- [5] K. Ishiwata, T. Ito, and T. Moroi, Long-Lived Unstable Superparticles at the LHC , Phys.Lett. B669 (2008) 28-33, [ arXiv:0807.0975 ].
- [6] S. Chang and M. A. Luty, Displaced Dark Matter at Colliders , arXiv:0906.5013 .
- [7] P. Meade, S. Nussinov, M. Papucci, and T. Volansky, Searches for Long Lived Neutral Particles , JHEP 1006 (2010) 029, [ arXiv:0910.4160 ].
- [8] P. Meade, M. Reece, and D. Shih, Long-Lived Neutralino NLSPs , JHEP 1010 (2010) 067, [ arXiv:1006.4575 ].
- [9] S. Bobrovskyi, W. Buchmuller, J. Hajer, and J. Schmidt, Quasi-stable neutralinos at the LHC , JHEP 1109 (2011) 119, [ arXiv:1107.0926 ].
- [10] J. Heisig and J. Kersten, Production of long-lived staus in the Drell-Yan process , Phys.Rev. D84 (2011) 115009, [ arXiv:1106.0764 ].
- [11] P. Ghosh, D. E. Lopez-Fogliani, V. A. Mitsou, C. Munoz, and R. Ruiz de Austri, Probing the Îij-from-Ρ supersymmetric standard model with displaced multileptons from the decay of a Higgs boson at the LHC , Phys.Rev. D88 (2013) 015009, [ arXiv:1211.3177 ].
- [12] J. Heisig and J. Kersten, Long-lived staus from strong production in a simplified model approach , Phys.Rev. D86 (2012) 055020, [ arXiv:1203.1581 ].
- [13] P. W. Graham, D. E. Kaplan, S. Rajendran, and P. Saraswat, Displaced Supersymmetry , JHEP 1207 (2012) 149, [ arXiv:1204.6038 ].
- [14] J. Heisig, J. Kersten, B. Panes, and T. Robens, A survey for low stau yields in the MSSM , JHEP 1404 (2014) 053, [ arXiv:1310.2825 ].
- [15] M. Cirelli, F. Sala, and M. Taoso, Wino-like Minimal Dark Matter and future colliders , arXiv:1407.7058 .
- [16] J. L. Feng and B. T. Smith, Slepton trapping at the large hadron and international linear colliders , Phys.Rev. D71 (2005) 015004, [ hep-ph/0409278 ].
- [17] M. Hirsch, W. Porod, and D. Restrepo, Collider signals of gravitino dark matter in bilinearly broken R-parity , JHEP 0503 (2005) 062, [ hep-ph/0503059 ].
- [18] N. Arkani-Hamed and N. Weiner, LHC Signals for a SuperUnified Theory of Dark Matter , JHEP 0812 (2008) 104, [ arXiv:0810.0714 ].
- [19] L. J. Hall, K. Jedamzik, J. March-Russell, and S. M. West, Freeze-In Production of FIMP Dark Matter , JHEP 1003 (2010) 080, [ arXiv:0911.1120 ].
- [20] C. Cheung, G. Elor, L. J. Hall, and P. Kumar, Origins of Hidden Sector Dark Matter II: Collider Physics , JHEP 1103 (2011) 085, [ arXiv:1010.0024 ].
- [21] L. Covi, M. Olechowski, S. Pokorski, K. Turzynski, and J. D. Wells, Supersymmetric mass spectra for gravitino dark matter with a high reheating temperature , JHEP 1101 (2011) 033, [ arXiv:1009.3801 ].
- [22] M. Endo, K. Hamaguchi, and K. Nakaji, Probing High Reheating Temperature Scenarios at the LHC with Long-Lived Staus , JHEP 1011 (2010) 004, [ arXiv:1008.2307 ].
- [23] J. M. Lindert, F. D. Steffen, and M. K. Trenkel, Direct stau production at hadron colliders in cosmologically motivated scenarios , JHEP 1108 (2011) 151, [ arXiv:1106.4005 ].
- [24] M. Garny, A. Ibarra, D. Tran, and C. Weniger, Gamma-Ray Lines from Radiative Dark Matter Decay , JCAP 1101 (2011) 032, [ arXiv:1011.3786 ].
- [25] M. Garny, A. Ibarra, and S. Vogl, Dark matter annihilations into two light fermions and one gauge boson: General analysis and antiproton constraints , JCAP 1204 (2012) 033, [ arXiv:1112.5155 ].
- [26] M. Garny, A. Ibarra, and D. Tran, Constraints on Hadronically Decaying Dark Matter , JCAP 1208 (2012) 025, [ arXiv:1205.6783 ].
- [27] R. Barbier, C. Berat, M. Besancon, M. Chemtob, A. Deandrea, et al., R-parity violating supersymmetry , Phys.Rept. 420 (2005) 1-202, [ hep-ph/0406039 ].
- [28] A. Y. Smirnov and F. Vissani, Upper bound on all products of R-parity violating couplings lambda-prime and lambda-prime-prime from proton decay , Phys.Lett. B380 (1996) 317-323, [ hep-ph/9601387 ].
- [29] N. Fornengo, L. Maccione, and A. Vittino, Constraints on particle dark matter from cosmic-ray antiprotons , JCAP 1404 (2014) 003, [ arXiv:1312.3579 ].
- [30] A. Ibarra, A. S. Lamperstorfer, and J. Silk, Dark matter annihilations and decays after the AMS-02 positron measurements , Phys.Rev. D89 (2014) 063539, [ arXiv:1309.2570 ].
- [31] LAT Collaboration Collaboration, M. Ackermann et al., Fermi LAT Search for Dark Matter in Gamma-ray Lines and the Inclusive Photon Spectrum , Phys.Rev. D86 (2012) 022002, [ arXiv:1205.2739 ].
- [32] Fermi-LAT Collaboration Collaboration, M. Ackermann et al., Search for Gamma-ray Spectral Lines with the Fermi Large Area Telescope and Dark Matter Implications , Phys.Rev. D88 (2013) 082002, [ arXiv:1305.5597 ].
- [33] X. Chu, T. Hambye, and M. H. Tytgat, The Four Basic Ways of Creating Dark Matter Through a Portal , JCAP 1205 (2012) 034, [ arXiv:1112.0493 ].
- [34] X. Chu, Y. Mambrini, J. Quevillon, and B. Zaldivar, Thermal and non-thermal production of dark matter via Z'-portal(s) , JCAP 1401 (2014) 034, [ arXiv:1306.4677 ].
- [35] M. Klasen and C. E. Yaguna, Warm and cold fermionic dark matter via freeze-in , JCAP 1311 (2013) 039, [ arXiv:1309.2777 ].
- [36] M. Blennow, E. Fernandez-Martinez, and B. Zaldivar, Freeze-in through portals , arXiv:1309.7348 .
- [37] L. Covi, J. E. Kim, and L. Roszkowski, Axinos as cold dark matter , Phys.Rev.Lett. 82 (1999) 4180-4183, [ hep-ph/9905212 ].
- [38] J. L. Feng, A. Rajaraman, and F. Takayama, Superweakly interacting massive particles , Phys.Rev.Lett. 91 (2003) 011302, [ hep-ph/0302215 ].
- [39] J. L. Feng, A. Rajaraman, and F. Takayama, SuperWIMP dark matter signals from the early universe , Phys.Rev. D68 (2003) 063504, [ hep-ph/0306024 ].
- [40] M. Gomez, C. Jackson, and G. Shaughnessy, Dark Matter on Top , arXiv:1404.1918 .
- [41] J. Alwall, M. Herquet, F. Maltoni, O. Mattelaer, and T. Stelzer, MadGraph 5 : Going Beyond , JHEP 1106 (2011) 128, [ arXiv:1106.0522 ].
- [42] L. Covi and F. Dradi, Long-Lived stop at the LHC with or without R-parity , arXiv:1403.4923 .
- [43] CMS Collaboration Collaboration, S. Chatrchyan et al., The CMS experiment at the CERN LHC JINST , 3 (2008) S08004.
- [44] CMS Collaboration Collaboration, Search for long-lived particles decaying to final states that include dileptons , Tech. Rep. CMS-PAS-EXO-12-037, CERN, Geneva, 2014.
- [45] Search for long-lived, heavy particles in final states with a muon and a multi-track displaced vertex in proton-proton collisions at sqrt(s) = 8TeV with the ATLAS detector. , Tech. Rep. ATLAS-CONF-2013-092, CERN, Geneva, Aug, 2013.
- [46] S. Asai, K. Hamaguchi, and S. Shirai, Measuring lifetimes of long-lived charged massive particles stopped in LHC detectors , Phys.Rev.Lett. 103 (2009) 141803, [ arXiv:0902.3754 ].
- [47] W. Beenakker, R. Hopker, and M. Spira, PROSPINO: A Program for the production of supersymmetric particles in next-to-leading order QCD , hep-ph/9611232 .
- [48] ATLAS Collaboration Collaboration, G. Aad et al., Searches for heavy long-lived sleptons and R-Hadrons with the ATLAS detector in pp collisions at √ s = 7 TeV Phys.Lett. , B720 (2013) 277-308, [ arXiv:1211.1597 ].
- [49] CMS Collaboration Collaboration, S. Chatrchyan et al., Searches for long-lived charged particles in pp collisions at √ s =7 and 8 TeV , JHEP 1307 (2013) 122, [ arXiv:1305.0491 ].
- [50] ATLAS Collaboration Collaboration, G. Aad et al., Search for long-lived stopped R-hadrons decaying out-of-time with pp collisions using the ATLAS detector , Phys.Rev. D88 (2013) 112003, [ arXiv:1310.6584 ].
- [51] CMS Collaboration Collaboration, S. Chatrchyan et al., Search for stopped long-lived particles produced in pp collisions at √ s = 7 TeV JHEP , 1208 (2012) 026, [ arXiv:1207.0106 ].
- [52] ATLAS Collaboration Collaboration, G. Aad et al., Search for direct production of charginos, neutralinos and sleptons in final states with two leptons and missing transverse
momentum in pp collisions at √ s = 8 TeV with the ATLAS detector , JHEP 1405 (2014) 071, [ arXiv:1403.5294 ].
- [53] CMS Collaboration Collaboration, V. Khachatryan et al., Searches for electroweak production of charginos, neutralinos, and sleptons decaying to leptons and W, Z, and Higgs bosons in pp collisions at 8 TeV , arXiv:1405.7570 .
- [54] P. J. Fox, R. Harnik, J. Kopp, and Y. Tsai, Missing Energy Signatures of Dark Matter at the LHC Phys.Rev. , D85 (2012) 056011, [ arXiv:1109.4398 ].
|
10.1088/1475-7516/2014/10/063
|
[
"Giorgio Arcadi",
"Laura Covi",
"Federico Dradi"
] |
2014-08-05T15:31:52+00:00
|
2014-08-05T15:31:52+00:00
|
[
"hep-ph"
] |
LHC prospects for minimal decaying Dark Matter
|
We study the possible signals at LHC of the minimal models of decaying dark
matter. Those models are characterized by the fact that DM interacts with SM
particles through renormalizable coupling with an additional heavier charged
state. Such interaction allows to produce a substantial abundance of DM in the
early Universe via the decay of the charged heavy state, either in- or
out-of-equilibrium. Moreover additional couplings of the charged particle open
up decay channels for the DM, which can nevertheless be sufficiently long-lived
to be a good DM candidate and within reach of future Indirect Detection
observations. We compare the cosmologically favored parameter regions to the
LHC discovery reach and discuss the possibility of simultaneous detection of DM
decay in Indirect Detection.
|
1408.1006v2
|
## Topology and Self-Similarity of the Hofstadter Butterfly
Indubala I Satija
School of Physics and Astronomy and Computational Sciences, George Mason University, Fairfax, VA 22030 (Dated: December 16, 2014)
We revisit the problem of self-similar properties of the Hofstadter butterfly spectrum, focusing on spectral as well as topological characteristics. In our studies involving any value of magnetic flux and arbitrary flux interval, we single out the most dominant hierarchy in the spectrum, which is found to be associated with an irrational number ζ = 2 + √ 3 where nested set of butterflies describe a kaleidoscope. Characterizing an intrinsic frustration at smallest energy scale, this hidden quasicrystal encodes hierarchical set of topological quantum numbers associated with Hall conductivity and their scaling properties. This topological hierarchy maps to an integral Apollonian gasket nearD 3 symmetry, revealing a hidden symmetry of the butterfly as the energy and the magnetic flux intervals shrink to zero. With a periodic drive that induces phase transitions in the system, the fine structure of the butterfly is shown to be amplified making states with large topological invariants accessible experimentally.
PACS numbers: 03.75.Ss,03.75.Mn,42.50.Lc,73.43.Nq
## I. INTRODUCTION
Hofstadter butterfly[1], also known as Gplot[2] is a fascinating two-dimensional spectral landscape, a quantum fractal where energy gaps encode topological quantum numbers associated with the Hall conductivity[3]. This intricate mix of order and complexity is due to frustration, induced by two competing periodicities as electrons in a crystalline lattice are subjected to a magnetic field. The allowed energies of the electrons are discontinuous function of the magnetic flux penetrating the unit cell, while the gaps , the forbidden energies are continuous except at discrete points. The smoothness of spectral gaps in this quantum fractal may be traced to topology which makes spectral properties stable with respect to small fluctuations in the magnetic flux when Fermi energy resides in the gap. The Gplot continues to arouse a great deal of excitement since its discovery, and there are various recent attempts to capture this iconic spectrum in laboratory[4-6].
Fractal properties of the butterfly spectrum have been the subject of various theoretical studies[7-12]. However, detailed description quantifying self-similar universal properties of the butterfly and the universal fixed point butterfly function has not been reported previously. In contrast to earlier studies where self-similarity of the spectrum is studied for a fixed value of the magnetic flux such as the golden-mean, and thus focusing on certain isolated local parts of the spectrum, this paper presents self-similar butterfly that is reproduced at all scales in magnetic flux .
In this paper, we address following questions regarding the butterfly fractal: (1) How to describe self-similar fractal properties of the butterfly at any value of magnetic flux given arbitrary flux interval. We determine the recursion relation, for determining the magnetic flux interval from one generation to the next , so that one reproduces the entire butterfly structure . In other words, we seek the fixed point function that contains the entire Gplot as the magnetic flux and the energy scales shrinks to zero. We try to answer this question without confining to a specific magnetic flux value such as the goldenmean and obtain universal scaling and the fixed point butterfly function that describes the spectrum globally. (2) In addition to spectral scaling, we also address the question of scaling for the topological quantum numbers . (3) We briefly investigate butterfly fractal for special values of magnetic flux such as the golden and the silver-mean that has been the subject of almost all previous studies. (4) Finally, we present a mechanism for amplifying small gaps of the butterfly fractal, making them more accessible in laboratory.
Our approach here is partly geometrical and partially numerical. Using simple geometrical and number theoretical tools, we obtain the exact scaling associated with the magnetic flux interval. Here we address the question of both magnetic flux as well as topological scaling. The spectral gaps are labeled by two quantum numbers which we denote as σ and τ . The integer σ is the Chern number , the quantum number associated with Hall conductivity[3] and τ is an integer. These quantum numbers satisfy the Diophantine equation (DE)[13],
$$\rho = \phi \sigma + \tau \quad \quad \quad$$
where ρ is the particle density when Fermi level is in the gap and φ denotes the magnetic flux per unit cell. We obtain exact expressions describing scaling of these quantum numbers in the butterfly hierarchy. The spectral scaling describing universal scalings associated with the energy interval is obtained numerically. Our analysis is mostly confined to the energy scales near E = 0 , that is near half-filling. This is a reasonable choice for two reasons: firstly, simple observation of the butterfly spectrum shows gaps of the spectrum forming 4 -wing structures ( the butterflies) exist mostly near half-filling. Secondly, gaps away from half-filling appear to be continuation of the gaps that exist near E = 0 . We believe that although the gaps characterized by a fixed ( στ ) are discontinuous at rational values of the flux, , these gaps continue ( with same topological numbers ) after a break at rational flux values, with their derivatives w.r.t the magnetic flux continuous
## A. Summary of the main results
- · Given an arbitrary value of magnetic flux φ 0 and an arbitrary flux interval δφ , (no matter how small), there is
a precise rule for obtaining the entire butterfly in that interval, as described in section (3-A).
- · Simple number theory provides an exact scaling ratio between two successive generations of the butterfly and this scaling is universal, independent of the initial window for zooming , described in subsections (3-B) and (IV).
- · The hierarchy characterized by the irrational numbers whose tail exhibit period2 continued fraction expansion with entries 1 and 2 , which we denote as ζ 1 2 , emerges as the most 'dominant' hierarchy as is associated with the smallest scaling ratio that describes butterflies between two successive generations. Commonly studied hierarchies characterized by golden-tail, which we denote as ζ 1 ( set of irrationals who tail end exhibits integer 1 only in its continued fraction expanding) are of lower significance as they are characterized by larger scaling ratio. A comparison between different hierarchies is given in Table III.
- · The emergence of ζ 1 2 , class of irrationals with the universal butterfly and its topological hierarchy of quantum numbers reveals a hidden dodecagonal quasicrystalline symmetry[15] in the butterfly spectrum. These results also apply to other lattices such as graphene in a magnetic field.
- · The dominant hierarchy ζ 1 2 , maps to a geometrical fractal known as the Integral Apollonian gasket [16] that asymptotically exhibits near D 3 symmetry and the nested set of butterflies describe a kaleidoscope where two successive generations of butterfly are mirror images through a circular mirror. This is discussed in section V.
- · In our investigation of the fractal properties of the Hofstadter butterfly, one of the key guiding concepts is a corollary of the DE equation that quantifies the topology of the fine structure near rational fluxes . We show that, for every rational flux, an infinity of possible solutions of the DE, although not supported in the simple square lattice model , are present in close vicinity of the flux. ( See section (IV)). Consequently, perturbations that induce topological phase transitions can transform tiny gaps with large topological quantum numbers into major gaps. This might facilitate the creation of such states in an experimental setting. In section VII, we illustrate this amplification by periodically driving the system.
## II. MODEL SYSTEM AND TOPOLOGICAL INVARIANTS
Model system we study here consists of (spinless) fermions in a square lattice. Each site is labeled by a vector r = nx ˆ + my ˆ , where n m , are integers, ˆ x ( ˆ y ) is the unit vector in the x ( y ) direction, and a is the lattice spacing. The tight binding
Hamiltonian has the form
$$H = - J _ { x } \sum _ { r } | { \mathbf r }$$
Here, | r 〉 is the Wannier state localized at site r . J x ( J y ) is the nearest neighbor hopping along the x ( y ) direction. With a uniform magnetic field B along the z direction, the flux per plaquette, in units of the flux quantum Φ 0 , is φ = -Ba / 2 Φ 0 .
In the Landau gauge realized in experiments[17], the vector potential A x = 0 and A y = -φx , the Hamiltonian is cyclic in y so the eigenstates of the system can be written as Ψ n,m = e ik y m ψ n where ψ n satisfies the Harper equation[18]
$$\text{only} \quad e ^ { i k _ { x } } \psi _$$
Here n ( m ) is the site index along the x ( y ) direction, λ = J /J y x and ψ r n + q = ψ r n , r = 1 2 , , ...q are linearly independent solutions. In this gauge the magnetic Brillouin zone is -π/qa ≤ k x ≤ π/qa and -π ≤ k y ≤ π .
At flux φ = p/q , the energy spectrum has in general q -1 gaps. For Fermi level inside each energy gap, the system is in an integer quantum Hall state[3] characterized by its Chern number σ , the quantum number associated with the transverse conductivity C xy = σe /h 2 [3]. The σ and τ are two quantum numbers that label various gaps of the butterfly and are solutions of DE[13]. The possible values of these integers are,
$$( \sigma, \tau ) = ( \sigma _ { 0 } - n q$$
Here σ , τ 0 0 are any two integers that satisfy the Eq. (1) and n is an integer. The quantum numbers σ that determines the quantized Hall conductivity corresponds to the change in density of states when the magnetic flux quanta in the system is increased by one and whereas the quantum number τ is the change in density of states when the period of the potential is changed so that there is one more unit cell in the system.
For any value of the magnetic flux , the system described by the Hamiltonian (2), supports only n = 0 solution of Eq. (4) for the quantum numbers σ and τ . Absence of changes in topological states from n = 0 to higher n values is due to the absence of any gap closing and reopening that is essential for any topological phase transition. However, as shown later, the DE which relates continuously varying quantities ρ and φ with integers σ and τ , has some important consequences about topological changes in close vicinity of rational values of φ .
## III. BUTTERFLY FRACTAL
## A. Miniature Copies of the Butterfly Graph: Butterfly at Every Scale
Butterfly graph is a plot of possible energies of the electron for various values of φ which varies between [0 , 1] . To understand this graph, we begin with values of φ that are rational numbers, focusing on simple fractions. Figure (1) shows one such graph, a skeleton of the butterfly graph, obtained using few rational values. The permissible energies
FIG. 1. Figure shows the butterfly graph for values of φ = p/q with q max = 8 . For fixed p/q , the energy ( vertical axis) shows q bands and q -1 gaps for oddq case. For q -even, the two bands at the center touch and therefore there are only q -2 gaps.
are arranged in bands separated from forbidden values, the gaps . In general, for a fixed φ = p q , the graph consists of of q bands ( the dark regions) and ( q -1) gaps ( empty regions).
For q -even, the two central bands touch one another and therefore, we see only ( q -2) gaps. The graphs show important distinctions between the even and the odd-denominator fractions as shown in Fig.(2). As we look in the immediate vicinity of even-denominator flux values, the two touching bands begin to split, opening a gap at the center. Consequently, in the butterfly landscape, as we look both to the left and to the right of the even-denominator fraction, we see four gaps or swaths resembling the four-wings of a butterfly, all converging at the center. In contrast, near odd-denominator fractions, we see a proliferation of a set of discrete levels, that cluster around a single band, namely the central band corresponding to the odd-denominator fractions.
Therefore, every even-denominator fractional flux value forms the center of a butterfly .To find miniature butterfly ( centered at E = 0 ) in the butterfly graph near an arbitrary location in φ = φ 0 and with a scale, say δφ ,
- · Pick any irreducible fraction , say f c = p c q c ≈ φ 0 , where q c is even and ( q c ) 2 ≈ δφ .
- · In a Farey sequence F q c ( sequence that consists of all irreducible rationals with q max = q c ), locate the left and right Farey neighbors of f c which we denote as f L = p L q L and f R = p R q R . Simple number theoretical reasoning shows that for every given f c , there is a unique
FIG. 2. Graphs highlight the distinction between the even and the odd-denominator flux values by showing the central band of the butterfly spectrum in the immediate neighborhood of φ = 1 4 / ( left) and φ = 1 3 / ( right). This illustrates typical scenario where near even denominator fractions the fragmented band structure clusters into two distinct bands, touching at the center while for the odddenominator fractions, the fragmented structure clusters around a single band.
pair f L and f R which are Farey neighbors of f c . ( See Appendix )
- · Determine the widths of the central band ( located symmetrically about E = 0 ), corresponding to fractions f L and f R , denoted as ∆ E L and ∆ E R , by diagonalizing the Harper equation.
- · The miniature butterfly is sub-part of the butterfly graph, symmetrically located about E = 0 with butterfly center at ( φ = f , E c = 0) and its left and right boundaries confined between ( f L , ± ∆ E L 2 ) , and ( f R , ± ∆ E R 2 ) . In other words, near any evendenominator fraction, one can find a unique butterfly, with flux interval ∆ φ c = ( f R -f L ) = 1 q L R q ( horizon-
TABLE I. Given the center, locating the left and the right boundaries of the butterfly where the center and the boundaries are shown in bold.
| f c | Farey Sequence Needed | f L | f R |
|-------|-------------------------------------------------------------------------------------------------------|---------|-------|
| 1 / 4 | F 4 : 0 / 1 , 1 / 4 , 1 / 3 , 1 / 2 | 0 / 1 | 1 / 3 |
| 1 / 6 | F 6 : 0 / 1 , 1 / 6 , 1 / 5 , 1 / 4 , 1 / 3 , 2 / 5 , 1 / 2 , | 0 / 1 , | 1 / 5 |
| 3 / 8 | F 8 : 0 / 1 , 1 / 8 , 1 / 7 , 1 / 6 , 1 / 5 , 1 / 4 , 2 / 7 , 1 / 3 , 3 / 8 , 2 / 5 , 3 / 7 , 1 / 2 , | 1 / 3 | 2 / 5 |
| 1 / 8 | F 8 : 0 / 1 , 1 / 8 , 1 / 7 , 1 / 6 , 1 / 5 , 1 / 4 , 2 / 7 , 1 / 3 , 3 / 8 , 2 / 5 , 3 / 7 , 1 / 2 | 0 / 1 | 1 / 7 |
FIG. 3. (color on line) Level1 butterfly in φ intervals [1 / 3 -2 5] / where the horizontal bars show the φ -interval that is zoomed in the level2 . The magnitude of the Chern numbers for the central butterfly and its left and right harmonics are 1+3 n and | 1 -5 n | , n = 1 2 3 , , .. as shown are determined using DE equation.
Figures 3 and (4) and (5) show numerically obtained energy spectrum displaying four successive blowups of butterfly structures.
To describe hierarchical structure of the butterfly fractal, we introduce a notion of 'levels' (or generations), where higher levels ( generations) correspond to viewing the butterfly at smaller and smaller scale in E vs φ plot. At level0 we have the central butterfly in the φ -interval [0 , 1] with center at φ c = 1 2 / and colonies of butterflies to the left as well as to the right of φ = 1 2 / . ( See Fig. (3) The left colony, all sharing a common left boundary at φ = 0 are centered at 1 2 n .( Similarly, there is a right colony, centered at 1 -1 2 n , all sharing the right boundary at φ = 1 . Therefore, the boundaries of the central butterfly enclose the boundaries of the left and the right colony.
When we refer to a Butterfly in the Gplot, we mean a central tal scale) and bounded vertically by ∆ E L and ∆ E R on the left and on the right respectively.
- · Since f c is a Farey neighbors of both f L and f R , the Ford circles representing these fractions touch and such butterflies satisfy the condition, f c = f L ⊕ f R . We note that for a fixed f c , f L and f R , the Farey neighbors of f c are uniquely determined and there is no additional butterfly in the interval ∆ φ c for q < q c .
Table I illustrates the process of deterring the boundaries of the butterfly, once we choose its center.
At each level l , we label the rational flux values at the center, the left and the right boundaries as f c ( ) l = p c ( ) l q c ( ) l , f L ( ) = l p L ( ) l q L ( ) l and f R ( ) = l p R ( ) l q R ( ) l respectively.
butterfly and a set of left and a set of right colonies of butterfly that share respectively the left and the right boundary of the central butterfly as discussed below, this entire structure is reproduced at all length scales.
The levell + 1 butterfly resides in a smaller flux interval that is entirely contained in the fluxl interval of the levell butterfly. In other words, neither the left nor the right boundary points of levell + 1 overlap with the boundaries of the level- . We note that beyond levell 0 , butterflies do not exhibit reflection symmetry about their centers.
Fig. (5) suggests the existence of a fixed point butterfly as two successive levels overlay. We note that choice of any magnetic flux interval in the butterfly fractal leads to similar result as discussed later in the paper.
FIG. 4. (color on line) Level1 butterfly in φ intervals [1 / 3 -2 5] / where the horizontal bars show the φ -interval that is zoomed in the level2 . The magnitude of the Chern numbers for the central butterfly and its left and right harmonics are 1+3 n and | 1 -5 n | , n = 1 2 3 , , .. as shown.
## B. Recursion Relations for Magnetic Flux Interval
We will now describe the scaling of the magnetic flux intervals as one zooms into the butterfly fractal.
Aclose inspection of the Gplot reveals that Farey sequences are the key to systematically sub-divide the φ interval, where each new interval reproduces the entire butterfly. By Farey path, we mean a path in the Farey tree that leads from levell to level l + 1 , connecting the centers of the butterfly at two successive levels, through its boundaries . We want to emphasize that our reference to 'Farey tree' does not correspond to a path that connects rational approximants of an irrational number, it is a path that finds the entire butterfly ( its boundaries and center ) between two generations or levels of hierarchy. This Farey path described for various different parts of the Gplot , is encoded in the following recursive set of equations,
$$f _ { L } ( l + 1 ) & = f _ { L } ( l ) \$$
where the Farey sum, denoted by ⊕ between two rationals p 1 q 1 and p 2 q 2 is defined as p 1 q 1 ⊕ p 2 q 2 = p 1+ 2 p q 1+ 2 q . Since f L and f R are neighbors in the Farey tree ( see Appendix ), we have,
FIG. 5. (color on line) Figure displays an overlay of level2 and level 3 butterflies in φ intervals [4 / 11 7 , / 19] and [15 / 41 26 , / 71] giving energy scale ratio ≈ 10 as shown in red and blue respectively. The magnetic flux scaling is found to be ζ 2 as described later. This illustrates an approach to fixed point butterfly fractal as E and φ intervals shrink to zero. Note that the φ scale is shown only at the boundaries and the centers of the butterflies. The magnitude of topological numbers for the central butterflies for level2 and 3 are respectively given by (15 3) , and (56 20) , . The asymmetry of the butterfly about its center reflects the asymmetry in the Chern numbers of the left and the right harmonics which can be determined from Eq. (21).
$$p _ { R } q _ { L } - p _ { L } q _ { R } = 1$$
Simple calculations lead to following recursion relations for p x and q x where x = c, L, R :
$$\underset { l \text{ num-} } { \underset { \text{$p$\dots$} } { \text{$p$\dots$} } } \underset { \text{$p_{x}(\left|\text{$l+1})=4p_{x}(\left|l\right)=-p_{x}(\left|l\right)$} } { \text{$p_{x}(\left|l\right)=-l\right)$} } } \underset { \text{$p_{x}(\left|l\right)$} } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } \underset { \text{$p_{x}(\left|l\right)$} } { \text{$p_{x}(\left|l\right)$} } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$ } } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$}} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$}}} { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$} } } { \text{$p_{x}(\left|l\right)$}$$
We now define the ratio ζ ( ) = l q x ( +1) l q x ( ) l and Eq. (7) gives,
$$\zeta ( l ) = 4 - \frac { 1 } { \zeta ( l - 1 ) },$$
We now define ζ = lim l →∞ ζ ( ) l , where ζ satisfies the following equation,
$$( \zeta ) ^ { 2 } - 4 \zeta + 1 = 0, \ \zeta = 2 + \sqrt { 3 } \quad \quad ( 9 )$$
We can now calculate the scaling associated with the magnetic flux, the horizontal scale of the butterfly. At a given level l , the magnetic flux interval that contains the entire butterfly is,
TABLE II. Topological integers and the centers of the fixed point butterflies for four initial flux intervals, ∆ (1) φ , and following the Farey path 'LRL' between two successive levels. In all cases, the centers of the fixed point butterflies are found to be irrational numbers whose tails exhibit periodic continued fraction with entries 1 and 2 . Note that the sequence of topological integers shown here correspond to rational approximants of φ ∗ c . In all cases, ratio of the topological numbers converge to R σ = R τ .
| ∆ φ (1) | σ | τ = β - 1 2 | φ ∗ c |
|-----------------|---------------------------|------------------------|----------------------------------|
| [2 / 5 - 1 / 3] | (4 , 15 , 56 , 209 .. ) | (1 , 5 , 20 , 76 .. ) | [2 , 1 , 2 , 1 .. ] = √ 3 - 1 2 |
| [1 / 3 - 2 / 7] | (5 , 18 , 67 , 250 .. ) | (1 , 5 , 20 , 76 .. ) | [3 , 3 , 1 , 2 , 1 , 2 ... ] |
| [2 / 9 - 1 / 5] | (7 , 26 , 97 , 362 .. ) | (1 , 5 , 20 , 76 .. ) | [4 , 1 , 2 , 1 , 2 , 1 , 2 .. ] |
| [3 / 7 - 2 / 5] | (6 , 23 , 86 , 321 .... ) | (2 , 9 , 35 , 132 .. ) | [2 , 2 , 2 , 1 , 2 , 1 , 2 ... ] |
$$\Delta \phi ( l ) = f _ { R } ( l ) - f _ { L } ( l ) = \frac { 1 } { q _ { L } ( l ) q _ { R } ( l ) } \quad \ \ ( 1 0 ) \quad \Printer \text{bour}$$
Therefore, we obtain the scaling associated with φ , which we denote as R φ ,
$$R _ { \phi } = \lim _ { l \to \infty } \frac { \Delta \phi ( l ) } { \Delta \phi ( l + 1 ) } = \zeta ^ { 2 } \, \ \ \ \ ( 1 1 ) \, \quad \delta \rho )$$
To calculate ∆( ) l for a an arbitrary level- , l we use Ford circles ( see the Appendix) that provide a pictorial representations of fractions.
These solutions describe the fine structure of the butterfly near φ 0 . Consequently, the spectral gaps near φ = 1 /q have Chern numbers changing by a multiple of q . This suggests a semiclassical picture near φ = p/q in terms of an effective Landau level theory with cyclotron frequency renormalized by q . This is demonstrated in the Fig. (3) near φ = 0 and in Fig.(4) near φ = 1 3 / and 2 / 5 .
## IV. TOPOLOGICAL CHARACTERIZATION OF THE BUTTERFLY
We now calculate the quantum numbers ( στ ) associated with various gaps of the butterfly structure at all scales. These results are consequence of Diophantine equation, and are based on two corollaries, C 1 and C 2 as described below.
[ C 1 ] The Chern number associated with the four dominant gaps that form the central butterfly are, σ c = ± q c 2 .
[ C 2 ] Chern numbers of a set of gaps that begin near the boundary are given by σ b = σ c ± nq b where q b is the denominator of the fractional flux at the boundary.
Proof C 1 : f L and f R are respectively the left and the right neighbors of f c and therefore,
$$p _ { L } q _ { c } - p _ { c } q _ { L } = - 1 \text{ \quad \ \ } ( 1 2 )$$
$$p _ { R } q _ { c } - p _ { c } q _ { R } = 1 \text{ \quad \ \ } ( 1 3 )$$
This implies that for any integer n ≥ 0 , p L + np c q L + nq c are a set of left neighbors of f c and similarly p R + np c q R + nq c are a set of right neighbors of f c in the Farey tree as,
$$\stackrel { \tilde { \wedge } \cdots } { \longrightarrow } \quad | p _ { c } ( q _ { x } + n q _ { c } ) - q _ { c } ( p _ { x } + n p _ { c } ) | = 1 ; \ x = L, R. \quad ( 1 4 )$$
We now calculate the Chern number near half filling for the neighbors p x + np c q x + nq c of f c , . This will correspond to r = ( q x + nq c ) / 2 -1 . Substituting in the DE equation, we obtain,
$$\sigma = \pm \frac { q _ { c } } { 2 } ; \ \tau = \frac { 1 \pm p _ { c } } { 2 }$$
We note that the central butterfly, characterized by four wings ( gaps ) is characterized by a unique pair of topological integers determined by the Eq. 15
Proof C 2 : Chern numbers of the set of gaps near the boundary are given by the infinity of solutions depicted in Eq.(4) reside in close proximity to the flux φ and label the fine structure of the butterfly in Gplot .
DE equations at φ 0 , ρ 0 and in its vicinity ( φ 0 + δφ ) , ( ρ 0 + δρ ) are,
$$\rho _ { 0 } = \phi _ { 0 } \sigma _ { 0 } + \tau _ { 0 } \text{ \quad \quad \ \ } ( 1 6 )$$
$$\rho _ { 0 } + \delta \rho = ( \phi _ { 0 } + \delta \phi ) ( \sigma _ { 0 } + \Delta \sigma ) + ( \tau _ { 0 } + \Delta ) \quad ( 1 7 )$$
Keeping terms linear in δρ and δφ , we get
$$\rho _ { 0 } + \delta \rho = \phi _ { 0 } \sigma _ { 0 } + \delta \phi \sigma _ { 0 } + \Delta \sigma \phi _ { 0 } + \tau _ { 0 } + \Delta \tau \quad ( 1 8 )$$
Using (16), Eq. (18) reduces to,
$$\delta \rho = \delta \phi \sigma _ { 0 } + \Delta \sigma \phi _ { 0 } + \Delta \tau \text{ \quad \ \ } ( 1 9 )$$
Key observation from Eq. (19) is that unlike δρ and δφ which are can chosen to be infinitesimally small, ∆ σ and ∆ τ are integers and therefore, for small δρ and δφ we get,
$$\phi _ { 0 } \Delta \sigma + \Delta \tau = 0 ; \ \frac { \Delta \sigma } { \Delta \tau } = - \frac { q _ { 0 } } { p _ { 0 } } \quad \quad ( 2 0 )$$
Since both ∆ σ and ∆ τ are integers and p 0 and q 0 are relatively prime, the simplest solutions of Eq. (20) are,
$$\Delta \sigma = \pm n q _ { 0 } ; \ \Delta \tau = \mp n p _ { 0 } ; \ n = 0, 1, 2, \dots. \quad ( 2 1 )$$
Topological Scaling
Equation (15) relating the denominators of the fraction and Equations from Chapter II that gives recursions from the numerator and denominator of the fractions, lead to the following recursion relations for topological integers,
$$\sigma ( l + 1 ) = 4 \sigma ( l ) - \sigma ( l - 1 ) \text{ \quad \ \ } ( 2 2 )$$
$$\beta ( l + 1 ) = 4 \beta ( l ) - \beta ( l - 1 ) \text{ \quad \ \ } ( 2 3 )$$
The Eq. (23) results in fixed point solution of the ratio of integers at two successive levels,
$$\frac { \sigma ( l + 1 ) } { \sigma ( l ) } = R _ { \sigma } ( l ) = 4 - \frac { 1 } { R _ { \sigma } ( l - 1 ) } \quad \quad ( 2 4 )$$
$$\frac { \beta ( l + 1 ) } { \beta ( l ) } = R _ { \beta } ( l ) = 4 - \frac { 1 } { R _ { \beta } ( l - 1 ) } \, \quad \ ( 2 5 )$$
$$R _ { \beta } = \lim _ { l \to \infty } R _ { \beta } ( l ) = 2 + \sqrt { 3 } \quad \quad ( 2 7 ) \quad \text{the} \quad \text{the}$$
$$\mu ( \iota ) & \quad \ n _ { \beta } ( \iota - \iota ) \\ R _ { \sigma } & = \lim _ { l \to \infty } R _ { \sigma } ( l ) = 2 + \sqrt { 3 } & \quad \ ( 2 6 ) & \quad \ \ 1$$
The irrational number ζ has a continued fraction expansion, ζ = [3 1 2 1 2 1 2 , , , , , , .... ] , given by,
We will refer this irrational number as diamond mean . It is instructive to consider a somewhat general case where a butterfly have left (right) boundary located at φ L R ( ) = 1 (2 n +1) . The fixed points of the centers of these butterfly and their cor-
The topological and magnetic flux scaling of the butterfly fractal is related to Apollonian gaskets. The Appendix provides a brief introduction to these fractals and discusses its properties that are key to understand the relationship between the butterfly fractal and the Apollonian gaskets.
The central concept that links Apollonian gaskets and the butterfly is hidden in the pictorial representations of fractions using Ford circles. We now discuss this relationship. Brief introduction to Ford circles is given in the Appendix.
## A. Ford Circles, Apollonian Gasket and the Butterfly
As stated in the Appendix, a fraction f = p q can be represented by a circle ( Ford circle ) , with curvature ( inverse radius) κ = 2 q 2 . These Ford circles also provide a pictorial representation of the size of the magnetic flux intervals. Using the Eq. (7), responding Chern numbers are given by,
$$\phi _ { c _ { R ( L ) } } ^ { * } = \frac { 1 } { 2 [ n + \alpha _ { L ( R ) } ] }$$
$$\Phi \Phi ^ { \Phi \Phi ^ { \prime } } & \Phi ^ { \Phi \Phi ^ { \prime } } _ { L ( R ) \rfloor } \\ \alpha _ { R } & = \frac { \sqrt { 3 } - 1 } { 2 } = \frac { 1 - \zeta } { 2 } \\ \Phi \Phi ^ { \Phi \Phi ^ { \prime } } _ { 1 + \zeta }$$
$$\alpha _ { L } = 1 - \alpha _ { R } = \frac { 1 + \zeta } { 2 } \quad \quad \ ( 3 1 ) \\ \psi \psi \psi \psi \psi$$
$$\sigma _ { R } ( l ) = 1 + 3 n$$
$$\sigma _ { L } ( l ) = 2 + 3 n$$
This illustrates the asymmetry of the universal butterfly as the gaps on the right have smaller Chern numbers compared to the gaps on the left. It is interesting to note that unlike σ , the quantum number τ are same for the left and the right colonies of butterfly. We emphasize that although the topological numbers depend upon the initial interval, the topological scaling ratio converges to the same universal value.
Asymptotically, σ l ( ) → ζ -l , τ → ζ -l and the underlying φ interval scales as, ∆ φ → ζ 2 l .
For the butterfly fractal shown in Fig. (3,4), the entire band spectrum is numerically found to scale approximately as, ∆ ( ) E l ≈ 10 -l . Although the precise value of quantum numbers ( and hence the universal butterfly fractals) depend upon φ , the scaling ratios between two successive levels is φ independent.
Comparing scaling exponents for the size of the butterfly, ( described by ∆ φ and ∆ E ) and the corresponding topological quantum numbers, we note that the topological variations occur at a slower rate than the corresponding spectral variations as one views the butterfly at a smaller and smaller scale.
## V. INTEGRAL APOLLONIAN GASKET AND THE BUTTERFLY TOPOLOGY
$$\sqrt { \kappa _ { c } ( l + 1 ) } = 4 \sqrt { \kappa _ { c } ( l ) } - \sqrt { \kappa _ { c } ( l - 1 ) } \quad \ ( 3 4 )$$
The Ford circles do not touch and are all tangent to the horizontal axis of the butterfly graph. Introducing a scale factor ζ ( ) l as,
$$\zeta ( l ) = \sqrt { \frac { \kappa _ { c } ( l + 1 ) } { \kappa _ { c } ( l ) } },$$
we obtain,
$$\zeta ( l ) = 4 - \frac { 1 } { \zeta ( l - 1 ) }$$
For large l , ζ ( ) l → ζ ( l +1) , which we denote as ζ ∗ , which satisfies the quadratic equation,
FIG. 6. Upper Figure shows the Ford circle representation of the center ( 3 8 / blue) and the left ( 1 3 / black) and the right ( 2 5 / black). These three circles along with the horizontal black line are mutually tangent . We note that the scaling ratio between the red and the blue circles, corresponding to two consecutive generations of the butterfly asymptotically equals the ratio of the curvatures of the innermost ( solid blue) and the outermost circles ( with red dots) of the Apollonian gaskets with near D 3 symmetry ( Lower right). Lower left is the blowup of the upper figure with additional level of the hierarchy. The yellow circles ( big and small) are respectively the image of the horizontal axis reflected through the tangency points of the circles corresponding to 1 / , 3 2 / , 5 1 / 2 and 4 / 11 7 , / 19 3 , / 8 . We note that these two consecutive image circles also scale by the same ratio ( ζ ∗ ) 2 .
-
450
450
FIG. 7. Circles show sequential construction of Apollonian gaskets where the upper (lower) three circles show three generations beginning with ( -4 8 9 9) , , , ( ( -15 32 32 33) , , , ) corresponding to evenn ( oddn ) case. Note that the bounding circles with negative curvatures encode the chern numbers shown in the first row of Table I. Asymptotically, the Apollonian gasket exhibits D 3 symmetry. Also, note that all equivalent circles, asymptotically, scale by the same factor ζ 2 as shown by converging ratios in the figure and hence show the flux scaling intervals of the butterflies.
$$( \zeta ^ { * } ) ^ { 2 } - 4 \zeta ^ { * } + 1 = 0, \ \zeta ^ { * } = \lim _ { l \to \infty } \sqrt { \frac { \kappa _ { c } ( l + 1 ) } { \kappa _ { c } ( l ) } } = 2 + \sqrt { 3 } \quad \text{circle} \\ ( 3 7 )$$
Therefore, Ford circles corresponding to even-denominator fractions form a self-similar fractal consisting of circles whose curvatures scale asymptotically by ζ ∗ . Interestingly, starting with different even-denominator fraction, we get a different set, all exhibiting the same scaling. We note that,
450
$$\sqrt { \frac { \kappa _ { c } ( l ) } { \kappa _ { L } ( l ) } } \to 1 + & \sqrt { 3 }, \quad \sqrt { \frac { \kappa _ { c } ( l ) } { \kappa _ { R } ( l ) } } \to \frac { 1 + \sqrt { 3 } } { \sqrt { 3 } }, \quad \sqrt { \frac { \kappa _ { R } ( l ) } { \kappa _ { L } ( l ) } } \to \sqrt { 3 } \\ \text{Fixture } \, \text{6} \, \text{ shows the $Ford$ circle representation of the}$$
Figure (6) shows the Ford circle representation of the levels 0 and 1 of the butterfly centers f c and the boundaries, f L , f R . In general, any two successive levels l and l + 1 of the butterfly flux intervals, we have two configurations which we list below, of four mutually tangent circles with curvatures k , i i = 1 -4 representing fractions
(1) f c ( ) l , f L ( +1) l , f R ( +1) l and base line are mutually tangent where k 1 ≡ k c ( ) = 2 l q 2 c ( ) l , k 2 ≡ k L ( +1) = 2 l q 2 L ( +1) l , k 3 ≡ k R ( l +1) = 2 q R ( l +1) 2 and the base line with k 4 = 0 and
(2) f c ( l + 1) , f L ( l + 1) , f R ( l + 1) are mutually tangent where k 1 ≡ k c ( ) l = 2 q 2 c ( ) l , k 2 ≡ k L ( l + 1) = 2 q 2 L ( l + 1) , k 3 ≡ k R ( l +1) = 2 q R ( l +1) 2 and the base line with k 4 = 0
From Descartes's theorem ( Eq. (A1)), we obtain
$$k _ { c } ( \pm ) = k _ { R } ( l + 1 ) \pm k _ { L } ( l + 1 ), \text{ \quad \ \ } ( 3 9 )$$
where we can identify,
$$k _ { c } ( + ) = k _ { c } ( l + 1 ), \ \ k _ { c } ( - ) = k _ { c } ( l ) \quad \ ( 4 0 )$$
These two configurations describing butterfly fractal at two successive generations are in fact mirror image of each other, through a circle drawn from the tangency point of f L ( l +1) , f R ( l + 1) and the base line. In other words, the circles with curvatures k c ( ) l and k c ( l + 1) play the same role as the outermost and innermost circles of the Apollonian gasket in the configuration described on the right in Fig (6).
To see explicitly how the scaling ratio for the inner and the outermost radius of the Apollonian gasket is identical to that of the scaling ratio between the flux -intervals for two successive generations of the butterfly, we note that from Eq. (A4), the ratio of the outer bounding circle and the innermost circles ( See Figs. (7) ) as obtained from Eq. (A4) ) is,
$$\varsigma \cdot \cdot \cdot \cdot \cdot \\ \frac { k _ { 5 } } { k _ { 4 } } = \frac { 2 + \sqrt { 3 } } { 2 - \sqrt { 3 } } = 7 - 4 \sqrt { 3 } = \zeta ^ { 2 } \, \quad \quad ( 4 1 )$$
Therefore, the ratio of the bounding to the inner-most circle describe the scaling of the magnetic flux intervals of the butterfly. See Fig. (7).
$$\frac { k _ { 4 } ( + ) } { k _ { 4 } ( - ) } = \frac { 2 + \sqrt { 3 } } { 2 - \sqrt { 3 } }$$
In the case of butterfly, we have,
TABLE III. Comparing the scaling ratios R φ and R E for various irrational fluxes whose even denominator approximants form the fixed points of the centers of the butterfly. Each irrational value describes a set of numbers with same tail in the continued fraction expansion.
| φ ∗ c | Recursion relations for q x , p x ≡ s | R φ | ≈ R E | Farey path |
|--------------------------------------|-----------------------------------------|----------------------|---------|--------------|
| ζ 1 , 2 = [ ... 1 , 2 , 1 , 2 .... ] | s ( l +1) = 4 s ( l ) - s ( l - 1) | 2 + √ 3 = 3 . 73205 | 10 | LRL |
| ζ 1 = [ ... 1 , 1 , 1 , 1 .... ] | s ( l +1) = 4 s ( l ) + s ( l - 1) | 2 + √ 5 = 4 . 236068 | 14 | LRLRLR |
| ζ 2 = [ ... 2 , 2 , 2 , 2 ... ] | s ( l +1) = 6 s ( l ) - s ( l - 1) | 2 √ 2 +3 = 5 . 82843 | 38 | LRRL |
$$\sqrt { k _ { c } ( \pm ) } & = \sqrt { k _ { R } ( l + 1 ) } \pm \sqrt { k _ { L } ( l + 1 ) } = ( \sqrt { 3 } \pm 1 ) \sqrt { k _ { L } ( l + 1 ) }, \quad \text{that} \text{orrespo} \\ \cdots & \cdot \cdot \cdot$$
which gives, fractal closer to the Apollonian gasket. We finally remark that although the image circles of the horizontal line do not correspond to butterflies symmetric about E = 0 , their size scale by the same ratio ζ and may correspond to off-centered patterns that are beyond the subject of this paper.
$$\frac { k _ { c } ( + ) } { k _ { c } ( - ) } = \frac { k _ { c } ( l + 1 ) } { k _ { c } ( l ) } = \frac { 2 + \sqrt { 3 } } { 2 - \sqrt { 3 } } \quad \quad ( 4 4 ) \quad \begin{pmatrix} F \\ \text{hibi} \\ \text{num} \\ \text{[1 /^{\prime}} \end{pmatrix}$$
As described above, configuration of circles underlying the butterfly fractal appears to be a special case of a general construction involving four mutually tangent circles. However, we note that if we consider the mirror image of the horizontal (base) line through the tangency points of the Ford circles corresponding to f L ( l + 1) , f R ( l + 1) , f c ( ) l and f L ( l + 1) , f R ( l + 1) , f c ( l + 1) ( See Fig. (6)) , we obtain configuration involving four mutually tangent circles, each corresponds to non-zero curvature. This puts butterfly
## VI. GOLDEN AND SILVER MEAN HIERARCHIES
As discussed above, we have investigated the entire Gplot by zooming in the equivalent sets of butterflies and calculating the asymptotic scaling properties of the fixed point butterfly fractal. In contrast, earlier studies have explored the butterfly fractal by starting with a fixed irrational number. We briefly investigate this line of analysis of the butterfly fractal for the golden and the silver mean flux values. We follow the irrational magnetic flux by following a sequence of its rational approximants with even denominators where the relation between the boundaries and the center is always given by, f c ( ) = l f L ( ) l ⊕ f R ( ) l . However, the ordered set of three rationals f L , f c and f R , need not belong to the set of rational approximants of the irrational magnetic flux. √
For the golden-mean γ g = ( 5 -1) / 2 , its even denominator approximants (1 / 2 3 , / 8 13 , / 34 55 , / 144 233 , / 610 ... ) form the centers of the butterfly at E = 0 with Chern numbers (1 , 4 17 72 305 , , , .... ) . As shown in upper part of the Fig. (8), the centers do not form a monotonic sequence and therefore the equivalent set of butterflies correspond to the Farey path
Figure (7) shows the integral Apollonian gaskets, exhibiting hierarchical set of integers that describe quantum numbers of the butterfly obtained by zooming the interval [1 / 3 -2 / 5] . To obtain complete hierarchy, of topological integers, we begin with two sets of Apollonian gaskets with curvatures (2 , 2 3 , , -1) and (8 , 9 9 , , -4) , and use the recursion relation (7) for the negative curvatures. We obtain all the Chern numbers as listed in the first row of Table I , in fact all equivalent circles scale by ζ 2 suggests that the butterfly fractal characterized by the Farey path 'LRL' corresponding to the whole set of irrationals ζ 12 are described by the Apollonian gasket.
## LRLRLR or RLRLRL .
The silver-mean γ s = 2 -1 with rational approximants ( 1 / 2 2 , / 5 5 , / 12 12 , / 29 29 , / 70 70 , / 169 169 , / 408 .... ) result in an silver hierarchy with Chern numbers (1 , 6 35 204 , , ... ) and correspond to the Farey path LRRL .
√
Table III compares the three hierarchies which we will also refer as diamond, silver and golden hierarchy. Figure (9) shows the three generations of the butterflies and clearly illustrate the dominance of diamond hierarchy as the scale factors for both the φ and the E intervals are smaller than the corresponding scale factors for the silver and golden cases.
We also note that unlike diamond hierarchy, silver and golden mean hierarchies do not map to integral Apollonian gaskets.
As a final comparison between three irrationals, we look at the three Polygons, as shown in Fig. (11) whose angles relate to these irrationals. Mathematical simplicity of the diamond mean is rather intriguing and it remains to be seen if the area of dodecagonal being equal to 2 R 2 ( where R is the radius of the circle ) provides any further insight towards its relationship to D 3 symmetry of the Apollonian gasket.
FIG. 8. Upper graph is a schematically representation of magnetic flux intervals of the butterfly fractal for three levels of nesting showing ζ 12 , ζ 2 and ζ 1 hierarchies, all starting with the common level1 sub-interval 1 / 3 -2 / 5 . Lower graph shows the corresponding Fords circles ( from left-right ) for ζ 1 2 , , ζ 2 and ζ 1 hierarchies for level1 ( black), level2 (blue) and level3 (red). This pictorial representation of the center and the boundaries of the butterfly illustrate distinctions between the these three hierarchies where the golden-tail butterfly flips between two successive levels. The entire butterfly plot obtained numerically in shown in Fig. (9).
FIG. 9. Butterfly plots for three levels for ζ 12 ( red), ζ 2 ( blue ) and ζ 1 (green) hierarchies, all starting from the black butterfly at the top. The horizontal lines on the black plot show the appropriate window that is zoomed and displayed below. We note that golden-mean hierarchy has its butterfly flips between two successive generations.
## VII. PERIODIC DRIVING AND GAP AMPLIFICATION
We next address the question of physical relevance of states of higher topological numbers in view of the fact that size of the spectral gaps decreases exponentially with σ , as confirmed by our numerical study of the system described by Eq. (3). We now show that by perturbing such systems, we can induce quantum phase transitions to topological states with n > 0 given by (4) with dominant gaps characterized by higher Chern numbers. We study butterfly spectrum for a periodically kicked quantum Hall system[19] where J y is a periodic function of time t with periodT [19],
$$J _ { y } = \lambda \sum _ { n } \delta ( t / T - n )$$
FIG. 10. (color on line) Upper half of the butterfly spectrum for static ( J x = J y )(upper) and lower-half of the quasienergy spectrum for driven system with ¯ = J . 33 , λ ¯ = 1 1 . (lower). Dominance of higher Chern states in kicked system is due to phase transitions where Chern1 state for φ = 1 3 / is transformed into Chern -2(= 1 -3) state while Chern ( -2 , 1 ) states of φ = 2 5 / evolve into Chern 3(= -2 + 5) , -4(= 1 -5) states.
The time evolution operator of the system, defined by | ψ t ( ) 〉 = U t ( ) | ψ (0) 〉 , has the formal solution U t ( ) = T exp[ -i ∫ t 0 H t ( ′ ) dt ′ ] , where T denotes time-ordering and we set glyph[planckover2pi1] = 1 throughout. The discrete translation symmetry H t ( ) = H t ( + T ) leads to a convenient basis {| φ glyph[lscript] 〉} , defined as the eigenmodes of Floquet operator U T ( ) ,
$$U ( T ) | \phi _ { \ell } \rangle = e ^ { - i \omega _ { \ell } T } | \phi _ { \ell } \rangle.$$
We have two independent driving parameters, ¯ = J J T/ x glyph[planckover2pi1] and ¯ = λ λT/ glyph[planckover2pi1] . For rational flux φ = p/q , U is a q × q matrix with q quasienergy bands that reduce to the energy bands of the static system as T → 0 .
New topological landscape of the driven system as shown in the Fig. (10) can be understood by determining the topological states of flux values corresponding to simple rationals such as 1 3 / , 2 / 5 . In the Fig. (10), parameter values correspond to the case where the Chern1 gap associated with 1 / 3 has undergone quantum phase transition to a n = 1 solution of the DE (Eq. (4)) and the Chern--2 1 , states of 2 / 5 have also undergone transitions to Chern3 , -4 state. This almost wipes out the Chern1 state from the landscape, exposing the topological states of higher Cherns that existed in tiny gaps in the static system. We note that, topological invariants associated with irrational flux values can not change under driving and in view of infinite set of irrationals in the vicinity of every rational, the ordering of the Cherns as we vary the filling factor remains unchanged.
Gap amplifications of states in periodically driven quantum hall system may provide a possible pathway to see fractal aspects of Hofstadter butterfly and engineer states with large Chern numbers experimentally. Recent experiments with ultracold atoms[17] [20] and shaken optical lattices[21] offer a promising means to realize the butterfly and its transformation in driven systems.
## VIII. CONCLUSIONS AND OPEN QUESTIONS
The unveiling of a dodecagonal quasicrystal[15], also characterized by integral Apollonian gasket with D 3 symmetry that fully encodes the topological hierarchy of the butterfly fractal is the central result of this paper. However, the relationship between these two symmetries remain obscure. We note that these results also apply to other 2D lattices such as graphene in the magnetic field. The associated scaling for topological quantum numbers is universal, independent of lattice symmetry and perhaps indicates result of greater validity and significance. Why dodecagonal quasicrystals emerge as the dominant hierarchy remains an open question. The fact that only these symmetries map to integral Apollonian gaskets makes the puzzle deeper and more intriguing. Emergence of hidden symmetries, as energy scale approaches zero is reminiscent of phenomena such as asymptotic freedom in Quantum Chromodynamics.
Recently, there is renewed interest in quasiperiodic systems[22-24] due to their exotic characteristics that includes their relationship to topological insulators. Our findings about new symmetries and topological universality will open new avenues in the study of interplay between topology and self-similarity in frustrated systems.
## Appendix A: Geometrical Representation of Fractions: Farey Tree, Ford Circles and Descartes' Theorem
Farey Sequences of order n is the sequence of completely reduced fractions between 0 and 1 , which have denominators less than or equal to n, arranged in order of increasing size. We note that two neighboring terms say p 1 q 1 and p 2 q 2 in any Farey sequence satisfy the equation | q 1 p 2 -q 2 p 1 | = 1 and their difference | p 1 q 1 -p 2 q 2 | = 1 q 1 2 q . Of all fractions that are neighbors of p q ( q > 1 ) only two have denominators less than q . It is this property that associates a unique pair p L /q L and p R /q R with a fraction p /q c c when q c is even and q L < q c and q R < q c are odd[25].
Ford Circles ( See Fig. (12)) provide geometrical representation of fractions[25]. For every fraction p q (where p and q are relatively prime) there is a Ford circle, which is the circle with
A = 3R2 = 3[2 + VBja?
A 2vR? = 2(1 + V2)a?
A = 5 [0.4754]R? = 5 . (0.3441)02
FIG. 11. Polygons whose angles relate to the three irrational numbers. Here R is the radius of the circumscribed circle, a is the length of the side of the regular polygon and A is the area of the polygon. From left to right , the polygons with vortex angles 30 45 72 , , degrees or equivalently 2 π 12 , 2 π 8 and 2 π 5 radian reveal relationship between the three irrationals and the corresponding 12 8 5 , , fold symmetry. The area A reveals hidden mathematical simplicity of the diamond mean.
radius 1 2 q 2 and center at p q and tangent to the base line. Two Ford circles for different fractions are either disjoint or they are tangent to one another. In other words, two Ford circles never intersect. If 0 < p q < 1 , then the Ford circles that are tangent to the ford circle centered at p q , are precisely the Ford circles for fractions that are neighbors of p q in some Farey sequence.
Three mutually tangent Ford circles in the Fig. (12), along with the base line that can be thought of as a circle with infinite radius, are a special case of four mutually tangent circles such as those shown in Fig. (6 R). Relation between the radii of four such mutually tangent circles is given by Descartes's theorem .
Descartes's theorem states that if four circles are tangent to each other, and the circles have curvatures ( inverse of the radius) k i (for i = 1 , ..., 4 ), a relation between the curvatures k i of these circles is given by,
$$( k _ { 1 } + k _ { 2 } + k _ { 3 } + k _ { 4 } ) ^ { 2 } = 2 ( k _ { 1 } ^ { 2 } + k _ { 2 } ^ { 2 } + k _ { 3 } ^ { 2 } + k _ { 4 } ^ { 2 } ). \quad ( A 1 )$$
Solving for k 4 in terms of k i , i = 1 2 3 , , gives,
$$\mathfrak { l } _ { 4 } \quad \ k _ { 4 } = k _ { 1 } + k _ { 2 } + k _ { 3 } \pm 2 \sqrt { k _ { 1 } k _ { 2 } + k _ { 2 } k _ { 3 } + k _ { 3 } k _ { 1 } } \quad ( \mathbf A 2 )$$
The two solutions ± respectively correspond to the inner ( solid blue circle) in Fig. (6 R ) and the outer bounding circle
FIG. 12. Ford circles provide geometrical representation of Farey sequences. Figure where every fraction p/q is associated with a circle of radius 1 2 q 2 , describes a special case of Descartes's theorem as the straight line is a circle of zero curvature.
( circle with red dots ). The consistent solutions of above set of equations require that bounding circle must have negative curvature. Denoting the curvature of the inner circle as k 5 , it follows that
$$k _ { 4 } + k _ { 5 } = 2 ( k _ { 1 } + k _ { 2 } + k _ { 3 } ) \quad \quad ( A 3 ) \quad /$$
Important consequence of the Eq. (A3) is the fact that if k , i i = 1 -4 are integers, k 5 is also an integer.
Patterns obtained by starting with three mutually tangent circles and then recursively inscribing new circles in the curvilinear triangular regions formed between the circles are Known as the Apollonian gasket, or Curvilinear Sierpinski Gasket , as the three mutually tangent circles form a triangle in curved space. An Apollonian gasket describes a packing of circles arising by repeatedly filling the interstices between four mutually tangent circles with further tangent circles.
Integral Apollonian Gasket has all circles whose curvatures are integers. As described above, such a fractal made up integers alone can be constructed if the first four circles have integer curvatures. ( See figs (7)).
Apollonian gaskets with D 3 symmetry is a fractal with D 3 symmetry, which corresponds to three reflections along diameters of the bounding circle (spaced 120 degrees apart), along with three-fold rotational symmetry. Such a gasket can be constructed if the three circles with smallest positive curvature have the same curvature. Setting k 1 = k 2 = k 3 = k in Eq. (A2), we obtain
$$k _ { 4 } ( \pm ) = ( 3 \pm 2 \sqrt { 3 } ) k \quad \quad \quad ( A 4 )$$
FIG. 13. Upper graph illustrates reflection through a circle in a configuration involving four mutually tangent circles. Smallest ( dashed red) and largest ( red) circles are mirror image of each other through reflection through the yellow circle, a circle drawn through the tangency points of the three circle. The lower graphs illustrates this further, highlighting the preservation of tangency. The red circle (d) tangent to the three blue circles, a, b, c , and its reflection in the yellow circle, drawn at the tangency points of the circles a, b, c , producing the image ( dotted red) circle labeled d ′ . Create a circle through the intersection points of three mutually tangent circles. The three tangent circles are blue in the picture below.
The fact that the ratio k /k 4 is an irrational number means that no integral Apollonian circle packing possesses D 3 symmetry, although many packings come close. As illustrated in the Fig. (7), this symmetry is restored in the iterative process where the inner circle at level l becomes the outer circle at level l +1 .
Apollonian Gasket-Kaleidoscope Another remarkable property of the Apollonian gasket is that the whole Apollonian gasket is like a kaleidoscope where the image of the first four circles is reflected again and again through an infinite collection of curved mirrors.
This is illustrated in Fig. (13) using an operation called in-
- [1] D. R. Hofstadter, Physical Review B 14 , 2239 (1976).
- [2] D. Hopstadter, The world 'Gplot' was used by Hofstadter after a friend of his struck by infinitely many infinities of the plot called it a picture of Go: Godel Escher and Bach (Vintage Books Edition, 1989).
- [3] D. J. Thouless, M. Kohmoto, M. P. Nightingale, and M. den Nijs, Physical Review Letters 49 , 405 (1982).
- [4] C. R. Dean, L. Wang, C. Maher, Forsythe, Ghahari, F. Gao, J. Katoch, M. shigami, P. Moon, M. Koshino, T. Taniguchi, K. Watanabe, K. L. Shepard, H. J, and P. Kim, Nature 45 , 12186 (2013).
- [5] M. Aidelsburger, M. Atala, M. Lohse, J. T. Barreiro, B. Paredes, , and I. Bloch, Physical Review Letter 111 , 108301 (2013).
- [6] D. Weld, P. Medley, H. Miyake, D. Hucul, D. Pritchard, and W. Ketterle, Physical Review Letters 111 , 108302 (2013).
- [7] G. H. Wannier, Phys Status Solidi B 88 , 757 (1978).
- [8] A. MacDonald, Physical Review B 88 , 6713 (1983).
- [9] M. Wilkinson, Journal of Physics A, Math. Gen. 20 , 4337 (1987).
- [10] M. Wilkinson, Journal of Physics A, Math. Gen. 21 , 8123 (1994).
- [11] M.-C. Chang and Q. Niu, Physical Review B 53 , 7010 (1996).
- [12] A. MacDonald, Journal of Physics, B 42 , 055302 (2009).
- [13] Z. J. Dana I, Avron Y, Journal of Physics, C, Solid State 18 ,
version , a classic tool to understand configurations involving mutually tangent circles, which is can be thought of as a reflection through a circle. The key feature of the inversion that maps circles to circles, is that it preserves tangency as both the circle and its reflected image are tangent to same set of circles as illustrated in Fig. (13).
L679 (1985).
- [14] Hardy and Wright, Introduction to theory of Numbers , 1st ed. (Oxford University Press, 1979).
- [15] A. MacDonald, Physical Review B 39 , 10519 (1989).
- [16] B. Mandelbrot, Fractal Geometry of Nature (Cambridge University Press, 1982).
- [17] Y.-J. Lin, R. L. Compton, K. Jimnez-Garca, J. V. Porto, and I. B. Spielman, Nature 462 , 628 (2009).
- [18] P. G. Harper, Proc of Physical Society of London A68 , 874 (1955).
- [19] M. Lababidi, I. I. Satija, and E. Zhao, Physical Review Letters 112 , 026805 (2014).
- [20] E. Zhao, N. Bray-Ali, C. J. Williams, I. B. Spielman, and I. I. Satija, Phys. Rev. A 84 , 063629 (2011).
- [21] W. Zheng and H. Zhai, arXiv:1402.4034 [cond-mat] (2014), arXiv: 1402.4034.
- [22] Y. E. Kraus and O. Zilberberg, Physical Review Letters 109 , 116404 (2012).
- [23] I. I. Satija and G. Naumis, Physical Review B 88 , 054204 (2013).
- [24] I. Dana, Phys. Rev. B 89 , 205111 (2014).
- [25] L. R. Ford, The American Mathematical Monthly 39 , 586 (1938).
| null |
[
"Indubala Satija"
] |
2014-08-05T15:34:12+00:00
|
2014-12-14T02:32:05+00:00
|
[
"cond-mat.dis-nn"
] |
Topology and Self-Similarity of the Hofstadter Butterfly
|
We revisit the problem of self-similar properties of the Hofstadter butterfly
spectrum, focusing on spectral as well as topological characteristics. In our
studies involving any value of magnetic flux and arbitrary flux interval, we
single out the most dominant hierarchy in the spectrum, which is found to be
associated with an irrational number $\zeta=2+\sqrt{3}$ where nested set of
butterflies describe a kaleidoscope. Characterizing an intrinsic frustration at
smallest energy scale, this hidden quasicrystal encodes hierarchical set of
topological quantum numbers associated with Hall conductivity and their scaling
properties. This topological hierarchy maps to an {\it integral Apollonian
gasket} near-$D_3$ symmetry, revealing a hidden symmetry of the butterfly as
the energy and the magnetic flux intervals shrink to zero. With a periodic
drive that induces phase transitions in the system, the fine structure of the
butterfly is shown to be amplified making states with large topological
invariants accessible experimentally.
|
1408.1007v1
|
## Coherent transient transport of interacting electrons through a quantum waveguide switch
Nzar Rauf Abdullah, 1, ∗ Chi-Shung Tang, 2 Andrei Manolescu, 3 and Vidar Gudmundsson 1, †
1
Science Institute, University of Iceland, Dunhaga 3, IS-107 Reykjavik, Iceland 2 Department of Mechanical Engineering, National United University, 1, Lienda, Miaoli 36003, Taiwan 3 Reykjavik University, School of Science and Engineering, Menntavegur 1, IS-101 Reykjavik, Iceland
We investigate coherent electron-switching transport in a double quantum waveguide system in a perpendicular static or vanishing magnetic field. The finite symmetric double waveguide is connected to two semi-infinite leads from both ends. The double waveguide can be defined as two parallel finite quantum wires or waveguides coupled via a window to facilitate coherent electron inter-wire transport. By tuning the length of the coupling window, we observe oscillations in the net charge current and a maximum electron conductance for the energy levels of the two waveguides in resonance. The importance of the mutual Coulomb interaction between the electrons and the influence of two-electron states is clarified by comparing results with and without the interaction. Even though the Coulomb interaction can lift two-electron states out of the group of active transport states the length of the coupling window can be tuned to locate two very distinct transport modes in the system in the late transient regime before the onset of a steady state. A static external magnetic field and quantum-dots formed by side gates (side quantum dots) can be used to enhance the inter-waveguide transport which can serve to implement a quantum logic device. The fact that the device can be operated in the transient regime can be used to enhance its speed.
## I. INTRODUCTION
Various schemes associated with quantum computing have been proposed for quantum information storage and transfers, such as superconducting coplanar waveguide resonators and dual waveguide devices [1, 2]. A waveguide can be defined as a quantum wire in which the electron wave propagates in quantized modes without losing phase coherence at low temperature. Various proposals have been suggested to implement a semiconductor qubit [3-5]. Among these, an idea is to use parallel quantum waveguides with a coupling-window placed between them [6]. Tuning the window-coupling allows an electron wave interference between the waveguides and can switch the electron motion from one waveguide to the other [7].
which an external magnetic field is applied to switch electron motion from one quantum-waveguide to the other [7, 10]. Another suggestion for a switching-qubit is to vary the length of the coupling window (CW) L CW . By varying the length of the CW, a maximum inter-waveguide tunneling can be found that increases the efficiency of the device [7]. Another approach for a switching-qubit is to define a saddle potential in the coupling window instead of a hard-walled well potential. The saddle-potential makes a soft barrier between the two waveguides that washes out resonance peaks, but decreases the efficiency of the qubit [11].
The interference in the coupling-region specifies the possible types of qubit-operations: The electron wave can be transferred through either waveguide only (pass gateoperation). It can travel through both waveguides, split equally, implementing the so called square-root-of-Not ( √ NOT) operation [8]. It can switch totally from one waveguide to the other (Not-operation) [9]. Or it might switch from the first waveguide to the second one and then re-enter the first waveguide (CNot-operation) [10]. There are several parameters and phenomena that can modify the efficiency of the kind of qubit-operation that are mentioned above such as external magnetic field, the Coulomb interaction, electrostatic gate voltage, and the geometry of the coupling window between the waveguides.
Ferry's group proposed magnetically switching transport in an asymmetric double quantum waveguide, in
∗ [email protected]
† [email protected]
In the present work, we consider a double quantum waveguide (DQW) system with a CW between the waveguides in a static perpendicular magnetic field. Both ends of the DQW are connected to semi-infinite leads with an applied source-drain bias. The coherent electron transport is investigated in the system by using a non-Markovian master equation [12-15]. Smooth Gaussian potentials are used to define the barrier between the waveguides and the saddle-like CW potential. In addition, two dots are embedded in the waveguides close to the CW to enhance inter-waveguide transport. We will investigate coherent electron transport in the system by varying the CW length.
Our results reveal the following conclusions: First, we predict a maximum conductance at resonant energylevels. At a resonance energy, inter-waveguide scattering is enhanced and the charge density splits equally between the waveguides in the absence of the Coulomb interaction as two-electron states are activated in the transport. In this case, the DQW essentially works as a √ NOToperation qubit. In the presence of the Coulomb interaction, the charge current density does not split equally between the waveguides anymore at the resonance since two-electron states are blocked for the bias range used.
The strength of the blocking can be varied by the geometry or the material of the system. The transient transport through two capacitively coupled parallel quantum dots has been investigated by Moldoveanu et al. in a lattice model for a weaker Coulomb interaction than is assumed here [16]. Second, an external static magnetic field can increase inter-waveguide backward scattering leading to a decrease in the efficiency of the qubit. These results make clear the difference to previous schemes where variable magnetic field was used to control the switching between asymmetric waveguides [6, 11]. Third, the influence of the side quantum dots (QDs) can be to induce more energy-levels into resonance, and consequently increase the coupling between the waveguides and the interwaveguide transport mechanism. In the presence of side QDs, a Not-operation qubit is realized. The efficiency is slightly decreased in the presence of side QDs because inter-waveguide backward scattering is enhanced. At high magnetic field such that the effective magnetic length is comparable to the radius of the dots, variation of length of the CW does not influence the electron transport characteristics anymore because the electrons are localized in the side dots. Thus, the net charge current is extremely suppressed.
The paper is organized as follows: In Sec. II, we present the model describing the window-coupled double waveguide system based on the quantum master equation (QME) approach. Section III presents our numerical results and discussion. Concluding remarks are addressed in Sec. IV.
## II. MODEL AND THEORY
In this section, we present a method for calculating the transient ballistic transport in the coupled-window quantum waveguides without and with side QDs. We consider two laterally coupled waveguides connected to two semi-infinite wires from both ends. The lower waveguide (control) and the upper waveguide (target) are coupled through a CW with length L CW as is schematically shown in Fig. 1. We consider a control-QD (QD C ) and a target-QD (QD T ) embedded, respectively, in the controland the target-waveguide near the CW. The total system is exposed to an external magnetic field in the z -direction Bˆ z. The system is designed such that the electrons are injected from the left lead to only the control-waveguide (red arrow).
Our aim here is to study the charging of, and the output from each of the waveguides by varying the CW length L CW (blue dashed arrow). In addition, the effects of the Coulomb interaction, the external magnetic field, and the side dots on the conductance will be explored.
The DQW is a two dimensional system with hard-wall confinement in the contact area to the external leads at x = ± L / x 2 with L x the length of the waveguides in the transport direction and parabolic confinement in the y -direction. The single-electron Hamiltonian of the DQW
FIG. 1. (Color online) Schematic diagram depicting the double waveguide with a CW of length L CW (blue dashed arrow). A quantum dot QD C (QD ) is embedded in the control (tarT get) waveguide. An electron from left side enters the control waveguide (red arrow). Asymmetric contacts are indicated by the light red rectangles marked L and R.
including side QDs in an external magnetic field is
$$\text{metic} _ { \text{on of} } \quad h _ { S } = \frac { ( p + e \mathbf A _ { \text{ext} } ) ^ { 2 } } { 2 m ^ { * } } + V _ { c } ( x, y ) + V _ { \text{DW} } ( \mathbf r ) + V _ { \text{QD} s } ( \mathbf r ) \,, \ ( 1 )$$
where p indicates the canonical momentum, A ext = (0 , -By, 0) is vector potential of the external magnetic field, and m ∗ is the effective mass of an electron. The confining potential of the DQW is V c ( x, y ) = V c ( x ) + V c ( y ), where V c ( x ) stands for a hard-wall confining potential and V c ( y ) = 1 2 m ∗ Ω 2 0 y 2 refers to a parabolic confining potential with characteristic energy /planckover2pi1 Ω . The Gaussian po0 tential barrier between the two waveguides is presented by V DW ( r ) and the side QDs are defined by V QDs ( r ). We diagonalize the Hamiltonian of the central system presented in Eq. (1) using a basis [15] built with the eigenfunctions of the first two terms of (1) to find the singleelectron energy spectrum E n , where n is the composite quantum number of a single-electron state.
Below, we shall demonstrate how the DQW is connected to the leads and show the time evolution of electron in the system.
## A. DQW connected to leads
We connect the DQW to two leads that act as electron reservoirs via a contact region. The central system and the leads have the same confinement energy and the same strength of the perpendicular magnetic field. The total Hamiltonian of the system is then
$$\text{guide} \quad & \text{Hamiltonian of the system is then} \\ \text{$1$ the} \quad & H ( t ) = \sum _ { n } E _ { n } d _ { n } ^ { \dagger } d _ { n } + \frac { 1 } { 2 } \sum _ { i j r s } ( V _ { \text{Coul} } ) d _ { i } ^ { \dagger } d _ { j } ^ { \dagger } d _ { s } d _ { r } \\ \text{effects} \quad & + \sum _ { l = L, R } \int d q \, \epsilon ^ { l } ( q ) c _ { q } ^ { l } \, ^ { \dagger } c _ { q } ^ { l } \\ \text{$l-wall} \\ \text{ds at} \\ \text{n the} \quad & \text{$l-L,R}$} \quad \chi ^ { l } ( t ) \sum _ { n } \int d q \, \left [ c _ { q } ^ { l } \, ^ { \dagger } T _ { q } ^ { l } d _ { n } + d _ { n } ^ { \dagger } ( T _ { n } ^ { l } q ) ^ { * } c _ { q } ^ { l } \right ]. \\ \text{he $y$} \quad & \text{The first term describes the central system, where $E_{n}$ is} \quad. \quad. \quad. \quad. \quad.$$
The first term describes the central system, where E n is the single-electron energy of state | n 〉 of the central sys-
tem, and d † n ( d n ) denotes the electron creation (annihilation) operator respectively. The second term is the mutual Coulomb interaction between electrons in the central system. The third term of the Hamiltonian denotes the semi-infinite l th lead, where l refers to the left L or the right R lead with the dummy index q representing the 'momentum' of an electron in the lead and its subband index [15], /epsilon1 l ( q ) is the single-electron energy spectrum in lead l , and the electron creation and annihilation operator in lead l are c l q † and c l q , respectively.
The last term is the time-dependent coupling Hamiltonian that connects the central system, the DQW to the external leads via a coupling tensor T l q n , and χ l ( ) t dictates the time-dependence of the coupling. The coupling Hamiltonian describes the transfer of an electron between the single-electron state of the central system | n 〉 and the single-electron state of the leads | q 〉
$$T _ { q n } ^ { l } = \int d { \mathbf r } d { \mathbf r } ^ { \prime } \psi _ { q } ^ { l } ( { \mathbf r } ^ { \prime } ) ^ { * } g _ { q n } ^ { l } ( { \mathbf r }, { \mathbf r } ^ { \prime } ) \psi _ { n } ^ { S } ( { \mathbf r } ). \quad ( 3 ) \quad \text{into} \\ \infty \,. \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots \quad \, \dots.$$
ψ S n ( r ) and ψ l q ( r ′ ) are the corresponding single-electron wave functions of the DQW system and the leads. g l q n ( r r , ′ ) is a nonlocal coupling kernel
$$g _ { q } ^ { l } ( \mathbf r, \mathbf r ^ { \prime } ) & = g _ { 0 } ^ { l } \exp \left [ - \delta _ { x } ^ { l } ( x - x ^ { \prime } ) ^ { 2 } - \delta _ { y } ^ { l } ( y - y ^ { \prime } - \alpha ) ^ { 2 } \right ] \quad \text{the c} \, \\ & \times \exp \left ( - \Delta _ { n } ^ { l } ( q ) / \Delta \right ), \quad \text{$(4)} \, \\ \text{where $g_{0}$ is the coupling strength, $\delta_{x}^{l}$ and $\delta_{y}^{l}$ are the cou-}.$$
where g 0 is the coupling strength, δ l x and δ l y are the coupling range parameters in the x - and y -direction, respectively, ∆ controls together with ∆ l n ( q ) = | E n -/epsilon1 l ( q ) | the energy affinity between states in the leads and the central system [13], and α is a skewing parameter that shifts the weight of the coupling to the lower waveguide for an appropriate value stated below.
## B. QME and Charge current density
In this subsection, we present how the time evolution of the open system is calculated in order to study time-dependent transport properties within the QME approach [17]. The time evolution of our system obeys the quantum Liouville-von Neumann equation
$$\frac { d W ( t ) } { d t } = - \frac { i } { \hslash } \left [ H ( t ), W ( t ) \right ], \text{ \quad \ \ } ( 5 )$$
where W t ( ) is the density operator of the total system.
In order to focus our attention on the open central system we introduce a reduced density operator (RDO) with respect to it, by taking the trace over the Fock space with respect to the leads ρ t ( ) = Tr L Tr R W t ( ) with the initial condition W t < t ( 0 ) = ρ L ρ R S ρ , where ρ L and ρ R are the density operators of the isolated left and the right leads, respectively. The density operator of the l th lead can be defined as ρ l = e -β H ( l -µ N l l ) / Tr l { e -β H ( l -µ N l l ) } , where β = 1 ( / k B T ) and k B is the Boltzmann constant, and N l is the number of electrons in the leads. The density operator of the central system at t = t 0 being ρ t ( 0 ) = ρ S [18] allows us to obtain the equation of motion for the RDO in the following form [19]
$$\frac { d \rho ( t ) } { d t } & = - i \mathcal { L } _ { \text{S} } \rho ( t ) + \int _ { t _ { 0 } } ^ { t } d t ^ { \prime } \mathcal { K } ( t, t ^ { \prime } ) \rho ( t ^ { \prime } ), \quad ( 6 ) \\ \cdots & \cdots$$
where L · S = [ H , S · ] / /planckover2pi1 is the Liouvillian with respect to the time-independent Hamiltonian H S of the central system and K ( t, t ′ ) is an integral kernel with its roots in the dissipative time-dependent coupling to the leads [15, 19].
In order to explore the switching processes between the waveguides, we define the net charge current as
$$I _ { \mathrm Q } ( t ) = I _ { \mathrm L } ( t ) - I _ { \mathrm R } ( t ), \text{ \quad \ \ } ( 7 )$$
where I L ( ) refers to the partial current from the left lead t into the control-waveguide and I R ( ) indicates the partial t current into the right lead from both waveguides [20]. In Eq. (7) the negative sign in front of I R is since a positive charge current is defined as the current into the central system, but the currents I L and I R are defined positive from left to right.
To monitor the dynamic evolution of the electrons in the central system, we calculate the expectation value of charge current density operator
$$\mathbf J ( \mathbf r, t ) = \text{Tr} \left ( \hat { \rho } ( t ) \hat { \mathbf J } ( \mathbf r ) \right ), \quad \quad ( 8 ) \\ \text{charge current density operator is defined } \mathbf h v$$
where the charge current density operator is defined by
$$\Lambda ( t ) | \text{ the } \stackrel { \sim } { \text{ $ \text{ $ \text{ $ \text{ $ \ell$} } } } \Lambda ( \mathbf r ) & = \sum _ { n n ^ { \prime } } \left ( \frac { \ e \hbar } { 2 m ^ { * } i } \left [ \psi _ { n } ^ { S * } ( \mathbf r ) ( \nabla \psi _ { n ^ { \prime } } ^ { S } ( \mathbf r ) ) - ( \nabla \psi _ { n } ^ { S * } ( \mathbf r ) ) \psi _ { n ^ { \prime } } ^ { S } ( \mathbf r ) \right ] \\ & \quad + \frac { \ e ^ { 2 } } { m ^ { * } } \mathbf A _ { \text{ext} } ( \mathbf r ) \psi _ { n } ^ { S * } ( \mathbf r ) \psi _ { n ^ { \prime } } ^ { S } ( \mathbf r ) \right ) d _ { n } ^ { \dagger } d _ { n ^ { \prime } }. \\ & \quad \quad \quad$$
In the following, we shall demonstrate the influence of the length of the CW, external magnetic field, and the side QDs on the coherent electron transport through the system and the charge current density in the central system in order to investigate inter-waveguide forward and backward scattering processes.
## III. RESULTS
In this section, we shall discuss our numerical results of the ballistic transport properties through the double waveguide system made of a GaAs semiconductor with the electron effective mass m ∗ = 0 067 . m e . The length of the central system is L x = 300 nm which is assumed to be much smaller than the phase coherent length L φ . At low temperature T ∼ 0 1 . -2 0 K, the phase coher-. ent length of semiconductor based (GaAs/Al 1 -x Ga As) x electron waveguide is ∼ (30 -40) × 10 3 nm [21]. The coherence in the electron transport in the double waveguide
is an essential requirement for constructing a qubit. For the system placed in a static or vanishing external magnetic field, its length scales can be characterized by the effective magnetic length a w = { /planckover2pi1 / m ( ∗ √ ( ω 2 c +Ω )) 2 0 } 1 / 2 , where the cyclotron frequency ω c = eB/m ∗ and the transverse confinement energy is /planckover2pi1 Ω 0 = 1 0 meV. Nu-. merically the effective magnetic length can be expressed as
$$a _ { w } & = \left ( \frac { \hbar } { m ^ { * } \Omega _ { 0 } } \right ) ^ { 1 / 2 } \left ( \frac { 1 } { 1 + ( e B / ( m ^ { * } \Omega _ { 0 } ) ) ^ { 2 } } \right ) ^ { 1 / 4 } \quad \text{where} \\ & = \frac { 3 3. 7 4 } { \sqrt { 1 + 2. 9 8 2 [ B ( T ) ] ^ { 2 } } } \, \text{nm}. \quad \text{(10)} \\ \text{For the case of a vanishing magnetic field we assign to it} \quad \text{CW} \\ \text{a finite very small value in order to avoid numerical prob}$$
For the case of a vanishing magnetic field we assign to it a finite very small value in order to avoid numerical problems caused by exact degeneration of states with opposite spin.
We assume the central system is connected to the external leads acting as electron reservoirs with chemical potential in the left (right) lead µ L = 4 0 . meV ( µ R = 3 0 meV), implying an applied potential difference . or bias window ∆ µ = eV bias = 1 0 meV. Furthermore, the . temperature of the leads is fixed at T = 0 5 K. .
The potential of the double quantum waveguides is shown in Fig. 2 with the side QDs located near the CW in order to influence the inter-waveguide transport. In
FIG. 2. (Color online) The potential (in meV) defining the DQW central system with the CW and the side QDs. The parameters are /planckover2pi1 Ω 0 = 1 0 meV, . B = 0 0 T, . a w = 33 72 nm, . L CW = 100 nm, and R Dot = 25 nm.
the following, we shall explore the influence of the CW and the side QDs on the charging of and the transport through the system.
## A. DQW without side QDs
We start by considering an ideal window-coupled DQW without side QDs. The injected electrons are assumed to come from the left lead only into the controlwaveguide. This asymmetry in the injection is accomplished with the choice of the skewing parameter α =
4 0 . a w defined in Eq. (4) and only used for the left lead. In addition, we assume the two quantum waveguides to be of the same width. They are separated by a Gaussian potential barrier, in which the control- and targetwaveguide interact with each other via a CW. The DQW system is described by
$$V _ { D Q W } ( { \mathbf r } ) = V _ { \mathrm B } ( y ) + V _ { \mathrm C W } ( x, y ), \quad \ \ ( 1 1 )$$
where V B ( y ) is the barrier potential between the quantum waveguides defined as
$$V _ { B } ( y ) = V _ { 0 } \exp ( - \beta _ { 0 } ^ { 2 } y ^ { 2 } ), \text{ \quad \ \ } ( 1 2 )$$
with V 0 = 18 0 meV and . β 0 = 0 03 nm . -1 , as well as a CW potential
$$V _ { \text{CW} } ( x, y ) = - V _ { 0 } \exp ( - \beta _ { x } ^ { 2 } x ^ { 2 } - \beta _ { y } ^ { 2 } y ^ { 2 } ), \quad ( 1 3 )$$
with V 0 = 18 0 meV, and . β y = 0 03 nm . -1 implying a barrier width W B /similarequal 66 5 nm for the first subband which . can prevent electron tunneling between the waveguides through the barrier. β x is a parameter that determines the CW length. Thus, the length of the CW can be defined as L CW = 2 /β x .
We begin to analyze the results by presenting Fig. 3 which shows the net charge current I Q as a function of the CW length L CW at t = 200 ps and B = 0 0 T for non-. interacting electrons (solid blue) and for Coulomb interacting electrons (dashed red). The Coulomb interaction reduces the overall current slightly.
The oscillations in the net charge current are indicative of a possible charge transport between the controlwaveguide and the target-waveguide. The oscillations give rise to a peak and a dip in the net charge current at L CW /similarequal 40 and 110 nm, respectively. Similar oscillation features have been found by Zibold et al . and Gong et al . in the absence of the Coulomb interaction, and put in relation to energy-dependent inter-waveguide transmission in the quantum regime. They pointed out that the inter-waveguide transmission can be enhanced when the energy levels of the coupled-waveguide system achieve a resonance conditions for a specific CW length [22, 23].
To demonstrate the resonant energy levels in our system, we present Fig. 4 which shows the many-electron (ME) energy spectra for the DQW system as a function of the CW length at B = 0 0 T including one-electron . states (1ES, red dots) and two-electron states (2ES, blue dots). The ME spectrum without the mutual Coulomb interaction is displayed in Fig. 4(a), but the interacting spectrum in Fig. 4(b). The only difference between the energy spectra is that the Coulomb interaction has raised most of the 2ES well above the bias window in Fig. 4(b). The 2ES do not contribute to the electron transport in the presence of the Coulomb interaction because they are far from the chosen bias window, leading to a decrease in the net charge current.
In the case of no coupling window the two waveguides are only coupled through tunneling and the Coulomb interaction [16] leading to nearly degenerate states. When
FIG. 3. (Color online) The net charge current I Q versus CW length L CW without (w/o) (blue solid) and with (w) (red dashed) electron-electron (e-e) interaction for magnetic field B = 0 0 T The chemical potentials are . µ L = 4 0 meV and . µ R = 3 0 meV implying ∆ . µ = 1 0 meV. .
the length of the CW L CW is increased we thus find: First, the energy spectrum decreases monotonically [22]. Second, the near degeneration of the energy-levels is lifted. As a result, this leads to a crossover of energy levels (resonance) between the control- and targetwaveguide at L CW /similarequal 40 nm (left green rectangular) which corresponds to the current-peak in Fig. 3. The resonance between the waveguides enhances the interwaveguide electron transport and increases the net charge current. The splitting of the energy levels increases at a higher window-coupling length such as L CW /similarequal 110 nm (right green rectangular) indicating weaker resonances and more back-scattering inter-waveguide transport.
To further explain the peak and the dip in the net charge current, we present the charge current density at B = 0 0 T and . t = 200 ps in Fig. 5 for the noninteracting system (left panels) and the interacting one (right panels) for the peak and the dip net charge current at L CW /similarequal 40 nm (a-b) and 110 nm (c-d), respectively, shown in Fig. 3.
The dynamical evolution occurring in the CW implements different types of quantum logic gates in our system. In Fig. 5(a) the charge current density in the absence of the Coulomb interaction reveals the following electron motion in the waveguides at L CW /similarequal 40 nm: The charge is injected into the control input, then it exits equally from both control and target output. In the DQWthe charge exhibits partial inter-waveguide forward scattering which is similar to the condition of having equal electron transmission ratio in the double waveguides at a resonance [22]. The DQW here can be defined as a beam splitter or a √ NOT-operation quantum logic gate ( √ NOT-operation qubit) [8].
Figure 5(c) shows the charge current density in the current-dip at L CW /similarequal 110 nm. The charge current den-
FIG. 4. (Color online) ME energy spectrum in the central system as a function of CW length L CW without (a), and with (b) the electron-electron Coulomb interaction including one-electron states (1ES, red dots) and two-electron states (2ES, blue dots) at B = 0 0 T. The one-electron states . in the left green rectangle are close to be in resonance, but the one-electron states in the right green rectangle are not. The chemical potentials are µ L = 4 0 meV and . µ R = 3 0 meV . (black) implying ∆ µ = 1 0 meV. .
sity switches from the control- to the target-waveguide with a small ratio of inter-waveguide backward scattering. The charge switching property of the system here can be introduced as a Not-operation quantum logic gate [11]. The net charge current is suppressed here due to the presence of an inter-waveguide backward scattering that forms the current-dip. The cause of the interwaveguide backward scattering is the geometry of the system, the symmetric waveguides, while in the asymmetric system the back scattering can be avoided [6].
With the Coulomb interaction the charge current density is very similar for the dip near 110 nm (comparing Fig's 5(c) and (d)), but different for the peak at 40 nm, where very little scattering into the target waveguide is seen (comparing Fig. 5(a) and (b)). Without the Coulomb interaction 2ES contribute up to 1/3 of the charge current, but the interaction reduces this with Coulomb blocking to a negligible quantity. The forward inter-waveguide scattering in Fig. 5(a) is facilitated by two-electron processes and states. In the present system the Coulomb interaction blocks these processes to a large extent.
The effects of a static external magnetic field are presented in Fig. 6(a) which shows the net charge current I Q as a function of the CW length L CW at t = 200 ps in the presence of the Coulomb interaction for different values of the magnetic field B = 0 0 T (blue solid), 0 1 T . . (green dashed), and 0 2 T (red dashed). For comparison . we display in Fig. 6(b) the left and the right currents I L and I R at B = 0 0 T and the same point in time. . The comparison shows that at t = 200 ps the system is still in
FIG. 5. (Color online) Charge current density in the DQW at t = 200 ps without (left panels) with (right panels) electronelectron Coulomb interaction in the peak and the dip net charge current at L CW /similarequal 40 nm (a)-(b) and 110 nm (c)-(d), respectively, shown in Fig. 3 in the case of B = 0 0 T (blue solid). . The effective magnetic length is a w = 33 72 nm .
the charging phase, but the overall shape of the currents is similar.
The role of the external magnetic field in our system is different from the role it plays in many models where researchers have applied a variable magnetic field to switch between different processes in the waveguides. The external magnetic field has been used to switch an electron current from one quantum-waveguide to the other one in a window coupled double waveguide [6]. In our model, we observe switching characteristic in absence of a magnetic field B = 0 0 T. In the presence of stronger mag-. netic field, the electron motion is affected by the Lorentz force in which the electrons tend to a circular motion in both controland target-waveguides increasing the electron dwell-time in the system. The inter-waveguide backward scattering is enhanced while a suppression in the inter-waveguide forward scattering is observed in the presence of strong magnetic field. Therefore, the net charge current decreases as shown in Fig. 6(a) in the case of B = 0 1 T (green dashed) and 0 2 T (red dotted). . .
To understand the reasons for the suppression of the net charge current in the presence of a higher magnetic field, we show in Fig. 7 the charge current density at magnetic field B = 0 2 T in the peak at . L CW /similarequal 40 nm (a) and the dip at L CW /similarequal 110 nm (b) of the net charge current shown in Fig. 6(a) (red dotted). An obvious explanation is: The perpendicular magnetic field reduces
FIG. 6. (Color online) (a) The net charge current I Q versus CW length L CW in the presence of the Coulomb interaction for different values of the magnetic field B = 0 0 T (blue . solid), 0 1 T (green dashed), and 0 2 T (red dotted) at time . . t = 200 ps. (b) The current from the left lead I L (pink dotted), the current into the right lead I R (blue dashed), and the net charge current I Q (blue solid) for B = 0 0 T versus the . CW length. The chemical potentials are µ L = 4 0 meV and . µ R = 3 0 meV implying ∆ . µ = 1 0 meV. .
FIG. 7. (Color online) Charge current density in the DQW at t = 200 ps in the presence of the Coulomb interaction in the peak at L CW /similarequal 40 nm (a), and the dip at L CW /similarequal 110 nm (b) net charge current of Fig. 6(a) in the case of B = 0 2 T (red . dotted). The effective magnetic length is a w = 32 76 nm. .
the effective barrier height and increases the splitting of the energy levels at the crossover energy or resonance energy (not shown). Therefore the charge current density does not equally split between the waveguides or localize in only control-waveguide, but rather an inter-waveguide backward scattering is observed due to the Lorentz force, consequently the net charge current is suppressed and we observe edge states forming in each of the waveguides. In this case the double-waveguide system does neither play the role of a √ NOT-operation nor a Not-operation qubit in the presence of a higher magnetic field.
In the following, we shall study the influence of the side QDs in the DQW for different values of the magnetic field.
## B. DQW with side QDs
In this section, we consider a side dot embedded in each waveguide to enhance the inter-waveguide transport as is shown in Fig. 2. The dots to the sides of the CW may be expected to increase the dwell time in the coupling region and change the resonance condition between the two waveguides. The window-coupling DQW potential with side QDs is described by
$$V _ { D Q W } ( { \mathbf r } ) = V _ { \mathrm B } ( y ) + V _ { C W } ( x, y ) + V _ { \text{Dot} } ( x, y ), \quad ( 1 4 )$$
where V Dot ( x, y ) is the side dot potential that is defined as
$$V _ { \text{Dot} } ( x, y ) = - V _ { \text{D} } \exp ( - \gamma _ { x } ^ { 2 } x ^ { 2 } - \gamma _ { y } ^ { 2 } ( y - y _ { 0 } ) ^ { 2 } ), \quad ( 1 5 )$$
with V D = 8 0 meV, . y 0 = 5 6a . w , and γ x = γ y = 0 04 . nm -1 implying the radius of each side-dot R Dot ≈ 25 nm.
Figure 8(a) shows the net charge current as a function of the CW length at t = 200 ps in the presence of the Coulomb interaction for different values of the magnetic field B = 0 0 T (blue solid), 0 1 T (green dashed), . . and 0 2 T (red dashed). An oscillation in the net charge . current is again established in a vanishing magnetic field 0 0 T, while at higher magnetic field values . 0 1 and 0 2 T . . some extra fluctuations in the current oscillation characteristic are observed. The existence of fluctuations correlates with dynamic motion of the charge in the system in which electrons participate in an inter-waveguide backward or forward scattering at different CW length. If the magnetic field is increased to B = 0 9 T, such that . the effective magnetic length is close to the dot radius a w = R Dot the net charge current is suppressed to vanishing values (not shown). In that case, varying of the CW length does not affect the electron transport in the system. There are several reasons that lead to the almost vanishing net charge current at high magnetic field such as edge effects, inter-waveguide backward scattering and localized electrons in the dots as the effective magnetic length is close to the dot radius.
We present Fig. 8(b) to compare the left partial current (pink dotted), the right partial current (blue dashed) to the net charge current (blue solid) at B = 0 0 T. It can be . seen that the partial currents have the same oscillation features and develop a peak and a current dip at the same CW lengths.
In order to understand the characteristics of net charge current, we refer to the energy spectrum and the charge current density. Figure 9 shows the ME-energy versus the window-coupling length at B = 0 0 T including one-. electron states (1ES, red dots) and two electron states (2ES, blue dots). It should be noted that more energy levels get into resonance at L CW /similarequal 40 nm (left green rectangular) in the presence of the side-gate dots which indicates a stronger coupling between the controland the target-waveguide. At larger CW length such as L CW /similarequal 110 nm (right green rectangular), the energy splitting increases resulting in different electron motion in the system.
The electron motion in the waveguides is completely changed in the presence of the side QDs. Figure 10 displays the charge current density at B = 0 0 T and . L CW /similarequal 40 nm (a) and current-dip at L CW /similarequal 110 nm (b) shown in Fig. 8(a) (blue solid). In Fig. 10(a) the charge current density is displayed for L CW /similarequal 40 nm. The
FIG. 8. (Color online) The net charge current I Q versus the window-coupling length L CW in the presence of the Coulomb interaction for different magnetic field B = 0 0 T (blue solid), . 0 1 T (green dashed), 0 2 T (red dotted) ,0 9 T (pink dotted) . . . at time t = 200 ps. (b) The current from the left lead I L (pink dotted), the current into the right lead I R (blue dashed), and the net charge current I Q (blue solid) for B = 0 0 T versus . the CW length. The chemical potentials are µ L = 4 0 meV . and µ R = 3 0 meV implying ∆ . µ = 1 0 meV. .
FIG. 9. (Color online) The interacting ME energy spectrum of the central system with side QDs as a function of the CW length L CW including one-electron states (1ES, red dots) and two electron states (2ES, blue dots) at B = 0 0 T. The one-. electron states in the left green rectangular are close to be in resonance, but the one-electron states in the right green rectangular are not. The chemical potentials are µ L = 4 0 meV . and µ R = 3 0 meV (black) implying ∆ . µ = 1 0 meV. .
charge in the control-waveguide is switched to the targetwaveguide, thus a Not-operation is realized. Comparing to the system without side-gate dots in Fig. 5(b), the inter-waveguide transport is enhanced due to a stronger coupling between the waveguides at the resonant-energy levels at L CW /similarequal 40 nm. Therefore, the side QDs play an essential role to convert the qubit operation in the system. In addition, the inter-waveguide backward scattering here reduces the efficiency of the qubit. As we mention previously, the existence of inter-waveguide backward scattering might be due to symmetric waveguides while the backward-scattering in an asymmetric waveguides can be avoided or reduced [24]. Figure 10(b) shows the charge current density at the current-dip L CW /similarequal 110 nm. The incoming charge into the controlwaveguide encounters the control-QD leading to electron back-scattering in the control-waveguide, thus the net charge current is suppressed forming a current-dip.
We can say that the side-dots play a scattering role in the waveguides in such way that the electrons partially get forwardand backward-scattered in and into both waveguides.
FIG. 10. (Color online) Charge current density in the DQW at t = 200 ps in the presence of the Coulomb interaction in the peak L CW /similarequal 40 nm (a), and the dip L CW /similarequal 110 nm (b) of the net charge current of Fig. 8(a) in the case of B = 0 0 T. . Other parameters are a w = 33 72 nm, and . R Dot ≈ 25 nm.
By tuning the external magnetic field to a higher value,
B = 0 2 T, the electrons manifest different . dynamical motion in the waveguides. In Fig. 11 the charge current density is shown in this stronger magnetic field B = 0 2 T . at the CW length L CW /similarequal 40 (a) and 110 nm (b). The graphs show circular motion, edge states, in each waveguide.
In Fig. 11(a) is shown the current charge density at L CW /similarequal 40 nm, the electrons in the control-waveguide do not totally switch to the target-waveguide, because of increased energy level splitting at the resonance-energy in the presence of higher magnetic field. In addition, charge is partially localized in the control-QD as the effective magnetic length approaches the radius of the quantum dot a w /similarequal R Dot .
FIG. 11. (Color online) Charge current density in the DQW at t = 200 ps in the presence of the Coulomb interaction in the peak L CW /similarequal 40 nm (a), and the dip L CW /similarequal 110 nm (b) net charge current of Fig. 8(a) in the case of B = 0 2 T. Other . parameters are a w = 32 78 nm, and . R Dot ≈ 25 nm.
In Fig. 11 (b), the electrons are mostly localized in the control-QD corresponding to the current-dip at L CW /similarequal 110 nm. This electron localization is expected because at a larger CW length the splitting of energy levels increases and leads to suppression the inter-waveguide scattering. Therefore, the electrons move along the controlwaveguide and get localized in the control-QD at this higher magnetic field.
## IV. CONCLUDING REMARKS
We have presented a model for a window coupled double quantum waveguide in an external perpendicular magnetic field. The DQW is weakly connected to two leads with different chemical potentials in which a non-Markovian master equation is utilized to investigate coherent switching in electron transport between the waveguides for the implementation of quantum logic gates.
A comparison between the charge current and the left and right currents in Figures 6 and 8 at the time point in time t = 200 ps shows that the system has not completely reached a steady state yet, but the sought after function of the DQW is already present in the late transition regime. This is important. The exact steady state takes a long time to reach due to a low rate of charging for 2 and 3 electron states. Valuable time can be gained by operation of a device in the transition regime.
By tuning the length of the CW, we have demonstrated two important physical characteristics of the waveguide system: resonant energy-levels and oscillations in the net charge current. The resonant energy levels indicate a strong coupling between the DQW and the oscillations denote that the resonance conditions are governed by the length of the CW.
In the case of an ideal non-Coulomb-interacting window-coupled DQW system, the charge current density splits equally between the waveguides at resonant energies, therefore, the net charge current reaches it's maximum value and the waveguide system works as a √ NOT-operation qubit. In the presence of strong magnetic field, the inter-waveguide backward scattering is enhanced, and the efficiency of the logic gate decreases.
This inter-waveguide forward scattering at resonant levels without a magnetic field is not seen without including two-electrons states in the model. It is strongly reduced by the Coulomb interaction that lifts the 2ES out of the group of active states in the transport. This aspect can though be controlled by the exact system size and design as the Coulomb blocking is reduced in a larger system, or by using a higher bias window or different material. Our models only includes sequential tunneling of electrons from the leads to the system, but the addition of higher order tunneling would only increase the importance of the many-electron structure and the Coulomb interaction on the operation of the system.
Two parallel waveguides or dots can be coupled in many distinct ways. Moldoveanu[16] et al. attain it capacitively, while Zibold[22] et al. fine tune the system by tunneling coupling. Both groups stress the strong dependence of the device on its exact geometry and the structure and correlations of many-electron states that are active in the transport. Our investigation with a different coupling scheme for the waveguides confirms this sensitive dependence.
In the presence of side QDs, more energy-levels of the two waveguides are brought into resonance which can
lead to stronger coupling between them. Effectively, the dots can increase the density of states around the resonant transport states. The electrons from the controlwaveguide switch to the target-waveguide indicating a Not-operation qubit. But on the other hand, the sidedots can also increase the inter-waveguide backward scattering and reduce the qubit efficiency. In a stronger external magnetic field, the electrons get localized in the dots as the effective magnetic length is comparable to the radius of the side dot. In this case, the geometry of the CW does not affect the electron transport anymore.
The external magnetic field here is considered static, but one more variable to influence the transport is the shape of the side QDs in the waveguides. We show that
- [1] P. Zhang, Z.-L. Xiang, and F. Nori, Phys. Rev. B 89 , 115417 (2014).
- [2] R. Ionicioiu, G. Amaratunga, and F. Udrea, Int. J. of Mod. Phys. B 15 , 125 (2001).
- [3] J. M. Martinis, S. Nam, J. Aumentado, and C. Urbina, Phys. Rev. Lett. 89 , 117901 (2002).
- [4] P. Bordone, A. Bertoni, M. Rosini, and C. Jacoboni, Semicond.Sci.Technol. 19 , 412 (2004).
- [5] L. E. Reichl and M. G. Snyder, Phys. Rev. A 72 , 032330 (2005).
- [6] M. J. Gilbert, R. Akis, and D. K. Ferry, Appl. Phys. Lett. 81 , 4284 (2002).
- [7] J. Harris, R. Akis, and D. K. Ferry, Appl. Phys. Lett. 79 , 2214 (2001).
- [8] M. G. Snyder and L. E. Reichl, Phys. Rev. A 70 , 052330 (2004).
- [9] A. Marchi, A. Bertoni, S. Reggiani, and M. Rudan, IEEE Transaction on nanotechnology 3 , 129 (2003).
- [10] A. Bertoni, P. Bordone, R. Brunetti, C. Jacoboni, and S. Reggiani, Phys. Rev. Lett. 84 , 5912 (2000).
- [11] A. Ramamoorthy, R. Akis, and J. P. Bird, IEEE Transaction on nanotechnology 5 , 712 (2006).
- [12] V. Moldoveanu, A. Manolescu, and V. Gudmundsson, New Journal of Physics 11 , 073019 (2009).
- [13] V. Gudmundsson, C. Gainar, C.-S. Tang, V. Moldoveanu, and A. Manolecu, New J. Phys.
the varying of the CW length between the waveguides can specify the type of the qubit logic gate to be implemented by the system.
## ACKNOWLEDGMENTS
This work was financially supported by the Icelandic Research and Instruments Funds, the Research Fund of the University of Iceland, the Nordic High Performance Computing facility in Iceland, and the National Science Council in Taiwan through Contracts No. NSC1002112-M-239-001-MY3 and No. MOST103-2112-M-239001-MY3.
- 11 , 113007 (2009).
- [14] V. Moldoveanu, A. Manolescu, C.-S. Tang, and V. Gudmundsson, Phys. Rev. B 81 , 155442 (2010).
- [15] N. R. Abdullah, C.-S. Tang, and V. Gudmundsson, Phys. Rev. B 82 , 195325 (2010).
- [16] V. Moldoveanu, A. Manolescu, and V. Gudmundsson, Phys. Rev. B 82 , 085311 (2010).
- [17] H.-P. Breuer and F. Petruccione, The Theory of Open Quantum Systems (Oxford University Press, Oxford, 2002).
- [18] F. Haake, Phys. Rev. A 3 , 1723 (1971).
- [19] F. Haake, Quantum Statistics in Optics and Solid-state Physics , edited by G. Hohler and E.A. Niekisch, Springer Tracts in Modern Physics Vol. 66 (Springer, Berlin, Heidelberg, New York, 1973, p. 98.).
- [20] N. R. Abdullah, C. S. Tang, A. Manolescu, and V. Gudmundsson, Journal of Physics:Condensed Matter 25 , 465302 (2013).
- [21] S. Datta, Electronic Transport in Mesoscopic system (Cambridge University Press, Cambridge, 1995).
- [22] T. Zibold, P. Vogl, and A. Bertoni, Phys. Rev. B 76 , 195301 (2007).
- [23] J. Gong, F.-H. Yang, and S.-L. Feng, Chin. Phys. Lett. 24 , 2383 (2007).
- [24] G. B. Akguc, L. E. Reichl, A. Shaji, and M. G. Snyder, Phys. Rev. A 69 , 042303 (2004).
|
10.1088/0953-8984/27/1/015301
|
[
"Nzar Rauf Abdullah",
"Chi-Shung Tang",
"Andrei Manolescu",
"Vidar Gudmundsson"
] |
2014-08-05T15:45:21+00:00
|
2014-08-05T15:45:21+00:00
|
[
"cond-mat.mes-hall"
] |
Coherent transient transport of interacting electrons through a quantum waveguide switch
|
We investigate coherent electron-switching transport in a double quantum
waveguide system in a perpendicular static or vanishing magnetic field. The
finite symmetric double waveguide is connected to two semi-infinite leads from
both ends. The double waveguide can be defined as two parallel finite quantum
wires or waveguides coupled via a window to facilitate coherent electron
inter-wire transport. By tuning the length of the coupling window, we observe
oscillations in the net charge current and a maximum electron conductance for
the energy levels of the two waveguides in resonance. The importance of the
mutual Coulomb interaction between the electrons and the influence of
two-electron states is clarified by comparing results with and without the
interaction. Even though the Coulomb interaction can lift two-electron states
out of the group of active transport states the length of the coupling window
can be tuned to locate two very distinct transport modes in the system in the
late transient regime before the onset of a steady state. A static external
magnetic field and quantum-dots formed by side gates (side quantum dots) can be
used to enhance the inter-waveguide transport which can serve to implement a
quantum logic device. The fact that the device can be operated in the transient
regime can be used to enhance its speed.
|
1408.1008v1
|
## Entanglement dynamics in a quantum-classical hybrid of two q-bits and one oscillator
## L. Fratino , A. Lampo a b and H.-T. Elze c, 1
- a SEPnet & Hubbard Theory Consortium, Department of Physics, Royal Holloway, University of London, Egham, Surrey TW20 0EX, UK
- b ICFO, Mediterranean Technology Park, Av. C.F. Gauss 3, 08860 Castelldefels (Barcelona), Spain
- c Dipartimento di Fisica 'Enrico Fermi', Largo Pontecorvo 3, 56127 Pisa, Italia
E-mail: [email protected], [email protected], [email protected]
Abstract. We investigate new features, especially of entanglement dynamics, which arise in a quantum-classical hybrid. As a model, we study the coupling between two quantum mechanical two-level systems, i.e. two q-bits, and a classical harmonic oscillator. Their interaction is described by a hybrid coupling, in accordance with a recently developed quantum-classical hybrid theory. We discuss various situations in which entanglement of the q-bits does (not) evolve. Furthermore, we point out an experimental application in a hybrid cooling scheme and indicate topics for future study.
Keywords: quantum-classical hybrid theory; optomechanical experiments; q-bits; quantum control; quantum decoherence
## 1. Introduction
Microscopic and macroscopic objects are commonly described by means of two different kinds of theories: the former through quantum mechanics and the latter by classical mechanics. Despite its successes, this 'brutal' division confines these two different kinds of objects in two noncommunicating 'worlds', the boundaries of which are still not clearly delineated. In fact, it has been unknown since the inception of quantum theory whether a borderline between the classical and quantum worlds exists or not. In order to unify quantum and classical descriptions, we make use of a recent generalization of these theories called quantum-classical hybrid theory [1, 2, 3, 4]. This approach builds a bridge between classical and quantum mechanics, proposing a consistent mathematical framework for the description of a quantum-classical mixed reality and a tool for future applications, e.g., in measurement or quantum information protocols that incorporate classical control of or read-out from a quantum object [5]. Hybrid theory and its applications will be relevant here, since laboratory quantum devices, at some stage, incorporate a quantum-classical interface.
Presently, we study one of the most peculiar properties of multipartite quantum states, their entanglement , in situations involving quantum-classical hybrid systems.
1 Corresponding author.
A state is entangled , if and only if it is not separable into states that describe independent subsystems of a composite. A composite system is disentangled or separable, in turn, if and only if its state can be factorized. The main question that motivated this work is: What kind of new features arise in this context in the presence of a quantum-classical hybrid coupling? In order to answer this, we consider a model that describes the coupling between two quantum-mechanical two-level (spin) systems , i.e. two q-bits, and a classical harmonic oscillator .
We base our analysis on the quantum-classical hybrid approach developed in Refs. [1, 6, 7], which satisfies a large list of consistency requirements discussed in the recent literature, as summarized in Ref. [8] .
The model of a harmonic oscillator coupled to two q-bits, when treated fully quantum mechanically, exhibits a great richness of entanglement dynamics, see, for example, Refs. [9, 10, 11, 12, 13]. Presently, we consider the case that the oscillator part of this composite system is strictly classical and, therefore, its coupling to the q-bits must be of hybrid nature.
The Hamiltonian of this system consists of a classical, a quantum mechanical, and a hybrid part, respectively (choosing units such that ¯ = 1): h
$$H _ { c l } \ \coloneqq \ \frac { p ^ { 2 } } { 2 m } + \frac { m \omega ^ { 2 } x ^ { 2 } } { 2 }$$
$$\hat { H } _ { q m } \ \coloneqq \ \omega _ { 0 } \hat { S } _ { z }$$
$$\hat { \mathcal { I } } _ { h y b } \ \coloneqq \ \beta \omega \sqrt { m \omega } \, x \hat { S } _ { x } = \beta \Omega \, x \hat { S } _ { x }$$
where the two-q-bit (' A ' and ' B ') operators are ˆ S x ≡ (ˆ σ A x + ˆ σ B x ) / 2 and ˆ S z ≡ (ˆ σ A z + ˆ σ B z ) / 2, and where we introduced Ω ≡ ω √ mω β ; is a dimensionless constant determining the coupling strength, which can be expressed in terms of microscopic physical parameters depending on the physical realization of the model (cf. Section 2). We recall that, with appropriate substitutions, the contribution H cl , i.e. the classical Hamiltonian, can describe a single mode of a classical electric field.
Following hybrid theory [1], we proceed to construct the full hybrid Hamiltonian function from the above contributions. In particular, introducing the oscillator expansion of a generic state | ψ 〉 in the eigenbasis of ˆ S x [6, 14]:
$$| \psi \rangle = \sum _ { \alpha = 1 } ^ { 4 } \frac { X _ { \alpha } + \imath P _ { \alpha } } { \sqrt { 2 } } | \phi _ { \alpha } \rangle$$
where the | φ α 〉 denote, respectively, the states | + + , 〉 | - -〉 , ( | + -〉 + | -+ ) 〉 / √ 2, ( | + -〉 - | -+ ) 〉 / √ 2, with ˆ σ x |±〉 = ±|±〉 . The 'coordinates' and 'momenta' introduced here have to satisfy a constraint, which stems from the wave function normalization, 〈 ψ t ( ) | ψ t ( ) 〉 = 1, i.e.:
$$\mathcal { C } ( X _ { \alpha }, P _ { \alpha } ) \coloneqq \sum _ { \alpha = 1 } ^ { 4 } ( X _ { \alpha } ^ { 2 } + P _ { \alpha } ^ { 2 } ) = 2 \.$$
Based on this expansion, contributions to the full Hamiltonian function from eqs. (2)-(3) become:
$$H _ { q m } \ \coloneqq \ \langle \psi | \hat { H } _ { q m } | \psi \rangle = \frac { \omega _ { 0 } } { 2 } \sum _ { \alpha, \beta = 1 } ^ { 4 } ( X _ { \alpha } + \imath P _ { \alpha } ) S _ { \alpha \beta } ( X _ { \beta } - \imath P _ { \beta } )$$
$$\mathcal { I } _ { h y b } \ \coloneqq \ \langle \psi | \hat { \mathcal { I } } _ { h y b } | \psi \rangle = \frac { \beta \Omega } { 2 } x \sum _ { \alpha = 1 } ^ { 4 } E _ { \alpha } ( X _ { \alpha } ^ { 2 } + P _ { \alpha } ^ { 2 } )$$
introducing matrix notation through ( α, β = 1 , . . . , 4):
$$\hat { S } \equiv ( S _ { \alpha \beta } ) \coloneqq \left ( \begin{array} { c c c c } 0 & 0 & \sqrt { 2 } & 0 \\ 0 & 0 & \sqrt { 2 } & 0 \\ \sqrt { 2 } & \sqrt { 2 } & 0 & 0 \\ 0 & 0 & 0 & 0 \end{array} \right ) \, \ \ ( E _ { \alpha } ) \coloneqq \left ( \begin{array} { c c c c } 1 \\ - 1 \\ 0 \\ 0 \end{array} \right ) \.$$
Having a Hamiltonian function in hand, given by H Σ = H cl + H qm + I hyb , the equations of motion are obtained through the Poisson brackets with respect to all canonical coordinates and momenta ( X ,P ,x,p α α ), as usual [1, 14]:
$$\dot { x } = \{ x, H _ { \Sigma } \} \ = \ \frac { p } { m }$$
$$\dot { p } = \{ p, H _ { \Sigma } \} \ = \ - m \omega ^ { 2 } x - \frac { \beta \Omega } { 2 } \sum _ { \alpha = 1 } ^ { 4 } E _ { \alpha } ( X _ { \alpha } ^ { 2 } + P _ { \alpha } ^ { 2 } )$$
$$\dot { X _ { \alpha } } = \{ X _ { \alpha }, H _ { \Sigma } \} \ = \ \frac { \omega _ { 0 } } { \sqrt { 2 } } \sum _ { \beta = 1 } ^ { 4 } S _ { \alpha \beta } P _ { \beta } + \frac { \beta \Omega } { \sqrt { 2 } } x P _ { \alpha }$$
$$\dot { P _ { \alpha } } = \{ P _ { \alpha }, H _ { \Sigma } \} \ = \ - \frac { \omega _ { 0 } } { \sqrt { 2 } } \sum _ { \beta = 1 } ^ { 4 } S _ { \alpha \beta } X _ { \beta } - \frac { \beta \Omega } { \sqrt { 2 } } x X _ { \alpha }$$
see also Section 3 for explicit applications of the Poisson brackets.
In the following, we will employ the convenient substitution z α := X α + ıP α , α = 1 , . . . , 4 , which replaces eq. (10) by:
$$\dot { p } = - m \omega ^ { 2 } x - \frac { \beta \Omega } { 2 } ( z _ { 1 } ^ { * } z _ { 1 } - z _ { 2 } ^ { * } z _ { 2 } )$$
for example, while eq. (9) is unaffected.
Furthermore, by suitable linear combinations of the equations for the quantum sector (variables X ,P α α ), eqs. (11)-(12), the above substitution yields the coupled evolution equations:
$$\left ( \begin{array} { c } \dot { z } _ { 1 } \\ \dot { z } _ { 2 } \\ \dot { z } _ { 3 } \\ \dot { z } _ { 4 } \end{array} \right ) = \imath \frac { \omega _ { 0 } } { \sqrt { 2 } } \left ( \begin{array} { c c c c } 0 & 0 & 1 & 0 \\ 0 & 0 & 1 & 0 \\ 1 & 1 & 0 & 0 \\ 0 & 0 & 0 & 0 \end{array} \right ) \left ( \begin{array} { c } z _ { 1 } \\ z _ { 2 } \\ z _ { 3 } \\ z _ { 4 } \end{array} \right ) + \imath \frac { \beta \Omega } { \sqrt { 2 } } x \left ( \begin{array} { c c c c } 1 & 0 & 0 & 0 \\ 0 & - 1 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \end{array} \right ) \left ( \begin{array} { c } z _ { 1 } \\ z _ { 2 } \\ z _ { 3 } \\ z _ { 4 } \end{array} \right ) \.$$
Thus, we find that z 4 is presents a constant of motion.
We remark that the formal solution for the classical sector (variables x, p ) can be obtained by the method of Green's function. However, since the full dynamics of our model is not amenable to straightforward analytical techniques, we resort to numerical solutions of the coupled equations of motion (9)-(12).
## 2. Dynamics of the hybrid model
We have solved the equations of motion numerically with the help of programs provided in Wolfram Mathematica 8.0 . In the following subsection, we present solutions of the full dynamics of the system, which is the quantum-classical hybrid analogue of a generalization of the Rabi model called the Tavis-Cummings model [15, 16]; cf. Refs. [11, 12, 13] for the all-quantum results.
We have chosen a parameter regime in which the oscillator frequency ω is much larger than the characteristic frequency of the q-bit system ω 0 which, in turn, is of the same order of magnitude as the effective coupling λ = βω . The reason for this choice is that recent experiments have shown clear spectroscopic evidence that a Cooper-pair box, or Josephson charge q-bit, coupled to a superconducting transmission line behaves analogously to an atom in a cavity (as in the Tavis-Cummings model). In these cases, the dipole coupling between the two-level systems is β ≈ 10 -3 [17].
Capacitive or inductive couplings offer the possibility of still larger coupling strengths than those possible with dipole coupling, even at large detuning between the fundamental frequencies of the oscillator and q-bit system. Some results from a flux-based, inductively coupled system give preliminary evidence for coupled quantum behaviour and entanglement between the twolevel system and the oscillator [18].
In Figures 1, 2, we present various examples of the numerical solutions obtained for the variables z α , α = 1 , . . . , 4 , which characterize the quantum sector of the hybrid model; we recall that z 4 remains constant in time and, therefore, is not represented here. Various combinations of initial q-bit states and model parameters are shown, with full (dashed) lines showing the real (imaginary) parts of the z α . The classical oscillator initial conditions always have been x 0 = 0 , p 0 = 1 ; since explicit behaviour of the oscillator is not relevant for the following, we skip it here. However, various aspects of the oscillator evolution will be discussed in Sections 4 and 5.
Also shown in these figures is the concurrence, which will be related to a measure of entanglement of the two q-bits next. - We considered a large set of initial conditions and always have found similarly complex behaviour of the numerical solutions, no matter whether the initial q-bit states are entangled or not.
## 2.1. Entanglement
For the following, we have constructed the density matrix of the pure state q-bit subsystem as ˆ ρ ≡ | ψ 〉〈 ψ | , with | ψ 〉 obtained by employing eq. (4) and X α = Re( z α ), P α = Im( z α ), based on the numerical solutions of the full set of equations of motion, for the z α and x, p , in particular.
With the two-q-bit density matrix as input, we calculate the concurrence as a measure of their entanglement [19, 20].
Let us recall the essential aspects of the notion of concurrence needed here. - Given the density matrix ˆ of ρ a pair of quantum systems A and B , we define the entanglement (of formation) for a pure state ψ of the bipartite system as the entropy E pure of either of the subsystems A,B [21]:
$$\mathcal { E } _ { p u r e } ( \psi ) \coloneqq - \text{tr} ( \hat { \rho } _ { A } \log _ { 2 } \hat { \rho } _ { A } ) = - \text{tr} ( \hat { \rho } _ { B } \log _ { 2 } \hat { \rho } _ { B } )$$
where we have introduced the reduced density matrices ˆ ρ A,B , defined by ˆ ρ A := tr B ˆ and ρ ˆ ρ B := tr A ˆ. ρ This definition can be extended for the case of mixed states. In this case, the entanglement of formation is the average entanglement of the pure states of a decomposition, minimized over all decompositions of ˆ: ρ
$$\mathcal { E } _ { m i x e d } ( \hat { \rho } ) \coloneqq \min \left [ \sum _ { i } p _ { i } \mathcal { E } _ { p u r e } ( \psi _ { i } ) \right ] \.$$
The goal now is to replace the right-hand side of this formal definition by an explicit function of ˆ that allows us to evaluate it. ρ
We begin with the 'spin-flipped' transform of a state that we denote by a tilde, | ˜ . ψ 〉 This transformation is the standard time reversal operation for a 1 2 -spin particle. For a pure state
| ψ 〉 of two q-bits, it is given by:
$$| \tilde { \psi } \rangle = ( \sigma _ { y } ^ { ( A ) } \otimes \sigma _ { y } ^ { ( B ) } ) | \psi ^ { * } \rangle$$
with | ψ ∗ 〉 the conjugate of | ψ 〉 . - For a generic state of two q-bits, the obvious generalization is:
$$\tilde { \rho } = ( \sigma _ { y } ^ { ( A ) } \otimes \sigma _ { y } ^ { ( B ) } ) \tilde { \rho } ^ { * } ( \sigma _ { y } ^ { ( A ) } \otimes \sigma _ { y } ^ { ( B ) } )$$
with ˆ ρ ∗ the conjugate of ˆ. ρ
Next, we define the concurrence C pure for a pure state | ψ 〉 by:
$$C _ { p u r e } ( \psi ) \coloneqq | \langle \psi | \tilde { \psi } \rangle | \.$$
The 'spin-flip' transformation of a pure product state takes the state of each q-bit to an orthogonal state, which implies C pure ( ψ ) = 0. Instead, a generic state, as for example an entangled Bell state, can be invariant except for a global phase factor, which implies C ψ ( ) = 1. One immediately sees that 0 ≤ C ≤ 1.
Most importantly, it can be shown that eq. (16), for a pure state, reduces to [19]:
$$\mathcal { E } _ { p u r e } ( \psi ) = h \left ( \frac { 1 + \sqrt { 1 - C ^ { 2 } } } { 2 } \right ) = \colon \mathcal { E } ( C ) \$$
with:
$$h ( x ) \coloneqq - x \log _ { 2 } x - ( 1 - x ) \log _ { 2 } ( 1 - x ).$$
Thus, the entanglement of formation reduces to an explicit function of the concurrence.
Furthermore, a generalization of the concurrence for mixed states has been given in Ref. [20]:
$$C _ { m i x e d } ( \rho ) \coloneqq \max [ 0, \sqrt { \lambda _ { 1 } } - \sqrt { \lambda _ { 2 } } - \sqrt { \lambda _ { 3 } } - \sqrt { \lambda _ { 4 } } ],$$
in terms of the eigenvalues λ , i i = 1 , . . . , 4 , with λ i ≥ λ i +1 , of the non-Hermitian matrix ˆ := ˆ˜, With this definition of the concurrence, R ρρ entanglement of formation of two q-bits can be expressed similarly as above:
$$\mathcal { E } _ { m i x e d } ( \rho ) = \mathcal { E } ( C )$$
namely, it becomes again a function of the concurrence. Thus, the concurrence C will be considered as an indicator for entanglement in the following; the entanglement of formation, E ( C ), could be further evaluated, as we have discussed here.
In Figures 1, 2, in the lower right quadrant, respectively, is shown the concurrence C for the evolving two-q-bit states under consideration. We find from the numerical solutions of the equations of motion that C is constant. In the following Section 3, we will demonstrate this result also analytically. The result of the calculations here, therefore presents a nontrivial check for the accuracy of our numerical method.
## 3. Analytical and semiclassical aspects of concurrence
Instead of the numerical results presented so far, we will analyze the dynamics of entanglement in the hybrid model analytically here in terms of the concurrence [19].
If the initial two-q-bit state is a generic pure state | ξ (0) , the oscillator expansion (4) yields: 〉
$$| \xi ( 0 ) \rangle = \sum _ { \beta = 1 } ^ { 4 } \frac { X _ { \beta } ( 0 ) + \imath P _ { \beta } ( 0 ) } { \sqrt { 2 } } | \phi _ { \beta } \rangle \.$$
Figure 1. Evolution of z 1 ( ) (upper left), t z 2 ( ) (upper right), t z 3 ( ) (lower left), and concurrence t C (lower right) for the initial state 1 √ 2 ( | ++ + 〉 | - -〉 ) and parameters ω = 0 03, . ω 0 = 0 15 . ω , β = 0 2; full (dashed) lines represent real (imaginary) parts. .
Figure 2. Evolution of z 1 ( ) (upper left), t z 2 ( ) (upper right), t z 3 ( ) (lower left), and concurrence t C (lower right) for the initial state 1 √ 2 ( | ++ + 〉 ı | - -〉 ) and parameters ω = 0 03, . ω 0 = 0 15 . ω , β = 0 16; full (dashed) lines represent real (imaginary) parts. .
The equations of motion, eqs. (11)-(12) in particular, determine the evolution of this state. Correspondingly, we obtain the time dependent density matrix:
$$\hat { \rho } ( t ) = \sum _ { \alpha, \beta = 1 } ^ { 4 } \frac { ( X _ { \alpha } ( t ) - \imath P _ { \alpha } ( t ) ) ( X _ { \beta } ( t ) + \imath P _ { \beta } ( t ) ) } { 2 } | \phi _ { \beta } \rangle \langle \phi _ { \alpha } |$$
the input for the evaluation of the concurrence, as described in Section 2.1 .
Next, we need the spin-flipped version of the density matrix, ˜( ) = ( ρ t σ y ⊗ σ y )ˆ( ) ρ t ∗ ( σ y ⊗ σ y ). In the basis presently chosen together with the oscillator expansion, we have:
$$\sigma _ { y } \otimes \sigma _ { y } = \left ( \begin{array} { c c c c } 0 & - 1 & 0 & 0 \\ - 1 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & - 1 \end{array} \right )$$
,
and, consequently:
$$\text{lently} \colon \quad & \tilde { \rho } ( t ) = \left ( \begin{array} { c c c c } | z _ { 2 } | ^ { 2 } & z _ { 1 } ^ { * } z _ { 2 } & - z _ { 2 } z _ { 3 } ^ { * } & z _ { 2 } z _ { 4 } ^ { * } \\ z _ { 1 } z _ { 2 } ^ { * } & | z _ { 1 } | ^ { 2 } & - z _ { 1 } z _ { 3 } ^ { * } & z _ { 1 } z _ { 4 } ^ { * } \\ - z _ { 2 } ^ { * } z _ { 3 } & - z _ { 1 } ^ { * } z _ { 3 } & | z _ { 3 } | ^ { 2 } & - z _ { 3 } z _ { 4 } ^ { * } \\ z _ { 2 } ^ { * } z _ { 4 } & z _ { 1 } ^ { * } z _ { 4 } & - z _ { 3 } ^ { * } z _ { 4 } & | z _ { 4 } | ^ { 2 } \end{array} \right ) \. \quad & \text{(26)} \\ \cdot \quad & \quad \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \, - \end{array}$$
Finally, we have to find the square roots of the eigenvalues of the matrix ˆ := ˆ˜, in order to R ρρ calculate the concurrence. - To this end, we note that the rank of the density matrix ˆ is equal ρ to one, because it is a projector for a pure state. Then, the rank of R is the same as that of ˆ ρ and the only non-trivial eigenvalue λ is given by the trace of ˆ . R Thus, we obtain from eq. (26) and recalling z α := X α + ıP α , α = 1 , . . . , 4 :
$$\lambda \ = \ [ 2 ( X _ { 1 } - \imath P _ { 1 } ) ( X _ { 2 } - \imath P _ { 2 } ) - ( X _ { 3 } - \imath P _ { 3 } ) ^ { 2 } + ( X _ { 4 } - \imath P _ { 4 } ) ^ { 2 } ]$$
$$\times [ 2 ( X _ { 1 } + \imath P _ { 1 } ) ( X _ { 2 } + \imath P _ { 2 } ) - ( X _ { 3 } + \imath P _ { 3 } ) ^ { 2 } + ( X _ { 4 } + \imath P _ { 4 } ) ^ { 2 } ]$$
which is real and non-negative, as it should be.
A straightforward, even if lengthy, calculation shows that indeed, in the present case, the concurrence given by C ρ ( ) = √ λ is constant in time. We obtain explicitly, using the equations of motion (11)-(12), that d C ρ / ( ) d t = d λ/ d t = 0. This confirms the results obtained in Section 2 for pure initial states of the two q-bits by numerically solving the equations of motion.
Naturally, since the equations of motion are based on the suitably generalized Poisson bracket algebra for quantum-classical hybrids [1, 6], as we discussed, we can employ the Poisson bracket between the Hamiltonian, H Σ = H cl + H qm + I hyb , and λ directly, in order to demonstrate that λ is a constant of motion. Thus, we find:
$$\left \{ H _ { \Sigma } ( x, p, X _ { \alpha }, P _ { \alpha } ), \lambda ( X _ { \alpha }, P _ { \alpha } ) \right \} = \left \{ [ H _ { q m } ( X _ { \alpha }, P _ { \alpha } ) + \mathcal { I } _ { h y b } ( x, X _ { \alpha }, P _ { \beta } ) ], \lambda ( X _ { \alpha }, P _ { \beta } ) \right \} \\ = \sum _ { \gamma = 1 } ^ { 4 } \left ( \frac { \partial ( H _ { q m } + \mathcal { I } _ { h y b } ) } { \partial X _ { \gamma } } \frac { \partial \lambda } { \partial P _ { \gamma } } - \frac { ( \partial H _ { q m } + \mathcal { I } _ { h y b } ) } { \partial P _ { \gamma } } \frac { \partial \lambda } { \partial X _ { \gamma } } \right ) = 0$$
which independently confirms the previous result. Here the classical part of the bracket does not contribute, since λ does not depend on the classical variables x, p .
In order to find possibly another invariant, we have evaluated also the Poisson bracket between the constraint C ( X ,P α α ), eq. (5), which is an invariant, and the present invariant λ X ,P ( α β ). Unfortunately, since the bracket between these two quantities vanishes, {C , λ } = 0, the resulting invariant is a trivial one. We have not found any other conserved quantity for our hybrid model.
Finally, we observe that the classical equations of motion (9)-(10) represent a harmonic oscillator affected by an 'external force', which describes the back reaction of the q-bits on the harmonic oscillator. Calculating the Poisson bracket between this force:
$$F ( X _ { 1 }, P _ { 1 }, X _ { 2 }, P _ { 2 } ) \coloneqq \frac { \beta \Omega } { 2 } \left ( X _ { 1 } ^ { 2 } + P _ { 1 } ^ { 2 } - X _ { 2 } ^ { 2 } - P _ { 2 } ^ { 2 } \right )$$
and the Hamiltonian, we obtain:
$$\left \{ & H _ { \Sigma } ( x, p, X _ { \alpha }, P _ { \alpha } ), F ( X _ { 1 }, P _ { 1 }, X _ { 2 }, P _ { 2 } ) \right \} \\ & = \left \{ \left [ H _ { q m } ( X _ { \alpha }, P _ { \alpha } ) + \mathcal { I } _ { h y b } ( x, X _ { \alpha }, P _ { \alpha } ) \right ], F ( X _ { 1 }, P _ { 1 }, X _ { 2 }, P _ { 2 } ) \right \} \\ & = \ 2 \omega _ { 0 } \beta \Omega \left ( X _ { 3 } ( P _ { 2 } - P _ { 1 } ) + P _ { 3 } ( X _ { 1 } - X _ { 2 } ) \right ) \.$$
This result, which generally does not vanish, demonstrates that the 'external force' F , due to the genuine quantum-classical hybrid coupling in our model, is not a constant of motion, as could be expected. Its effect on energy conservation will be discussed in Section 4.
## 3.1. Concurrence in presence of a semi-classical system
Our results for the concurrence, in particular its constancy, are to some extent expected. In fact, if the quantum sector of our model is appropriately coupled with another quantum system, such as the uncoupled q-bits interacting with a quantum harmonic oscillator, like in the generalized Tavis-Cummings model [11, 12, 13, 15, 16], then entanglement can be established between them. If we have only an interaction between a quantum and a classical subsystem, they cannot become entangled and the quantum state is confined to evolve in the subspace of states maintaining the initial quantum correlations.
More specifically, our study based on hybrid theory [1], confirms a recent analysis of two q-bits coupled to a largej integral spin, which latter presents a semi-classical subsystem [22]. We have included figures, from the cited article, which illustrate the trend of the concurrence in this case. In the limit j → ∞ , the integral spin subsystem becomes analogous to our classical harmonic oscillator. From Figure 3 and Figure 4, the trend is clear: in the classical limit, the concurrence becomes constant in time.
In Section 5, we study further related features in our hybrid model, namely when sharp classical oscillator initial conditions are replaced by a phase space distribution.
## 3.2. Concurrence in presence of external perturbations
We may wonder how the hybrid system is affected, if we switch on an external perturbation acting on a single q-bit of this general form:
$$\mathcal { H } _ { p e r t } ^ { ( 1 ) } \coloneqq \mathbf I \otimes ( \omega _ { 1 } \hat { \sigma } _ { x } + \omega _ { 2 } \hat { \sigma } _ { y } + \omega _ { 3 } \hat { \sigma } _ { z } )$$
with ω 1 2 3 , , given external parameters. In our chosen basis, cf. Section 1, we have:
$$\mathbf I \otimes \hat { \sigma } _ { x } = \left ( \begin{array} { c c c c } 1 & 0 & 0 & 0 \\ 0 & - 1 & 0 & 0 \\ 0 & 0 & 0 & - 1 \\ 0 & 0 & - 1 & 0 \end{array} \right ) \, \ \mathbf I \otimes \hat { \sigma } _ { y } = \left ( \begin{array} { c c c c } 0 & 0 & \imath & \imath \\ 0 & 0 & - \imath & \imath \\ - \imath & \imath & 0 & 0 \\ - \imath & - \imath & 0 & 0 \end{array} \right )$$
$$\mathbf I \otimes \hat { \sigma } _ { z } = \left ( \begin{array} { c c c c } 0 & 0 & 1 & 1 \\ 0 & 0 & 1 & - 1 \\ 1 & 1 & 0 & 0 \\ 1 & - 1 & 0 & 0 \end{array} \right ) \.$$
Figure 3. The concurrence of two q-bits coupled by an integral spin-j intermediary; for separable initial state | ++ 〉| -j 〉 of the tripartite system (left) and for entangled initial state (1 / √ 2)( | - -〉 + ++ ) | 〉 | -j 〉 (right). Figures from [22].
Figure 4. The concurrence (left) and its variance (right) of two q-bits coupled by an integral spin-j intermediary, representing an average over 103 initial states of the tripartite system. Figures from [22].
These matrices have to be inserted in between a generic state of the two-q-bit system and its adjoint, in order to obtain the corresponding contribution to the hybrid Hamiltonian, cf. Section 1.
Then, incorporating the oscillator expansion for the chosen basis, we may calculate the Poisson bracket between this perturbation and the concurrence (for pure initial states, as before in Section 3):
$$\ddot { \mathcal { H } } _ { p e r t } ^ { ( 1 ) }, \lambda \Big \} = \sum _ { \gamma = 1 } ^ { 4 } \left ( \frac { \partial \mathcal { H } _ { p e r t } ^ { ( 1 ) } } { \partial X _ { \gamma } } \frac { \partial \lambda } { \partial P _ { \gamma } } - \frac { \partial \mathcal { H } _ { p e r t } } { \partial P _ { \gamma } } \frac { \partial \lambda } { \partial X _ { \gamma } } \right ) = 0 \ \.$$
We note that this result is independent of ω 1 2 3 , , , which implies that even with this kind of perturbation the entanglement of the q-bits does not evolve. Thus, our results about the evolution of entanglement hold also for inhomogeneous q-bits .
This is obviously true for any perturbation of the classical sector as well, because the concurrence determined by λ depends explicitly only on the quantum canonical variables X ,P α α . A different situation arises for perturbations acting on both q-bits:
$$\mathcal { H } _ { p e r t } ^ { ( 2 ) } \coloneqq \omega _ { 1 } \hat { \sigma } _ { x } \otimes \hat { \sigma } _ { x } + \omega _ { 2 } \hat { \sigma } _ { y } \otimes \hat { \sigma } _ { y } + \omega _ { 3 } \hat { \sigma } _ { z } \otimes \hat { \sigma } _ { z } \.$$
In the chosen basis, we obtain the corresponding matrix representation:
$$\hat { \sigma } _ { x } \otimes \hat { \sigma } _ { x } = \left ( \begin{array} { c c c c } 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & - 1 & 0 \\ 0 & 0 & 0 & - 1 \end{array} \right ) \, \ \hat { \sigma } _ { y } \otimes \hat { \sigma } _ { y } = \left ( \begin{array} { c c c c } 0 & - 1 & 0 & 0 \\ - 1 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & - 1 \end{array} \right )$$
$$\hat { \sigma } _ { z } \otimes \hat { \sigma } _ { z } = \left ( \begin{array} { c c c c } 0 & 1 & 0 & 0 \\ 1 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & - 1 \end{array} \right ) \. \quad \ \quad \ \ ( 3 8 ) \\, \ \cdot \ \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \, \end{array}$$
Calculating the Poisson bracket between this kind of perturbation and the concurrence, we find:
glyph[negationslash]
$$\left \{ \mathcal { H } _ { p e r t } ^ { ( 2 ) }, \lambda \right \} \neq 0 \.$$
This implies that the entanglement evolves in time as, for example, shown in Figure 5. Furthermore, this result is independent of ω 1 , which means that perturbations of the form ˆ σ x ⊗ ˆ σ x have no effect on the entanglement, in the chosen basis. However, looking at the figure, we see that - independently of whether the initial state is entangled or not - in general, the two-q-bit perturbations affect the entanglement.
Figure 5. Evolution of the concurrence in presence of a two-q-bit perturbation for the uncorrelated initial state | + -〉 (left) and entangled initial state 1 √ 2 ( | ++ + 〉 | --〉 ) (right), with ω = 0 1, . ω 0 = 0 15 . ω , ω 1 = ω 2 = 0, ω 3 = 0 015 . ω , and β = 0 2. .
We also find that the Poisson bracket between the two-q-bit perturbation and the 'external force', eq. (30), which describes the back reaction of the q-bits on the harmonic oscillator, vanishes for certain combinations of the parameters. We have:
$$\{ \mathcal { H } _ { p e r t }, F ( X _ { 1 }, P _ { 1 }, X _ { 2 }, P _ { 2 } ) \} = 2 ( \omega _ { 2 } - \omega _ { 3 } ) ( X _ { 2 } P _ { 1 } - X _ { 1 } P _ { 2 } )$$
which is zero for ω 2 = ω 3 and, in particular, if only ω 1 is nonzero. In these cases, the dynamics of the quantum sector, especially of the entanglement, may change; however, the force exerted by the q-bits on the classical oscillator retains its form. Nevertheless, since its time dependence changes, the classical oscillator generally will be sensitive to the presence of perturbations acting only on the q-bits.
## 4. Quantum-classical hybrid cooling
In this section, we would like to point out an interesting practical aspect of the quantum-classical hybrid coupling. Following hybrid theory, as we have described, it could be employed to transfer energy from the quantum to the classical sector ( 'hybrid cooling' ), or vice versa.
While the total energy of the hybrid system is a constant of motion, the energy of the classical and quantum sectors separately changes in time, due to the coupling. Furthermore, we have seen in Section 3 that the concurrence is conserved. This means that we can cool (or heat) a sector of the system without changing the entanglement of the two-q-bit states.
For illustration, we have selected states that show a strong back reaction on the classical harmonic oscillator. Other choices are possible. Figures 6, 7 present the energy as a function of time for components of the quantum-classical oscillator/two-q-bits hybrid, cf. eqs. (1)-(3). Note that the total energy is given by H Σ = H cl + H qm + I hyb (see Section 1) and includes the contribution from the hybrid interaction (not shown in the figures).
√
Figure 6. Time dependence of energies, for the initial state ( | + -〉 + | -+ ) 〉 / 6 + ( | - -〉 + | + + ) 〉 / √ 3 and parameters ω = 0 03, . ω 0 = 0 16 . ω , β = 0 2: . on the left, total hybrid energy dashed line (blue) and classical energy of eq. (1) full line (green); on the right, quantum energy expectation of eq. (6) (blue).
Figure 7. Time dependence of energies, for the initial state ( -| ++ 〉-|--〉 + + | -〉 + | -+ ) 2 〉 / and parameters ω = 0 03, . ω 0 = 0 16 . ω , β = 0 02: . on the left, total hybrid energy dashed line (blue) and classical energy of eq. (1) full line (green); on the right, quantum energy expectation of eq. (6) (blue).
We see that the oscillations of the energies are wave packet like. The maximal amplitude of the classical energy coincides with the minimal amplitude of the quantum energy, with the hybrid interaction energy (not shown) carrying the remainder of the conserved total energy.
The figures particularly suggest that - if we can switch on the hybrid coupling for a limited period of time - we can manipulate the energies of the quantum and classical sectors. This effect could have practical applications and, indirectly, test the hybrid theory. A topic for further investigations is how to optimize the desired control of the respective energies.
To illustrate this effect, we choose a time-dependent coupling described by a Gaussian:
$$\Omega ( t ) = A \exp \frac { ( t - t _ { 0 } ) ^ { 2 } } { 2 \sigma ^ { 2 } }$$
with the parameter t 0 denoting the instant of maximal hybrid coupling, σ determining its temporal width and A its amplitude. - Since such coupling requires an interaction of the quantum-classical hybrid with yet another external system, the total energy of the hybrid is not necessarily conserved here.
Figure 8. Evolution of energies for the time dependent hybrid coupling shown in the upper left graph ( A = 10, σ = 1000, t 0 = 10 , cf. 4 eq. (41)); shown are the hybrid energy (top right) of eq. (7), the classical energy (bottom left) of eq. (1), and the quantum energy (bottom right) of eq. (6). The initial state is the same as in Figure 6 and ω = 0 03, . ω 0 = 0 15 . ω , β = 0 2. .
Depending on the choice of parameters of the hybrid coupling, the amount of cooling (and, analogously, of heating) effected on the two-q-bit subsystem can be varied as desired. The hybrid interaction, however, does not modify the entanglement present or absent in the initial state. Thus, it may offer an interesting new tool for controlling the energy of a quantum (sub)system, while leaving characteristic quantum features intact, as we have shown in the present example.
## 5. Quantum-classical hybrid evolution for a distribution of classical initial states
We turn to the more realistic situation of having to deal with a distribution of initial conditions instead of a sharp one (i.e. a phase space point) for the classical harmonic oscillator subsystem. Incorporating this into hybrid initial states, we may ask how the average over such a distribution modifies the properties of the two-q-bit subsystem, which we have studied.
We choose Gaussian distributions of the initial positions and momenta of the oscillator:
$$x _ { 0 } = \frac { 1 } { \sigma \sqrt { 2 \pi } } e ^ { ( - \frac { ( x - x ) ^ { 2 } } { 2 \sigma ^ { 2 } } ) } \, \ \ p _ { 0 } = \frac { 1 } { \sigma \sqrt { 2 \pi } } e ^ { ( - \frac { ( p - \bar { p } ) ^ { 2 } } { 2 \sigma ^ { 2 } } ) } \.$$
In the Figure 9, we demonstrate for a typical example that a probabilistic distribution of the (classical) initial states leads effectively to turning the pure initial (quantum) state into a mixed one and to loosing entanglement of the two q-bits. We recall here qualitatively similar results obtained in Ref. [22], which we have discussed in Section 3.1 .
Figure 9. Effects of distributed oscillator initial conditions, cf. eqs. (42), with parameters ¯ = 0, x ¯ = 10, p σ = 1. Shown is the evolution of the linear entropy of the two-q-bit subsystem (left figures) and of its concurrence (right figures), for couplings ω = 0 03, . ω 0 = 0 15 . ω , β = 0 2; . the initial states are ( ++ | 〉 + | --〉 ) / √ 2 (top figures) and ( + | -〉 + | -+ ) 〉 / √ 2) (bottom figures).
In order to obtain these results, we have used the so-called Monte-Carlo method with direct sampling [23]. - We randomly select M configurations, { ( x 0 ,i , p 0 ,i ) , i = 1 , . . . , M } , for the initial conditions of the classical sector, which are distributed, however, with bias according to eqs. (42). Then, we solve the whole set of hybrid equations of motion M times, once for each configuration i ; this concerns the differential equations (9)-(12). With the help of these solutions, we correspondingly obtain M density matrices representing the two-q-bit subsystem, as before. The average of these matrices (for M sufficiently large) provides an estimate of the limiting density matrix incorporating the distributed oscillator initial conditions:
$$\hat { \rho } \approx \frac { 1 } { M } \sum _ { i = 1 } ^ { M } \hat { \rho } ( x _ { 0, i }, p _ { 0, i } )$$
in accordance with the central limit theorem.
We have performed numerical simulations for M = 10 30 50 80 100 and observed that the , , , , results stabilize for M > ∼ 50. Therefore, in Figure 9, only results for M = 100 are presented. Following Ref. [23], an error analysis could be performed. We leave this for future work, including the development of a more powerful numerical scheme.
Having performed a series of simulations of this kind, we find that the time in which a maximally entangled state passes from the initially maximal to zero concurrence, is quite sensitive to the width of the Gaussian initial state distribution. Indeed, choosing σ = 1, we have assumed a rather narrow initial distribution corresponding roughly to a phase space cell of size ¯ h (in physical units). To complement the present study, it will be interesting to carry out more extensive (and computationally more demanding) numerical simulations with wider initial phase space distributions for the classical oscillator, e.g., of nanometer to micrometer size and momenta corresponding to small thermal energies relevant for optomechanical experiments.
## 6. Conclusions
We have studied in this work the quantum-classical hybrid version of the versatile all-quantum model of one oscillator coupled to two q-bits [11, 12, 13, 15, 16]. In the present case, the oscillator is considered as classical, while the q-bits retain their full quantum features. The coupling between quantum and classical sectors has been constructed following the hybrid theory developed in Refs. [1, 6, 7], a topic that has found increasing attention recently, see, for example, Refs. [2, 3, 4, 5, 24, 25].
We haven chosen the hybrid theory developed in Refs. [1, 6, 7] for the present work, since earlier proposals have been beset with problems, such as, for example, violation of energy conservation or of unitarity, which has led to various 'no-go' theorems. An extensive list of such issues has been discussed and overcome [1, 6, 7]. The only other consistent formulation, as far as we know, is the theory developed by Buric and collaborators [2, 3], the formal results of which are, however, equivalent to the hybrid theory used here. - The configuration space theory of Hall and Reginatto [4] might provide an alternative consistent framework for the description of quantum-classical hybrids under certain circumstances. However, it has been shown recently that it leads to nonlocal signaling [26] and nonseparability of degrees of freedom that are classically and quantum mechanically separable [27], due to the intrisically nonlinear formulation. Since these features can interfere with entanglement, we have not consulted this theory here.
The present hybrid model may be adapted to experiments, in order to test the basic tenet of hybrid dynamics, namely that the direct coupling of quantum mechanical to classical degrees of freedom can be consistently formulated - be it with orientation towards approximation schemes for fully quantum mechanical yet complex systems or because of considerations of foundational issues. One may speculate that a quantum-classical hybrid coupling could be of fundamental character, when microscopic quantum degrees of freedom interact directly with macroscopic classical degrees of freedom, e.g. elementary particles, atoms, or molecules interacting with a gravitational field (if it is and remains classical) or quantum mechanical systems undergoing a measurement by a classical apparatus (Copenhagen interpretation).
By evaluating the concurrence, we have found that entanglement present in the two-q-bit system is invariant under evolution according to the hybrid dynamics. This result may have been expected but has not been studied in a consistent dynamical theory before, except Ref. [28]. A possible application could be in quantum information protocols transferring information coherently between q-bits, yet under the controlling influence of distinct classical degrees of freedom embedded in the environment.
Furthermore, we have shown that the model suggests a hybrid cooling scheme, such that the energy (temperature) of the q-bit subsystem can be selectively lowered without affecting the quantum correlations between its components.
An important question that remains to be studied further is whether other interesting aspects of a quantum object can be addressed by monitoring classical degrees of freedom coupled directly to it. The extension of our study allowing for mixed quantum states should be carried out and, more generally, incorporating decoherence and dissipation due to the environment.
On the other hand, we have seen that if (for example, the initial conditions of) such classical degrees of freedom are distributed statistically, correspondingly averaged density (sub)matrices for the q-bits show impurity and disentanglement increasing with time. This appears much like in the all-quantum version of the model and might mask desperately searched for quantum effects such as collapse or (gravitationally induced) spontaneous localization of the wave function, see, for example, Refs. [28, 29, 30] and further references therein.
Certainly, the study of quantum-classical hybrid systems has only begun and much more can be envisaged, be it for technologically interesting situations or be it as far-reaching as to wonder about the role of hybrids under biologically relevant circumstances.
## Acknowledgments
We thank N.Buric, G. Cella, and A. Posazhennikova for discussions. H.-T. E. wishes to thank A.Khrennikov in particular for the invitation to organize a special session on quantum-classical hybrid systems during the Vaxjoe meeting (QTAP, June 2013) and for his kind hospitality there. L.F. and A.L. gratefully acknowledge support through Phd programs of their institutions; A.L. has been supported by an ERC AdG OSYRIS fellowship.
## References
- [1] H.-T. Elze, Linear dynamics of quantum-classical hybrids , Phys. Rev. A 85, 052109 (2012)
- [2] M. Radonji´ c, S. Prvanovi´ and N. Buri´, c c Hybrid quantum-classical models as constrained quantum systems , Phys. Rev. A 85, 064101 (2012)
- [3] N. Buri´ c, I. Mendas, D. B. Popovi´ c, M. Radonji´ and S. Prvanovi´, c c Statistical ensembles in Hamiltonian formulation of hybrid quantum-classical systems , Phys. Rev. A 86, 034104 (2012)
- [4] M.J.W. Hall and M. Reginatto, Interacting classical and quantum ensembles , Phys. Rev. A 72, 062109 (2005)
- [5] L. Di´si, o Hybrid quantum-classical master equations , to appear in this volume (2014) [arXiv:1401.0476]
- [6] H.-T. Elze, Four questions for quantum-classical hybrid theory , J. Phys.: Conf. Ser. 361 (2012) 012004
- [7] H.-T. Elze, Proliferation of observables and measurement in quantum-classical hybrids , Int.J.Qu.Inf.(IJQI) 10, No. 8 (2012) 1241012
- [8] H.-T. Elze, Quantum-classical hybrid dynamics - a summary , J. Phys.: Conf. Ser. 442 (2013) 012007
- [9] C.J. Hood, T.W. Lynn, A.C. Doherty, A.S. Parkins and H.J. Kimble, The atom-cavity microscope: single atoms bound in orbit by single photons , Science 287 (2000) 1447
- [10] J.M. Raimond, M. Brune and S. Haroche, Manipulating quantum entanglement with atoms and photons in a cavity , Rev. Mod. Phys. 73 (2001) 565
- [11] E.K. Irish, J. Gea-Banacloche, I. Martin and K.C. Schwab, Dynamics of a two-level system strongly coupled to a high-frequency quantum oscillator , Phys. Rev. B 72, 195410 (2005)
- [12] E.K. Irish, Generalized rotating-wave approximation for arbitrarily large coupling , Phys. Rev. Lett. 99 (2007) 173601
- [13] S. Agarwal, S.M. Hashemi Rafsanjani and J.H. Eberly, Tavis-Cummings model beyond the rotating wave approximation: Quasidegenerate q-bits , Phys. Rev. A 85 (2012) 043815
- [14] A. Heslot, Quantum mechanics as a classical theory , Phys. Rev. D 31 (1985) 1341
- [15] M. Tavis and F.W. Cummings, Exact solution for an N-molecule-radiation-field Hamiltonian , Phys. Rev. 170 (1968) 379
- [16] M. Tavis and F.W. Cummings, Approximate solutions for an N-molecule-radiation-field Hamiltonian , Phys. Rev. 188 (1969) 692
- [17] A. Wallraff, D.I. Schuster, A. Blais, L. Frunzio, R.-S. Huang, J. Majer, S. Kumar, S.M. Girvin and R.J. Schoelkopf, Strong coupling of a single photon to a superconducting qubit using circuit quantum electrodynamics , Nature (London) 431(2004) 162
- [18] P. Forn-D´ ıaz, J. Lisenfeld, D. Marcos, J.J. Garc´ ıa-Ripoll, E. Solano, C.J.P.M. Harmans and J.E. Mooij, Observation of the Bloch-Siegert shift in a qubit-oscillator system in the ultrastrong coupling regime , Phys. Rev. Lett. 105 (2010) 237001
- [19] C.H. Bennett, D.P. Di Vincenzo, J.A. Smolin and W.K. Wootters, Entanglement of a pair of quantum bits , Phys. Rev. Lett. 78 (1997) 5022
- [20] W.K. Wootters and P. Delsing, Entanglement of formation of an arbitrary state of two qubits , Phys. Rev. Lett. 80 (1998) 2245
- [21] C.H. Bennett, H.J. Bernstein, S. Popescu and B. Schumacher, Concentrating partial entanglement by local operations , Phys. Rev. A 53 (1996) 2046
- [22] J. Dajka, Disentanglement of qubits in classical limit of interaction , Int. J. Theor. Phys. 53 (2014) 870
- [23] W. Krauth, Statistical Mechanics: Algorithms and Computations (Oxford Univ. Press, Oxford UK, 2006)
- [24] L. Di´si, N. Gisin and W.T. Strunz, o Quantum approach to coupling classical and quantum dynamics , Phys. Rev. A 61 (2000) 022108
- [25] A. Peres and D.R. Terno, Hybrid classical-quantum dynamics , Phys. Rev. A 63 (2001) 022101
- [26] M.J.W. Hall, M. Reginatto and C.M. Savage, Nonlocal signaling in the configuration space model of quantumclassical interactions , Phys. Rev. A 86 (2012) 054101
- [27] F. Di Lauro, Mixing quantum and classical dynamics , Tesi di Laurea Triennale (advisor H.-T. Elze), University of Pisa (2011), unpublished
- [28] A. Lampo, L. Fratino and H.-T. Elze, The Marshall et al. optomechanical experiment in quantum-classical hybrid theory , submitted to Phys. Rev. A (July 2014)
- [29] W. Marshall, C. Simon, R. Penrose and D. Bouwmeester, Towards quantum superpositions of a mirror , Phys. Rev. Lett. 91 (2003) 130401
- [30] M. Schlosshauer, Decoherence and the quantum-to-classical transition (Springer, Berlin, 2007)
|
10.1088/0031-8949/2014/T163/014005
|
[
"L. Fratino",
"A. Lampo",
"H. -T. Elze"
] |
2014-08-05T15:56:34+00:00
|
2014-08-05T15:56:34+00:00
|
[
"quant-ph"
] |
Entanglement dynamics in a quantum-classical hybrid of two q-bits and one oscillator
|
We investigate new features, especially of entanglement dynamics, which arise
in a quantum-classical hybrid. As a model, we study the coupling between two
quantum mechanical two-level systems, i.e. two q-bits, and a classical harmonic
oscillator. Their interaction is described by a hybrid coupling, in accordance
with a recently developed quantum-classical hybrid theory. We discuss various
situations in which entanglement of the q-bits does (not) evolve. Furthermore,
we point out an experimental application in a hybrid cooling scheme and
indicate topics for future study.
|
1408.1009v1
|
## Gravitational resonance spectroscopy with an oscillating magnetic field gradient in the GRANIT flow through arrangement.
G. Pignol, 1, ∗ S. Baeßler, 2 V. V. Nesvizhevsky, 3 K. Protasov, 1 D. Rebreyend, 1 and A. Yu. Voronin 4
1 LPSC, Universit´ e Grenoble-Alpes, CNRS/IN2P3, Grenoble, France 2 Physics Department, University of Virginia 382 McCormick Road, Charlottesville, VA 22904, U.S.A. 3 Institut Max von Laue - Paul Langevin, 71 av. des Martyrs, 38000 Grenoble, France 4 Lebedev Institute 53 Leninskii pr., 119991 Moscow, Russia (Dated: August 6, 2014)
Gravitational resonance spectroscopy consists in measuring the energy spectrum of bouncing ultracold neutrons above a mirror by inducing resonant transitions between different discrete quantum levels. We discuss how to induce the resonances with a flow through arrangement in the GRANIT spectrometer, excited by an oscillating magnetic field gradient. The spectroscopy could be realized in two distinct modes (so called DC and AC) using the same device to produce the magnetic excitation. We present calculations demonstrating the feasibility of the newly proposed AC mode.
## I. INTRODUCTION
Ultracold neutrons bouncing over a horizontal mirror are used to probe quantum effects of a particle in the gravitational field [1]. The vertical motion of such neutrons bouncing at sub-millimeter distance from the mirror has discrete energy spectrum that can be calculated from the stationary Schr¨dinger equation o where glyph[epsilon1] n = [2 338 4 088 5 521 6 787 . , . , . , . , · · · ] is the series of the negative zeros of the Airy function. To perform the spectroscopy of the bouncing neutron, the interaction that couples different quantum states can be a vibration of the bottom mirror, or an oscillating magnetic field gradient [8]. The former has been used by the QBounce collaboration [7], the latter will be used in the GRANIT spectrometer [9].
$$\frac { \hbar { ^ { 2 } } } { 2 m } \frac { d ^ { 2 } } { d z ^ { 2 } } \psi _ { n } + m g z \, \psi _ { n } = E _ { n } \psi _ { n }, \quad \quad ( 1 ) \quad \text{me} \\ \text{free}$$
where m is the neutron mass, g = 9 81 m/s . 2 is the local gravitational acceleration, ψ n ( z ) n =1 2 , , ··· are the stationary wavefunctions with associated energy E n . The existence of the quantization of the vertical motion was demonstrated a decade ago [2], profiting from the relatively large spatial extension of the ground state wavefunction characterized by z 0 = (¯ h / m g 2 2 2 ) 1 / 3 ≈ 5 87 . µ m.
Precision study of the quantum states is motivated by their sensitivity to extra short range interactions (see [3] and references therein) in particular those induced by Chameleon Dark Energy [4-6]. In addition, the neutron quantum states provide a unique test of the weak equivalence principle in a quantum regime, since the inertial and gravitational masses in eq. (1) do not cancel.
High precision measurements can be achieved with the gravitational resonance spectroscopy technique, where transitions between quantum states are induced by a periodic excitation [7]. The characteristic frequencies of the transitions, as low as f 0 = mgz / πh 0 2 ¯ ≈ 145 Hz, are accessible by electrical as well as mechanical oscillators. More precisely, the quantum frequencies for a transition n → m from the state of energy E n to the state of energy E m is given by f nm = ( E n -E m ) / 2 πh ¯. Solving the problem (1) one can show that
$$f _ { n m } = f _ { 0 } ( \epsilon _ { n } - \epsilon _ { m } ), \quad \quad ( 2 ) \quad 0. \Sigma$$
∗ Electronic address: [email protected]
As a first step of GRANIT, a flow through measurement of the resonant transitions magnetically excited between the first three quantum states will be realized as first proposed in [10]. According to this proposal, a space periodic (but static) magnetic field gradient will be generated at the surface of the bottom mirror. The frequency of the excitation seen by a neutron will thus vary according to its horizontal velocity. A detailed analysis of this scheme is provided in [9]. Another mode of operation (the AC mode) can be implemented with the same setup, consisting in generating a homogeneous gradient, but oscillating in time, that should allow a more direct probe of the resonances.
The paper is organized as follows: in section 2 we present the flow through setup, in section 3 we describe in some details the magnetic excitation, in section 4 we calculate the conditions for the adiabaticity of spin transport and in section 5 we present a theoretical description of the magnetically induced transitions in the AC mode.
## II. THE GRANIT FLOW THROUGH SETUP
The sketch of the flow through setup is shown in fig. 1. Ultracold neutrons are produced from a dedicated superthermal source installed at a cold beamline of the high flux reactor of the Institut Laue Langevin. The source relies on down-scattering of neutrons with a wavelength of 0 89 nm in a superfluid helium bath cooled down to 0.8 K . [11, 12]. According to a preliminary measurement of the UCN velocity spectrum [13], we expect the v x velocity along the beam to be distributed with a mean value of v 0 = 4 m/s and a standard deviation of 1 5 m/s. . In the
FIG. 1: Sketch of the flow through setup. Ultracold neutrons enter from the left, they go through the step to depopulate the ground quantum state (1), 16 cm long transition region (2), 9 cm long analyzer (3) and detector (4).
remaining of this article we will assume a Gaussian profile for the v x distribution with those parameters. UCNs are extracted using a narrow slit to accept only those with practically no vertical velocity. The spectroscopy is performed with four steps: (1) state preparation, (2) resonant transition, (3) state analysis, (4) detection of transmitted flux.
- · UCNs are first prepared in an excited state by going down a step (1) of height 15 µ m. The populations p n of the quantum states after the step are expected to be about p 1 = 0 02 . , p 2 = p 3 = p 4 = 0 3. . Thus, the population of the ground quantum state is suppressed as compared to the populations of excited states.
- · Next, transitions between quantum states are induced with a periodic magnetic field gradient. The length of the transition region is L = 16 cm, corresponding to an average passage time t 0 = 40 ms. Two different schemes could be implemented in principle: the AC excitation and the DC excitation. In the DC mode, the field gradient is static and spatially oscillating in the x direction with a period of d = 1 cm. In this case only neutrons with specific horizontal velocities meet the resonance condition. The deexcitation 2 → 1 is expected to be induced by an excitation frequency f 21 = 254 Hz, corresponding to the resonant horizontal velocity of v 21 = d f 21 = 2 54 m/s, for the 3 . → 1 case we expect f 31 = 462 Hz and v 31 = 4 62 m/s. . In the AC mode, the field gradient is spatially uniform, oscillating in time. One would then find the resonances by directly scanning the excitation frequency.
- · A second horizontal mirror above the main mirror serves as a state analyzer. For a slit opening of about 25 µ m, only ground state neutrons are accepted, higher quantum states are rejected. The length of the analyzer in the x direction is 9 cm.
- · Finally, neutrons are detected at the exit of the analyzer. In AC mode, the flux of transmitted neutrons should display a resonance pattern as a function of the excitation frequency. In DC mode, one has to measure the horizontal velocity of the transmitted neutrons to deduce the resonant frequency. This is achieved by measuring the height of the neutrons
with a position sensitive detector after a free fall distance of 30 cm.
Before developing the details, let us estimate the strength of the needed oscillating magnetic field gradient. The interaction of a neutron with a magnetic field glyph[vector] B is described by the operator -ˆ glyph[vector] µ · glyph[vector] B , where ˆ glyph[vector] µ is the neutron magnetic moment observable acting on the internal spin degree of freedom. When the variation of the magnetic field is slow compared to the Larmor frequency, the spin will follow the direction of the magnetic field. This adiabaticity condition will be addressed in section 4. In this case the neutron trajectory and the spin dynamics are effectively decoupled. Then the motion of the neutron is determined by the potential sµ glyph[vector] | B | where µ = 60 3 neV/T is the magnetic moment of the neutron . and s = 1 for 'spin up' neutrons and s = -1 for 'spin down' neutrons. Classically, a vertical force is applied on the neutron by the field gradient ∂ z | glyph[vector] B | . Let us now assume a magnetic excitation of the form | glyph[vector] B | = βz cos( ωt ). The quantum mechanical excitation potential reads
$$\hat { V } = s \mu \ \beta \ \hat { z } \ \cos ( \omega t ).$$
When the excitation frequency is close to a resonance, ω ≈ 2 πf nm , a Rabi oscillation between states | n 〉 and | m 〉 will take place at the angular frequency
$$\Omega _ { n m } = \frac { \mu } { \hbar } \langle n | \hat { z } | m \rangle \beta,$$
where the matrix elements of ˆ can be expressed as (see z e.g. [14])
glyph[negationslash]
$$\langle n | \hat { z } | m \rangle \ & = \ \frac { 2 z _ { 0 } } { ( \epsilon _ { n } - \epsilon _ { m } ) ^ { 2 } } \quad ( n \neq m ), \quad \quad ( 5 ) \\ \langle n | \hat { z } | n \rangle \ & = \ \frac { 2 } { 3 } \ z _ { 0 } \ \epsilon _ { n }.$$
To maximize the transition probability at resonance, the excitation strength β should verify Ω t 0 = π , where t 0 is the excitation time. This condition can be expressed using eq. (4) and (5) as
$$\beta _ { \text{needed} } = \frac { \pi } { 2 } \frac { \hbar } { \mu z _ { 0 } } \left ( \frac { f _ { n m } } { f _ { 0 } } \right ) ^ { 2 } \, \frac { 1 } { t _ { 0 } }. \quad \quad ( 6 )$$
One finds a needed field gradient of β = 0 22 T/m to . induce the 2 → 1 transition and β = 0 74 T/m to induce . the 3 → 1 transition.
## III. THE MAGNETIC EXCITATION
The magnetic field excitation will be generated by an array of 128 copper wires with square section arranged as shown in fig. 1. In practice the system is constituted of four modules, each one holding 32 adjacent wires. A wire has a section of 1 mm 2 and a length of 30 cm
FIG. 2: Magnetic field gradient ∂ z | B | produced at the surface of the main mirror, without any external magnetic field.
in the y direction. Adjacent wires are separated by a gap of 0 25 mm. . Electrical connectors are arranged so that the following 8-periodic pattern current could be applied I 1 , I 2 , I 3 , I 4 , -I 1 , -I 2 , -I 3 , -I 4 , I 1 · · · . Thus the magnetic field will be 1-cm periodic. The system will be placed above the horizontal mirror in the transition region as shown in fig. 1 at a distance of 0 8 mm from the . mirror.
The magnetic field generated by a single infinitely long square wire can be calculated analytically, the corresponding formulas are reported in the appendix. The magnetic field components B x ( x , B ) z ( x ) and gradients ∂ B z x ( x , ∂ ) z B z ( x ) at the surface of the mirror are obtained by summing the corresponding quantities for each of the 128 wires weighted by the individual currents. Then the field gradient is calculated according to
$$\partial _ { z } | B | = \frac { B _ { x } \partial _ { z } B _ { x } + B _ { z } \partial _ { z } B _ { z } } { | B | }. \quad \quad ( 7 ) \quad \min _ { \text{and} }$$
It is possible to tune the currents to obtain a homogeneous gradient at the surface of the mirror. We show in fig. 2 the result for I 1 = I 4 = 1 4 A and . I 2 = I 3 = 3 5 A . where a field gradient of 0 52 T/m is obtained, as needed . to induce resonant transitions between quantum states. This will be the benchmark configuration for the rest of the article. The residual 'noise' seen in fig 2 has an amplitude of 0 02 T/m; this noise would increase for a . wire array closer to the mirror. The frequency of this noise seen by a neutron passing at 4 m/s is about 2 kHz, much higher than the frequencies of interest for resonant transitions between low lying quantum states.
Note that the result shown in fig. 2 assumes that no external field is applied. By applying an external field glyph[vector] B 0 with for example B 0 ,x = B 0 ,z = 1 5 mT and . B 0 ,y = 0, the situation changes dramatically as shown in fig. 3. With a strong external field applied, a gradient oscillating in the x direction with a period of 1 cm is generated.
FIG. 3: Magnetic field gradient ∂ z | B | produced at the surface of the main mirror with external magnetic field applied in the x, z plane.
FIG. 4: Magnetic field generated at the surface of the main mirror assuming the benchmark configuration I 1 = I 4 = 1 4 A . and I 2 = I 3 = 3 5 A. .
As a result, the wire array is a versatile device to generate the field gradient that can be used for the AC excitation mode as well as for the DC excitation mode. In the DC mode, we apply DC current in the wire array and apply a strong external field in the x, z direction. The vertical force exerted by the field gradient on the neutron will oscillate in space. In the AC mode, we apply AC current in the wire array and a small external field B 0 ,y to satisfy the adiabaticity of spin transport, as detailed in the next section.
## IV. THE ADIABATICITY OF SPIN TRANSPORT
The magnetic field gradient produced by the array of wires described in the previous section will exert a force
on the passing neutrons. The sign of the force depends on the relative orientation between the neutron spin and the magnetic field. To induce resonant transitions between quantum states, a neutron should feel a well defined oscillating vertical force. Thus, one must make sure that the magnetic field is strong enough to hold the neutron spins parallel or antiparallel to the magnetic field at any time. If the adiabaticity condition for spin transport along the wire array is fulfilled, then the spin dynamics and the neutron trajectory are decoupled.
In the AC mode of excitation, the input current in the wire array is oscillating with driving frequency f . The magnetic field amplitude created by the wire array at the surface of the mirror is given by the pattern shown in fig. 4, with the whole pattern oscillating in time at the driving frequency f . Thus the magnetic field generated by the wire array will not be sufficient to hold the neutron spin, since the magnitude of the field crosses zero at a frequency f . To maintain a nonzero value of the field magnitude at anytime, a static, homogeneous external field B 0 ,y is applied in the transverse y direction. The purpose of this section is to calculate the minimum B 0 ,y field to apply in order to guarantee the adiabaticity of neutron spin transport when passing by the magnetic excitation.
Here we calculate the spin dynamics only, assuming that a neutron pass below the wire array in a straight horizontal line trajectory, at the surface of the bottom mirror. A neutron with velocity v along the x direction sees a time-dependent magnetic field in its rest frame given by:
$$B _ { x } ( t ) \ = \ B _ { 1 } \cos ( 2 \pi f t + \phi ) \sin ( 2 \pi v t / d ), \quad ( 8 ) \quad \text{the f}$$
$$B _ { y } ( t ) \ = \ B _ { 0, y }, & & ( 9 ) \quad \text{the f}$$
$$B _ { z } ( t ) \ = \ - B _ { 1 } \cos ( 2 \pi f t + \phi ) \cos ( 2 \pi v t / d ). \ \ ( 1 0 ) \quad \text{field}$$
It results from the combination of the oscillation of the field in space with period d and the oscillation in time with frequency f . We will set B 1 = 0 8 mT for the . benchmark wire currents described in the previous section.
The spin dynamics is given by the Bloch equation for the polarization vector Π: glyph[vector]
$$\frac { d \vec { \Pi } } { d t } = \gamma \ \vec { \Pi } \times \vec { B } ( t ), \quad \quad ( 1 1 ) \quad \text{th} \quad \text{th} \quad \text{gru}$$
where γ = 2 µ/h ¯ = 183 kHz/mT is the neutron gyromagnetic ratio.
We have solved numerically the Bloch equation using a Runge Kutta solver. The initial condition for the polarization vector Π(0) was set to the unit vector aligned glyph[vector] with glyph[vector] B (0), describing a 'spin up' neutron. We define the spin-flip probability at time t as p t ( ) = (1 -glyph[vector] Π · glyph[vector] B/ glyph[vector] | B / | ) 2. With this definition p t ( ) = 0 if the spin is aligned with the magnetic field at time t and p t ( ) = 1 if the spin has reversed its direction relative to the magnetic field at time t . As an adiabaticity criterion we take p max , the maximumspin-flip probability during the passage of a neutron
FIG. 5: Numerically calculated neutron spin-flip probability during the passage below the wire array in the benchmark configuration I 1 = I 4 = 1 4 A and . I 2 = I 3 = 3 5 A, as a . function of the driving frequency, for different values of the external B 0 ,y field.
below the wire array of duration L/v . For a given set of parameters B 0 ,y , f, v, φ the criterion p max was numerically calculated. The result was then averaged over the phase φ and the velocity spectrum v . The final result is presented in fig. 5 as a function of the driving frequency f , for different values of the external field B 0 ,y .
As a conclusion of the numerical study, the value of the external field B 0 ,y = 0 3 mT is sufficient to hold the . neutron spin with an accuracy better than one percent, in the frequency range of interest between 0 and 300 Hz. In the following calculations the external holding transverse field will thus be set to B 0 ,y = 0 3 mT. .
## V. THE RESONANT TRANSITIONS IN THE AC MODE
We have now defined a magnetic configuration for the AC mode with the wire array (oscillating currents at variable frequency f and fixed amplitude I 1 = I 4 = 1 4 A, . I 2 = I 3 = 3 5 A) and the external field ( . B 0 ,y = 0 3 mT) . that (i) holds the neutron spin (ii) generate an oscillating gradient with the required amplitude to induce resonant transitions between quantum states. Note that the time dependent gradient β t ( ) seen by the neutrons is not perfectly harmonic. Following eq. (7) the expression of the time dependent gradient is
$$\text{ed} \\ \text{he} \\ / 2. \\ \cdot.$$
where f is the driving frequency of the current in the wire array, B 1 = 0 8 mT and . ˆ = 0 52 T/m. β . We plot β t ( ) in fig. 6, where it is apparent that the excitation frequency (the frequency of the β t ( ) excitation) is twice
FIG. 6: Time dependent gradient β t ( ) for a driving frequency of f = 100 Hz, corresponding to an excitation frequency of 200 Hz. Solid black line: eq. (12), dashed red line: first order Fourier expansion eq. (16).
the driving frequency f (the frequency of the oscillating currents in the wire array).
We will now simulate the transition probabilities that could be observed in the GRANIT flow through setup as described in section 2, assuming the benchmark magnetic configuration resulting in the excitation (12). The problem consists in calculating the time evolution of a neutron quantum state
$$\left | \psi ( t ) \right \rangle = \sum _ { n } a _ { n } ( t ) \ | n \rangle, \quad \ \ ( 1 3 ) \quad \text{wh} _ { \text{sta} }$$
which is the solution of the time dependent Schr¨dinger o equation
$$i \frac { d a _ { n } } { d t } = E _ { n } / \hbar { \ } a _ { n } + \sum _ { m } \frac { s } { 2 } \gamma \beta ( t ) \langle n | \hat { z } | m \rangle a _ { m }. \quad ( 1 4 ) \quad \text{(} \quad \text{(} \quad \text{(} \quad \text{$\dots}$}$$
Here we assume that the adiabaticity condition is fulfilled. The initial condition of the state is chosen to be | ψ (0) 〉 = 2 | 〉 immediately after the preparation step. Then we solved eq. (14) with a Runge-Kutta algorithm for a given set of parameters f, v, s, φ , where the sum is restricted to the first four quantum states. For a given horizontal velocity v , the probability | a 1 ( L/v ) | 2 for the neutron to be detected in the ground state at the exit of the magnetic excitation is calculated. The result is then averaged over the excitation phase φ , the spin state s = ± 1 and the velocity spectrum. We plot in fig. 7 the transition probability as a function of the driving frequency f .
According to this numerical calculation, we expect to see two resonances in the transmitted UCN flux associated with the 2 → 1 transition, with frequencies
$$f ^ { + } = 1 4 1. 5 \, H z \ \text{ and } \ f ^ { - } = 1 1 3. 5 \, H z, \quad ( 1 5 ) \quad \text{that}$$
FIG. 7: Numerical solution of the Schr¨dinger equation (14). o The probability of the 2 → 1 transition is plotted as a function of the driving frequency.
where f + and f -are the maxima of the resonance curve corresponding to spin 'up' neutrons ( s = 1) and spin 'down' neutrons ( s = -1) respectively.
The splitting of the resonances depending on the spin state could be interpreted as a 'Stern-Gerlach' split due to a constant magnetic field gradient. To see this, it is useful to perform the Fourier expansion of the gradient excitation β t ( ) given by (12)
$$\beta ( t ) = \beta _ { 0 } + \beta _ { 1 } \cos ( 4 \pi f t + 2 \phi ) + \cdots \quad ( 1 6 )$$
where β 0 = 0 289 T/m and . β 1 = 0 228 T/m. . The constant term β 0 of the excitation should be thought of as an spin-dependent effective modification of g . This is done by identification of the total constant vertical force (gravity plus constant gradient) to an effective gravitational force:
$$\ m g ^ { \pm } = m g \pm \mu \beta _ { 0 }. \text{ \quad \ \ } ( 1 7 )$$
Thus we expect a spin-dependent resonance frequency of the n → m transition given by
$$\ t \ v o \\ \ t \text{ep.} \\ \text{htm is} \\ \text{ven} \quad \text{when } \tilde { \epsilon } \quad \text{is to the number should nothing to come} \ \text{Tho}.$$
where f nm is the unperturbed transition frequency. The splitting of the resonances obtained by the numerical resolution of the full Schr¨dinger equation (14) is in quano titative agreement with eq. (18).
In addition, eq. (18) motivates a combination of the two resonant frequencies f + and f -to extract the unperturbed transition frequency f 21 , namely
$$f _ { 1 2 } = \left ( \frac { ( 2 f ^ { + } ) ^ { 3 / 2 } + ( 2 f ^ { - } ) ^ { 3 / 2 } ) } { 2 } \right ) ^ { 2 / 3 }. \quad ( 1 9 )$$
Remember that f + and f -refer to driving frequencies, that correspond to excitation frequencies of 2 f + and 2 f -.
When applied to the maxima of the curve shown in fig. 7, one extract f 21 = 255 8 Hz which differs from the true . resonance frequency f 21 true , = 253 8 Hz given by eq. . (2) by 2 Hz. In fact there are several features of the full problem (given by eq. (14)), that the simple estimate (18) does not catch. The simplified formula (18) can be obtained by assuming a two level system (states | 1 〉 and | 2 ) excited by an harmonic force given by (16) that cou-〉 ple states | 1 〉 and | 2 〉 and neglecting the self couplings of the type 〈 1 ˆ 1 | z | 〉 and 〈 2 ˆ 2 . | z | 〉 The full calculation (14) takes into account the non-harmonic excitation given by (12) that couples all states | 1 , 〉 | 2 , 〉 | 3 〉 and | 4 〉 including self couplings. All these complications are potential sources of frequency shifts of the resonance line. We then conclude that these shifts are below the percent level.
## VI. CONCLUSION
In the GRANIT flow through arrangement, two possible modes to induce resonant transitions between the quantum states could be used. In the DC mode, where the magnetic field gradient oscillates spatially along the x direction, the excitation frequency is controlled by the horizontal neutron velocity. In the AC mode, where the gradient is homogeneous in space and oscillates in time, the excitation frequency is selected directly by the frequency of the current driving the magnetic excitation. We have shown that the condition of adiabaticity of spin transport can also be fulfilled in the AC mode using a moderate horizontal magnetic field normal to the neutron propagation axis. Finally a calculation of the expected resonance line for the 2 → 1 transition indicates that a measurement of the transition frequency at a precision better than a percent is possible. A detailed comparison describing the relative merits of the two methods with the associated systematic effects is left for a future work.
## Appendix A: Magnetic field of a square wire
Here we provide formulas for the magnetic field generated by an infinitely long square wire with current I
- [1] V. V. Nesvizhevsky, Physics-Uspekhi 53 (7): 645-675 (2010).
- [2] V. V. Nesvizhevsky et al. , Nature 415 , 297 (2002).
- [3] I. Antoniadis et al. , Compt. Rend. Phys. 12 , 755 (2011).
- [4] P. Brax and G. Pignol, Phys. Rev. Lett 107 , 111301 (2011).
- [5] P. Brax, G. Pignol and D. Roulier Phys. Rev. D 88 , 083004 (2013).
- [6] T. Jenke et al , Phys. Rev. Lett 112 , 151105 (2014).
- [7] T. Jenke, P. Geltenbort, H. Lemmel and H. Abele, Nat. Phys. 7 , 468 (2011).
- [8] V. V. Nesvizhevsky and K.V. Protasov, In: Trends in Quantum Gravity Research, 65-107 Nova Science, New
flowing uniformly in the wire in the y direction. The magnetic field lies in the ( x, z ) plane. We assume that ( x = 0 , y = 0) corresponds to the center of the wire. The length of the square is denoted by c ( c = 1 mm for the purpose of this article). Formulas are valid outside the wire. We define the following quantities
$$\begin{array} { r c l }. \ v e c u e \, \text{me in moving quanuues} \\ \\ x _ { m } \ = x - c / 2, & z _ { m } \ = z - c / 2, \\ x _ { p } \ = x + c / 2, & z _ { p } \ = z + c / 2, \\ L _ { 1 } \ = \ \ln \left ( \frac { ( x _ { m } ^ { 2 } + z _ { m } ^ { 2 } ) ( x _ { p } ^ { 2 } + z _ { p } ^ { 2 } ) } { ( x _ { p } ^ { 2 } + z _ { m } ^ { 2 } ) ( x _ { m } ^ { 2 } + z _ { p } ^ { 2 } ) } \right ), \\ \\ L _ { 2 } \ = \ \ln \left ( \frac { ( x _ { m } ^ { 2 } + z _ { m } ^ { 2 } ) ( x _ { p } ^ { 2 } + z _ { m } ^ { 2 } ) } { ( x _ { m } ^ { 2 } + z _ { p } ^ { 2 } ) ( x _ { p } ^ { 2 } + z _ { p } ^ { 2 } ) } \right ), \\ \\ L _ { 3 } \ = \ \ln \left ( \frac { ( x _ { m } ^ { 2 } + z _ { m } ^ { 2 } ) ( x _ { m } ^ { 2 } + z _ { p } ^ { 2 } ) } { ( x _ { p } ^ { 2 } + z _ { m } ^ { 2 } ) ( x _ { p } ^ { 2 } + z _ { p } ^ { 2 } ) } \right ). \end{array}$$
$$L _ { 3 } \, = \, \ln \left ( \frac { ( x _ { m } ^ { 2 } + z _ { m } ^ { 2 } ) ( x _ { m } ^ { 2 } + z _ { p } ^ { 2 } ) } { ( x _ { p } ^ { 2 } + z _ { m } ^ { 2 } ) ( x _ { p } ^ { 2 } + z _ { p } ^ { 2 } ) } \right ).$$
$$\mathrm e ^ { \quad } A _ { m m } \ = \arctan ( x _ { m } / z _ { m } ), \ A _ { p m } \ = \arctan ( x _ { p } / z _ { m } ), \\ \mathrm e ^ { \quad } A _ { m p } \ = \arctan ( x _ { m } / z _ { p } ), \ A _ { p p } \ = \arctan ( x _ { p } / z _ { p } ).$$
By integrating the Biot-Savart law over the volume of the wire we find the following expressions for the field components
$$B _ { x } ( x, z ) \, = \, - \frac { \mu _ { 0 } I } { 4 \pi c ^ { 2 } } \, \left [ x L _ { 1 } - \frac { c } { 2 } L _ { 2 } +$$
$$\mathbf n \quad & \quad \cdots \quad \cdots \quad \mathbf r \cdots \quad \mathbf r \cdots \quad \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \\ \mathbf d \quad & \quad B _ { z } ( x, z ) \ = \ \frac { \mu _ { 0 } I } { 4 \pi c ^ { 2 } } \, \left [ z L _ { 1 } - \frac { c } { 2 } L _ { 3 } + \quad \cdots \quad \cdots \quad \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \mathbf r \cdots \\ \mathbf a \quad & \quad 2 x _ { m } \, ( A _ { m p } - A _ { m m } ) + 2 x _ { p } \, ( A _ { p m } - A _ { p p } ) \, \right ].$$
$$\begin{array} { c c c } \text{ion.} \\ \text{spin} \\ \text{tag a} \\ \text{trend} \\ \text{tate} \end{array} \quad B _ { x } ( x, z ) \ & = \ - \frac { \mu _ { 0 } I } { 4 \pi c ^ { 2 } } \, \left [ x L _ { 1 } - \frac { c } { 2 } L _ { 2 } + \quad ( A 1 ) \\ \quad 2 z _ { m } \, ( A _ { m m } - A _ { p m } ) + 2 z _ { p } \, ( A _ { p p } - A _ { m p } ) \right ], \\ \quad \text{tend} \\ \quad \text{at a}$$
By taking the derivative we find the following expressions for the gradients
$$\partial _ { z } B _ { x } \, = \, \frac { \mu _ { 0 } I } { 2 \pi c _ { \tau } ^ { 2 } } \, \left [ A _ { m p } - A _ { m m } + A _ { p m } - A _ { p p } \right ], \ ( A 3 )$$
$$\partial _ { z } B _ { z } \, = \, \frac { \mu _ { 0 } I } { 4 \pi c ^ { 2 } } \, L _ { 1 }.$$
York (2006).
- [9] S. Baessler et al. , Compt. Rend. Phys. 12 , 707 (2011).
- [10] M. Kreuz et al. , Nucl. Instr. and Meth. A611 , 326 (2009).
- [11] P. Schmidt-Wellenburg et al. , Nucl. Inst. Meth A 611 , 267 (2009).
- [12] O. Zimmer, F. M. Piegsa and S. N. Ivanov, Phys. Rev. Lett. 107 , 134801 (2011).
- [13] D. Roulier et al , Adv. High En. Phys., this issue.
- [14] C. Caudau et al. , Nucl. Instr. and Meth. A677 , 10 (2012).
|
10.1155/2014/628125
|
[
"G. Pignol",
"S. Baessler",
"V. V. Nesvizhevsky",
"K. Protasov",
"D. Rebreyend",
"A. Yu. Voronin"
] |
2014-08-05T16:02:19+00:00
|
2014-08-05T16:02:19+00:00
|
[
"quant-ph"
] |
Gravitational resonance spectroscopy with an oscillating magnetic field gradient in the GRANIT flow through arrangement
|
Gravitational resonance spectroscopy consists in measuring the energy
spectrum of bouncing ultracold neutrons above a mirror by inducing resonant
transitions between different discrete quantum levels. We discuss how to induce
the resonances with a flow through arrangement in the GRANIT spectrometer,
excited by an oscillating magnetic field gradient. The spectroscopy could be
realized in two distinct modes (so called DC and AC) using the same device to
produce the magnetic excitation. We present calculations demonstrating the
feasibility of the newly proposed AC mode.
|
1408.1010v2
|
## Inverted Singlet-Triplet Qubit Coded on a Two-Electron Double Quantum Dot
Sebastian Mehl 1, 2, ∗ and David P. DiVincenzo 1, 2
1 Peter Grünberg Institute (PGI-2), Forschungszentrum Jülich, D-52425 Jülich, Germany 2 JARA-Institute for Quantum Information, RWTH Aachen University, D-52056 Aachen, Germany (Dated: November 27, 2014)
The s z = 0 spin configuration of two electrons confined at a double quantum dot (DQD) encodes the singlet-triplet qubit (STQ). We introduce the inverted STQ (ISTQ) that emerges from the setup of two quantum dots (QDs) differing significantly in size and out-of-plane magnetic fields. The strongly confined QD has a two-electron singlet ground state, but the weakly confined QD has a two-electron triplet ground state in the s z = 0 subspace. Spin-orbit interactions act nontrivially on the s z = 0 subspace and provide universal control of the ISTQ together with electrostatic manipulations of the charge configuration. GaAs and InAs DQDs can be operated as ISTQs under realistic noise conditions.
## I. INTRODUCTION
Encoded spin qubits in a two-electron configuration have become popular since the seminal experiment by Petta et al. 1 Single electrons are trapped using gate-defined quantum dots (QDs) in semiconducting nanostructures. 2 The spin is used as the information carrier. 3 We consider the qubit encoding using the s z = 0 spin subspace of two electrons. 4-6 The passage between different charge configurations realizes single-qubit control electrostatically. Applying voltages at metallic gates close to the structure enables the transfer of electrons between the QDs. The (1 1) , configuration labels separated electrons on the two QDs; two electrons occupy a single QD in (2 0) , and (0 , 2) .
can be corrected using refocusing techniques. 10,11 In particular, the ISTQ is superior to the two-electron encoding that uses the singlet state | S 〉 and the s z = 1 triplet state | T + 〉 . 12-15 There is also an energy degeneracy of | S 〉 and | T + 〉 in this setup, but hyperfine interactions induce noise with larger weights at higher frequencies. 11
In this paper, we explore a two-electron double quantum dot (DQD) under the influence of magnetic fields and spin-orbit interactions (SOIs). The qubit is encoded in the s z = 0 subspace of two electrons using the singlet | S 〉 and spinless triplet | T 〉 states, similarly to common singlet-triplet qubits (STQs). 4-6 Our setup has an energy degeneracy of | S 〉 and | T 〉 in (1 1) , that is a consequence of the competition between the confining potential and the Coulomb interactions. In the absence of SOIs, the orbital contributions from the out-of-plane magnetic fields favor triplets, while the confining potential favors singlets. We call this qubit inverted STQ (ISTQ) because it differs from normal STQs by the occurrence of a singlettriplet inversion. We realize an ISTQs with one strongly confined QD and one weakly confined QD. | T 〉 is the ground state in s z = 0 for one QD when it is doubly occupied, but the other QD has a singlet ground state. SOIs couple | S 〉 and | T 〉 . In contrast to the setup with two QDs differing significantly in size, it was argued that SOIs act trivially on the s z = 0 subspace for two identical QDs. 7,8
The encoding in the s z = 0 subspace is optimal because the qubit encoding is protected from hyperfine interactions. Nuclear spins generate local magnetic field fluctuations B hyp . Mainly the component B glyph[parallelshort] hyp parallel to the external magnetic field B influences the s z = 0 subspace. 9 Fluctuations in B glyph[parallelshort] hyp are low frequency and
The organization of this paper is as follows. Sec. II introduces the model to construct ISTQs and describes the qubit encoding. Sec. III characterizes SOIs as a source to influence the s z = 0 subspace. We describe different possibilities to manipulate the ISTQ in Sec. IV and discuss its performance in Sec. V.
The main purpose of this paper is to explore the ISTQ encoding. We show that SOIs act nontrivially on the s z = 0 subspace. The influence of SOIs can be described by an effective magnetic-field difference between the QDs. The effective local magnetic field depends on the confining potential of the wave functions. ISTQs are controlled using electrostatic voltages, which tune the DQD between different charge configurations. DQDs that consist of QDs with different sizes realize ISTQs that can be operated in the presence of realistic noise sources. A DQD that is coded using two distinct QDs gives also other perspectives: A strongly confined QD is favorable for the initialization and the readout of STQs. A weakly confined QD may be favorable for qubit manipulations. 16 We are convinced that this setup is likely to be explored as the search for alternative spin qubit designs continues. 17-19 Operating STQs coded using two QDs with different sizes as ISTQs is achieved by applying sufficiently large outof-plane magnetic fields.
## II. MODEL
Our study includes the orbital Hamiltonian H 0 , external magnetic fields H 1 , and SOIs H 2 . The orbital Hamiltonian for two electrons in gate-defined lateral DQDs is described by:
$$\mathcal { H } _ { 0 } = \sum _ { i = 1, 2 } \left [ \frac { \wp _ { i } ^ { 2 } } { 2 m } + V \left ( x _ { i } \right ) \right ] + V \left ( x _ { 1 }, x _ { 2 } \right ). \quad \left ( 1 \right )$$
The orbital contributions of the magnetic-field component perpendicular to the lateral direction (called the zdirection) are included by the kinematic momentum operator ℘ = glyph[planckover2pi1] i ∇ + e A . e > 0 is the electric charge, m is the effective mass, and A = B z 2 ( -y, x, 0) T describes orbital effects from the out-of-plane magnetic-field component B z in the symmetric gauge. Orbital contributions from in-plane magnetic fields are weak for strong confining potentials in the z-direction. V ( x ) is the single-particle potential that includes external electric fields. Two QDs are present at the positions ( ± a, 0 0) , T . V ( x 1 , x 2 ) is the Coulomb interaction. Magnetic fields couple directly to the spins through the Zeeman Hamiltonian:
$$\mathcal { H } _ { 1 } = \frac { g \mu _ { B } } { 2 } B \cdot \sum _ { i = 1, 2 } \sigma _ { i }. \quad \ \ \ ( 2 ) \, \stackrel { \mu \nu _ { 0 } } { \text{grc} } \\ \text{by} \\ \text{out}$$
σ = ( σ , σ x y , σ z ) T is the vector of Pauli matrices, B is the magnetic field, g is the g-factor, and µ B is the Bohr magneton.
glyph[negationslash]
We include two orbitals at each QD: the single-dot ground { | L , 〉 | R 〉 } and the single-dot excited states {∣ ∣ L 〉 , ∣ ∣ R 〉 } . We consider only the s z = 0 subspace, since states with s z = 0 are far away in energy during qubit manipulations. The wave functions of the singlet state (S) and spinless triplet state (T) of different charge configurations ( n , n L R ) are
$$| S _ { 1, 1 } \rangle = \frac { 1 } { \sqrt { 2 } } \left ( c _ { L, \uparrow } ^ { \dagger } c _ { R, \downarrow } ^ { \dagger } - c _ { L, \downarrow } ^ { \dagger } c _ { R, \uparrow } ^ { \dagger } \right ) | 0 \rangle \,, \quad \ \ ( 3 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
$$| S _ { 2, 0 / 0, 2 } \rangle & = \left ( c _ { L / R, \uparrow } ^ { \dagger } c _ { L / R, \downarrow } ^ { \dagger } \right ) | 0 \rangle \,, & ( 4 ) & \quad \Omega _ { ( 2, 0 ) }$$
$$| T _ { 1, 1 } \rangle = \frac { \stackrel { \cdot } { 1 } } { \sqrt { 2 } } \left ( c _ { L, \uparrow } ^ { \dagger } c _ { R, \downarrow } ^ { \dagger } + c _ { L, \downarrow } ^ { \dagger } c _ { R, \uparrow } ^ { \dagger } \right ) | 0 \rangle \,, \quad \ \ ( 5 ) \quad \text{are t}$$
$$| T _ { 2, 0 / 0, 2 } \rangle = \frac { \ddot { 1 } } { \sqrt { 2 } } \left ( c _ { L / R, \uparrow } ^ { \dagger } c _ { \overline { L } / \overline { R }, \downarrow } ^ { \dagger } + c _ { L / R, \downarrow } ^ { \dagger } c _ { \overline { L } / \overline { R }, \uparrow } ^ { \dagger } \right ) | 0 \rangle \,, \quad \text{the s}$$
where | 0 〉 is the vacuum state and c † iσ is the creation operator of an electron in orbital i with spin σ . We use a Hubbard model to describe the (2 , 0) , (1 1) , , and (0 2) , configurations. 20,21 The electrons are on separate QDs in (1 , 1) . The orbital ground states are filled with two electrons for the singlets | S 2 0 , 〉 and | S 0 2 , 〉 ; the Pauli exclusion principle requires that electrons fill different orbitals for | T 2 0 , 〉 and | T 0 2 , 〉 . Orbital effects of H 0 and H 1 are described by
$$\mathcal { H } & = \left ( \begin{array} { c c c c } 0 & t _ { s } ^ { L } & t _ { s } ^ { R } \\ t _ { s } ^ { L } & U _ { L } + \epsilon & 0 \\ t _ { s } ^ { R } & 0 & U _ { R } + \Omega _ { ( 0, 2 ) } - \epsilon \end{array} \right ) & \text{$QD_{L}$} \\ & \oplus \left ( \begin{array} { c c c c } 0 & t _ { t } ^ { L } & t _ { t } ^ { R } \\ t _ { t } ^ { L } & U _ { L } + \Omega _ { ( 2, 0 ) } + \epsilon & 0 \\ t _ { t } ^ { R } & 0 & U _ { R } - \epsilon \end{array} \right ). \quad ( 7 ) & \text{$of$-pli$} \\ & \text{is rot} \\ & \text{the r} \\ \end{array} \right ) \quad \text{$\sim} \quad. \quad \dots \quad. \quad. \quad.$$
Eq. (7) is written in the basis {| S 1 1 , 〉 , | S 2 0 , 〉 , | S 0 2 , 〉 , | T 1 1 , 〉 , | T 2 0 , 〉 , | T 0 2 , 〉} . The real constants t L,R s,t characterize the spin-conserving hopping processes of electrons from (1 1) , towards two electrons on the same QD. The relative energies of (2 0) , , (1 1) , , and (0 , 2) are tunable by voltages at gates near the left and right QD; we model their influence as a modification of the addition energies U L → U L + glyph[epsilon1] and U R → U R -glyph[epsilon1] . The left QD is doubly occupied for glyph[epsilon1] →-∞ (and, similarly, the right QD for glyph[epsilon1] →∞ ). The electrons are separated on different QDs for glyph[epsilon1] ∼ 0 .
As above, one needs to overcome the charging energies U L of the left QD or U R of the right QD to add two electrons to the same QD. One QD (e.g., QD L ) is in the normal configuration and has a singlet ground state, but | T 0 2 , 〉 is the ground state of QD R . The singlet is the ground state in the absence of magnetic fields. 22 Doubly occupied QDs with E T < E S are obtained at finite out-of-plane magnetic fields also for s z = 0 . 23,24 Finite values of B z decrease the sizes of the orbital wave functions and raise the Coulomb repulsions between the electrons. Electrons prefer to minimize the Coulomb repulsion, which makes triplets favorable. The inversion from a singlet to a triplet ground state was experimentally detected at B z = 1 5 . T in elongated GaAs QDs. 25 A theoretical study predicts an orbital singlet-triplet inversion at B z = 0 5 . T in weakly confined, circular GaAs QDs. 8 However, ISTQs only require a triplet ground state for one of the two QDs, which is realized for one strongly confined QD and one weakly confined QD (cf. Fig. 1). 26 U L +Ω (2 0) , and U R +Ω (0 2) , are the energies to reach the first excited, doubly occupied states,
$$\Omega _ { ( 2, 0 ) } = E _ { T _ { 2, 0 } } - E _ { S _ { 2, 0 } } > 0, \ \Omega _ { ( 0, 2 ) } = E _ { S _ { 0, 2 } } - E _ { T _ { 0, 2 } } > 0,$$
are the energy differences of the doubly occupied states. We neglect matrix elements between (2 0) , and (0 2) , of the same spin 20 because their contributions are weak.
Figure 1. STQ coded on an asymmetric DQD. Each QD contains one electron. Modifications of the confining potentials allow an electron transfer to reach a doubly occupied QD. QD L has a two-electron singlet ground state; for QD R the spinless triplet state is lower in energy. The out-of-plane component B z of the magnetic field favors triplets, but the confining energy favors singlets. The magnetic field B is tilted by the angle φ from the dot-connection axis e x and by the outof-plane angle θ from e z . The [1 , 0 0] , -direction of the lattice is rotated by the angle ξ from e x . We introduce additionally the rotation angle ρ between B and the [1 , 0 0] , -direction.
Fig. 2 shows a typical energy diagram of the s z = 0
subspace in the charge configurations (2 0) , , (1 , 1) , and (0 , 2) . Eq. (7) describes a state crossing of the singlet | S 〉 and s z = 0 triplet state | T 〉 . | S 2 0 , 〉 is the ground state deep in (2 0) , , while | T 0 2 , 〉 is the ground state in (0 2) , . | S 1 1 , 〉 and | S 2 0 , 〉 have the same orbital energies at glyph[epsilon1] = -U L ; | T 1 1 , 〉 and | T 0 2 , 〉 are at equal energies at glyph[epsilon1] = U R . Similarly, there is a state degeneracy of | T 1 1 , 〉 and | T 2 0 , 〉 at glyph[epsilon1] = -( U L +Ω (2 0) , ) . | S 1 1 , 〉 and | S 0 2 , 〉 have the same energy at glyph[epsilon1] = U R +Ω (0 2) , . Electron tunneling between the QDs hybridizes states of different charge configurations. The singlet ground state | S 〉 is degenerate with s z = 0 triplet state | T 〉 in (1 1) , because the tunnel couplings t L,R s,t are smaller than U L , U R , Ω (2 0) , , and Ω (0 2) , . We label this point glyph[epsilon1] ∗ . The next section describes SOIs, which couple | S 〉 and | T 〉 by ∆ so at glyph[epsilon1] ∗ .
Figure 2. Energy diagram of a STQ as a function of the electrostatic bias glyph[epsilon1] according to Eq. (7) and Eq. (9). The blue and red lines describe the energies of the lowest singlet E S and spinless triplet E T ; black lines show excited states. The left dotted line labels the charge transition point at glyph[epsilon1] = -U L , where E S 1 1 , and E S 2 0 , have the same energies (similarly E T 1 1 , and E T 0 2 , have equal energies at glyph[epsilon1] = U R ). We obtain a (2 0) , singlet ground state at glyph[epsilon1] < 0 , while glyph[epsilon1] > 0 favors the (0 2) , triplet. E S and E T cross at glyph[epsilon1] ∗ . SOIs couple E S and E T . The inset shows the region around glyph[epsilon1] ∗ . The dashed curves are energy levels in the absence of SOIs.
## III. CALCULATION OF ∆ so
We consider QDs fabricated in the crystal's (0 0 1) , , plane. The strong confining potential in the z-direction causes interactions between the electron spins and the in-plane momentum components. SOIs are described by:
$$\mu _ { 1 } & = & \frac { \alpha } { \hslash } \sum _ { i = 1, 2 } [ \sigma _ { x ^ { \prime } } \wp _ { y ^ { \prime } } - \sigma _ { y ^ { \prime } } \wp _ { x ^ { \prime } } ] _ { i } & & \text{are} \, \\ & \quad + \frac { \beta } { \hslash } \sum _ { i = 1, 2 } [ - \sigma _ { x ^ { \prime } } \wp _ { x ^ { \prime } } + \sigma _ { y ^ { \prime } } \wp _ { y ^ { \prime } } ] _ { i } \,. \quad \text{and} \, ( 9 ) \quad \text{the} \, \quad \text{posi} \\ & \quad - \langle L$$
The first term, which is called the Rashba SOI, 27 is caused by the broken structure inversion symmetry from the confining potential in the z-direction. 28 The second term, called the Dresselhaus SOI, 27 is present for a crystal lattice without inversion symmetry. 29 x ′ and y ′ label the [1 , 0 0] , -direction and [0 , 1 0] , -direction of the lattice. [1 , 0 0] , is rotated by the angle ξ from e x , which is the vector connecting the QD centers (cf. Fig. 1). Large spin-orbit (SO) effects are expected when electrons are free to move, which is possible between the QDs in the e x -direction. We consider only the SO contributions that involve the momentum component in the e x -direction ( ℘ x ) and extract from Eq. (9) ˜ H 2 = Ξ glyph[planckover2pi1] · ∑ i =1 2 , [ ℘ x σ ] i , with Ξ = ( -β cos (2 ξ ) , -α -β cos (2 ξ ) , 0) T . Additional contributions from the in-plane momentum component perpendicular to e x are discussed in Appx. A.
H 0 from Eq. (1) dominates over the SO contributions. We apply a unitary transformation U = e i ( S 1 + S 2 ) , with S i = mx i glyph[planckover2pi1] 2 Ξ · σ i . 30-32 U was introduced to remove SOIs to second order for confined systems. This transformation turns out to be useful because the transformed Hamiltonian is only position dependent. Note that the equivalent transformation was used in Refs. [7,8] to show that SOIs act trivially on the s z = 0 subspace for a highly symmetric DQD. The transformed Hamiltonian reads:
$$\Big \| _ { \infty } \quad \begin{array} { c } \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \text{$\dots } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \, \stackrel { \dots } { \sim } \,
\infty \quad \begin{array} { c } \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \end{array} } \Big \| _ { \infty } \quad \begin{array} { c } \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \end{array} } \Big \| _{ \infty } \quad \begin{array} { c } \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \\ \mathcal { U } \left ( \mathcal { H } _ { 0 } + \mathcal { H } _ { 1 } + \widetilde { \mathcal { H } _ { 2 } \right ) U ^ { \dagger } = \mathcal { H } _ { 0 } - \frac { m } { \hbar { 2 } ^ { m } } \left | \Xi ^ { 2 } \right | ^ { 2 } \\ \mathcal { U } \left ( \mathcal { H } _ { 1 } + \widetilde { \mathcal { H } _ { 3 } \right ) \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \dots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \quad \end{array} \Big \| _ { \infty } \quad \begin{array} { c } \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \end{array} \Big \| _ { \infty } \quad \begin{array} { c } \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \end{array} \Big \| _ { \infty } \quad \begin{array} { c } \dots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \end{c } \Big \| _ { \infty } \quad \begin{array} { c } \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \\ \mathcal { U } \left ( \mathcal { H } _ { 1 } + \widetilde { \mathcal { H } _ { 2 } \right ) U ^ { \dagger } = \mathcal { H } _ { 0 } - \frac { m } { \hbar { 2 } ^ { m } } \left | \Xi ^ { 2 } \right | ^ { 2 } \right | ^ { 2 } \left | \Xi ^ { 2 } \right | ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 r } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \\ \left | \Xi ^ { 2 } \left | \Xi ^ { 2 } \left | \Xi ^ { 2$$
$$\underset { \underset { \underset { \underset { \underset { \underset { \underset { \text{e}} { v } } } { \text{e}} } } } { \ e } } B _ { \text{eff} } ^ { [ j ] } ( x ) \equiv \frac { 1 } { j! } \left ( \frac { 2 m } { \hslash ^ { 2 } } x \right ) ^ { j } [ ( \dots ( \begin{array} { c } B \underbrace { \times \Xi ) \times \dots ) \times \Xi } _ { j \text{ times} } \end{array} ].$$
$$\text{$i$formed Hamiltonian reads} \quad \\ \Big ) \mathcal { U } ^ { \dagger } = & \mathcal { H } _ { 0 } - \frac { m } { \hslash ^ { 2 } } \left | \Xi \right | ^ { 2 } \quad ( 1 0 ) \\ & + \frac { g \mu _ { B } } { 2 } \sum _ { \substack { i = 1, 2 \\ j \in \mathbb { N } } } B _ { \text{eff} } ^ { [ j ] } ( x _ { i } ) \cdot \sigma _ { i }, \\ \frac { m } { \hslash ^ { 2 } } x \right ) ^ { j } \left [ ( \dots ( B \underbrace { \times \Xi ) \times \dots ) \times \Xi } _ { j \text{ times} } \right ]. \\ \Big | y \text{ unchanged}. \ Besides the constant}$$
H 0 remains formally unchanged. Besides the constant energy shift -m glyph[planckover2pi1] 2 | Ξ | 2 , there are only position dependent terms (note the restriction to the x -direction). Eq. (10) couples only states of the same charge sector because the orbital states are strongly confined at the QD's position. We restrict the discussion to the contribution in (1 1) , . Contributions from (2 , 0) and (0 , 2) are negligible, as described in Appx. B. The charge configuration is confined to a small area compared to the SO scale ( glyph[planckover2pi1] 2 2 m | Ξ | ) 2 , with the result that terms in Eq. (11) with higher order in j are less important.
The external magnetic field is rotated by the polar angle θ from the [0 , 0 1] , -direction and the azimuthal angle φ from e x (cf. Fig. 1). We fix the spin quantization axis parallel to B . The components of Eq. (11) that are parallel to the external magnetic field ( B [ j ] eff ) glyph[parallelshort] couple | S 1 1 , 〉 and | T 1 1 , 〉 , while the perpendicular components couple subspaces of different s z . We assume that the states | L 〉 and | R 〉 are strongly confined at the QD position, with 〈 L x R | | 〉 = 〈 L x ∣ ∣ 2 ∣ ∣ R 〉 = 0 , 〈 R x R | | 〉 = -〈 L x L | | 〉 = a . We introduce the variances of the or-
bitals 〈 L ∣ ∣ ∣ ( x -a ) 2 ∣ ∣ ∣ L 〉 = var L and 〈 R ∣ ∣ ∣ ( x + ) a 2 ∣ ∣ ∣ R 〉 = var R . Note that the transformation U in Eq. (10) modifies also the definitions of the basis states | L 〉 and | R 〉 . The effective Hamiltonian in (1 , 1) , including SOIs to second order, is written in the basis | S 1 1 , 〉 from Eq. (3), | T 1 1 , 〉 from Eq. (5), ∣ ∣ T + 1 1 , 〉 = c † L ↑ c † R ↑ | 0 〉 , and ∣ ∣ T -1 1 , 〉 = c † L ↓ c † R ↓ | 0 〉 :
$$\mathcal { H } _ { ( 1, 1 ) } & = E _ { Z } \left ( \frac { 0 } { h ^ { 2 } } \left ( \frac { 0 } { 0 } \right ) \right ) - i \sqrt { 2 } E _ { Z } \frac { 2 m a } { h ^ { 2 } } \Xi _ { \perp } \left ( \frac { 0 } { h ^ { 2 } } \left ( \frac { 0 } { 0 } \right ) \right ) - E _ { \perp } \left ( 2 m \right ) ^ { 2 } \Xi _ { \perp } ^ { 2 } \left ( \tilde { \tau } _ { L } - \tilde { \tau } _ { R } \right ) \tilde { \tau } _ { R } \right ) \left ( \frac { 0 } { h ^ { 2 } } \left ( \frac { 1 } { 0 } \right ) \right ) \\ + \sqrt { 2 } E _ { Z } \left ( \frac { 2 m } { h ^ { 2 } } \right ) ^ { 2 } \Xi _ { \perp } \Xi _ { \perp } \times \left ( \frac { 0 } { h ^ { 2 } } \right ) \right ) - \frac { \tilde { \tau } _ { L } - \tilde { \tau } _ { R } } { 2 } \frac { \tilde { \tau } _ { L } + \tilde { \tau } _ { R } } { 2 } - 2 a ^ { 2 } \right ) \left ( \frac { \tilde { \tau } _ { L } + \tilde { \tau } _ { R } } { 2 } - 3 a ^ { 2 } \right ) } { \frac { \tilde { \tau } _ { L } + \tilde { \tau } _ { R } } { 2 } \frac { \tilde { \tau } _ { L } + \tilde { \tau } _ { R } } { 2 } - 3 a ^ { 2 } \right ) } \right ), \quad \text{(12)} \\ & \coprod$$
with the Zeeman energy E Z = gµ B 2 | B | . Ξ glyph[parallelshort] = | Ξ | cos [ glyph[measuredangle] ( Ξ , B )] is the component of Ξ parallel to B , and Ξ ⊥ = | Ξ | sin [ glyph[measuredangle] ( Ξ , B )] is the component of Ξ perpendicular to B [all components are determined by the angle glyph[measuredangle] ( Ξ , B ) between the vectors Ξ and B , cf. Fig. 1].
The first term in Eq. (12) represents the Zeeman interaction that shifts ∣ ∣ T + 1 1 , 〉 and ∣ ∣ T -1 1 , 〉 relative to the s z = 0 energy levels. This term dominates over all SO contributions. The second term in Eq. (12) couples | S 1 1 , 〉 with ∣ ∣ T + 1 1 , 〉 and | S 1 1 , 〉 with ∣ ∣ T -1 1 , 〉 . This term was discussed in great detail in Refs. [7,8]. It does not couple | S 1 1 , 〉 and | T 1 1 , 〉 . Note that the coupling to the triplet states does not cause an energy shift of | S 1 1 , 〉 in second-order Schrieffer-Wolff perturbation theory 33 because the couplings between | S 1 1 , 〉 and ∣ ∣ T + 1 1 , 〉 and between | S 1 1 , 〉 and ∣ ∣ T -1 1 , 〉 cancel each other.
The dominant SO contribution on the s z = 0 subspace is obtained from the third term of Eq. (12). This term represents the component of the effective magnetic field parallel to B , which is second order in the SOI: ( B [2] eff ) glyph[parallelshort] =
-B 2 ( 2 m glyph[planckover2pi1] 2 ) 2 Ξ 2 ⊥ x 2 [cf. Eq. (11)]. ( B [2] eff ) glyph[parallelshort] realizes a direct coupling between | S 1 1 , 〉 and | T 1 1 , 〉 : ∆ so ≈ E Z var R -var L l 2 so . We introduce the length scale l so = glyph[planckover2pi1] 2 2 m | Ξ ⊥ | . The fourth term in Eq. (12) gives small corrections to ∆ so . Appx. A describes the angular dependency of ∆ so and extends the analysis of SOIs using all terms of Eq. (9).
The smallest possible values for l so are on the order of the Rashba and Dresselhaus spin precession lengths l α so and l β so . Typically, GaAs heterostructures have spin precession lengths l α so , l β so glyph[greaterorsimilar] 1 µ m for the Rashba and the Dresselhaus SOIs (cf. Appx. C). The variances of the orbital wave functions can be approximated using the noninteracting descriptions of electrons that are confined at QDs. The Fock-Darwin states are the solutions of the noninteracting eigenvalue problem of two-dimensional circular QDs. 34,35 The variances of these wave functions are directly related to the confining potentials as var ≈ l 2 0 , when assuming a harmonic confining potential that has
$$\stackrel { - r } { y } \, \text{the} \, \underset { \text{ig. 1} } { \text{the magnitude} } \, \hbar { \omega } _ { 0 } \, \text{with} \, l _ { 0 } = \sqrt { \frac { \hbar } { m \omega _ { 0 } } } \left [ 1 + \left ( \frac { e B _ { z } } { 2 m \omega _ { 0 } } \right ) ^ { 2 } \right ] ^ { - 1 / 4 }.$$
Normal values for strongly confined QDs in GaAs are glyph[planckover2pi1] ω 0 = 3 meV and l 0 = 20 nm. 20 Weakly confined QDs in GaAs of glyph[planckover2pi1] ω 0 = 0 1 . meV have l 0 = 100 nm. We obtain, for l so = 1 µ m and B = 500 mT, ∆ so = 0 1 . µ eV ( ∆ so /h ≈ 25 MHz).
Small-band-gap materials tend to have stronger SOIs. SOIs are, for example, by one order of magnitude larger in InAs than in GaAs ( l α so = 1 1 . µ m for GaAs and l α so = 0 14 . µ m for InAs; cf. Appx. C). Furthermore, the variances of the wave functions of InAs QDs are potentially larger than of GaAs QDs due to the smaller effective mass. It should therefore be possible to reach values of ∆ so ≈ 1 µ eV ( ∆ so /h ≈ 250 MHz).
The coupling between | S 〉 and | T 〉 at glyph[epsilon1] ∗ can be approximated by ∆ so ≈ E Z var R -var L l 2 so , as one can see from Eq. (7). The state coupling is determined by the weights of | S 1 1 , 〉 in | S 〉 and | T 1 1 , 〉 in | T 〉 at glyph[epsilon1] ∗ . glyph[epsilon1] ∗ is close to the center of (1 1) , because t L,R s,t are much smaller than U L , U R , Ω (2 0) , , and Ω (0 2) , . Therefore | S 1 1 , 〉 and | T 1 1 , 〉 have weights close to unity.
In summary, SOIs couple | S 〉 and | T 〉 via their state contributions in (1 1) , . There is a second-order coupling through SOIs, describing an effective magnetic field parallel to the external magnetic field B glyph[parallelshort] eff at the QDs. The magnitude of B glyph[parallelshort] eff depends on the sizes of the wave functions. ∆ so is caused by an effective magnetic-field gradient across the DQDs generated from SOIs.
## IV. QUBIT MANIPULATIONS
An ISTQ encodes a qubit similar to a normal STQ. We identify the singlet state | S 〉 with the logical '1' and the s z = 0 triplet state | T 〉 with the logical '0'. Pauli operators are used to describe interactions on the qubit subspace: From this point onward, σ x = | S 〉 〈 T | + | T 〉 〈 S | , σ y = - | i S 〉 〈 T | + i | T 〉 〈 S | , and σ z = | S 〉 〈 S | - | T 〉 〈 T | . A complete set of single-qubit gates together with one maximally entangling two-qubit gate are convenient for universal quantum computation. 36 Fig. 3 shows an energy diagram of the qubit levels as a function of the bias parameter glyph[epsilon1] , which is extracted from Fig. 2. We identify three points that are favorable for qubit manipulations. The qubit states are coupled by a transverse Hamiltonian H glyph[epsilon1] ∗ = ∆ so σ x at glyph[epsilon1] ∗ . | S 〉 and | T 〉 are energy eigenstates far from the anticrossing. We label one point in (2 0) , as glyph[epsilon1] (2 0) , with H glyph[epsilon1] (2 0) , = -Ω (2 0) , σ z [and, similarly, glyph[epsilon1] (0 2) , in (0 , 2) with H glyph[epsilon1] (0 2) , = Ω (0 2) , σ z ].
Figure 3. Sketch of the energy levels | S 〉 and | T 〉 that encode the ISTQ. The energy levels are shifted compared to Fig. 2, while the energy differences between E S and E T remain unchanged at each glyph[epsilon1] . | S 〉 and | T 〉 have equal orbital energies at glyph[epsilon1] ∗ . SOIs lift the degeneracy and cause an anticrossing ∆ so . | S 〉 is the ground state for glyph[epsilon1] < glyph[epsilon1] ∗ , but | S 〉 is the excited state for glyph[epsilon1] > glyph[epsilon1] ∗ . We label one point deep in (2 , 0) by glyph[epsilon1] (2 0) , with the energy splitting Ω (2 0) , [similarly, | S 〉 and | T 〉 have the energy splitting Ω (0 2) , at glyph[epsilon1] (0 2) , in (0 2) , ].
## A. Single-Qubit Gates
The ISTQ provides different approaches for singlequbit manipulations. The effective Hamiltonian on the qubit subspace can be tuned using electric gates. Gate manipulations rotate the direction of an effective magnetic field. A magnetic field in the z -direction is applied at glyph[epsilon1] (2 0) , in (2 0) , and glyph[epsilon1] (0 2) , in (0 , 2) . Ω (2 0) , and Ω (0 2) , correspond to the energy differences of | S 〉 and | T 〉 . The effective magnetic-field direction is tilted to the x -axis in (1 1) , . It points exactly along e x at glyph[epsilon1] ∗ and has a magnitude ∆ so . Rotations around the z -axis and x -axis can be generated when the qubit is tuned fast between glyph[epsilon1] (2 0) , , glyph[epsilon1] (0 2) , , and glyph[epsilon1] ∗ . The qubit manipulation time τ must be diabatic with the SOI, but adiabatic to the orbital Hamiltonian: h/ ∆ so glyph[greatermuch] τ glyph[greatermuch] h/ Ω (2 0) , , h/ Ω (0 2) , . 37 The time scale of single-qubit gates is determined by h/ Ω (2 0) , , h/ Ω (0 2) , , and h/ ∆ so ; it should be in the range of 10 MHz to a few GHz. Larger values make the gates too fast to be controlled by electronics. Smaller values require long gate times.
Rabi driving becomes impractical for small ∆ so because the gate times increase. We propose another possibility of driven gates that are described by the LandauZener (LZ) model. 13,40,41 Traversing the anticrossing in a time similar to τ = h/ ∆ so generates single-qubit rotations. For large transition amplitudes, as for the sweep from glyph[epsilon1] (2 0) , to glyph[epsilon1] (0 2) , , the time evolution, 13,40,41
We describe two other possibilities for single-qubit control that are practical if ∆ so is either very large or very small. A large value of ∆ so permits resonant Rabi driving, which has already been successful for a qubit encoded in triple QDs. 38,39 The effective Hamiltonian at glyph[epsilon1] ∗ is H = Ω( ) glyph[epsilon1] σ z + ∆ so σ x . Transitions are driven by Ω( ) = 2Ω glyph[epsilon1] 0 cos (2∆ so t/ glyph[planckover2pi1] + ψ ) . If Ω 0 glyph[lessmuch] ∆ so , then one obtains after a rotating wave approximation the static Hamiltonian H ′ = Ω [ 0 -σ y sin ( ψ ) + σ z cos ( ψ )] . A universal set of single-qubit gates can be generated when the phase ψ is adjusted.
$$\mathcal { U } _ { L Z } = e ^ { - i \zeta _ { R } \sigma _ { z } } e ^ { - i \gamma \sigma _ { y } } e ^ { - i \zeta _ { L } \sigma _ { z } }, \text{ \quad \ \ } ( 1 3 )$$
is decomposed into phase accumulations (through ζ R and ζ L ) and one rotation around an orthogonal axis. The phase accumulations ζ R and ζ L are determined by the adiabatic evolution under the energy splitting Ω( ) t σ z and the Stückelberg phase. The essential part is the rotation around the y -axis by the angle γ = γ LZ + π/ 2 , with sin ( γ LZ ) = √ P LZ , P LZ = e -2∆ 2 so glyph[planckover2pi1] v . v = dE/dt | glyph[epsilon1] ∗ is the linearized velocity at glyph[epsilon1] ∗ . For example, the state | 0 〉 is transferred to an equal superposition of | 0 〉 and | 1 〉 for P LZ = 1 2 .
## B. Two-Qubit Gates
Two-qubit gates can be realized using Coulomb interactions between two ISTQs. 5 We consider a linear arrangement of four QDs and label the two DQDs by ( L ) and ( R ) (cf. Fig. 4). QD ( L ) R and QD ( R ) L are closest to each other, and the electron configurations n ( L ) R at QD ( L ) R and n ( R ) L at QD ( R ) L dominate the Coulomb coupling between the ISTQs: 39,42 H int = e 2 4 πglyph[epsilon1] 0 glyph[epsilon1] r d n ( L ) R n ( R ) L . d is the distance between QD ( L ) R and QD ( R ) L , glyph[epsilon1] 0 is the dielectric constant, and glyph[epsilon1] r is the relative permittivity. H int leaves the spin at ISTQ ( L ) and the spin at ISTQ ( R ) unchanged and can
only cause the effective interaction C σ (1) z σ (2) z up to local energy shifts. 43 C has finite values only when ∣ ∣ S ( L ) 〉 has a charge configuration that differs from that of ∣ ∣ T ( L ) 〉 and ∣ ∣ S ( R ) 〉 has a charge configuration that differs from that of ∣ ∣ T ( R ) 〉 [cf. Eq. (7)]:
$$\mathcal { C } = & \frac { e ^ { 2 } } { 1 6 \pi \epsilon _ { 0 } \epsilon _ { r } d } \left [ \left \langle S ^ { ( L ) } \left | n _ { R } ^ { ( L ) } \right | S ^ { ( L ) } \right \rangle - \left \langle T ^ { ( L ) } \left | n _ { R } ^ { ( L ) } \right | T ^ { ( L ) } \right \rangle \right ] \quad \text{thos} \\ & \times \left [ \left \langle S ^ { ( R ) } \left | n _ { L } ^ { ( R ) } \right | S ^ { ( R ) } \right \rangle - \left \langle T ^ { ( R ) } \left | n _ { L } ^ { ( R ) } \right | T ^ { ( R ) } \right \rangle \right ]. \ ( 1 4 ) \quad \text{tain} \quad \text{on} \quad \text{th}.$$
We discuss C , with STQ ( L ) and STQ ( R ) at glyph[epsilon1] ∗ , as an example. QD ( L ) R has a higher occupation in ∣ ∣ T ( L ) 〉 than in ∣ ∣ S ( L ) 〉 because the doubly occupied triplet in (0 2) , ( L ) is favored over the doubly occupied singlet. The opposite effect is true for QD ( R ) L , with a higher electron configuration at QD ( R ) L for ∣ ∣ S ( R ) 〉 than for ∣ ∣ T ( R ) 〉 . The magnitude of C strongly depends on the material and the DQD setup. Two electrons with the distance d = 200 nm interact with C ≈ 100 µ eV for GaAs and InAs heterostructures ( glyph[epsilon1] r = 12 9 . for GaAs and glyph[epsilon1] r = 15 2 . for InAs 44 ). C is by orders of magnitudes smaller for ISTQs. We assume that C / ∆ so = 1 10 can be reached.
Figure 4. Two DQDs [labeled by ( L ) and ( R ) ] encode two ISTQs, which are coupled using Coulomb interactions. QD ( L ) R and QD ( R ) L are closest to each other and the electron configurations at these QDs ( n ( L ) R and n ( R ) L ) dominate the interaction between the qubits [cf. Eq. (14)].
We construct an entangling gate for ISTQs that is similar to common STQs. 45 Both STQs are pulsed to the transition region of (1 , 1) and (0 , 2) with an effective Hamiltonian H = Ω ( L ) σ ( L ) z + Ω ( R ) σ ( R ) z + C σ ( L ) z σ ( R ) z . A CPHASE gate is generated after the waiting time t = h 8 C . This description is valid away from glyph[epsilon1] ∗ . Directly at glyph[epsilon1] ∗ , driven entangling operations are permitted through the Hamiltonian H = Ω ( L ) σ ( L ) z + ∆ ( L ) so σ ( L ) x + Ω ( R ) σ ( R ) z + ∆ ( R ) so σ ( R ) x + C σ ( L ) z σ ( R ) z . For ∣ ∣ ∣ ∆ ( L ) so -∆ ( R ) so ∣ ∣ ∣ glyph[greatermuch] C , one possible two-qubit gate is obtained when qubit ( L ) is driven with the frequency 2∆ ( R ) so /h . These driven gates are popular for superconducting qubits. 46-49 The requirement is again that ∆ ( L ) so and ∆ ( R ) so reach magnitudes of µ eV to obtain fast gate operations.
## V. DISCUSSION AND CONCLUSION
An ISTQ with a finite ∆ so provides universal control of the s z = 0 subspace. Operations mainly at glyph[epsilon1] ∗ in (1 , 1) and glyph[epsilon1] (0 2) , in (0 , 2) are very favorable because the qubit is protected from small fluctuations in glyph[epsilon1] . Ω (0 2) , /h in (0 , 2) should not exceed a few GHz to control phase accumulations at glyph[epsilon1] (0 2) , . Note that the out-of-plane magnetic-field component B z determines the magnitude of Ω (0 2) , . Obtaining large ∆ so is most critical. The size of ∆ so depends on the confining energies of the QDs and the magnitude of the SOIs. Values of ∆ so ≈ µ eV will be needed for driven Rabi gates. We showed that these magnitudes are obtained for two QDs differing strongly in size. This setup is also promising due to other reasons. Strongly confined QDs are ideal for the initialization and readout of STQs. A weakly confined QD can be very useful for qubit manipulations (cf. also Ref. [16]).
Hyperfine interactions influence qubits in the ISTQ encodings. Nuclear spins couple to the electrons that are confined at QDs by creating local magnetic-field fluctuations B hyp . B hyp can be considered as static during one measurement because the nuclear magnetic-field fluctuations are low frequency, but B hyp gives random contributions between successive measurements. 9,11 An approximation for the rms of the component parallel to the external magnetic field that couples to an electron at a QD is δB glyph[parallelshort] hyp ( QD ) = ∑ ν B ν √ I ν ( I ν +1) √ N . 37,50 ν labels the different nuclear spin isotopes of the semiconductor, which have the spin I . B contains material-dependent coupling constants of the isotope, and N is the number of nuclei interacting with an electron that is confined at a QD. If QD L and QD R have different components of B hyp parallel to the external magnetic field, then for ISTQs the states | S 1 1 , 〉 from Eq. (3) and | T 1 1 , 〉 from Eq. (5) are coupled equivalently to ∆ so by gµ B [ B glyph[parallelshort] hyp ( QD L ) -B glyph[parallelshort] hyp ( QD R ) ] . In the analysis of many measurements, the rms values of the fluctuations δB glyph[parallelshort] hyp ( QD L ) and δB glyph[parallelshort] hyp ( QD R ) will be detected when assuming independent fluctuations of the magnetic fields at QD L and QD R . We arrive at an effective coupling element between | S 1 1 , 〉 and | T 1 1 , 〉 :
$$\underset { \text{to} \\ \text{to} \\ \text{ve} \\ \text{.}$$
An electron at a GaAs QD typically interacts with 10 6 nuclear spins, which gives δB glyph[parallelshort] hyp ( QD L ) = δB glyph[parallelshort] hyp ( QD R ) ≈ 5 mT and ∆ hyp ≈ 100 neV for a symmetric DQD. 51
A weakly confined QD has a smaller uncertainty δB glyph[parallelshort] hyp ( QD because the electron wave function interacts ) with more nuclear spins. This size effect will, however, not affect ∆ hyp significantly for the ISTQ, because the size of only one of the QDs will increase compared to a normal DQD setup, and δB glyph[parallelshort] hyp ( QD ) changes only by N -1 2 . InAs QDs have larger ∆ hyp than GaAs QDs.
Indium isotopes are spin-9/2 nuclei, in contrast to Ga and As nuclei that are spin-3/2. Because of the equivalent influences of hyperfine interactions and SOIs, ∆ so should be significantly larger than ∆ hyp to allow high fidelity qubit gates. Our estimates of ∆ so = 1 µ eV and C / ∆ so = 1 10 suggest that ∆ hyp and C have the same order of magnitude for uncorrected nuclear magnetic fields. Fortunately, many methods are known to reduce the uncertainty of the nuclear magnetic-field distributions by orders of magnitude. 52,53 Additionally, refocusing techniques can be applied to correct for small ∆ hyp because the magnetic field fluctuations are low frequency. 11
Charge noise is another source of decoherence. The filling and unfilling of charge traps cause fluctuating electric fields at the positions of the DQDs. If the qubit is operated as a charge qubit, then charge noise dephases the ISTQ. 9,16,54 Charge fluctuations are dominantly low frequency and lead typically to energy shifts δE C = µeV between different charge states. 55,56 The phase coherences between charge states are lost within a few ns. The most significant influence of charge noise can be described by small fluctuations in glyph[epsilon1] . 56 Charge noise is less important at glyph[epsilon1] ∗ , glyph[epsilon1] (2 0) , , and glyph[epsilon1] (0 2) , because small fluctuations in glyph[epsilon1] do not dephase the qubit.
fluctuations in ∆ hyp are low frequency. Charge noise dephases the qubit in the transition region between different charge sectors. Charge noise will be dealt with most efficiently if the ISTQ is operated only at glyph[epsilon1] ∗ and deep in (0 2) , . All qubit operations require fast manipulation periods between different charge configurations, which has been achieved in previous experiments. 17,18 Motivated by the search for alternative spin qubit designs, 17-19 we are hopeful that DQDs are explored where the QDs differ significantly in size. Realizing an ISTQ in a DQD of two different QDs will be possible by simply tilting the magnetic field out of plane. The perspective of universal electrostatic control which uses only a static SO-induced anticrossing should further motivate the exploration of this setup.
Acknowledgments We are grateful for support from the Alexander von Humboldt foundation.
## Appendix A: Full Calculation of ∆ so from SOIs
In summary, we have discussed a two-electron qubit encoding in the s z = 0 subspace for an ISTQ. The outof-plane magnetic field is used to generate a level crossing of | S 〉 and | T 〉 that is not present for normal STQs. SOIs couple | S 〉 and | T 〉 if the sizes of the QDs differ. Different variances of the wave functions of the QD orbitals cause an effective magnetic-field difference across the DQD. A DQD that consists of two unequal QDs can be a promising spin qubit also for other reasons. It has one QD with a large singlet-triplet splitting and one QD with a small singlet-triplet splitting already without external magnetic fields. The strongly confined QD is ideal for the qubit initialization and the readout, while the weakly confined QD is suitable for the qubit manipulations. We suggest ISTQs in GaAs and InAs because they provide sufficiently large ∆ so .
This section extends the calculation of ∆ so from to the main text. Here we take into account that the DQD system is not only one dimensional. Besides ˜ H 2 = Ξ glyph[planckover2pi1] · ∑ i =1 2 , [ ℘ x σ ] i , with Ξ = ( -β cos (2 ξ ) , -α -β sin (2 ξ ) , 0) T , describing the momentum component connecting the QDs, there is also the in-plane perpendicular momentum component ˜ ˜ H 2 = Ψ glyph[planckover2pi1] · ∑ i =1 2 , [ ℘ y σ ] i , with Ψ = ( α -β sin (2 ξ ) , β cos (2 ξ ) , 0 ) T . ˜ ˜ H 2 matters for QDs, in which the electrons have space to move in the y -direction. Now, we discuss the extreme case of circular QDs. We assume, additionally to the properties of | L 〉 and | R 〉 that were introduced in Sec. III, 〈 L y L | | 〉 = 〈 R y R | | 〉 = 0 , 〈 L y ∣ ∣ 2 ∣ ∣ L 〉 = var L , 〈 R y ∣ ∣ 2 ∣ ∣ R 〉 = var R , and that | L 〉 and | R 〉 are separable into an x-part and y-part.
Hyperfine interactions and charge noise dephase ISTQs. Hyperfine interactions cause dephasing mainly in (1 , 1) through low-frequency magnetic-field fluctuations. Nuclear spins and SOIs couple to ISTQs in the same way. It is very important to fabricate ISTQs, where ∆ so is larger than the fluctuation ∆ hyp from nuclear spins. Nuclear spin noise can be refocused for ISTQs because
We apply the transformation U = e i ( S 1 + S 2 ) , with S i = m glyph[planckover2pi1] 2 [ ℘ x Ξ + ℘ y Ψ ] i · σ i . The transformed Hamiltonian U ( H 0 + H 1 + ˜ H 2 + ˜ ˜ H 2 ) U † contains similar terms as in Eq. (10). Formally, H 0 remains unchanged, and there is an overal energy shift -m glyph[planckover2pi1] 2 ( | Ξ | 2 + | Ψ | 2 ) . H 1 from Eq. (2) gives a position-dependent magnetic field,
$$\mathcal { U H } _ { 1 } \mathcal { U } ^ { t } = & \frac { g \mu _ { B } } { 2 } \sum _ { \substack { i = 1, 2 \\ j \in N } } B _ { \text{eff} } ^ { [ j ] } ( x _ { i } ) \cdot \sigma _ { i }, & & ( A 1 ) \\ & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & && & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &
& & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & - & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & \\ & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &
& & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & ; & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & ; & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & \\ & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &$$
$$B _ { \text{eff} } ^ { [ j ] } \left ( x \right ) \equiv \frac { 1 } { j! } \left ( \frac { 2 m } { \hbar { ^ { 2 } } } \right ) ^ { j } \left [ \left ( \dots \underbrace { \left ( B \, \underbrace { \left ( \Xi x + \Psi y \right ) \right ) \dots \right ) \times \left ( \Xi x + \Psi y \right ) } _ { j \text{ times} } \right ].$$
We extract from Eq. (A2) the effective magnetic-field component parallel to B in second order of the SOIs:
$$\left ( B _ { \text{eff} } ^ { [ 2 ] } \right ) _ { \text{i} } \approx - \frac { B } { 2 } \left ( \frac { 2 m } { \hslash ^ { 2 } } \right ) ^ { 2 } \left ( \Xi _ { \perp } ^ { 2 } x ^ { 2 } + \Psi _ { \perp } ^ { 2 } y ^ { 2 } \right ). \quad ( \text{A3} ) \quad ( \text{cf.} \.$$
Eq. (A3) neglects mixed terms in the position operators ( ∼ xy ) and couples | S 1 1 , 〉 and | T 1 1 , 〉 by ∆ Z so = E Z var R -var L ( l Z so ) 2 with l Z so = glyph[planckover2pi1] 2 2 m √ Ξ 2 ⊥ +Ψ 2 ⊥ (which we call the Zeeman spin precession length). Ξ ⊥ = | Ξ | sin [ glyph[measuredangle] ( Ξ , B )] and Ψ ⊥ = | Ψ | sin [ glyph[measuredangle] ( Ψ , B )] are the components of Ξ and Ψ perpendicular to the external magnetic field (cf. Fig. 1). Note that l Z so is on the order of the Rashba and Dresselhaus spin precession length, which is smaller than the confining radius of the QD wave functions.
The transformation of ˜ H 2 + ˜ ˜ H 2 adds additional contributions, dominated by:
$$\frac { m \Xi \times \Psi } { \hbar { ^ { 3 } } } \cdot \sum _ { i = 1, 2 } \sigma _ { i } \left [ ( l _ { z } ) _ { i } - \frac { m \omega _ { c } } { 2 } \left ( x _ { i } ^ { 2 } + y _ { i } ^ { 2 } \right ) \right ], \quad ( A ) ^ { \text{$\max_{\text{tio} }$} }$$
with l z = p x y -p x y , and ω c = eB z m . Especially the second term in Eq. (A4) couples | S 1 1 , 〉 and | T 1 1 , 〉 directly by an effective magnetic field parallel to B :
$$\left ( B _ { \text{eff}, o } ^ { [ 2 ] } \right ) _ { \text{i} } = - \frac { \hbar { \omega } _ { c } / 8 } { \frac { g \mu _ { B } } { 2 } } \left ( \frac { 2 m } { \hbar { ^ { 2 } } } \right ) ^ { 2 } ( \Xi \times \Psi ) _ { \text{i} } \left ( x ^ { 2 } + y ^ { 2 } \right ). \ ( A 5 ) \quad \text{Occup} \quad \text{at } ( a,$$
( Ξ × Ψ ) glyph[parallelshort] is the component parallel to B , which can be positive or negative. ( B [2] eff ,o ) glyph[parallelshort] is determined by the orbital contribution of the magnetic field glyph[planckover2pi1] ω c instead of the Zeeman energy E Z = gµ B 2 | B | . It describes the magnetic field produced by the orbital motion of electrons. We introduce the orbital spin precession length ( l o so ) 2 = ( glyph[planckover2pi1] 2 2 m ) 2 1 ( Ξ × Ψ ) glyph[parallelshort] , with which we write ∆ o so = glyph[planckover2pi1] ω c 2 var L -var R ( l o so ) 2 .
Eq. (A3) and Eq. (A5) couple | S 1 1 , 〉 and | T 1 1 , 〉 by a magnetic-field gradient across the DQD, similarly to the consideration in the main text:
$$\Delta _ { s o } = ( \text{var} _ { R } - \text{var} _ { L } ) \left ( \frac { E _ { Z } } { \left ( l _ { s o } ^ { Z } \right ) ^ { 2 } } + \frac { \hbar { \omega } _ { c } / 2 } { \left ( l _ { s o } ^ { o } \right ) ^ { 2 } } \right ). \quad ( \text{A} 6 ) \quad \mathcal { U } \left ( \mathcal { U }$$
Whether the Zeeman contribution E Z ( l Z so ) 2 or the orbital contribution glyph[planckover2pi1] ω / c 2 ( l o so ) 2 dominates Eq. (A6) depends in detail on the DQD. The orbital contribution should be dominant if the QDs are circular because glyph[planckover2pi1] ω c is usually larger than E Z : ∣ ∣ ∣ glyph[planckover2pi1] ω c E Z ∣ ∣ ∣ ≈ 135 for GaAs and ∣ ∣ ∣ glyph[planckover2pi1] ω c E Z ∣ ∣ ∣ ≈ 12 for InAs (cf. Appx. C). If the DQD setup prefers one spatial direction, then the Zeeman contribution dominates.
We analyze the angular dependencies of ∆ so , which are influenced by the direction of the magnetic field B , the orientation of the crystal lattice, and the dot connection axis e x (cf. Fig. 1). The Zeeman spin precession length gives ( l Z so ) -2 ∝ Ξ 2 ⊥ + Ψ 2 ⊥ = 2 ( α 2 + β 2 ) + sin 2 ( θ ) { α 2 + β 2 +2 αβ sin [2 ( φ -ξ )] } . SO contributions are maximal for out-of-plane magnetic fields, but they can vanish for in-plane magnetic fields. This is exactly the case if there is no coordinate of the SO field perpendicular to the magnetic field. ( l o so ) -2 ∝ ( Ξ × Ψ ) glyph[parallelshort] = ( α 2 -β 2 ) cos ( θ ) is independent of the orientation of the crystal lattice. Orbital effects are maximal for out-ofplane magnetic fields, but they vanish for in-plane orientations.
## Appendix B: Doubly Occupied Single QDs
This section describes the influence of the SOIs in the (2 0) , and (0 , 2) configurations when one QD is doubly occupied. A doubly occupied single QD with the center at ( a, 0 0) , T is described by
$$\underset { \underset { \underset { \underset { \underset { \underset { \underset { \underset { \pi } { \pi } { \pi } } } } { \text{or} } } } { \text{be} } \quad \underbrace { \sum _ { i = 1, 2 } \left [ \frac { \wp _ { i } ^ { 2 } } { 2 m } + V \left ( x _ { i } \right ) \right ] + V \left ( x _ { 1 }, x _ { 2 } \right ) } } _ { \mathcal { H } _ { 0 } } + \underbrace { \frac { g \mu _ { B } } { 2 } B \cdot \sum _ { i = 1, 2 } \sigma _ { i } } _ { \mathcal { H } _ { 1 } },$$
and ˜ H 2 = Ξ glyph[planckover2pi1] · ∑ i =1 2 , [ ℘ x σ ] i . We apply a unitary transformation U = e i ( S 1 + S 2 ) , with S i = m glyph[planckover2pi1] 2 ( x i -a ) Ξ · σ i . -ma glyph[planckover2pi1] 2 Ξ · σ i generates a constant, position-dependent phase shift of the transformed states. The transformed Hamiltonian,
$$\mathcal { U } \left ( \mathcal { H } _ { 0 } + \mathcal { H } _ { 1 } + \mathcal { H } _ { 2 } \right ) \mathcal { U } ^ { \dagger } = & \mathcal { H } _ { 0 } - \frac { m } { \hslash ^ { 2 } } \left | \Xi \right | ^ { 2 } \\ & + \frac { g \mu _ { B } } { 2 } \sum _ { i = 1, 2 } B _ { \text{eff} } \left ( x _ { i } \right ) \cdot \boldsymbol \sigma _ { i }, \ ( \mathbb { B } 2 )$$
describes a position-dependent magnetic field:
$$B _ { \text{eff} } \left ( x \right ) \equiv \left \{ \left ( \begin{array} { c } \sin \left ( 2 \rho \right ) \left \{ \frac { 1 } { 2 } - \cos \left [ \frac { 2 m \Xi } { h ^ { 2 } } \left ( x - a \right ) \right ] \right \} \\ \sin \left ( \rho \right ) \sin \left [ \frac { 2 m \Xi } { h ^ { 2 } } \left ( x - a \right ) \right ] \end{array} \right ) \, + \, \left ( \begin{array} { c } 0 \\ 0 \\ 1 \end{array} \right ) \left [ \sin ^ { 2 } \left ( \rho \right ) \cos \left ( \frac { 2 m \Xi } { h ^ { 2 } } \left ( x - a \right ) \right ) + \cos ^ { 2 } \left ( \rho \right ) \right ] \right \}, \quad ( B 3 )$$
where ρ is the rotation angle between B and Ξ , Ξ = | Ξ | . Note that there is a simple geometric relation between the angle ρ and the angles θ , φ , and ξ (cf. Fig. 1) B eff ( x )
does not couple | S 〉 and | T 〉 below the quadratic order in the position. Here, a different spread of the singlet and triplet wave functions will be seen. We can neglect these contributions to ∆ so because | S 〉 and | T 〉 have low weights in (2 0) , and (0 2) , at glyph[epsilon1] ∗ .
## Appendix C: Spin-Orbit Parameters
This section describes SOIs for typical semiconductor materials to build QDs. We introduce the Rashba and Dresselhaus SOIs for following Refs. [33,57,58]. Rashba SOI is caused by the broken structure inversion symmetry through the confining potential. The Rashba parameter α is determined by the confining electric field E z and a material constant α R : α = α R z E . 57 Typical values for E z are 0 1 . mV nm -1 . We introduce the Rashba spin precession length l α so = glyph[planckover2pi1] 2 2 mα . Dresselhaus SOI is present for a semiconducting lattice without inversion symmetry. The Dresselhaus parameter β is determined by a band parameter β D and the size of the wave function in the z-direction 〈 k 2 z 〉 : β = β D 〈 k 2 z 〉 . Typical values are 〈 k 2 z 〉 = (10 nm ) -1 . We introduce the Dresselhaus spin precession length l β so = glyph[planckover2pi1] 2 2 mβ .
- ∗ [email protected]
- 1 J. R. Petta, A. C. Johnson, J. M. Taylor, E. A. Laird, A. Yacoby, M. D. Lukin, C. M. Marcus, M. P. Hanson, and A. C. Gossard, Science 309 , 2180 (2005).
- 2 R. Hanson, L. P. Kouwenhoven, J. R. Petta, S. Tarucha, and L. M. K. Vandersypen, Rev. Mod. Phys. 79 , 1217 (2007).
- 3 D. Loss and D. P. DiVincenzo, Phys. Rev. A 57 , 120 (1998).
- 4 J. Levy, Phys. Rev. Lett. 89 , 147902 (2002).
- 5 J. M. Taylor, H.-A. Engel, W. Dür, A. Yacoby, C. M. Marcus, P. Zoller, and M. D. Lukin, Nat. Phys. 1 , 177 (2005).
- 6 R. Hanson and G. Burkard, Phys. Rev. Lett. 98 , 050502 (2007).
- 7 F. Baruffa, P. Stano, and J. Fabian, Phys. Rev. Lett. 104 , 126401 (2010).
- 8 F. Baruffa, P. Stano, and J. Fabian, Phys. Rev. B 82 , 045311 (2010).
- 9 W. A. Coish and D. Loss, Phys. Rev. B 72 , 125337 (2005).
- 10 H. Bluhm, S. Foletti, I. Neder, M. Rudner, D. Mahalu, V. Umansky, and A. Yacoby, Nat. Phys. 7 , 109 (2011).
- 11 I. Neder, M. S. Rudner, H. Bluhm, S. Foletti, B. I. Halperin, and A. Yacoby, Phys. Rev. B 84 , 035441 (2011).
- 12 J. R. Petta, H. Lu, and A. C. Gossard, Science 327 , 669 (2010).
- 13 H. Ribeiro, J. R. Petta, and G. Burkard, Phys. Rev. B 82 , 115445 (2010).
- 14 H. Ribeiro, G. Burkard, J. R. Petta, H. Lu, and A. C. Gossard, Phys. Rev. Lett. 110 , 086804 (2013).
- 15 H. Ribeiro, J. R. Petta, and G. Burkard, Phys. Rev. B 87 , 235318 (2013).
- 16 S. Mehl and D. P. DiVincenzo, Phys. Rev. B 88 , 161408
Typical parameters for GaAs, Si, and InAs are summarized in Tab. I. Conduction band electrons in Si have weak SOIs. Electrons in GaAs heterostructures have micrometer spin precession lengths. SOIs are by one order of magnitude larger in InAs than in GaAs because InAs has a much smaller band gap.
| | g | m/m e | α [ meV nm ] | l α so [ µ m ] | β [ meV nm ] | l β so [ µ m ] |
|------|----------|---------|----------------|------------------|----------------|------------------|
| GaAs | - 0 . 44 | 0 . 067 | 0 . 52 | 1 . 1 | 0 . 28 | 2 . 0 |
| Si | 2 | 0 . 19 | 0 . 01 | 20 | - | - |
| InAs | - 14 . 9 | 0 . 023 | 11 . 7 | 0 . 14 | 0 . 27 | 6 . 1 |
Table I. Parameters for the Rashba ( α ) and the Dresselhaus ( β ) SOIs, as described in the main text. The effective mass for the conduction band electron m (compared to the free electron mass m e ) and the g-factor are taken from Refs. [57,58,59]. The following band parameters are used: α GaAs R = 5 2 . eÅ , 2 α Si R = 0 11 . eÅ , 2 α InAs R = 117 1 . eÅ , 2 β GaAs R = 5 2 . eÅ , and 2 β InAs R = 117 1 . eÅ . 2 33,57,60 We introduce the Rasba spin precession length l α so = glyph[planckover2pi1] 2 2 mα and the Dresselhaus spin precession length l β so = glyph[planckover2pi1] 2 2 mβ . Si crystals have a center of inversion, which excludes the Dresselhaus SOI.
(2013).
- 17 Z. Shi, C. B. Simmons, D. R. Ward, J. R. Prance, X. Wu, T. S. Koh, J. K. Gamble, D. E. Savage, M. G. Lagally, M. Friesen, S. N. Coppersmith, and M. A. Eriksson, Nat. Commun. 5 , 3020 (2014).
- 18 D. Kim, Z. Shi, C. B. Simmons, D. R. Ward, J. R. Prance, T. S. Koh, J. K. Gamble, D. E. Savage, M. G. Lagally, M. Friesen, S. N. Coppersmith, and M. A. Eriksson, Nature (London) 511 , 70 (2014).
- 19 A. P. Higginbotham, F. Kuemmeth, M. P. Hanson, A. C. Gossard, and C. M. Marcus, Phys. Rev. Lett. 112 , 026801 (2014).
- 20 G. Burkard, D. Loss, and D. P. DiVincenzo, Phys. Rev. B 59 , 2070 (1999).
- 21 X. Hu and S. Das Sarma, Phys. Rev. A 61 , 062301 (2000).
- 22 E. Lieb and D. Mattis, Phys. Rev. 125 , 164 (1962).
- 23 U. Merkt, J. Huser, and M. Wagner, Phys. Rev. B 43 , 7320 (1991).
- 24 M. Wagner, U. Merkt, and A. V. Chaplik, Phys. Rev. B 45 , 1951 (1992).
- 25 D. M. Zumbühl, C. M. Marcus, M. P. Hanson, and A. C. Gossard, Phys. Rev. Lett. 93 , 256801 (2004).
- 26 The singlet-triplet inversion of gate-defined, asymmetric DQDs was proven in numerical simulations of GaAs heterostructures, T. Hiltunen et al., to be published (2014).
- 27 The Rashba constant α and the Dresselhaus constant β have the units Jm. l α so = glyph[planckover2pi1] 2 2 mα and l β so = glyph[planckover2pi1] 2 2 mβ are the Rashba and Dresselhaus spin precession lengths that have the units m.
- 28 E. I. Rashba, Sov. Phys. Solid State 2 , 1224 (1960).
- 29 G. Dresselhaus, Phys. Rev. 100 , 580 (1955).
- 30 A. V. Khaetskii and Y. V. Nazarov, Phys. Rev. B 61 , 12639
(2000).
- 31 I. L. Aleiner and V. I. Fal'ko, Phys. Rev. Lett. 87 , 256801 (2001).
- 32 L. S. Levitov and E. I. Rashba, Phys. Rev. B 67 , 115324 (2003).
- 33 R. Winkler, Spin-Orbit Coupling Effects in TwoDimensional Electron and Hole Systems , Springer Tracts in Modern Physics Vol. 191 (Springer, Berlin, 2010).
- 34 V. Fock, Z. Phys. 47 , 446 (1928).
- 35 C. G. Darwin, Proc. Cambridge Philos. Soc. 27 , 86 (1931).
- 36 A. Barenco, C. H. Bennett, R. Cleve, D. P. DiVincenzo, N. Margolus, P. Shor, T. Sleator, J. A. Smolin, and H. Weinfurter, Phys. Rev. A 52 , 3457 (1995).
- 37 J. M. Taylor, J. R. Petta, A. C. Johnson, A. Yacoby, C. M. Marcus, and M. D. Lukin, Phys. Rev. B 76 , 035315 (2007).
- 38 J. Medford, J. Beil, J. M. Taylor, E. I. Rashba, H. Lu, A. C. Gossard, and C. M. Marcus, Phys. Rev. Lett. 111 , 050501 (2013).
- 39 J. M. Taylor, V. Srinivasa, and J. Medford, Phys. Rev. Lett. 111 , 050502 (2013).
- 40 S. N. Shevchenko, S. Ashhab, and F. Nori, Phys. Rep. 492 , 1 (2010).
- 41 N. V. Vitanov and B. M. Garraway, Phys. Rev. A 53 , 4288 (1996).
- 42 A. Pal, E. I. Rashba, and B. I. Halperin, Phys. Rev. X 4 , 011012 (2014).
- 43 SOIs mix the spin part and the orbital part of the wave functions, and they also enable effective two-qubit interactions other than σ (1) z σ (2) z . We neglect SO contributions for the construction of two-qubit interactions because H 2 from Eq. (9) is weak.
- 44 Ioffe Institute, Electronic archive: New Semiconductor Materials. Characteristics and Properties (http://www.ioffe.ru/SVA/NSM/, 2001).
- 45 M. D. Shulman, O. E. Dial, S. P. Harvey, H. Bluhm, V. Umansky, and A. Yacoby, Science 336 , 202 (2012).
- 46 G. S. Paraoanu, Phys. Rev. B 74 , 140504 (2006).
- 47 C. Rigetti and M. Devoret, Phys. Rev. B 81 , 134507
(2010).
- 48 J. M. Chow, A. D. Córcoles, J. M. Gambetta, C. Rigetti, B. R. Johnson, J. A. Smolin, J. R. Rozen, G. A. Keefe, M. B. Rothwell, M. B. Ketchen, and M. Steffen, Phys. Rev. Lett. 107 , 080502 (2011).
- 49 J. M. Chow, J. M. Gambetta, A. D. Córcoles, S. T. Merkel, J. A. Smolin, C. Rigetti, S. Poletto, G. A. Keefe, M. B. Rothwell, J. R. Rozen, M. B. Ketchen, and M. Steffen, Phys. Rev. Lett. 109 , 060501 (2012).
- 50 I. A. Merkulov, A. L. Efros, and M. Rosen, Phys. Rev. B 65 , 205309 (2002).
- 51 L. V. C. Assali, H. M. Petrilli, R. B. Capaz, B. Koiller, X. Hu, and S. Das Sarma, Phys. Rev. B 83 , 165301 (2011).
- 52 H. Bluhm, S. Foletti, D. Mahalu, V. Umansky, and A. Yacoby, Phys. Rev. Lett. 105 , 216803 (2010).
- 53 M. D. Shulman, S. P. Harvey, J. M. Nichol, S. D. Bartlett, A. C. Doherty, V. Umansky, and A. Yacoby, Nat. Commun. 5 , 5156 (2014).
- 54 X. Hu and S. Das Sarma, Phys. Rev. Lett. 96 , 100501 (2006).
- 55 K. D. Petersson, J. R. Petta, H. Lu, and A. C. Gossard, Phys. Rev. Lett. 105 , 246804 (2010).
- 56 O. E. Dial, M. D. Shulman, S. P. Harvey, H. Bluhm, V. Umansky, and A. Yacoby, Phys. Rev. Lett. 110 , 146804 (2013).
- 57 T. Ihn, Semiconductor Nanostructures: Quantum States and Electronic Transport (Oxford University Press, New York, 2010).
- 58 F. A. Zwanenburg, A. S. Dzurak, A. Morello, M. Y. Simmons, L. C. L. Hollenberg, G. Klimeck, S. Rogge, S. N. Coppersmith, and M. A. Eriksson, Rev. Mod. Phys. 85 , 961 (2013).
- 59 M. Thalakulam, C. B. Simmons, B. M. Rosemeyer, D. E. Savage, M. G. Lagally, M. Friesen, S. N. Coppersmith, and M. A. Eriksson, Appl. Phys. Lett. 96 , 183104 (2010).
- 60 C. Tahan and R. Joynt, Phys. Rev. B 71 , 075315 (2005).
|
10.1103/PhysRevB.90.195424
|
[
"Sebastian Mehl",
"David P. DiVincenzo"
] |
2014-08-05T16:02:24+00:00
|
2014-11-26T14:22:27+00:00
|
[
"cond-mat.mes-hall",
"quant-ph"
] |
Inverted Singlet-Triplet Qubit Coded on a Two-Electron Double Quantum Dot
|
The $s_z=0$ spin configuration of two electrons confined at a double quantum
dot (DQD) encodes the singlet-triplet qubit (STQ). We introduce the inverted
STQ (ISTQ) that emerges from the setup of two quantum dots (QDs) differing
significantly in size and out-of-plane magnetic fields. The strongly confined
QD has a two-electron singlet ground state, but the weakly confined QD has a
two-electron triplet ground state in the $s_z=0$ subspace. Spin-orbit
interactions act nontrivially on the $s_z=0$ subspace and provide universal
control of the ISTQ together with electrostatic manipulations of the charge
configuration. GaAs and InAs DQDs can be operated as ISTQs under realistic
noise conditions.
|
1408.1011v3
|
## Non-hierarchical Structures: How to Model and Index Overlaps?
Faegheh Hasibi Norwegian University of Science and Technology [email protected]
Svein Erik Bratsberg Norwegian University of Science and Technology [email protected]
## ABSTRACT
Overlap is a common phenomenon seen when structural components of a digital object are neither disjoint nor nested inside each other. Overlapping components resist reduction to a structural hierarchy, and tree-based indexing and query processing techniques cannot be used for them. Our solution to this data modeling problem is TGSA (Tree-like Graph for Structural Annotations), a novel extension of the XML data model for non-hierarchical structures. We introduce an algorithm for constructing TGSA from annotated documents; the algorithm can efficiently process non-hierarchical structures and is associated with formal proofs, ensuring that transformation of the document to the data model is valid. To enable high performance query analysis in large data repositories, we further introduce an extension of XML pre-post indexing for non-hierarchical structures, which can process both reachability and overlapping relationships.
## 1. INTRODUCTION
Texts, data, objects (digital or otherwise) can be viewed from different perspectives. No perspective depends upon another and each exists separate from the other. By attempting to encode different perspectives that are independent of each other, one can encounter'overlapping elements'. The most common example is a document with two distinct structural views, when the logical view is section/subsection/ paragraph and the physical view is page/column . Each single structural view of this document is a hierarchy and the components are either disjoint or nested inside each other. The overlapping issue arises when one structural element cannot be neatly nested into others. For instance, when a paragraph starts in one page and terminates in the next page. Similar situations can appear in videos and other multimedia contents, where temporal or spatial constituents of a media file may overlap each other [25].
The most used model for expressing structure of documents is based on hierarchies, which ensures that each region is nested within another and the regions can be accessed
Balisage: The Markup Conference. August 2014, Washington, DC, USA.
.
by use of parent-child or ancestor-descendant relationships. This tree data structure requires organizing structural information of digital objects in a single tree, which is not applicable for overlapping structures. In other words, treebased markup languages -such as XML- and the corresponding indexing and retrieval techniques are not sufficient for documents with non-hierarchical structures [14].
Scholars have introduced several solutions for overlapping problem. TEI (Text Encoding Initiative) consortium [8] suggests several methods to deal with non-hierarchical structures in SGML or XML context. However, these methods are just syntactical solutions to represent non-hierarchical structures and non of them are based on a well-defined data model. On the other hand, there are some non-XML markup languages such as SGML CONCUR [26], LMNL [23] and TexMECS [15], in which TexMECS is based on a data welldefined model called GODDAG.
In order to efficiently query over overlapping structures in large data repositories, the structural elements of an overlapping data model should be efficiently summarized and indexed. To this end, we consider documents with continuous, ordered structural elements (as common features of nonhierarchical structures) and introduce a data model together with an XML-compatible parsing and indexing method. In other words the scope of this paper is overlap-only marked up documents, such that the order of appearance of document elements is unique. Overlap-only documents are also addressed in [21] and [20], where the authors mostly focused on serializability of the data model.
The contributions of this paper are three-fold: 1) We formally define TGSA 1 (Tree-like Graph for Structural Annotations) as a new data model for overlapping structures, which preserves simplicity while being expressive. 2) We introduce an efficient algorithm to map annotated documents to the TGSA data model. Our algorithm is compatible with XML parsing algorithms and can efficiently generate overlapping data model (TGSA) comparing to [27]. 3) We extend XML pre-post indexing method to the TGSA data model, which can represent parent-child and ancestor-descendant relationships as well as overlaps.
This paper is organized as follows. Section 2 motivates this research by providing use cases and applications of nonhierarchical structures. Section 3 presents the background and related works. Modelling and indexing methods are represented in Section 4, where subsection 4.1, 4.2 and 4.3 describe the TGSA data model, the TGSA construction al-
1 TGSA pronounced as 'tegsa'.
gorithm and indexing methods, respectively. Section 5 concludes the paper.
## 2. USE CASES AND APPLICATIONS
Overlapping is a situation that is more common than it may be thought of. In general, annotating several aspects of an object (i.e. stream of data) implies overlapping structures. Lots of scholars encounter overlaps in the area of computational linguistic, speech and complex text analysis. Some of these situations are as follows:
Content analysis of textual data. Many search applications use a pipeline process for analysing input data. For instance, FSIS (FAST search for internet sites), which is a Microsoft search platform, uses such pipeline to detect content of unstructured or semi-structured documents [12, 1]. As shown in Figure 1, this pipeline consists of several components, where each component analyses one aspect of data and adds annotations to the input. These annotations can be structural information such as sentences, paragraphs and links or extracted entities such as addresses, locations and names. Since each component extracts annotations independent of other components, some annotations may overlap each other.
Question Answering systems. QA systems need to search among a large collection of annotated documents to find the answer for a natural language query. These systems, mostly have access to a rich linguistic and semantic annotation of data, where overlapping is very likely to happen. As it is discussed in [22, 5], in order to get effective results for complex queries, a QA system needs to handle overlaps for both indexing and query processing.
Annotating non-textual objects. Overlapping structures can be seen not only in documents, but also while encoding other kinds of digital data. For instance, to classify a movie database based on genres, the time interval of each genre should be annotated by movie analysers. Among these annotations, there might be scenes that are annotated for more than one genres. Example 1 illustrates this scenario, where 'dream' scene overlaps both 'romance' and 'musical' genres. Handling overlapping annotations, allows classical structural information retrieval systems to answer the queries like:
- · Find all scenes that only contain romance content (i.e. no overlaps with other genres).
- · Find all scenes that have both musical and romance contents (i.e. scenes that overlap two genres).
```
contents (i.e. scenes that overlap two genres).
Example 1.
<sample>
<video>
<scene id="love" s="00:00" e="00:38"/>
<scene id="dream" s="00:39" e="01:44"/>
</video>
<genres>
<genre id="romance" s="00:00" e="00:45"/>
<genre id="musical" s="00:46" e="01:34"/>
<genre id="romance" s="01:00" e="01:44"/>
</genres>
</sample>
```
</sample>
As shown in Example 1, video annotations are encoded using stand-off annotations, where the original data is separated from structural views. Stand-off annotation is a natural technique to encode annotations of non-textual data and also linguistic annotations of text documents [3, 22].
Change tracking of documents. Text change tracking is a feature provided by text editors such as Microsoft Office to allow users to revise a document. Tracking of text changes is a known challenging topic, since this information needs to be stored in a markup language and the new structure of text may conflict with the former one. In other words, whenever a change affects the structure of a text (such as merging two paragraphs or splitting a paragraph), the notations of this change break the hierarchical structure of current text encoding. Example 2 demonstrates a case where the text change overlaps existing structure of document [11].
Example 2.
```
a
<p>The end of one paragraph <change-start change-id= "1"/> </p> </change-end <change-end change-id="1"/> <!-- --> | \
- + > } .
: & ( ) /
```
Digitizing old manuscripts. This task is another domain that needs dealing with non-hierarchical structures. Building electronic version of old manuscripts requires encoding massive amounts of information, such as textual content, physical location of texts, linguistic information, visibility of characters and information about damages of manuscript [9].
## 3. BACKGROUND AND RELATED WORK
## 3.1 Overlapping Data Models
The difficulty of handling overlaps is that overlapping structures are not hierarchies and the popular markup languages, such as XML are based on hierarchies. As a result of lacking an adequate overlapping data model, several data structures have been proposed to describe overlapping structures. In the following, we discuss these data structures and their abilities to model different types of overlaps.
Multiple Hierarchies (SGML CONCUR). The most straightforward model for the overlapping problem is to keep multiple hierarchies in a single document. This model is captured by the CONCUR feature of SGML, which maintains multiple structural views of a document. It actually extends the SGML/XML data model to a model, where multiple trees (with the same frontier) can be encoded within a single document.
The CONCUR model is represented as a part of SGML and consequently it is a legitimate and maintainable approach for overlapping problem. However, this model is not widely implemented as a solution of overlaps. Here is a list of CONCUR drawbacks [27, 10, 4]:
- 1. The model is not able to constrain relations among DTDs. For example a quote element may overlap paragraph elements but not chapter elements in the other DTD [10].
- 2. CONCUR does not provide self-overlaps, where two elements, with the same name, coinciding each other.
Figure 1: Data processing pipeline in large scale search engines
In such cases one element have to be moved to another hierarchy, which means CONCUR should support unpredictable DTDs.
ing from one hierarchy to another hierarchy. XIRAF's query approach is based on Burkowski's [7] work, which adds four new overlapping operations to XPath queries.
- 3. Concur cannot simply model the deletion, insertion, duplication, or reordering of data in the various views [27].
MCT. The Multi-Coloured Tree (MCT) [18] data structure was introduced for storing multiple hierarchies. In MCT, different coloured hierarchies are built on top nodes, where they can share the same nodes. As discussed in [17], MCTs are not suitable for representing multiple hierarchies in document-centric XML. One important issue is that there is no global order for children in different hierarchies.
LMNL. LMNL(Layered Markup and Annotation Language) is a data model associated with a markup language [23, 24]. LMNL data model is based on layers rather than hierarchies.It represents documents without forcing elements into a hierarchy and can capture overlaps.
GODDAGs. GODDAG (Generalized Ordered Descendant Directed Acyclic Graph) [27] is a well-defined data model, which is introduced to represent documents with overlapping structures. Principally, GODDAG is a directed acyclic graph (DAG), where each non-terminal node has ordered descendants.
GODDAG has two variations: restricted and generalized GODDAG. Restricted GODDAG is capable of representing overlaps; however its constraints rules out the possibility of modelling non-contiguous elements. Here are three constraints of restricted GODDAG, which are relaxed in generalized GODDAG: 1) Leaf nodes are ordered. 2) Each nonterminal dominates a contiguous subsequence of leaves. 3) No two nodes dominate the same subsequence of the frontier. By relaxing these constraints, generalized GODDAG can represent discontinuous elements.
Sperberg-McQueen et al. introduced and algorithm to convert MECS encoded documents to restricted GODDAG [27], while no algorithms is suggested for generalized GODDAG. In this paper, we show that the TGSA data model and its corresponding algorithm outperforms GODDAGs in efficiency.
## 3.2 Querying of Overlapping Structures
Querying over non-hierarchical structures requires a mechanism that relates structural regions to each other. Iacob et al. [16] extended XPath as EXPath to query overlaps over GODDAG structures. XIRAF [3] is another system that allows querying over overlapping annotations by mov-
It should be noted that all of these query approaches are developed for domain specific applications. However, query processing for large data repositories needs investigations on indexing structures.
## 3.3 Structural Indexing
Overlapping data structures can be modelled by either graphs or tree-like structures, such as GODDAGs. To the best of our knowledge, there is no research directly investigating on the overlapping indexing, however there has been large number of studies on XML and graph indexing.
According to [13], there are two main classes of structural indexing for XML data: numbering schemes and index graph schemes. The former is used for path joining, while the latter is for path selection in answering XML queries. Zhang et al. [29] proposed a numbering scheme for XML documents, called PrePost encoding. This model is capable of processing parent-child as well as ancestor-descendant relationships. Dewey coding [28] is another famous numbering scheme, which can be maintained easier than PrePost method. Jin [19] introduced a 3-hop indexing scheme, which is targeted for directed graphs with high edge-vertex density.
## 4. MODELLING AND INDEXING
In this section, we divide the problem of handling nonhierarchical structures into three issues and we dig into each issue in the following subsections. The first issue is which data model can represent hierarchies as well as overlaps? Although graphs can represent any relationships, we need less generalized but expressive enough data model that can represent overlapping structures efficiently. The second is how to parse and convert encoded documents to the data model? and the third is how to index these structures to efficiently process structural queries?
## 4.1 TGSA:Tree-like Graph for Structural Annotations
In general, structural annotations of documents are represented with graphs. However, for well-formed XML documents, these graphs are basically tree data structures, where edges represent parent-child relationships that are explicitly encoded in the document. In order to extend this data model to capture overlapping annotations, we introduce a new type of edge that represents overlapping relationship. In addition, we allow nodes to have more than one parent. As a result, the new data model would be a graph that satisfies some properties of trees. We call this new data model
Figure 2: An example of the TGSA data model. Element 'c' overlaps elements 'b, f' , and 'g' , while precedes elements 'f' and 'g' . Element 'f' overlaps elements 'b,c,d,e,g' , while it precedes element 'g' . (a) XML encoded document. The start and end of overlapping elements are indicated with the type attribute. (b)The TGSA data model. Solid arcs and dashed arcs represent 'P-C' (parent-child) and 'O' (overlapping) relationships, respectively.
TGSA, which stands for Tree-like Graph for Structural Annotations.
TGSA is basically a directed graph that has a set of nodes and arcs. The nodes correspond to the structural elements or text-nodes. Each arc in TGSA graph is labelled with either 'P-C' (parent-child) or 'O' (overlapping). Examples of the data model are shown in Figure 2 and Figure 3 (d), in which overlapping and parent-child relationships are shown using dashed and solid arcs, respectively. We consider TGSA as a two relational directed graph, defined as follows:
Definition 1. Two-relational graph G = ( V, E ) has a set of vertices V and a set of arcs E : V × V denoted by triples of form ( v , l, v i j ), where l ∈ { ' P -C , ' ' O ' } .
Based on this definition, we define TGSA as follows:
Definition 2. Let G = ( V, E ) be a two-relational directed graph. G is TGSA iff :
- 1. G is connected and has no loops.
- 2. All modes have atleast one incoming edge except one root node that has no incoming edge.
- 3. Given nodes v i and v j , e = ( v , i ' P -C ,v ' j ) iff v i is parent of v j .
- 4. Given nodes v i and v j , e = ( v , i ' O ,v ' j ) iff v i precedes v j AND v i and v j are overlapping.
- 5. Given nodes v i and v j , if there is an indirect path from v i to v j , then the arc e = ( v , i ' P -C ,v ' j ) cannot exist.
Paths in TGSA are defined whenever two nodes are connected via ' P -C ' arcs. In other words, ' O ' arcs do not create paths between nodes.
Definition 3. Apath in TGSA is a sequence of { x , x 1 1 , ..., x n } such that ( x , 1 ' P -C ,x ' 2 ) , ( x , 2 ' P -C ,x ' 3 ) , ..., ( x n -1 , ' P -C ,x ' n ) are graph arcs of the type ' P -C ' and the x i are distinct.
Paths in TGSA are transitive relations, in which given a path from u to v and v to w , there is always a path from u to w . Unlike paths, overlapping relations are not transitive.
Axiom 1. Given a TGSA graph G = ( V, E ) and arcs e 1 = ( v , i ' O ,v ' j ) and e 2 = ( v , i ' O ,w ' j ), does not imply that v i 'overlaps' w j .
Based on the above definitions, TGSA can represent elements that overlap each other in very complex ways. Moreover, TGSA does not constrain relationships between annotations across the document; except that each annotation is related to a continuous piece of text and the document elements are ordered.
## TGSA vs. GODDAG.
TGSA is similar to restricted GODDAG in three aspects: 1) Cycles are not allowed in both data models. 2) The constraint of being only one path between two nodes is relaxed in both of them, which is one of the main properties of trees. This implies that each node can have multiple parents and in TGSA, each node can overlaps multiple nodes. 3) Each node dominates or overlaps a contiguous set of nodes, which means discontinuous elements can not be modelled.
On the other hand, unlike restricted GODDAG, TGSA mark overlapping elements using a different arcs type. By introducing overlapping arcs, TGSA can represent order of overlapping nodes, which is not possible in restricted GODDAG. TGSA can represent continuous, ordered elements, while
GODDAGs can handle disordered elements (e.g. two parents with the same order but in different order). Such constraints together with introducing overlapping relations, makes TGSA a data model that be efficiently indexed, queried and converted from structural annotations.
## 4.2 Constructing TGSA from Annotated Documents
We now present an algorithm for building TGSA from annotated text documents, as shown in Algorithm 1. The input to the algorithm is an in-line tagged document, where start and end of each node is labelled in the text. Such documents ensure that the document elements are ordered. The algorithm is independent of encoding format of input, in which any encoding method or markup language can be used to represent overlaps. The only requirement is that, the start and end delimiters should be specified among the text. It must be stated that stand-off annotations, where the start and end of nodes are specified by positional values are not supported by this algorithm. However, if the elements of a document are continuous and ordered, they can be converted to TGSA, even if they are encoded by stand-off annotations. For the sake of simplicity, we consider in-line tagged documents as the input of Algorithm 1 and show how document elements are mapped to TGSA.
To construct the TGSA graph G = ( V, E ), we need to initialize a list called L to keep track of open elements. A node will be added to and removed from this list when an open and end tag is seen, respectively. As shown in Figure 3 (b), we first assume that the document is hierarchical and every two adjacent open nodes have parent-child relationship. Whenever an end tag is seen, we need to check weather the node is overlapping with other nodes or not. If the node is the latest element added to the list L , then it is not overlapping, otherwise the node is overlapping with the nodes added after it. The situation is demonstrated in Figure 3 (c), where node p overlaps node s . In such cases, four changes should be applied to the TGSA graph: 1) Remove 'P-C'arcs between p and s . 2) Add'O'arcs from p to s . 3) Define a new parent for node s by adding 'P-C' arcs from parent of node p to node s . 4) Define new children for node p by adding 'P-C' relation from node p to the children of node s .
One of the challenges of the TGSA construction algorithm is when a new node should be added as a child of an overlapping node (the last step in the above paragraph). In this step, we need to be sure adding this relationship is legal and satisfies the last item of Definition 2.
To clarify, let us consider two overlapping nodes d and f in Figure 2. When the algorithm reaches the end tag of node d , children of node f are 'could be' and 'increased'. Here we should only add 'increased' as a child of node d , since there is a path from d to 'could be' through node e . The straightforward solution to overcome this problem is to check all possible paths from node d to all children of node f and add ones that are not descendants of node d . This solution is very costly and will increase time complexity of the algorithm drastically.
We efficiently address this issue by introducing a new set for each node in the list of open nodes L . This set keeps the descendants of each node that are assigned to the node due to overlapping relations. For each node in the list L , we initialize this empty set, which is called D in algorithm
## Algorithm 1 TGSA Construction
Require:
Annotated document D
Ensure:
TGSA graph G = ( V, E )
## begin
```
The
```
end
```
node
```
Figure 3: Constructing the TGSA data model from a non-hierarchical annotated document. L represents the list of open nodes in Algorithm 1. (a) XML encoded document. Attributes sID and eID indicate the start and end of each node. (b) Parsing line 5. (c) Parsing line 6. (d) Parsing line 10.
- 1. Function AddParentChildRelation describes how we use this set to add new children to each overlapping node. Given node n and node c as a child of node y , the function first checks whether c is member of n.D or not. If not, a new arc will be added from node n to c . In addition, node c will be added to the D set of all nodes added before n to the list L . The following proof describes how this solution will guarantee that every node will be added to the TGSA graph according to its definition.
Lemma 1. Given TGSA graph G = ( V, E ) and a list of open elements L = { a, b, ..., d, e, ..., n, ... } , if node e overlaps node n and x is a child of node n , then there is an indirect path from each precedents of node e in the list L to node x .
Proof. To prove that there exists indirect path between two nodes, we need to proves that two nodes have ancestordescendant relationship (and not parent-child relationship). Node e overlaps node n implies that node d is parent of node e . Therefore, node d is ancestor of node x .
All other precedents of node e in list L are either ancestor/parent of node d or overlaps node d . Assuming node b as one of precedents of node d :
- 1. If node b is ancestor/parent of node d , then node b is ancestor of node x .
- 2. If node b overlaps node e , then children of node d are assigned as children of node b . Since node d is ancestor of node x , then node b is also ancestor of node x .
The TGSA construction algorithm is a very efficient way of processing and representing overlapping structures, which is also compatible to hierarchical structures. Given n as number of nodes in an annotated document, time complexity of this algorithm for hierarchical structures is O n ( ). For nonhierarchical structures, we do not consider lookup costs of function addParentChildRelation , since the algorithms only needs to test that a node is not definitely in a set and this can be efficiently done by bloom filters [6].
Comparing the TGSA algorithm to the algorithm represented in [27], our algorithm has less time complexity than GODDAG construction algorithm. GODDAG algorithm needs to remove illegal arcs from GODDAG after the whole graph is build, which means the algorithm requires to find all possible paths in the graph, and remove redundant ones. Since there are multiple paths between two nodes, this process has large affect on the performance of algorithm.
## 4.3 Indexing Method for TGSA
Indexing is an essential method for providing high-performance query processing for large data repositories. Like tree-based data, non-hierarchical data needs efficient methods of storage to summarize the structural view of a document. Without indexing, one should traverse the data model forward and backward to find structural relationships.
Motivated by this, indexing method for non-hierarchical structures can be inspired by XML indexing methods. XML indexes include value and structural indexes, where the former is to index string literals and the later is to index structure of documents. In this paper, we extend XML pre-post indexing to non-hierarchical structures, as it provides more efficient support for finding relationships between nodes.
Pre-post indexing labels each node with two numbers, ( start, end ), which corresponds to the position of start and end tag of each element. By the help of ( start, end ), we can easily find the ancestor-descendant and overlapping relationships relationships. However, this information is not enough to determine parent-child relationships. This relationship can be supported by only adding parent(s) of each to the ( start, end ) pair. Figure 4 illustrates an example of pre-post indexing for non-hierarchical structures. It should be stated that unlike XML, each entry in the structural index can have multiple parents, which is the intrinsic feature of TGSA. For hierarchical structures, this indexing method only contains one parent for each node. The following properties show how different relationships are supported by this indexing method.
Property 1. (Overlapping relationship) Node a precedes and overlaps node b iff a.start < b.start AND a.end < b.end
Figure 4: Constructing the TGSA data model and inverted index from a non-hierarchical annotated document. (a) XML encoded document. Attributes sID and eID indicate the start and end of each node. (b) TGSA data model. The numbers at the left and right side of each node indicate the position of start and end tags of the element, respectively. (c) Inverted list of text index. (d) Inverted list of element index.
Property 2. (Ancestor-Descendant relationship) Node a is ancestor of node b iff a.start < b.start AND b.end < a.end
Property 3. (Parent-Child relationship) Node a is parent of node b iff b.parent = a.start
Another directions of research we hope to pursue in near future is investigating on serializability between TGSA and annotated documents and supporting stand-off annotated documents in the TGSA construction algorithm.
Interval indexing requires much less storage space than storing the whole TGSA data model. Moreover, it provides very efficient support of different relationships by few numerical comparison. On the other hand, the drawback of this indexing method is the difficulty in maintaining the index, since updating the data will result major changes in the start and end position of each node.
## 5. CONCLUSION AND FUTURE WORK
While XML is the dominant format for storing structural information, studies show that there are a series of structural data that cannot be presented and stored in hierarchical formats. In order to query these non-hierarchical structures, the first need is an expressive and simple data model that can represent different structural relationships.
In this paper, we considered documents with continuous, ordered structural elements, and we formally introduced a novel data model called TGSA, which can handle overlapping and parent-child relationships. The main features of this data model are its compatibility with hierarchical data and handling overlapping elements. Converting structural elements to the TGSA data model is the second step on handling overlapping structures. To this end, we introduced an efficient algorithm for constructing TGSA, which is validated with formal proofs. We also proposed an extension of XML pre-post indexing method to address high performance query analysis for large data repositories. This indexing method can represent parent-child, ancestor descendant and overlapping relationships between two nodes.
The focus of this paper was on theoretical aspects of nonhierarchical structures. The on-going research is likely to focus on experimental aspects of these structures using real world data sets and to adapt hierarchical query processing methods to non-hierarchical structures.
## 6. REFERENCES
- [1] Fast search server 2010 for internet sites. http://sharepoint.microsoft.com/en-us/Pages/Videos. aspx?VideoID=26, 2010.
- [2] W. Alink, R. A. F. Bhoedjang, P. A. Boncz, and A. P. De Vries. Xiraf - xml-based indexing and querying for digital forensics. Digit. Investig. , 3:50-58, Sept. 2006.
- [3] W. Alink, R. A. F. Bhoedjang, P. A. Boncz, and A. P. De Vries. Xiraf - xml-based indexing and querying for digital forensics. Digit. Investig. , 3:50-58, Sept. 2006.
- [4] D. Barnard, R. Hayter, M. Karababa, G. Logan, and J. McFadden. Sgml-based markup for literary texts: Two problems and some solutions. Computers and the Humanities , 22:265-276, 1988. 10.1007/BF00118602.
- [5] M. W. Bilotti, P. Ogilvie, J. Callan, and E. Nyberg. Structured retrieval for question answering. In Proceedings of the 30th annual international ACM SIGIR conference on Research and development in information retrieval , SIGIR '07, pages 351-358, New York, NY, USA, 2007. ACM.
- [6] B. H. Bloom. Space/time trade-offs in hash coding with allowable errors. Communications of the ACM , 13(7):422-426, 1970.
- [7] F. J. Burkowski. Retrieval activities in a database consisting of heterogeneous collections of structured text. In Proceedings of the 15th annual international ACM SIGIR conference on Research and development in information retrieval , SIGIR '92, pages 112-125, New York, NY, USA, 1992. ACM.
- [8] T. Consortium, L. Burnard, and S. Bauman. TEI P5: Guidelines for electronic text encoding and interchange . TEI Consortium, 2014.
- [9] A. Dekhtyar and I. E. Iacob. A framework for management of concurrent xml markup. Data and Knowledge Engineering , 52(2):185 - 208, 2005.
- [10] S. DeRose. Markup overlap: a review and a horse. Interchange (UK) , 11(1):16 - 29, 2005/03/.
- [11] A. Di Iorio, S. Peroni, and F. Vitali. A semantic web approach to everyday overlapping markup. Journal of the American Society for Information Science and Technology , 62(9):1696-1716, 2011.
- [12] F. ESP. Powering enterprise search with fast esp R © , 2008.
- [13] G. Gou and R. Chirkova. Efficiently querying large xml data repositories: A survey. Knowledge and Data Engineering, IEEE Transactions on , 19(10):1381 -1403, oct. 2007.
- [14] F. Hasibi. Indexing and querying overlapping structures. In Proceedings of the 36th International ACM SIGIR Conference on Research and Development in Information Retrieval , page 1144, 2013.
- [15] C. Huitfeldt and C. Sperberg-McQueen. Texmecs: An experimental markup meta-language for complex documents. http://www.hit.uib.no /claus/mlcd/papers/texmecs., 2001.
- [16] I. Iacob and A. Dekhtyar. Queries over overlapping xml structures. Technical report, Technical Report TR 438-05, U. of Kentucky, CS Dept, 2005.
- [17] I. E. Iacob. Algorithms for management of document-centric xml data . PhD thesis, University of Kentucky, Lexington, KY, USA, 2005. AAI3201068.
- [18] H. V. Jagadish, L. V. S. Lakshmanan, M. Scannapieco, D. Srivastava, and N. Wiwatwattana. Colorful xml: one hierarchy isn't enough. In Proceedings of the 2004 ACM SIGMOD international conference on Management of data , SIGMOD '04, pages 251-262, New York, NY, USA, 2004. ACM.
- [19] R. Jin, Y. Xiang, N. Ruan, and D. Fuhry. 3-hop: a high-compression indexing scheme for reachability query. In Proceedings of the 2009 ACM SIGMOD International Conference on Management of data , SIGMOD '09, pages 813-826, New York, NY, USA, 2009. ACM.
- [20] C. Marcoux, Yves. Sperberg-McQueen and C. Huitfeldt. Modeling overlapping structures: Graphs and serializability. Proceedings of Balisage: The Markup Conference 2013. Balisage Series on Markup Technologies , 10(2013), August 2013.
- [21] Y. Marcoux. Graph characterization of overlap-only texmecs and other overlapping markup formalisms. In Proceedings of Balisage: The Markup Conference 2008. Balisage Series on Markup Technologies , 1 (2008), August 2008.
- [22] P. Ogilvie. Retrieval using Document Structure and Annotations . PhD thesis, Carnegie Mellon University, 2010.
- [23] W. Piez. Lmnl in miniature, an introduction. http:// www.piez.org/wendell/LMNL/Amsterdam2008/presentation-slides.html, 2008.
- [24] W. Piez. Luminescent: parsing lmnl by xslt upconversion. Proceedings of Balisage: The Markup Conference 2012. Balisage Series on Markup Technologies , 8, August 2012.
- [25] P. Salembier and A. B. Benitez. Structure description tools. Journal of the American Society for Information Science and Technology , 58(9):1329-1337, 2007.
- [26] C. Sperberg-McQueen and C. Huitfeldt. Concurrent document hierarchies in mecs and sgml. Literary and Linguistic Computing , 14(1):29-42, 1999.
- [27] C. Sperberg-McQueen and C. Huitfeldt. Goddag: A data structure for overlapping hierarchies. In P. King and E. Munson, editors, Digital Documents: Systems and Principles , volume 2023 of Lecture Notes in Computer Science , pages 606-630. Springer Berlin- Heidelberg, 2004.
- [28] I. Tatarinov, S. D. Viglas, K. Beyer, J. Shanmugasundaram, E. Shekita, and C. Zhang. Storing and querying ordered xml using a relational database system. In Proceedings of the 2002 ACM SIGMOD international conference on Management of data , SIGMOD '02, pages 204-215, New York, NY, USA, 2002. ACM.
- [29] C. Zhang, J. Naughton, D. DeWitt, Q. Luo, and G. Lohman. On supporting containment queries in relational database management systems. In Proceedings of the 2001 ACM SIGMOD international conference on Management of data , SIGMOD '01, pages 425-436, New York, NY, USA, 2001. ACM.
| null |
[
"Faegheh Hasibi",
"Svein Erik Bratsberg"
] |
2014-08-05T16:07:11+00:00
|
2016-10-08T06:46:54+00:00
|
[
"cs.DB",
"cs.DS",
"cs.IR",
"H.3.1; H.2.1"
] |
Non-hierarchical Structures: How to Model and Index Overlaps?
|
Overlap is a common phenomenon seen when structural components of a digital
object are neither disjoint nor nested inside each other. Overlapping
components resist reduction to a structural hierarchy, and tree-based indexing
and query processing techniques cannot be used for them. Our solution to this
data modeling problem is TGSA (Tree-like Graph for Structural Annotations), a
novel extension of the XML data model for non-hierarchical structures. We
introduce an algorithm for constructing TGSA from annotated documents; the
algorithm can efficiently process non-hierarchical structures and is associated
with formal proofs, ensuring that transformation of the document to the data
model is valid. To enable high performance query analysis in large data
repositories, we further introduce an extension of XML pre-post indexing for
non-hierarchical structures, which can process both reachability and
overlapping relationships.
|
1408.1012v1
|
## Geometric dependence of Nb-Bi Te -Nb topological 2 3 Josephson junction transport parameters
## CG Molenaar , DP Leusink , XL Wang 1 1 2 and A Brinkman 1
1
Faculty of Science and Technology and MESA+ Institute for Nanotechnology, University of Twente, 7500 AE Enschede, The Netherlands 2 Institute for Superconducting and Electronic Materials, University of Wollongong, Wollongong, NSW, 2522, Australia
E-mail: [email protected]
Abstract. Superconductor-topological insulator-superconductor Josephson junctions have been fabricated in order to study the width dependence of the critical current, normal state resistance and flux periodicity of the critical current modulation in an external field. Previous literature reports suggest anomalous scaling in topological junctions due to the presence of Majorana bound states. However, for most realised devices, one would expect that trivial 2 π -periodic Andreev levels dominate transport. We also observe anomalous scaling behaviour of junction parameters, but the scaling can be well explained by mere geometric effects, such as the parallel bulk conductivity shunt and flux focusing.
## 1. Introduction
Topological insulator (TI) superconductor (S) hybrids are potential systems for realizing p-wave superconductivity and hosting Majorana zero-energy states [1-8]. The common singlet s-wave pairing from a nearby superconductor is predicted to induce a spinless p-wave superconducting order parameter component in a topological insulator [4, 8] because of the spin-momentum locking of the surface states in a topological insulator [9-11]. In a Josephson junction between two s-waves superconductors with a topological insulator barrier (S-TI-S), Majorana Bound States (MBS) can occur with a sin( φ/ 2) current-phase relationship [4]. Contacts between superconductors and 3D topological insulators were realised on exfoliated flakes and films of Bi Te , 2 3 Bi Se , and strained 2 3 HgTe [12-18], and the Josephson behaviour was investigated by measuring Fraunhofer patterns in the presence of an applied magnetic field and Shapiro steps due to microwave radiation [13-15, 17, 18]. Despite the presence of conductivity shunts through bulk TI, the Josephson current was found to be mainly carried by the topological surface states [13,17,18].
Pecularities in the Fraunhofer diffraction patterns have been found for topological Josephson junctions [15, 19], including non-zero minima in the Fraunhofer patterns and periodicities which do not correspond to the junction size. In junctions with a varying width the characteristic energy I R c N was reported to scale inversely with the junction width [15]. This observation has been phenomenologically attributed to the width dependence of the Majorana modes contributing to a highly distorted currentphase relationship [15]. The Majorana modes have also been held responsible for the unexpectedly small flux periodicities in the I c ( B ) Fraunhofer pattern of the same junctions [15].
However, only one mode out of many channels is a MBS. For all non-perpendicular trajectories a gap appears in the Andreev Bound state spectra, giving trivial 2 π periodic bound states [20]. For typical device sizes fabricated so far, the number of channels is estimated to be large (a width of the order of a few 100 nm and a Fermi wavelength of the order of k -1 F = 1 nm already gives a few 100 modes). The Majorana signatures are, therefore, expected to be vanishingly small.
To understand the I c ( B ) periodicity as well as the scaling of I R c N with width, we have realized S-TI-S topological Josephson junctions with varying width. We also observe a non-trivial scaling of the critical current, normal state resistance and magnetic field modulation periodicity. However, a detailed analysis shows that all scaling effects can be explained by mere geometric effects of trivial modes. The dominance of trivial Andreev modes is supported by the absence of 4 π periodicity signatures in the Shapiro steps under microwave irradiation.
## 2. Expected Majorana related modifications of the critical current modulation by magnetic field and microwaves
Screening external flux from a superconducting junction results in the characteristic Fraunhofer pattern in Josephson junctions due to the DC Josephson effect. The critical current is modulated by the magnetic flux with a periodicity of the superconducting flux quantum, Φ 0 = h/ e 2 , threading the junction, due to the order-parameter being continuous around a closed contour. If the current-phase relationship is changed from sin( φ ) to sin( φ/ 2) in a topologically non-trivial junction the periodicity is expected to become h/e .
Additionally, for junctions where MBS are present it has been proposed that the minima in the Fraunhofer pattern are non-zero [21]. The current at the minima is predicted to be approximately equal to the supercurrent capacity of a single channel, I M ≈ ∆ Φ . / 0
Applying an AC bias on top of the DC bias will create a frequency to voltage conversion, the AC Josephson effect. In the voltage state of the junction, at DC voltages equal to k Φ 0 f = khf/ e 2 , with k an integer and f the frequency in Hz, there will be a current plateau with zero differential resistivity at fixed finite voltages. The presence of these Shapiro steps in a superconducting junction is one of the hallmarks of the Josephson effect. In contrast to a Fraunhofer pattern, it does not depend on the geometry of the junction, but on the current-phase relationship of the junction. A sum of different current-phase relationships [22-27] I = A 1 sin φ/ 2+ A 2 sin φ + A 3 sin 2 φ + . . . will result in current plateaus at V 1 ,l = lhf/e , V 2 ,m = mhf/ e 2 , V 3 ,n = nhf/ e 4 , etc. For a pure sin( φ/ 2) relationship one expects steps only at khf/e . The actual width and the modulation as function of applied RF power of the current plateaus depends on the ratio between the applied RF frequency and the I R c N product of the junctions. This can be numerically obtained by solving the Resistively Shunted Josephson (RSJ) [28] junction model.
## 3. Sample layout and fabrication
Devices were designed with junctions of constant electrode separation and varying width. The fabrication is similar to the method used by Veldhorst et al. [13], but has been modified to reduce the number of fabrication steps and increase the number of usable devices available on one chip. Exfoliated flakes are transferred to a Si/SiO 2 substrate. E-beam lithography, with 300 nm thick PMMA resist, is used to define junctions and contacts in two different write fields, eliminating the photo-lithography step.
In figure 1 the contact pads, written with a coarse write field, and the structure on the Bi Te 2 3 flake, written with a smaller and accurate write field, are visible. The smaller write field increases the resolution possible. An overlap of the structures was used in areas where the dose or write field was changed. These overlaps will cause overexposure and are only possible when the resolution is not critical. The 80nm
Figure 1. Scanning electron microscope images of a typical device. The white bar is 200 µ m, 5 µ mand 100nm wide in the three consecutive images respectively, and the white rectangles in the two left images mark the location of the image to the right. The first image show the Nb contact pads (dark grey) written with a large write field. The middle image shows the flake (trapezoid with bright edges and a step edge diagonal across) with leads (dark grey) leading to the junctions. The junctions are visible as faint white breaks in the leads. Along the top-row, left to right, are a 100, 250 and 500 nm junction. On the bottom-row are a 750, 1000 and 2000 nm junction. The 5 µ m on the right hand side of the flake is overexposed. The rightmost image is a close-up of a 250 nm by 150 nm designed junction.
niobium superconducting film and a 2.5 nm capping layer of palladium are sputter deposited. The flake is Ar-ion etched at 50 eV for 1 min prior to deposition resulting in transparent contacts.
The edge of the flake with the substrate provides a step edge for the Nb from the substrate to the flake, and it is advisable to keep the thickness of the flake comparable or less than the thickness of the Nb layer. The thickness of the sputter deposited Nb is limited by the thickness of the e-beam resist layer. For flakes of usable lateral dimensions, the thickness is generally in the order of 100 nm. The contacts for the voltage and current leads are split on the flake: if a weak link occurs on the lead transition from flake to substrate this will not influence the measured current-voltage ( IV ) characteristic of the junction.
Structures with a disadvantageous aspect ratio (junction width to electrode separation), such as wider junctions, are prone to overexposure. This increases the risk of the junction ends not being separated. For wider junctions a slightly larger separation has been used. Overexposure will decrease the actual separation. Actual dimensions are verified after fabrication as in figure 1.
The junctions are characterised in a pumped He cryostat with mu-metal screening and a superconducting Nb can surrounding the sample. The current and voltage leads are filtered with a two stage RC filter. A loop antenna for exposure to microwave radiation is pressed to the backside of the printed circuit board (PCB) holding the device. A coil perpendicular to the device surface is used to apply a perpendicular magnetic field. For different values of the applied microwave power or magnetic field current-voltage traces are recorded.
Table 1. Junction characteristics. The critical current I C and normal state resistance R N are given at 1.6 K. The measured junction separation is 140 nm. The 750 nm and 5000 nm wide junctions have shorted junctions due to e-beam overexposure, and the 2000 nm wide junction had a non-ohmic contact, caused by a break at the edge of the Bi 2 Te 3 flake.
| Junction width (nm) | Critical current ( µ A) | Normal state resistance (Ω) |
|-----------------------|---------------------------|-------------------------------|
| 100 | 0.2 | 1.5 |
| 250 | 3.5 | 1.14 |
| 500 | 13.5 | 0.84 |
| 1000 | 16 | 0.64 |
## 4. Measured scaling of transport parameters
## 4.1. Junction overview
The devices are characterised by measuring their IV curves at 1.6 K under different magnetic fields and microwave powers. The microwave frequency of 6 GHz is chosen for maximum coupling as determined by the maximum suppression of the supercurrent at the lowest power. The main measured junction parameters are given in table 1.
Both the magnetic and microwave field response has been studied for all junctions. Results for the 250, 500 and 1000nm wide junctions are shown in figure 2. The supercurrent for the 100 nm wide junction was suppressed in a magnetic and microwave field without further modulation. In the response to the microwave field a sharp feature is visible starting at 200 µ A and -10 dBm for the 1000 nm junction. This is likely the result of an unidentified weak link in one of the leads. The fainter structures in the 250 nm junction starting at 43 and 60 µ A and -40 dBm are reminiscent of an echo structure described by Yang et al. [14] for Pb-Bi Se -Pb Josephson junctions. Measuring 2 3 the microwave response at the minima of the Fraunhofer pattern [21] yielded no Shapiro features.
## 4.2. Scaling of I C and R N
In general, the normal state resistance, R N , of a lateral SNS junction [29,30] is expected to scale inversely with junction width, whereas I c is expected to be proportional to the width, such that the I R c N product is constant [28]. Josephson junctions on topological insulators are similar to SNS junctions with an induced proximity effect by superconducting leads into a TI surface state. For junctions on Bi Te 2 3 the transport was found to be in the clean limit, with a finite barrier at the interface between the superconductor and the surface states [13]. The supercurrent for ballistic SNS junctions with arbitrary length and barrier transparency is given by [31], which was found to fit the data of Veldhorst et al. well [13]. The normal state resistance in Bi Te 2 3 is complicated by the diffusive bulk providing an intrinsic shunt. The leads on the Bi Te 2 3 flake leading up to the junction also contribute towards a normal state conductivity shunt without
Figure 2. Magnetic field and microwave power dependence. The top row figures show the dynamic resistance of the 250 nm wide junction, the middle row figures correspond to the 500 nm wide junction and the bottom row figure to the 1000 nm wide junction. The left column shows the reaction to an applied magnetic field, the right column the reaction to microwave power at 6 GHz. The horizontal line at ∼ 19 mT is an artifact of the magnet current source. IV curves for these junctions are shown in figure 4.
carrying supercurrent. This results in current paths not only directly between the two electrodes but also through and across the whole area of the flake to the left and right of the electrodes.
The scaling of I c and R N with junction width is shown in Figure 3. In the junctions with an aspect ratio (width of the junction divided by electrode separation) greater than 5, the I R c N product is approximately 11 µ V, similar to junctions where the the length was varied instead of the width [13,32]. Below this, the I R c N product falls sharply. To verify whether this can be due to Majorana modes we estimate the number of conducting channels in these small junctions. The number of channels in a junction is related to the width of the junction: M = k F × W π ‡ . For a 100 nm junction this means that there are more than 60 channels active in the junction, and a MBS will not dominate transport properties.
Rather, due to the open edges of the junctions, the scaling of the normal state resistance is not directly proportional to the junction width, but offset due to the whole flake providing a current shunt. This is similar to an infinite resistor network [33] providing a parallel resistance to the resistance due to the separation of the two leads. Taking this into account, the resistance between the two leads takes the form R = ( ρ W R parallel ) / WR ( parallel + ρ W ), where rho W /W gives the junction resistance without a current shunt and R parallel is the resistance due to the current shunt through the flake. This equation does not disentangle the surface and bulk contributions but treats them as scaling the same. In the zero-width limit, the resistance is cut-off and does not diverge to infinity. The resulting I R c N product can then be well explained by the scaling of R N (including the shunt) and the usual scaling of I c with width ( I c being directly proportional to the number of channels, given by the width of the junctions with respect to the Fermi wavelength). Note, that the expected scaling of I c contrasts previous observations of inverse scaling [15].
## 4.3. Periodicty in Φ 0
The critical current of Josephson junctions oscillate in an applied magnetic field due to a phase difference induced across the junction. The magnetic flux in the junction area is the product of the area of weak superconductivity between the two electrodes and flux density in this area. The area of the junction is given by W × ( l +2 λ L ), where W l , and λ L are the width, length and London penetration depth respectively. The investigated junctions are smaller than or comparable to the Josephson penetration depth, λ J = √ Φ 0 / (2 πµ d J 0 ′ C ), where d ′ is the largest dimension (corrected by the London penetration depth) of the junction and J C has been estimated using the bulk mean free path of Bi Te crystals, 22 nm [13], which allows us to ignore the field produced 2 3 by the Josephson current. For the 80 nm thick Nb film used we use the bulk London penetration depth, 39 nm [34,35].
‡ The wave vector for linear dispersion is given by k F = E F ¯ hv F ≈ 2 × 10 9 m -1 . The Fermi energy is taken as 150 meV and the Fermi velocity is in the order of 1 × 10 5 m/s.
Figure 3. Scaling of the critical current, normal resistance and I c R N product. In the left panel the measured critical currents (black squares) and normal state resistances (red circles) are plotted. The critical current is assumed to be linear (black dashes). The resistance is modelled as a width resistivity ρ w = 7.9 × 10 -7 Ωm in parallel with a constant shunt resistance R parallel = 1.9 Ω. The I R c N product is plotted in the right panel, with the dashed line as the product of the fits in the left panel. The fit approaches the I R c N product of 10 to 15 µ Vfound in junctions of varying width [13].
The superconducting leads may be regarded as perfect diamagnets. This leads to flux lines being diverted around the superconducting structure. This causes flux focussing in the junctions, as more flux lines pass through the junctions due to their expulsion from the superconducting bulk. We estimate the amount of flux focussing by considering the shortest distance a flux line has to be diverted to not pass through the superconducting lead. In a long lead, half are passed to the one side and half to the other side. At the end of the lead the flux lines are diverted into the junction area. The flux diverted is (( W/ 2 -λ L ) 2 × B , see also the inset of Figure 4. This occurs at both electrodes, and is effectively the same as increasing the junction area by 2 × ( W/ 2 -λ L ) 2 . Without flux focussing, the expected magnetic field periodicity is given by the dashed line in figure 4(a). Correcting for flux focussing and taking λ L = 39nm results in the solid line, closely describing the measured periods.
The colour graphs in figure 2 show the modulation of the critical current with microwave power. In figure 4(b) IV traces for different applied powers are plotted. The steps all occur at multiples of Φ 0 f =12.4 µ V. A 4 π periodic Josephson effect will result in steps only at even multiples of Φ 0 f . Shapiro steps are not geometry dependent: in combination with the previously introduced geometry corrected magnetic field periodicity this illustrates the 2 π periodic Josephson effect in these junctions.
Figure 4. Behaviour of Fraunhofer oscillation frequency and Shapiro steps. In the left panel the modulation period of the Josephson current as a function of the external field is plotted as a function of the inverse junction width. The dashed line is the expected period for a rectangular junction. The solid line takes into account flux focussing, as presented in the inset. The flux incident on the grey areas A and B is diverted to the sides of the junction lead, while the red area C is added to the effective junction area between the two leads. The dimensions of the superconductor have been reduced by the London penetration depth, since flux can penetrate this area. The right panel shows IV characteristics under 6 GHz microwave irradiation. The line graphs are at -40, -30, -20, -10 and 0 dBm powers for the 250, 500 and 1000 nm wide junctions, and are offset in current for clarity. All current plateaus are at multiples of Φ 0 f =12.4 µ V.
## 5. Conclusion
We investigated superconducting junctions, coupling Nb leads on the surface of a Bi Te 2 3 flake, by varying the junction width. The critical current and normal state resistance decrease and increase respectively with reduced junction width. However, the I R c N product is found to be geometry dependent, as the normal state resistance does not diverge for zero width. The decreasing I R c N product with reduced junction width is understood when taking into account the resistance due to the entire flake surface. The I R c N product becomes of the order of 10 to 15 µ V for wide junctions, similar to previous junctions [32]. The junctions are found to be periodic with Φ 0 in a magnetic field when flux focussing is taken into account. Microwave irradiation results in steps at voltages at k Φ 0 f , which is to be expected from junctions with ten to hundred conducting channels contributing to the coupling between the superconducting leads.
Using topological insulators with reduced bulk conductivity should result in increased I R c N products and allow for electrostatic control of the Fermi energy. With similar junction geometries this will allow for reduction and control of the number of superconducting channels. This step will allow the behaviour of a possible MBS to be uncovered and separated from geometric effects which affect all conducting channels in S-TI-S junctions.
## Acknowledgments
This work is supported by the Netherlands Organization for Scientific Research (NWO), by the Dutch Foundation for Fundamental Research on Matter (FOM) and by the European Research Council (ERC).
## References
- [1] Read N and Green D 2000 Phys. Rev. B 61 (15) 10267-10297
- [2] Kitaev A Y 2001 Physics-Uspekhi 44 131
- [3] Ivanov D A 2001 Phys. Rev. Lett. 86 (2) 268-271
- [4] Fu L and Kane C L 2008 Phys. Rev. Lett. 100 (9) 096407
- [5] Nilsson J, Akhmerov A R and Beenakker C W J 2008 Phys. Rev. Lett. 101 (12) 120403
- [6] Tanaka Y, Yokoyama T and Nagaosa N 2009 Phys. Rev. Lett. 103 (10) 107002
- [7] Stanescu T D, Sau J D, Lutchyn R M and Das Sarma S 2010 Phys. Rev. B 81 (24) 241310
- [8] Potter A C and Lee P A 2011 Phys. Rev. B 83 (18) 184520
- [9] Hsieh D, Xia Y, Qian D, Wray L, Dil J H, Meier F, Osterwalder J, Patthey L, Checkelsky J G, Ong N P, Fedorov A V, Lin H, Bansil A, Grauer D, Hor Y S, Cava R J and Hasan M Z 2009 Nature 460 1101-1105 ISSN 0028-0836
- [10] Chen Y L, Analytis J G, Chu J H, Liu Z K, Mo S K, Qi X L, Zhang H J, Lu D H, Dai X, Fang Z, Zhang S C, Fisher I R, Hussain Z and Shen Z X 2009 Science 325 178-181
- [11] Hasan M Z and Kane C L 2010 Rev. Mod. Phys. 82 (4) 3045-3067
- [12] Sac´ ep´ B, Oostinga J B, Li J, Ubaldini A, Couto N J, Giannini E and Morpurgo A F 2011 e Nat Communications 2 575
- [13] Veldhorst M, Snelder M, Hoek M, Gang T, Guduru V K, Wang X L, Zeitler U, van der Wiel W G, Golubov A A, Hilgenkamp H and Brinkman A 2012 Nature Materials 11 417-421 ISSN 1476-1122
- [14] Yang F, Qu F, Shen J, Ding Y, Chen J, Ji Z, Liu G, Fan J, Yang C, Fu L and Lu L 2012 Phys. Rev. B 86 (13) 134504
- [15] Williams J R, Bestwick A J, Gallagher P, Hong S S, Cui Y, Bleich A S, Analytis J G, Fisher I R and Goldhaber-Gordon D 2012 Phys. Rev. Lett. 109 (5) 056803
- [16] Oostinga J B, Maier L, Sch¨ffelgen P, Knott D, Ames C, Br¨ne C, Tkachov G, Buhmann H and u u Molenkamp L W 2013 Phys. Rev. X 3 (2) 021007
- [17] Cho S, Dellabetta B, Yang A, Schneeloch J, Xu Z, Valla T, Gu G, Gilbert M J and Mason N 2013 Nat Commun 4 1689-
- [18] Galletti L, Charpentier S, Iavarone M, Lucignano P, Massarotti D, Arpaia R, Suzuki Y, Kadowaki K, Bauch T, Tagliacozzo A, Tafuri F and Lombardi F 2014 Phys. Rev. B 89 (13) 134512
- [19] Kurter C, Finck A D K, Ghaemi P, Hor Y S and Van Harlingen D J 2014 Phys. Rev. B 90 (1) 014501
- [20] Snelder M, Veldhorst M, Golubov A A and Brinkman A 2013 Phys. Rev. B 87 (10) 104507
- [21] Potter A C and Fu L 2013 Phys. Rev. B 88 (12) 121109
- [22] Kwon H J, Sengupta K and Yakovenko V 2003 Brazilian Journal of Physics 33 653
- [23] Kwon H J, Sengupta K and Yakovenko V M 2004 The European Physical Journal B-Condensed Matter and Complex Systems 37 349-361
- [24] Kwon H J, Yakovenko V and Sengupta K 2004 Low temperature physics 30 613 1-613 8
- [25] Fu L and Kane C L 2009 Phys. Rev. B 79 (16) 161408
- [26] Ioselevich P A and Feigel'man M V 2011 Phys. Rev. Lett. 106 (7) 077003
- [27] Pal A, Barber Z, Robinson J and Blamire M 2014 Nat Commun 5 3340
- [28] Tinkham M 2004 Introduction to Superconductiviy 2nd ed (Dover Publications)
- [29] Likharev K K 1979 Rev. Mod. Phys. 51 (1) 101-159
Geometric dependence of Nb-Bi Te -Nb topological Josephson junctions ... 2 3
- [30] Golubov A A, Kupriyanov M Y and Il'ichev E 2004 Rev. Mod. Phys. 76 (2) 411-469
- [31] Galaktionov A V and Zaikin A D 2002 Phys. Rev. B 65 (18) 184507
- [32] Veldhorst M, Molenaar C G, Wang X L, Hilgenkamp H and Brinkman A 2012 Applied Physics Letters 100 072602
- [33] Cserti J 2000 American Journal of Physics 68 896-906
- [34] Gubin A I, Il'in K S, Vitusevich S A, Siegel M and Klein N 2005 Phys. Rev. B 72 064503
- [35] Maxfield B W and McLean W L 1965 Phys. Rev. 139 A1515
|
10.1088/0953-2048/27/10/104003
|
[
"C. G. Molenaar",
"D. P. Leusink",
"X. L. Wang",
"A. Brinkman"
] |
2014-08-05T16:07:48+00:00
|
2014-08-05T16:07:48+00:00
|
[
"cond-mat.supr-con"
] |
Geometric dependence of Nb-Bi${_2}$Te${_3}$-Nb topological Josephson junction transport parameters
|
Superconductor-topological insulator-superconductor Josephson junctions have
been fabricated in order to study the width dependence of the critical current,
normal state resistance and flux periodicity of the critical current modulation
in an external field. Previous literature reports suggest anomalous scaling in
topological junctions due to the presence of Majorana bound states. However,
for most realised devices, one would expect that trivial $2\pi$-periodic
Andreev levels dominate transport. We also observe anomalous scaling behaviour
of junction parameters, but the scaling can be well explained by mere geometric
effects, such as the parallel bulk conductivity shunt and flux focusing.
|
1408.1014v2
|
## Composite fermion-boson mapping for fermionic lattice models
- J. Zhao , C. A. Jim´nez-Hoyos , G. E. Scuseria 1 e 1 1 2 , , D. Huerga , 3
- J. Dukelsky , S. M. A. Rombouts , and G. Ortiz 3 4 5
- 1 Department of Chemistry, Rice University, Houston, Texas 77005, USA
- 2 Department of Physics and Astronomy, Rice University, Houston, Texas 77005, USA
- 3 Instituto de Estructura de la Materia, CSIC, Serrano 123, E-28006 Madrid, Spain
- 4 Departamento de F´ ısica Aplicada, Universidad de Huelva, E-21071 Huelva, Spain
- 5 Department of Physics, Indiana University, Bloomington IN 47405, USA
Abstract. We present a mapping of elementary fermion operators onto a quadratic form of composite fermionic and bosonic cluster operators. The mapping is an exact isomorphism as long as the physical constraint of one composite particle per cluster is satisfied. This condition is treated on average in a composite particle mean-field approach, which consists of an ansatz that decouples the composite fermionic and bosonic sectors. The theory is tested on the one- and two-dimensional Hubbard models. Using a Bogoliubov determinant for the composite fermions and either a coherent or Bogoliubov state for the bosons, we obtain a simple and accurate procedure for treating the Mott insulating phase of the Hubbard model with mean-field computational cost.
## 1. Introduction
The Hubbard model is a prototypical example of a strongly correlated system characterized by the competition between the strong particle interaction ( U ) and the kinetic energy ( ). t It is exactly solvable in one dimension, where the ground state at half filling is a Mott insulator for any repulsive non-zero interaction [1]. Despite the simplicity of the model, the physics arising in dimensions higher than one remains poorly understood. Different many-body approximations have been applied along the years in different lattice geometries and coupling regimes ( U/t ). Among them, let us cite Quantum and Variational Monte Carlo calculations [2, 3, 4], Dynamical Mean-Field Theory [5], Density Matrix Renormalization Group [6] and, more recently, Density Matrix Embedding Theory [7, 8]. However, all these approaches have shown their limitations to describe the strongly correlated regime of the Hubbard model ( U/t /greatermuch 1), in spite of the significant computational cost of most of them. In order to overcome this limitation, DMFT theories have been extended to clusters, showing excellent convergence properties in one and two dimensional lattices [9, 10]. Alternative approaches, based on slave-particle methods, were developed in condensed matter
physics to model strongly correlated systems [11, 12, 13]. These methods, which treat a transformed boson-fermion Hamiltonian, may provide access to a much better approximation to the true ground-state and to physical processes that are otherwise difficult to build into the many-body wave function [14].
Slave particle mappings may not be canonical; i.e., they may not preserve the commutation properties of the mapped operators. Exact nonlinear mappings have been studied sparingly [15]. Even if the mapping is canonical, mixtures with unphysical states may appear in the wave function due to the mean-field approximations often employed to treat the slave-particle Hamiltonians. These states need to be removed, albeit only approximately in practice. There are numerous examples of slave-particle mappings in the literature. In this contribution, we explore a generalization to clusters of the Zou-Anderson (ZA) mapping [12, 16]. The many-body states of each cluster Fock space will be mapped onto composite boson and fermions operators. This composite character has been already revealed in previous cluster mappings of spin systems to composite bosons [17]. More recently, we have extended this mapping to lattice boson systems and applied it to the Bose-Hubbard Hamiltonian with good success [18]. As a precursor to the formalism that we present here, we mention the boson-fermion plaquette model approach of Altman and Auerbach [19] for the 2D Hubbard model. There, a few states of bosonic and fermionic character were selected ad hoc and then treated within the contractor renormalization group. On the contrary, the Composite Fermion-Boson (CFB) mapping that we present in this work takes into account all states of the cluster, recovering the ZA mapping in the one-site cluster limit.
## 2. Theory
## 2.1. Fermion-boson composite mapping
Let us start our derivation by decomposing the original lattice into a perfectly tiled cluster lattice that will be referred to as the superlattice. Preferably, the geometry of the clusters will be chosen such that they preserve as much as possible the symmetries of the lattice. The Fock space of the complete system F is a direct product of the Fock spaces of each cluster F R , where R denotes the position of the cluster in the superlattice. The states in F R contain the complete set of many-body states up to the maximum number of fermions that the cluster can accommodate. Each cluster Fock space is decomposed into two subspaces with odd and even number of fermions, which are denoted by F o R and F e R . Each state of F o R and F e R can be represented by the action of a composite fermion (CF) and a composite boson (CB) operators, respectively, over the corresponding vacuum,
$$| \mathbf R \alpha \rangle \rightarrow a _ { \mathbf R \alpha } ^ { \dagger } | 0 \rangle, \quad | \mathbf R \beta \rangle \rightarrow b _ { \mathbf R \beta } ^ { \dagger } | 0 \rangle,$$
where α ( β ) labels clusters states with an odd (even) number of fermions, respectively. The new composite particle Fock space is larger than the original physical space.
However, the subspace defined by all states having one-and-only-one composite particle on each superlattice site has a one-to-one correspondence with the original physical space. This is equivalent to require
$$\sum _ { \alpha } a _ { \mathbf R \alpha } ^ { \dagger } a _ { \mathbf R \alpha } + \sum _ { \beta } b _ { \mathbf R \beta } ^ { \dagger } b _ { \mathbf R \beta } = I, \quad \quad \quad$$
at each superlattice site R . This condition will be referred to as the physical constraint . The formal mapping which relates the physical fermionic operators with the new composite particles reads
$$c _ { j \sigma } ^ { \dagger } = \sum _ { \alpha \beta } & \langle \mathbf R \alpha | c _ { j \sigma } ^ { \dagger } | \mathbf R \beta \rangle \ a _ { \mathbf R \alpha } ^ { \dagger } b _$$
where the site j of the original lattice is contained within the cluster R after the tiling. Notice that α and β states in the previous matrix elements differ by just one electron. The composite fermion operators ( a † R α , a R α ) satisfy the anticommutation rules,
$$\{ a _ { \mathbf R \alpha } ^ { \dagger }, a _ { \mathbf R ^ { \prime } \alpha ^ { \prime } } \} = \delta _ { \alpha, \alpha ^ { \prime } } \delta _ { \mathbf R \mathbf R ^ { \prime } },$$
while the composite boson operators ( b † R β , b R β ) satisfy bosonic commutation rules
$$[ b _ { \mathbf R \beta }, b _ { \mathbf R ^ { \prime } \beta ^ { \prime } } ^ { \dagger } ] = \delta _ { \beta \beta ^ { \prime } } \delta _ { \mathbf R \mathbf R ^ { \prime } }, \quad [$$
The composite bosons and fermions commute with each other,
$$[ a _ { \mathbf R \alpha }, b _ { \mathbf R \beta } ^ { \dagger } ] = 0.$$
Let us now explore the conditions that should be fulfilled by transformation (3) in order to preserve the canonical fermionic anticommutation relations, { c iσ , c † jσ ′ } = δ ij δ σσ ′ . For i, j ∈ R , we insert the transformation (3) into the commutator and obtain, where we have used the commutation relations (4), (5), and applied the physical constraint (2). Noting that the complete set of bosonic and fermionic cluster states satisfy a resolution of the identity,
$$\mathfrak { r } \text{or} \, i, j \in \text{R}, \, \text{we insert the transformation } ( 3 ) \, \text{into the commutator and obtain}, \\ \{ c _ { i \sigma }, c _ { j \sigma ^ { \prime } } ^ { \dagger } \} & = \sum _ { \alpha \alpha ^ { \prime } } \sum _ { \beta ^ { \prime } } \langle \text{R} \alpha | c _ { i \sigma } | \text{R} \beta ^ { \prime } \rangle \langle \text{R} \beta ^ { \prime } | c _ { j \sigma ^ { \prime } } ^ { \dagger } | \text{R} \alpha ^ { \prime } \rangle \ a _ { \text{R} \alpha ^ { \prime } } ^ { \dagger } a _ { \text{R} \alpha ^ { \prime } } ^ { \dagger } \\ & \quad + \sum _ { \beta \beta ^ { \prime } } \sum _ { \alpha ^ { \prime } } \langle \text{R} \beta | c _ { i \sigma } | \text{R} \alpha ^ { \prime } \rangle \langle \text{R} \alpha ^ { \prime } | c _ { j \sigma ^ { \prime } } ^ { \dagger } | \text{R} \beta ^ { \prime } \rangle \ b _ { \text{R} \beta ^ { \prime } } ^ { \dagger } b _ { \text{R} \beta ^ { \prime } } ^ { \dagger } \\ & \quad + \sum _ { \alpha \alpha ^ { \prime } } \sum _ { \beta ^ { \prime } } \langle \text{R} \alpha | c _ { j \sigma ^ { \prime } } ^ { \dagger } | \text{R} \beta ^ { \prime } \rangle \langle \text{R} \beta ^ { \prime } | c _ { i \sigma } | \text{R} \alpha ^ { \prime } \rangle \ a _ { \text{R} \alpha ^ { \prime } } ^ { \dagger } a _ { \text{R} \alpha ^ { \prime } } ^ { \dagger } \\ & \quad + \sum _ { \beta \beta ^ { \prime } } \sum _ { \alpha ^ { \prime } } \langle \text{R} \beta | c _ { j \sigma ^ { \prime } } ^ { \dagger } | \text{R} \alpha ^ { \prime } \rangle \langle \text{R} \alpha ^ { \prime } | c _ { i \sigma } | \text{R} \beta ^ { \prime } \rangle \ b _ { \text{R} \beta ^ { \prime } } ^ { \dagger } b _ { \text{R} \beta ^ { \prime } } ^ { \dagger }. \\ \text{where we have } \, \text{used the commutation relations } ( 4 ), \, ( 5 ), \, \text{and applied the physical} \\ \text{onstraint} \, ( 2 ). \, \text{Nothing that the complete set of bosonic and fermionic cluster states}$$
$$\sum _ { \beta } | { \mathbf R }, \beta \rangle \langle { \mathbf R }, \beta | + \sum _ { \alpha } | { \mathbf R } \alpha \rangle \langle { \mathbf R } \alpha | = I,$$
and taking into account that the matrix elements 〈 R α c | † jσ | R α ′ 〉 , 〈 R α c | jσ | R α ′ 〉 , 〈 R α c | † jσ | R α ′ 〉 and 〈 R β c | jσ | R β ′ 〉 vanish, equation (7) reduces to
$$\vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \vdots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \vec { \slots } \end{cases} \\ \{ c _ { i \sigma }, c _ { j \sigma ^ { \prime } } ^ { \dagger } \} & = \sum _ { \alpha \alpha ^ { \prime } } \langle \mathbf R \alpha | \{ c _ { i \sigma }, c _ { j \sigma ^ { \prime } } ^ { \dagger } \} | \mathbf R \alpha ^ { \prime } \rangle a _ { \mathbf R \alpha } ^ { \dagger } a _ { \mathbf R \alpha ^ { \prime } } \\ & \quad + \sum _ { \beta \beta ^ { \prime } } \langle \mathbf R \beta | \{ c _ { i \sigma }, c _ { j \sigma ^ { \prime } } ^ { \dagger } \} | \mathbf R \beta ^ { \prime } \rangle b _ { \mathbf R \beta ^ { \prime } } ^ { \dagger } \\ & = \delta _ { i, j } \delta _ { \sigma, \sigma ^ { \prime } },$$
/negationslash where in the last step we have made use of the anticommutation relations of the physical fermions, the orthogonality of the cluster basis, and the physical constraint (2). Note that the result is equivalent to a direct mapping of { c iσ , c † jσ ′ } onto the composite particle space. The equivalence of both mappings is maintained in the physical space, characterized by the satisfaction of the physical constraint (2), when the complete Fock space is used in each cluster. We also point out that the anticommutation relation { c iσ , c † jσ ′ } = 0 for i ∈ R , j ∈ R ′ = R is trivially satisfied by using the commutation relations (4), (5) and (6).
As long as the complete set of bosonic and fermionic cluster states is used, one can map an arbitrary operator acting within a cluster R to a one-body composite operator of the general form
$$\hat { \mathcal { O } } _ { \mathbf R } & = \sum _ { \alpha \alpha ^ { \prime } } \langle \mathbf R \alpha | \hat { \mathcal { O } } _ { \mathbf R } | \mathbf R \alpha ^ { \prime } \rangle a _ { \mathbf R \alpha } ^ { \dagger } a _ { \mathbf R \alpha ^ { \prime } } + \sum _ { \beta \beta ^ { \prime } } \langle \mathbf R \beta | \hat { \mathcal { O } } _ { \mathbf R } | \mathbf R \beta ^ { \prime } \rangle b _ { \mathbf R \beta } ^ { \dagger } b _ { \mathbf R \beta ^ { \prime } } \\ & \quad + \sum _ { \alpha \beta } \langle \mathbf R \beta | \hat { \mathcal { O } } _ { \mathbf R } | \mathbf R \alpha \rangle b _ { \mathbf R \beta } ^ { \dagger } a _ { \mathbf R \alpha } + \sum _ { \alpha \beta } \langle \mathbf R \alpha | \hat { \mathcal { O } } _ { \mathbf R } | \mathbf R \beta \rangle a _ { \mathbf R \alpha } ^ { \dagger } b _ { \mathbf R \beta }. \\ \text{The first line applies if the operator } \hat { \mathcal { O } } _ { \mathbf R } \text{ preserves the number of fermions, or creates or}$$
The first line applies if the operator ˆ O R preserves the number of fermions, or creates or annihilates an even number of fermions (even number parity). The second line applies if it creates or annihilates an odd number of fermions (odd number parity). Equivalently, any algebraic operator acting on n different clusters will be mapped to a general n -body CFB operator. As an example, let us apply the mapping (3) to the Hubbard Hamiltonian in a hypercubic lattice with N sites in d dimensions,
$$\hat { H } = - t \sum _ { \langle i, j \rangle, \sigma } \left ( c _ { i, \sigma } ^ { \dagger } c _ { j, \sigma } + c _ { j, \sigma } ^ { \dagger } c _ { i, \sigma } \right ) + U \sum _ { i = 1 } ^ { N } n _ { i \uparrow } n _ { i \downarrow } - \mu \sum _ { i = 1 } ^ { N } n _ { i }. \quad \ \ ( 1 2 ) \\.$$
where c † iσ ( c iσ ) creates (annihilates) a fermion at lattice site i with spin σ = ↑ ↓ , , and n iσ = c † iσ c iσ is the number operator, n i = n i ↑ + n i ↓ . The first term accounts for the hopping of fermions to nearest neighbor sites with tunneling amplitude t . The second term accounts for the on-site interaction of strength U , and the third term regulates the density of the system via an external chemical potential µ . In what follows, we express all quantities in units of the hopping parameter t . In terms of the composite particles,
the mapped Hamiltonian is
$$\text{he mapped Hamiltonian is} \\ \hat { H } _ { C F B } & = \sum _ { \mathbf R } \left [ ( T _ { \mathbf R } ) ^ { \alpha } _ { \alpha ^ { \prime } } \, a ^ { \dagger } _ { \mathbf R \alpha } a _ { \mathbf R \alpha ^ { \prime } } + ( T _ { \mathbf R } ) ^ { \beta } _ { \beta ^ { \prime } } \, b ^ { \dagger } _ { \mathbf R \beta ^ { \prime } } b _ { \mathbf R \beta ^ { \prime } } \right ] \\ & \quad + \sum _ { \langle \mathbf R \mathbf R ^ { \prime } \rangle } \left [ ( V _ { \mathbf R \mathbf R ^ { \prime } } ) ^ { \alpha \beta } _ { \beta ^ { \prime } \alpha ^ { \prime } } \, a ^ { \dagger } _ { \mathbf R \alpha } b ^ { \dagger } _ { \mathbf R ^ { \prime } \beta } b _ { \mathbf R ^ { \prime } \alpha } \mathbf R ^ { \prime } \alpha ^ { \prime } + \mathbf H. c. \right ] \\ & \quad + \sum _ { \langle \mathbf R \mathbf R ^ { \prime } \rangle } \left [ ( V _ { \mathbf R \mathbf R ^ { \prime } } ) ^ { \alpha \alpha ^ { \prime } } _ { \beta \beta ^ { \prime } } \, a ^ { \dagger } _ { \mathbf R \alpha } a ^ { \dagger } _ { \mathbf R ^ { \prime } \alpha ^ { \prime } } b _ { \mathbf R \beta } b _ { \mathbf R ^ { \prime } \beta ^ { \prime } } + \mathbf H. c. \right ] \\ \text{where repeated Greek indices sum. The tensors $T$ and $V$, defined as}$$
where repeated Greek indices sum. The tensors T and V , defined as
$$( T _ { \mathbf R } ) ^ { \alpha } _ { \alpha ^ { \prime } } \quad & = \langle \mathbf R \alpha | \hat { H } ^ { \square } _ { \mathbf R } | \mathbf R \alpha ^ { \prime } \rangle \\ \not \, T _ { \mathbf R \cdot \vee } \not \, ^ { \beta } \quad & = \not \, \mathbf R \, \mathbf A | \hat { H } ^ { \square } | \mathbf R \, \mathbf A ^ { \prime } \langle \mathbf R \, \mathbf A ^ { \prime } \langle \cdot \vee }$$
$$( V _ { \text{RR} ^ { \prime } } ) ^ { \alpha \beta } _ { \beta ^ { \prime } \alpha ^ { \prime } } & = \langle \mathbf R \alpha, \mathbf R ^ { \prime } \beta | \hat { H } ^ { \times } _ { \text{RR} ^ { \prime } } | \mathbf R \beta ^ { \prime }, \mathbf R ^ { \prime } \alpha ^ { \prime } \rangle \\ \overbrace { \tau } \quad \langle \alpha \alpha ^ { \prime } } \quad \overbrace { \tau } \quad \mathbf R \tau \quad \mathbf R \tau \times \ \mathbf R \tau \circ \mathbf R \tau \times \ \overbrace { \tau } \quad \overbrace { \tau } \quad \overbrace { \tau } \quad \infty$$
$$( T _ { \mathbf R } ) ^ { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stackrel { \stack { \stack { \stack { \stack { \stack { \stack { \stack { \stack { \stack { \stack { \stack { \stack { \stack { \stack { \stack { \stack { \stack { \ \stack { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text } \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text { \ \text \text { \ \text { \ \text { \ \text { \ \text { \ \text \text { \ \text \text \text }
. - 1 + 1 + 2 1 - 2 + 3 0 3$$
$$( V _ { \text{RR} ^ { \prime } } ) _ { \beta \beta$$
contain all the information about the original Hamiltonian (12). Here, ˆ H /square R is the part of the Hubbard Hamiltonian acting within a cluster R , and the interaction between neighbor clusters ˆ H × RR ′ is exclusively due to the hopping term. Notice that due to the hermiticity of the interaction V , the tensors are symmetric under certain interchange of its indices,
$$( V _ { \text{R} ^ { \prime } } ) ^ { \beta \alpha$$
$$( V _ { \text{RR} ^ { \prime } } ) _ { \beta \beta$$
In the following, as we have only considered uniform, periodic superlattices with only nearest neighbour hopping, we will drop the R labels on the tensor T since it has the same value for all the clusters. In the same vein, the V tensor vanishes for all nonneighbour cluster pairs and takes the same value for all cluster pairs neighbouring in the same orientation. As a result, we shall drop the full RR ′ subscript and just subscript it with the neighbouring orientation. For instance, V x denotes the V tensor for clusters neighbouring in the x direction. Similarly to the ZA mapping, which is the one site cluster limit, the body-rank of hopping and on-site interaction terms are swapped. The one-body inter-cluster hopping is mapped into a two-body CFB term involving fermionboson interactions, while the original two-body on-site interaction is mapped into a sum of one-body composite fermion and one-body composite boson terms. As an illustration, this swapping effect is schematically shown in Figure 1 through an example of an intercluster hopping process and an on-site interaction on 2 × 2 neighbor clusters.
The Hamiltonian (13) is an exact isomorphism of the Hubbard Hamiltonian (12) within the physical subspace of the composite particle Fock space. Although the complexity of the original problem is not reduced, suitable approximations can be carried out with the advantage that short-range quantum correlations are included automatically within the collective structure of the composite particles.
Figure 1. Schematic representation of an inter-cluster hopping and an on-site interaction processes before (upper part of the figure) and after a mapping to 2 × 2 composite particles (lower part of the figures). The small yellow squares represent the original lattice sites. Green 2 × 2 clusters represents CF states, while the red clusters represent CB states. The figure illustrates how a inter-cluster hopping process maps onto a two-body CFB interaction while an on-site two-body interaction maps onto a one-body CB or CF term.
## 2.2. Mean-field solution
In order to proceed further, we assume that the system is translationally invariant and thus we perform a discrete Fourier transform of the composite operators,
$$\begin{array} {$$
where M is the number of sites of the hypercubic superlattice of clusters with size L d . The first Brillouin zone ( BZ ) is defined in the interval ( -π/L, π/L ] in each direction of the hypercubic reciprocal space. Under this transformation, the Hamiltonian can be
$$b _ { \mathbf R \beta } ^ { \dagger } = \frac { 1 }$$
written as
$$\text{written as} \hat { H } _ { C F B } = \sum _ { \mathbb { K } } \left ( T _ { \beta } ^ { \beta } b _ { \mathbb { K } \beta } ^ { \dagger } b _ { \mathbb { K } \beta ^ { \prime } } + T _ { \alpha } ^ { \alpha } a _ { \mathbb { K } \alpha } ^ { \dagger } a _ { \mathbb { K } \alpha ^ { \prime } } \right ) \\ + \frac { 1 } { M } \sum _ { u } \sum _ { \mathbb { K } _ { 1 } \mathbb { K } _ { 2 } Q } ( V _ { u } ) ^ { \beta } _ { \alpha ^ { \prime } \beta ^ { \prime } } \\ \times \left ( e ^ { - i Q _ { u } L } b _ { \mathbb { K } _ { 1 }, \beta } ^ { \dagger } a _ { \mathbb { K } _ { 2 } + Q, \alpha } ^ { \dagger } a _ { \mathbb { K } _ { 1 } + Q, \alpha ^ { \prime } } b _ { \mathbb { K } _ { 2 }, \beta ^ { \prime } } + H. c. \right ) \\ + \frac { 1 } { M } \sum _ { u } \sum _ { \mathbb { K } _ { 1 } \mathbb { K } _ { 2 } Q } ( V _ { u } ) ^ { \alpha \alpha ^ { \prime } } _ { \beta \beta ^ { \prime } } \\ \times \left ( e ^ { - i Q _ { u } L } a _ { \mathbb { K } _ { 1 }, \alpha } ^ { \dagger } a _ { \mathbb { K } _ { 2 } + Q, \alpha ^ { \prime } } b _ { \mathbb { K } _ { 1 } + Q, \beta } b _ { \mathbb { K } _ { 2 }, \beta ^ { \prime } } + H. c. \right ) \quad ( 2 2 ) \\ \text{where repeated Greek indices sum and } K _ { 1 }, K _ { 2 }, K, \, \text{and} \, Q \, \text{ are summed over all the} \\ \text{ectors in the Brillouin zone. } \, u \, \text{ is summed over the special dimensions of the system} \\ \text{nd} \, V \text{ denote the V} \, \text{tensor have another naive naihhaurin} \, \text{in the direction of} \, u$$
where repeated Greek indices sum and K 1 , K 2 , K , and Q are summed over all the vectors in the Brillouin zone. u is summed over the spacial dimensions of the system and V u denotes the V tensor for cluster pairs neighbouring in the direction of u .
The mean-field treatment of the Hamiltonian (22) assumes a decoupling of the composite fermion and boson spaces, | Ψ 〉 = Ψ | F 〉 ⊗ | Ψ B 〉 , yielding an effective meanfield Hamiltonian that is quadratic in both the fermionic and bosonic sectors and can be diagonalized by two independent fermionic and bosonic Bogoliubov transformations. The two sectors are coupled via self-consistent boson and fermion mean-fields. In other words, by taking partial variations of the expectation value of H CFB with respect to the fermion and boson wave functions, the stationary condition implies that | Ψ F 〉 and | Ψ B 〉 are eigenfunctions of the effective fermion and boson Hamiltonians,
$$\hat { H } ^ { F } & = \langle \Psi ^ { B } | \, \hat { H } _ { C F B } \, | \Psi ^ { B } \rangle \,, \\ \hat { u } ^ { B } & = \, / \mathfrak { d } ^ { F } | \, \hat { u } _ { \dashleftarrow } \, | \mathfrak { d } ^ { F } \, \cdot$$
$$\hat { H } ^ { B } = \langle \Psi ^ { F } | \, \hat { H } _ { C F B } \, | \Psi ^ { F } \rangle \,. \\ \cdot \quad \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \,.$$
More precisely, the fermionic sector reads
$$\text{for precisely, use requirements section reasons} \\ \hat { H } ^ { F } & = \sum _ { \mathbb { K } } T _ { \alpha ^ { \prime } } ^ { \alpha } a _ { \mathbb { K } _ { \alpha } } ^ { \dagger } a _ { \mathbb { K } _ { \alpha ^ { \prime } } } \\ & + \frac { 1 } { M } \sum _ { u } \sum _ { \mathbb { K } _ { \mathbb { Q } } } ( V _ { u } ) _ { \alpha ^ { \prime } \beta ^ { \prime } } ^ { \beta \alpha } \left ( e ^ { - i ( K _ { u } - Q _ { u } ) } a _ { \mathbb { K } _ { \alpha } } ^ { \dagger } a _ { \mathbb { K } _ { \alpha ^ { \prime } } } \langle b _ { \mathbb { Q }, \beta } ^ { \dagger } b _ { \mathbb { Q }, \beta ^ { \prime } } \rangle + H. c. \right ) \\ & + \frac { 1 } { M } \sum _ { u } \sum _ { \mathbb { K } _ { \mathbb { Q } } } ( V _ { u } ) _ { \beta \beta ^ { \prime } } ^ { \alpha \alpha ^ { \prime } } \left ( e ^ { - i ( K _ { u } - Q _ { u } ) } a _ { \mathbb { K } _ { \alpha } } ^ { \dagger } a _ { \mathbb { K } _ { \alpha ^ { \prime } } } ^ { \dagger } \langle b _ { \mathbb { Q }, \beta } b _ { \mathbb { Q }, \beta ^ { \prime } } \rangle + H. c. \right ), \\ \text{and the bosonic sector is obtained by a counterpart mean-field decoupling,}$$
and the bosonic sector is obtained by a counterpart mean-field decoupling,
$$\hat { H } ^ { B } & = \sum _ { K } T ^ { \beta } _ { \beta ^ { \prime } } b _ { \mathbf K \beta ^ { \prime } } ^ { \dagger } b _ { \mathbf K \beta ^ { \prime } } \\ & + \frac { 1 } { M } \sum _ { u } \sum _ { \mathbf K \mathbf Q } ( V _ { u } ) ^ { \beta \alpha ^ { \prime } } _ { \alpha ^ { \prime } \beta ^ { \prime } } \left ( e ^ { - i ( K _ { u } - Q _ { u } ) } b _ { \mathbf K, \beta ^ { \prime } } \langle a _ { \mathbf Q, \alpha } ^ { \dagger } a _ { \mathbf Q, \alpha ^ { \prime } } \rangle + \mathbf H. c. \right ) \\ & + \frac { 1 } { M } \sum _ { u } \sum _ { \mathbf K \mathbf Q } ( V _ { u } ) ^ { \alpha \alpha ^ { \prime } } _ { \beta \beta ^ { \prime } } \left ( e ^ { - i ( K _ { u } - Q _ { u } ) } b _ { \mathbf K, \beta ^ { \prime } } b _ { - \mathbf K, \beta ^ { \prime } } \langle a _ { \mathbf Q, \alpha } ^ { \dagger } a _ { \mathbf Q, \alpha ^ { \prime } } ^ { \dagger } \rangle + \mathbf H. c. \right ).$$
The physical constraint implies that one-and-only-one state is allowed per cluster, however, the mean-field decoupling of Eqs. (25) and (26) leads to wave functions | Ψ F 〉 and | Ψ B 〉 which do not preserve the local physical constraint (2) exactly. Therefore, we relax it fixing a global constraint of the total composite particle density,
$$& \frac { 1 } { M } \sum _ { \mathbb { R } } \left ( \sum _ { \alpha } \langle a ^ { \dagger } _ { \mathbb { R } \alpha } a _ { \mathbb { R } \alpha } \rangle + \sum _ { \beta } \langle b ^ { \dagger } _ { \mathbb { R } \beta } b _ { \mathbb { R } \beta } \rangle \right ) = 1. \quad & & ( 2 7 ) \\ \dots & \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \end{cases}$$
This latter condition is added to the effective bosonic and fermionic Hamiltonians via a unique Lagrange multiplier λ ,
$$\hat { F } ^ { F } & = \hat { H } ^ { F } - \lambda \sum _ { \mathbf K } \sum _ { \alpha } a ^ { \dagger } _ { \mathbf K \alpha } a _ { \mathbf K \alpha }, \\ \hat { F } ^ { B } & = \hat { H } ^ { B } - \lambda \sum _ { \beta } \sum _ { h ^ { \dagger } _ { \dots h _ { \nu, \vartheta } } } h ^ { \dagger } _ { \dots h _ { \nu, \vartheta } },$$
In this work, we have restricted ourselves to wave functions transforming as the totallysymmetric representation of the lattice translation group of the cluster superlattice. We would like to point out that, when the wave function does not break the translational symmetry of the cluster superlattice, the global imposition of the physical constraint would imply local on-average satisfaction of the physical constraint,
$$\hat { F } ^ { B } = \hat { H } ^ { B } - \lambda \sum _ { K } ^ { - - } \sum _ { \beta } ^ { - - } b _ { \mathbb { K } \beta } ^ { \dagger } b _ { \mathbb { K } \beta }. \\ \intertext { n \, } \text{$n \, } \text{this work\, we have restricted ourselves to wave functions transforming as the totally}$$
$$\sum _ { \alpha } \left \langle a _ { \mathbf R \alpha } ^ { \dagger } a _ { \mathbf R \alpha } \right \rangle + \sum _ { \beta } \left \langle b _ { \mathbf R \beta } ^ { \dagger } b _ { \mathbf R \beta } \right \rangle = 1, \\. \. \. \. \. \. \. \. \. \. \. \. \. \. \. \.$$
for all clusters R . Like in other slave particle approaches, fluctuations of the physical constraint induce mixtures with unphysical states. These unphysical processes are expected to decrease with increasing cluster sizes, approaching the exact physical eigenstate in the infinite size limit, in spite of the combinatorial increase of the computational cost.
In the same vein as our previous composite particle treatment of boson systems [18], one can consider several candidate mean-field reference states. The fermionic part of the product wave function may be treated within a Hartree-Fock (HF) or a HartreeFock-Bogoliubov (HFB) approximation, while the bosonic part may be treated within a Hartree-Bose (HB) or a Hartree-Bose-Bogoliubov (HBB) approximation. In this work, we present results for the Hubbard model in one and two dimensions obtained by diagonalizing the fermionic sector via a general HFB transformation, and by treating the bosonic sector in HB and HBB approximations. In the first approximation, we assume that the bosonic sector is described by a coherent state of CBs in the K = 0 , α = c mode, and neglect all bosonic fluctuations (HB). In the second approximation, the bosonic sector is also diagonalized by means of a Bogoliubov transformation (HBB).
Within the coherent approximation, we replace the condensate of CB by a c -number b ( † ) 0 , c → √ Mσ , where σ 2 is the CB condensate fraction . The coherent CB is a linear combination of CB configurations, i.e., | c 〉 = ∑ β U c β | β 〉 . Inserting this transformation into Eq. (29) and minimizing with respect to the variational amplitudes
U c leads to the Hartree-Bose eigensystem, where λ is the lowest eigenvalue, and U c its corresponding eigenvector,
$$\sum _ { \beta ^ { \prime } } h ^ { 0, 0 } _ { \beta \beta ^ { \prime } } U ^ { \mathfrak { c } } _ { \beta ^ { \prime } } & = \lambda U ^ { \mathfrak { c } } _ { \beta }, \\ f \, & \, \text{the Hartree matrix} \, h \, \text{for to incuraltice momentum} \, K - \mathfrak { n }$$
where the superindexes of the Hartree matrix h refer to superlattice momentum K = 0 . The value of the condensate fraction σ 2 is obtained through the physical constraint,
$$\sigma ^ { 2 } & = 1 - \frac { 1 } { M } \sum _ { \mathbb { K } \alpha } \langle a ^ { \dagger } _ { \mathbf K \alpha } a _ { \mathbf K \alpha } \rangle. \\. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &$$
A general symmetry-preserving Bogoliubov transformation for quadratic Hamiltonians has the form [20],
$$\gamma _ { \mathbf K _ { \eta } } & = \sum _ { \eta ^ { \prime } } \left ( X _ { \eta ^ { \prime } \eta } ^ { \mathbf K } \right ) ^ { * } \ p _ { \mathbf K _ { \eta ^ { \prime } } } ^ { \dagger } + \varepsilon \sum _ { \eta ^ { \prime } } \left ( Y _ { \eta ^ { \prime } \eta } ^ { - \mathbf K } \right ) ^ { * } \ p _ { - \mathbf K _ { \eta ^ { \prime } } }, \\ & \quad \quad \quad$$
where the operators ( p † K η , p K η ) refer to fermions ( ε = +1) or bosons ( ε = -1). Accordingly, the η labels either CF states α or CB states β . The amplitudes X and Y are obtained by solving the self-consistent matrix eigensystem of the form [20],
$$\left ( \begin{array} { c c } h ^ { K, K } & \Delta ^ { K, - K } \\ \varepsilon ( \Delta ^ { - K, K } ) ^ { * } & \varepsilon ( h ^ { - K, - K } ) ^ { * } \end{array} \right ) \left ( \begin{array} { c c } ( X ^ { K } ) ^ { * } & Y ^ { K } \\ ( Y ^ { - K } ) ^ { * } & X ^ { - K } \end{array} \right ) = \Omega ^ { K } \left ( \begin{array} { c c } ( X ^ { K } ) ^$$
where the positive eigenvalues of the diagonal matrix Ω K = Diag( -w K , w K ) will determine the fermionic (bosonic) quasi-particle excitation dispersions. The matrix elements are straightforwardly obtained by identifying the grand-canonical composite particle potentials given in (28) and (29) with the general expression
$$\hat { F } & = \sum _ { \mathbf K \eta \eta ^ { \prime } } \left ( h ^ { \mathbf K, \mathbf K } _ { \eta \eta ^ { \prime } } p _ { \mathbf K \eta ^ { \prime } } ^ { \dagger } p _ { \mathbf K \eta ^ { \prime } } \right ) + \frac { 1 } { 2 } \sum _ { \mathbf K \eta \eta ^ { \prime } } \left ( \Delta ^ { \mathbf K, - \mathbf K } _ { \eta \eta ^ { \prime } } p _ { \mathbf K \eta ^ { \prime } }$$
The block diagonal structure with respect to K stems from the fact that all normal density matrix elements 〈 p † K 1 η p K 2 η 〉 vanish except for K 1 = K 2 and the anomalous density matrix elements 〈 p K 1 η p K 2 η 〉 vanish except for opposite momentum K 1 = -K 2 when the wave function transforms according to an irreducible representation of the superlattice translation group. Upon inversion of the Bogoliubov transformation, we obtain the normal and pairing tensors
$$\langle p _ { \mathbf K \eta } ^ { \dagger } p _ { \mathbf K \eta ^ { \prime } } \rangle \ = \ \sum _ { \zeta } ( X _ { \eta ^ { \prime } \zeta } ^ { \mathbf K } ) ^ { * } X _ { \zeta \eta } ^ { \mathbf K }, \\ \quad _ { \nu _ { \mu } ^ { \dagger } \ \nu _ { \mu } ^ { \dagger } } \ \, \ \, - \quad \, \sim \, \sum _ { \nu ^ { \nu - \mathbf K } \vee \vee \nu$$
The convergence of the procedure is determined by the self-consistency between the fermion and boson Hamiltonians and their respective mean-field solutions. In the scheme where we diagonalize both sectors by coupling the fermionic and bosonic Bogoliubov
$$\langle p _ { \mathbf K \eta } ^ { \dagger } p _ { - \mathbf K \eta ^ { \prime } } ^ { \dagger } \rangle & = \stackrel { \cdot } { - \varepsilon } \sum _ { \zeta } ( Y _ { \eta ^ { \prime } \zeta } ^ { - \mathbf K } ) ^ { * } X _ { \zeta \eta } ^ { \mathbf K }. \\ \text{The convergence of the non-order is determined by the self consistency between the}$$
eigensystems self-consistently, the value of the Lagrange multiplier λ is varied smoothly until the density of composite particles equals one.
In both of the approximations here considered, the energy of the system can be readily computed taking the expectation value of the CFB Hamiltonian (22) with our ansatz | Ψ = Ψ 〉 | B 〉⊗| Ψ F 〉 . In particular, for the coherent approximation the total energy reads
$$\text{eads} \\ E = & \ M T _ { c } ^ { c } \sigma ^ { 2 } + \sum _ { \mathbf K } T _ { \alpha ^ { \prime } } ^ { \alpha } \langle \alpha _ { \mathbf K _ { \alpha } } ^ { \dagger } a _ { \mathbf K _ { \alpha ^ { \prime } } } \rangle \\ & + \sigma ^ { 2 } \sum _ { \mathbf K } \sum _ { u } \left ( V _ { u } \right ) _ { \alpha ^ { \prime } c } ^ { \alpha } \left ( e ^ { - i K _ { u } } \langle \alpha _ { \mathbf K _ { \alpha } } ^ { \dagger } a _ { \mathbf K _ { \alpha ^ { \prime } } } \rangle + \mathbf H. c. \right ) \\ & + \sigma ^ { 2 } \sum _ { \mathbf K } \sum _ { u } \left ( V _ { u } \right ) _ { \text{cc} } ^ { \alpha \alpha ^ { \prime } } \left ( e ^ { - i K _ { u } } \langle \alpha _ { \mathbf K _ { \alpha } } ^ { \dagger } a _ { - \mathbf K _ { \alpha ^ { \prime } } } ^ { \dagger } \rangle + \mathbf H. c. \right ). \\ \text{where repeated indices are summed and we have performed a contraction of the original} \\ \text{even} \, T \, \text{and} \, V \, \text{with} \, \, \text{$two} \, $even} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and} \, \text{and}$$
where repeated indices are summed and we have performed a contraction of the original tensors T and V with the boson condensate c . Notice that, as the mean-field treatment does not preserve the physical constraint exactly, the energy reached is not an upper bound to the exact ground state energy. Nevertheless, by performing a finite-size scaling analysis one can give a quantitative estimate of the exact result.
We are also interested in the double-occupation parameter, which gives a measure of on-site physical fermion correlations and provides a qualitative indicator of the Mott insulator transition. It is here computed in the CFB mean-field approximations as
$$D _ { j } & = \langle n _ { j \downarrow } n _ { j \uparrow } \rangle = \sum _ { \beta } \langle \mathbf R \beta | \, n _ { j \uparrow } n _ { j \downarrow } \, | \mathbf R \beta \rangle \, \langle b ^ { \dagger } _ { \mathbf R \beta } b _ { \mathbf R \beta } \rangle + \sum _ { \alpha } \langle \mathbf R \alpha | \, n _ { j \uparrow } n _ { j \downarrow } \, | \mathbf R \alpha \rangle \, \langle a ^ { \dagger } _ { \mathbf R \alpha } a _ { \mathbf R \alpha } \rangle, \, \left ( 3 9 \right ) \\ \dots & \, \cdot \quad \right )$$
## with j ∈ R .
As a short recapitulation, we have presented a canonical mapping which allows us to express a many-fermion lattice Hamiltonian as a many fermion-boson system on a superlattice of clusters. The composite boson and fermion cluster operators contain the exact quantum correlations inside the clusters. In order for the mapping to be an exact isomorphism, a physical constraint has to be satisfied. We proposed a mean-field ansatz for our mapped Hamiltonian that is a direct product of fermionic and bosonic states that preserve the superlattice translational symmetry. In the following, we present results obtained with this approach.
## 3. Results and discussion
We start the numerical analysis of the method by studying the ground-state energy E (38) and double-occupation D (39) of the Hubbard model at half-filling (12) by means of the HB approximation for the CB Hamiltonian (29), and HFB for the CF Hamiltonian (28). Half-filling can be achieved by setting the chemical potential µ to half the value of the Hubbard repulsion parameter U for particle-hole symmetry. In Figure 3, we present the energy per site (left panel) and double-occupation (right panel) for various values of U . We also include the exact energy per site obtained by the Bethe ansatz [1] and
/s32
/s32
/s32
/s32
Figure 2. Ground state energies per site and double-occupancies of the 1-D Hubbard model obtained from the CFB mean-field theory with various clusters sizes ( L = 2 4 6), , , the restricted and unrestricted Hartree-Fock (HF) approximation, and the exact Bethe anstaz (BA). The insets amplifies the differences in particular coupling regimes U .
the energy obtained by both the standard, symmetry-preserving and the unrestricted HF approximations of the original Hubbard Hamiltonian (12). The RHF mean-field methods is exact in the non-interacting limit, U = 0; however, it quickly deteriorates with increasing interaction. The ground state energy obtained with the CFB mean-field is in very good agreement with the exact result for large U . The UHF method is able to provide a qualitatively reasonable result for both small and large U . So it parallels the trend for the Bethe ansatz, but becomes higher in energy than the CFB meanfield energy for larger U values. The right inset shows how the CFB mean-field energy converges to the exact one as the size of the cluster is increased. In this way, a significant portion of the correlation energy relative to the RHF solution can be captured. However, at low U , the energy from the CFB mean-field theory starts to deviate from the exact Bethe ansatz value, and crosses the RHF energy at U ∼ 1 4 for the size-six cluster, as . it can be seen in the left inset.
By inspecting the double-occupation expectation value (39) shown in the right panel of Figure 3, we check that the method becomes more accurate for intermediate and strong repulsion by increasing the size of the cluster (see right panel inset). The deviation at weak coupling can in fact be understood by a careful scrutiny of the method and its strengths. The kinetic and interacting terms of the Hubbard Hamiltonian (12) are one- and two-body terms, respectively, when written in terms of the physical fermions. In the mapped Hamiltonian (13) part of the kinetic term is computed in mean-field (inter-cluster contribution) and part is computed exactly (intra-cluster contribution), while the on-site interaction is always computed exactly. As a result, when minimizing the effective mean-field energy, the CFB mean-field tends to underestimate the kinetic contribution, leading to a wave function with low double-occupancy at the expense of a higher kinetic energy. By increasing the cluster size, we are including more hopping
processes into the exact computation. In this way, the method is able to yield lower total energy by lowering the kinetic energy and adjusting the double-occupancy to higher values. This trend is exactly analogous to mean-field theory of physical fermions, where the HF method lowers the one-particle kinetic energy to a large extent but yields too high a repulsion energy and thus non-optimal total energy. In this way, it can be seen that the CFB mean-field theory is highly complementary to the mean-field theory based on physical fermions because they are treating different terms in a mean-field way. The physical fermion mean-field theory performs better in the small U regime, like in the Fermi-liquid phase, while the fermion-boson mean-field theory performs better in large U regimes like the Mott insulating phase. Also apparent in this figure is the systematic improvement of the method when the size of the cluster is increased.
/s32
/s32
/s32
/s32
Figure 3. Ground state energy per site versus the inverse of the clusters size L for the CFB approximations and exact diagonalization showing the convergence to the exact Bethe ansatz value at intermediate ( U = 4) and strong ( U = 8) Hubbard interactions.
In Figure 3 we show the convergence of the energy per site towards the exact Bethe ansatz result as a function of the reciprocal of the number of sites in the cluster, 1 /L , for U = 4 in the left panel and U = 8 in the right panel. Specifically we depict the energy of a single isolated cluster obtained from exact diagonalization of ˆ H /square R , together with the two CFB mean-field treatments described above: the bosonic coherent wave function (CFB Coherent) and the bosonic Bogoliubov wave function (CFB Bogoliubov). A clear improvement of the energy is seen in both panels as a function of 1 /L . In the limit where the cluster becomes the entire system, inter-cluster processes, which are here treated approximately in mean-field, are no longer present, the CFB Hamiltonian becomes doubly quadratic and represents the even and odd number parity sectors of the exact Hamiltonian. As a result, CFB mean-field becomes exact in this limit and simply corresponds to picking the exact ground state as the composite particle to occupy. As seen in Figure 3, in both approaches (Coherent and Bogoliubov), the energy extrapolates to the exact result. The CFB Bogoliubov treatment adds bosonic fluctuations to the coherent-state approximation. The improvement is apparent in the figure, despite the fact that for strong coupling ( U = 8) the method overbinds. This is due to violations
of the physical constraint (2) when treated on average (30), that induces mixtures with unphysical states, a point that will be addressed in future works. Notice that the computational cost associated with increasing the cluster size is combinatorial, limiting the size of clusters that can be used as building blocks for the mapping. An alternative way to improve the precision of the method is the use of more sophisticated many-body approaches to treat the CFB Hamiltonian (13).
We have also applied the CFB mean-field approach within the coherent approximation to the two-dimensional Hubbard model at half filling. Results for the ground state energies and double-occupancies are displayed in figure 4 using clusters of sizes 2 × 1, 2 × 2, and 2 × 3. We have been unable to reach a self-consistent solution for small U values, even with sophisticated convergence-acceleration techniques [21]. This might stem from the fact that larger spatial dimensionality implies larger weight of the kinetic energy term in the Hamiltonian. In this sense, when a larger part of the Hamiltonian is treated in mean-field, convergence can be harder to achieve. The twodimensional Hubbard model exhibits antiferromagnetic (AF) long-range order for all repulsive U [22]. Inspecting our converged density matrix for the 2 × 2 cluster, we find that one of the two possible AF boson states gets a significantly larger weight than the other. This symmetry breaking cannot be observed in the 2 × 1 and 2 × 3 cases because these lattices are not commensurate with AF long-range ordering. This observation rationalizes the significant energy difference between them and the reported relatively high stability of the 2 × 2 case. Also the symmetry-breaking has not been observed for the one-dimensional model, as expected.
A noticeable difference between the one-dimensional and two-dimensional cases is the relative energy between the CFB mean-field theory and HF of physical fermions. In 1D systems, starting from U = 1 4 for the size-six . cluster, U = 1 7 for size-four . cluster, and U = 2 2 for size-two cluster, HF yields a lower ground-state energy than . the composite mean-field approach. We can take this U crossover ratios as a rough estimate of the Mott transition point, which the exact solution predicts at U = 0 [1]. In the two-dimensional problem, recent DMFT results [23] place the Mott-insulator transition at approximately U = 5 3. . This is slightly higher than the crossing points between HF and our composite mean-field results.
## 4. Conclusion
In this work, we have proposed a mapping of fermion operators onto composite fermion and boson cluster operators in one-to-one correspondence to many-electron states. The mapping was used to transform the standard Hubbard model onto a new composite fermion-boson Hamiltonian acting on a cluster superlattice that can be treated by standard many-body approximations. The advantage of the cluster mapping is that the short-range interactions and the local quantum fluctuations are taken into account exactly from the onset. Specifically, on-site interactions and intra-cluster hopping processes are computed exactly by definition.
/s32
/s32
/s32
/s32
Figure 4. Ground state energies per site and double-occupancies of the 2-D Hubbard model obtained from the CFB mean-field theory with various clusters sizes, from the restricted and unrestricted Hartree-Fock approximation, and from an exact diagonalization of an isolated 4 × 4 cluster.
We have proposed a mean-field ansatz built as a tensor product of a fermionic and bosonic wave function. This ansatz in terms of composite particles can exactly handle single-cluster operators (such as the on-site repulsion), while inter-cluster scattering processes (such as hopping) are treated in a mean-field way. We showed that the ansatz performs well for large U at half-filling, and is fairly complementary to the Hartree-Fock approach of physical fermions, which treats hopping exactly at the expense of only a mean-field treatment of the on-site repulsion interaction. We plan to explore, in future work, the applicability of the ansatz to the doped phases of the Hubbard Hamiltonian.
## 5. Acknowledgments
This work was supported by the Department of Energy, Office of Basic Energy Sciences, Grant No. DE-FG02-09ER16053, the Welch Foundation (C-0036), DOE-CMCSN (DESC0006650), and by the Spanish Ministry of Economy and Competitiveness trough grants FIS2012-34479 and BES-2010-031607.
## References
- [1] E. H. Lieb and F. Y. Wu. Absence of Mott transition in an exact solution of the short-range, one-band model in one dimension. Phys. Rev. Lett. , 20:1445-1448, 1968.
- [2] S. Zhang, J. Carlson, and J. E. Gubernatis. Pairing correlations in the two-dimensional Hubbard model. Phys. Rev. Lett. , 78:4486-4489, 1997.
- [3] M. Guerrero, G. Ortiz, and J. E. Gubernatis. Correlated wave functions and the absence of long-range order in numerical studies of the hubbard model. Phys. Rev. B , 59:1706-1711, 1999.
- [4] S. Sorella. Wave function optimization in the variational Monte Carlo method. Phys. Rev. B , 71:241103, 2005.
- [5] A. Georges, G. Kotliar, W. Krauth, and M. J. Rozenberg. Dynamical mean-field theory of strongly correlated fermion systems and the limit of infinite dimensions. Rev. Mod. Phys. , 68:13-125, 1996.
- [6] U. Schollw¨ck. o The density-matrix renormalization group. Rev. Mod. Phys. , 77:259-315, 2005.
- [7] G. Knizia and G. K.-L. Chan. Density matrix embedding: A simple alternative to dynamical mean-field theory. Phys. Rev. Lett. , 109:186404, 2012.
- [8] I. W. Bulik, G. E. Scuseria, and J. Dukelsky. Density matrix embedding from broken symmetry lattice mean fields. Phys. Rev. B , 89:035140, 2014.
- [9] G. Kotliar, S. Y. Savrasov, G. P´lsson, and G. Biroli. a Cellular dynamical mean field approach to strongly correlated systems. Phys. Rev. Lett. , 87:186401, 2001.
- [10] M. Balzer, W. Hanke, and M. Potthoff. Mott transition in one dimension: Benchmarking dynamical cluster approaches. Phys. Rev. B , 77:045133, 2008.
- [11] G. Kotliar and A. E. Ruckenstein. New functional integral approach to strongly correlated fermi systems: The Gutzwiller approximation as a saddle point. Phys. Rev. Lett. , 57:1362-1365, 1986.
- [12] Z. Zou and P. W. Anderson. Neutral fermion charge-e boson excitations in the resonating-valencebond state and superconductivity in La CnO -based compounds. 2 4 Phys. Rev. B , 37:627, 1988.
- [13] S. Ostlund and M. Granath. ¨ Exact transformation for spin-charge separation of spin-1 / 2 fermions without constraints. Phys. Rev. Lett. , 96:066404, 2006.
- [14] P. A. Lee, N. Nagaosa, and X.-G. Wen. Doping a Mott insulator: Physics of high-temperature superconductivity. Rev. Mod. Phys. , 78:17-85, 2006.
- [15] K. Scharnhorst and J.-W. van Holten. Nonlinear Bogolyubov-Valatin transformations: Two modes. Ann. Phys. , 326(11):2868 - 2933, 2011.
- [16] P. Ribeiro, P.D. Sacramento, and M.A.N. Ara´ ujo. U(1) slave-particle study of the finitetemperature doped Hubbard model in one and two dimensions. Ann. Phys. , 326:1189-1206, 2011.
- [17] L. Isaev, G. Ortiz, and J. Dukelsky. Hierarchical mean-field approach to the J 1 -J 2 Heisenberg model on a square lattice. Phys. Rev. B , 79:024409, 2009.
- [18] D. Huerga, J. Dukelsky, and G. E. Scuseria. Composite boson mapping for lattice boson systems. Phys. Rev. Lett. , 111:045701, 2013.
- [19] E. Altman and A. Auerbach. Plaquette boson-fermion model of cuprates. Phys. Rev. B , 65:104508, 2002.
- [20] J.-P. Blaizot and G. Ripka. Quantum Theory of Finite Systems . The MIT Press, Cambridge, MA, 1986.
- [21] P. Pulay. Improved SCF convergence acceleration. J. Comput. Chem. , 3:556-560, 1982.
- [22] J. E. Hirsch and S. Tang. Antiferromagnetism in the two-dimensional Hubbard model. Phys. Rev. Lett. , 62:591-594, 1989.
- [23] G. Sordi, K. Haule, and A.-M. S. Tremblay. Mott physics and first-order transition between two metals in the normal-state phase diagram of the two-dimensional Hubbard model. Phys. Rev. B , 84:075161, 2011.
|
10.1088/0953-8984/26/45/455601
|
[
"Jinmo Zhao",
"Carlos A. Jimenez-Hoyos",
"Gustavo E. Scuseria",
"Daniel Huerga",
"Jorge. Dukelsky",
"Stefan. M. A. Rombouts",
"Gerardo Ortiz"
] |
2014-08-05T16:09:40+00:00
|
2014-09-06T14:45:37+00:00
|
[
"cond-mat.str-el"
] |
Composite fermion-boson mapping for fermionic lattice models
|
We present a mapping of elementary fermion operators onto a quadratic form of
composite fermionic and bosonic operators. The mapping is an exact isomorphism
as long as the physical constraint of one composite particle per cluster is
satisfied. This condition is treated on average in a composite particle
mean-field approach, which consists of an ansatz that decouples the composite
fermionic and bosonic sectors. The theory is tested on the one- and
two-dimensional Hubbard models. Using a Bogoliubov determinant for the
composite fermions and either a coherent or Bogoliubov state for the bosons, we
obtain a simple and accurate procedure for treating the Mott insulating phase
of the Hubbard model with mean-field computational cost.
|
1408.1015v3
|
## Inferring Synaptic Structure in presence of Neural Interaction Time Scales
Cristiano Capone 1 2 5 , , , ∗ , Carla Filosa , Guido Gigante 1 2 3 , , Federico Ricci-Tersenghi 1 4 5 , , , Paolo Del Giudice 2 5 ,
- 1 Physics Department, Sapienza University, Rome, Italy
- 2 Italian Institute of Health, Rome, Italy
- 3 Mperience s.r.l., Rome, Italy
- 4 IPCF-CNR, UOS Roma, Italy
- 5 INFN, Sezione di Roma 1, Italy
- ∗ E-mail: [email protected]
## Abstract
Biological networks display a variety of activity patterns reflecting a web of interactions that is complex both in space and time. Yet inference methods have mainly focused on reconstructing, from the network's activity, the spatial structure, by assuming equilibrium conditions or, more recently, a probabilistic dynamics with a single arbitrary time-step. Here we show that, under this latter assumption, the inference procedure fails to reconstruct the synaptic matrix of a network of integrate-and-fire neurons when the chosen time scale of interaction does not closely match the synaptic delay or when no single time scale for the interaction can be identified; such failure, moreover, exposes a distinctive bias of the inference method that can lead to infer as inhibitory the excitatory synapses with interaction time scales longer than the model's time-step. We therefore introduce a new two-step method, that first infers through cross-correlation profiles the delay-structure of the network and then reconstructs the synaptic matrix, and successfully test it on networks with different topologies and in different activity regimes. Although step one is able to accurately recover the delay-structure of the network, thus getting rid of any a priori guess about the time scales of the interaction, the inference method introduces nonetheless an arbitrary time scale, the time-bin dt used to binarize the spike trains. We therefore analytically and numerically study how the choice of dt affects the inference in our network model, finding that the relationship between the inferred couplings and the real synaptic efficacies, albeit being quadratic in both cases, depends critically on dt for the excitatory synapses only, whilst being basically independent of it for the inhibitory ones.
## Introduction
The attempt to infer synaptic connectivity from correlations between neural activities has a long history (see, e.g. , [1] and references therein; for recent developments, see, e.g. , [2]). In this context, a long recognized problem is that, since the real neuronal network is dramatically under-sampled by electrophysiology experiments, one cannot remove ambiguities as to whether observed correlations depend on direct synaptic connections or on indirect loops through unobserved components of the network. The advent of multi-electrode arrays bringing the number of simultaneously recordings to several tens (both in-vitro , e.g. MEA, ad in-vivo , e.g. Utah arrays) does not resolve the issue, since still a tiny fraction of the neural population can be sampled; however, it does offer an option for the reconstruction of some forms of effective synaptic connectivities and opens the possibility to address questions which were previously out of reach.
For instance, in the seminal work by Schneidman et al. [3], based on the so-called Inverse Ising inference method which is the substrate of the present work, it was possible to assess the share of network information accounted for by pairwise correlations. Such models were based on maximum entropy estimates which, under the assumptions of pairwise interactions, provide couplings and external fields for the Gibbs equilibrium distribution of an equivalent Ising model, and do not possess any inherent time scale. The obvious interest in relaxing the assumption of equilibrium for modeling neural data, led to the development of inference methods based on a kinetic formulation of the equivalent Ising system, that results in maximum-likelihood estimations of the transition probabilities between subsequent states of the system; in this way, one can account for non-stationary neural data, and non-symmetric synaptic couplings [4-6]. Compared to other methods to establish statistical models of multiple recordings, such inference methods, rooted in analogies with the statistical physics of complex systems, claim to afford an easier link with biologically meaningful quantities [5,7]. We also remark that Kinetic Ising inference methods can be seen as a special case of Generalized Linear Models [8-10], with a one-step time kernel.
We consider networks of spiking, integrate-and-fire neurons, sparsely coupled through excitatory and inhibitory synapses. The sampled spike trains were binarized by choosing a time-bin such that two or more spikes fell in the same time-bin with negligible probability. This we regard as a minimal requirement, not to loose information about correlations at the single spike level. Beyond this requirement, we study the impact of the chosen time-bin on the quality of the inference procedure to estimate synaptic couplings, by checking the results of an established method based on the Kinetic Ising Model, for a network of spiking neurons, and illustrate its limitations. We therefore introduce a two-step method to first estimate, for each sampled neurons pair, a characteristic time scale of their interaction (spike transmission), and then to use such estimated time scales as time-lags in the modified Kinetic Ising Model we propose. Finally we estimate analytically, and verify numerically, the relationship between true and inferred synapses, and its dependence on
the time-bin dt chosen to binarize the data.
## Results
We simulated sparsely coupled networks of spiking, integrate-and-fire neurons, interacting through excitatory and inhibitory synapses. The sampled spike trains were binarized by choosing a time-bin dt such that the probability two or more spikes falling in the same time-bin was negligible; apart from this requirement, at this stage the choice of the time-bin was arbitrary. We denote the spike train of neuron i by S i ( ), where t S i ( ) = 1 if t a spike is recorded in bin t and S i ( ) = 0 otherwise. t We assume that the data have been generated by the Ising model evolving in accordance with the Glauber dynamics [4,6,11], so that, at each time-step, S i ( t + dt ) is sampled according to the probability distribution:
$$P ( S _ { i } ( t + d t ) | \mathbf S ( t ) ) = \frac { \exp \left [ S _ { i } ( t + d t ) H _ { i } ( t ) \right ] } { 1 + \exp H _ { i } ( t ) } \quad \text{with } H _ { i } ( t ) = h _ { i } + \sum _ { j } J _ { i j } S _ { j } ( t ) \quad \ \ ( 1 )$$
that depends on the total 'field' H t i ( ) felt by neuron i , and generated by all the neurons in the network through the synaptic matrix J . Being H t i ( ) a function of the state of the system at time t only, the dynamics in Eq. (1) is Markovian. To infer the best parameters J and h , following [6] (see Methods), we resorted to a mean-field approximation of the gradient ∂L/∂J ij of the likelihood of the data being generated by the model; under this approximation, the equations ∂L/∂J ij = 0 are linear in J and are then easily invertible. It will be shown in Methods that the following relation holds:
$$J _ { I n f, i j } \simeq \frac { \langle S _ { i } | S _ { j } = 1 \rangle - m _ { i } } { m _ { i } \left ( 1 - m _ { i } \right ) \left ( 1 - m _ { j } \right ) }$$
where m i = 〈 S i 〉 and 〈 S S i | j = 1 〉 denotes the conditional probability that neuron i fires in the same time-bin dt in which it receives a spike from pre-synaptic neuron j .
## Role of the time-bin for the quality of inference
In Fig. 1a-c, we show the inference results for three choices of the time-bin dt (1 , 3 , 10 ms), for a (purely excitatory) network of N = 50 neurons with equal spike transmission delay δ = 3 ms for all connected neuron pairs. In each panel we plot the histograms of inferred synapses, separately for those corresponding to actually existing synaptic contacts (30% out of N N ( -1)), and those corresponding to unconnected pairs. It is seen that the expected separation between the peak around zero (corresponding to unconnected pairs) and the histogram corresponding to existing synapses is acceptable only for dt = 3 ms, that is when the time-bin dt equals the delay δ .
Fig. 1d gives a quantitative representation of the inference quality in terms of ROC curves for the three choices of the time-bin dt . The fraction of correctly identified existing
Fig. 1 Inference results for the Kinetic Ising Model with three choices of the time-bin dt (1 , 3 , 10 ms), on spike trains generated by a network of 50 neurons, sparsely connected with probability c = 0 3; actual synapses have all equal efficacy . J ij = 0 9 mV and equal spike transmission . delay δ = 3 ms. In panels a-c we plot the histograms of inferred synaptic couplings, for existing synaptic contacts (solid lines) and for those corresponding to unconnected pairs (dashed lines, roughly centered around 0). The separation between the two histograms is acceptable only for dt = δ = 3 ms (top-right panel); also note, for dt = 1 ms, how the solid line peaks around a negative inferred value: the excitatory synapses are in fact inferred as inhibitory - see text for a discussion of this puzzling result. In panel d the ROC curves corresponding to the three choices of dt are presented; the fraction of correctly identified existing synapses ('true positive' T P /P ) against the fraction of correctly identified unconnected pairs ('true negative' T N /N ) is plotted for moving discrimination threshold: the dashed line, corresponding to dt = 3 ms, clearly surmounts the other two, to all effects allowing for a perfect discrimination. Neurons fire at about 50 Hz; the total recording length is 500 s.
synapses ('true positive') against the fraction of correctly identified unconnected pairs ('true negative') is plotted parametrically varying a discrimination threshold: the dashed
line, corresponding to dt = 3 ms, clearly surmounts the other two, to all effects allowing for a perfect discrimination. A puzzling feature can be recognized when comparing the plots in Fig. 1: the histograms corresponding to existing (and always excitatory) synapses are centered around different values, depending on dt : in particular, for dt < 3 ms they appear to have been estimated as inhibitory synapses. We will come back to this point later, when explaining results in Fig. 2.
Fig. 2 Dependence of the inference quality on the choice of the time-bin dt , for a purely excitatory network (same parameters as in Fig. 1) with a distribution of synaptic delays δ between 0 1 . and 20 ms (see Methods). Panels a-c show the distribution of inferred synaptic couplings for three values of dt , including dt = 7ms glyph[similarequal] 〈 δ 〉 ; solid (dashed) lines are the distributions of inferred couplings for existing (non-existing) synapses. Panel d shows the inferred coupling J Inf against the associated delay δ , for dt = 7 ms. Neurons fire at about 50 Hz; the total recording length is 500 s.
We next considered the more interesting case in which delays δ were exponentially distributed between 1 and 20 ms (see Methods), for three values of the time-bin dt ,
including dt = 7ms glyph[similarequal] 〈 δ 〉 (Fig. 2a-c). The quality of the inference stays poor in all cases. Interestingly, a bimodal histogram appears for existing synapses for dt glyph[lessorsimilar] 〈 δ 〉 , the two peaks being associated with dt < δ (left peak, analogous to the one seen in Fig. 1a and dt > δ (right peak). To further investigate this, in Fig. 2d, we plot the inferred coupling J Inf,ij against the associated delay δ , for dt = 7 ms. A clear non-monotonic behavior is observed. The inferred J Inf is maximum when its associated delay δ equals dt . This can be understood observing that, for δ = dt , an (excitatory) pre-synaptic spike emitted during time-bin t will always reach neuron i in the following time-bin t + dt , thus maximally contributing to the conditional probability 〈 S i ( t + dt ) | S j ( ) = 1 t 〉 in Eq. (2), and therefore to J Inf ; for δ < dt such probability roughly scales with δ/dt , assuming uniform probability of pre-synaptic spike occurrence in each time-bin dt (see the initial rising ramp in Fig. 2d). Analogously, for δ = 2 dt = 14 ms, the conditional probability 〈 S i ( t + 2 dt ) | S j ( ) = 1 t 〉 will be maximum; since the probability for the post-synaptic neuron to fire in adjacent time-bins is negligible, this implies that 〈 S i ( t + dt ) | S j ( ) = 1 t 〉 will attain a minimum, thus explaining the negative peak at δ = 2 dt . The downward ramp for dt < δ < 2 dt linearly interpolates between the two peaks, confirming the expectation that the probability of firing two time-bins after the pre-synaptic spike roughly increases as ( δ -dt /dt ) , as the probability of firing in the previous bin decreases by the same amount. Beyond δ = 2 dt , the conditional probability 〈 S i ( t + dt ) | S j ( ) = 1 t 〉 stays below m i ( J Inf < 0), approaching asymptotically this value for δ greater than the average post-synaptic inter-spike interval (which is roughly 20 ms for the data shown in Fig. 2).
From the above discussion it is then clear that a positive real synaptic efficacy can result in a positive or negative inferred coupling depending on the relationship between δ and the time-bin dt ; this explains the already mentioned negative portion of the histograms shown in Fig. 1a-c.
## An extended Kinetic Ising Model with a distribution of synaptic delays
The poor results obtained in Fig. 2 motivated us to extend the model to account for a distribution of delays δ ij :
$$H _ { i } ( t ) = h _ { i } ( t ) + \sum _ { j } J _ { i j } S _ { j } ( t - \delta _ { i j } )$$
Instead of attempting to maximize the log-likelihood of the model on the data to infer δ ij , we devised an alternative way to estimate them from the observed neural activity, and insert them as fixed parameters in the maximum-likelihood inference of the couplings J ij .
The procedure is based on the intuition that the time-retarded cross-correlation, D ij ( τ ) (see Eq. (14)), between the activities of a given pair ij of connected neurons should peak at a time-lag τ close to the actual synaptic delay δ ij ; the peak is expected to be positive or negative depending on the synapse being excitatory or inhibitory, respectively. This
intuition is confirmed in Fig. 3, in which such correlation is reported for one pair of neurons connected by an excitatory synapse (black solid line), one pair of disconnected neurons (dashed line) and one pair of neurons connected by an inhibitory synapse (solid grey line); for both the existing synapses the delay is δ ij = 3 ms.
Fig. 3 Dependence of cross-correlation D ij on τ for a fixed pair of neurons ( i, j ). Excitatory synapse (thick solid line, synaptic delay δ = 3 ms), inhibitory synapse (thin solid line, synaptic delay δ = 3 ms), no synapse (dashed line), in a network with N E = 25 excitatory neurons, N I = 25 inhibitory neurons, connection probability c = 0 1, . J ij = ± 0 54 mV, . δ ij exponentially distributed from 1 ms to 20 ms (see Methods).
As expected, D ij ( τ ) fluctuates around zero for disconnected neurons, showing sharp positive or negative peaks at τ glyph[similarequal] 3 ms, for excitatory and inhibitory transmission, slowly decaying after the peak. This latter feature is explained considering that (taking the solid black curve as an example), for τ > δ ij an excitatory spike, which was ineffective to trigger a post-synaptic spike, still depolarized the membrane, thereby increasing the firing probability on a time scale comparable with the membrane time constant. On the other hand, for τ < δ ij , the 'absence' of the contributed excitatory pre-synaptic spike, lowers (a bit, but for a time of the order of the average inter-spike interval of the pre-synaptic neuron) the firing probability of the post-synaptic neuron. Including the analogous argument for inhibitory synapses, in summary one expects that for τ < δ ij the cross-correlation stays slightly above zero for inhibitory synapses, and slightly below zero for excitatory ones, both for a time comparable to the average inter-spike interval.
The above argument also helps explaining some highlighted features in Figs. 1 and 2. The standard procedure there used rests on the computation of the conditional probabilities 〈 S i ( t + dt ) | S j ( ) t = 1 , or equivalently (see Eq. (14)) of the time-retarded correlations 〉
D ij ( τ = dt ); with reference to the discussion at the end of the previous section, it is understood that, for a time-bin dt smaller than the actual delay, the negative correlation implied by the above discussion would result in an excitatory synapse being inferred as an inhibitory one.
In Fig. 4 we show results of the above two-steps procedure for a network with both excitatory and inhibitory synapses. In the left panel, we show inferred synaptic couplings, marking with different line styles the values inferred for existing synapses (solid black for excitatory, solid grey for inhibitory ones), and disconnected neuron pairs (dashed line). Two peaks, centered around positive and negative values respectively, emerge for existing excitatory and inhibitory synapses; both peaks have null or minimal overlap with the histogram for disconnected pairs. The latter give rise to two peaks almost symmetric around zero, contrary to what observed in Figs. 1 and 2. This feature derives from the procedure for inferring the synaptic delays, where we choose the lag τ corresponding to the largest absolute value of the time-retarded cross-correlation D ij ( τ ); such choice leads to the inference of statistically non-null synaptic couplings, even with the noisy cross-correlation found for unconnected neurons. However, the two peaks related to unconnected neuron pairs are expected to shift towards zero for longer recordings, according to the scaling appropriate for extreme value statistics.
Fig. 4 Inference results for the Kinetic Ising Model with time delays for a network of excitatory and inhibitory neurons (same parameters as in Fig. 3), obtained with time-bin dt = 1 ms. Left panel: histograms of inferred synaptic couplings, for excitatory (black solid line) and inhibitory (grey solid line) synaptic contacts, and for those corresponding to unconnected pairs (dashed line). Right panel: delays inferred from the cross-correlation peaks vs actual delays for existing synapses; the dashed line marks the identity line. Neurons fire at about 50 Hz; the total recording length is 500 s.
Fig. 4, right panel, shows the delays inferred from the cross-correlation peaks vs actual delays for existing synapses; a very good match is appreciable over the extended range covered by the exponential distribution (1 -20 ms). It is worth noticing that inferred
time delays are never smaller than true ones, while they can be slightly larger due to the noise in the data and their inherent discretization as multiples of the time-bin dt .
## Relation between true and estimated synaptic efficacy, and its dependence on the time-bin
We notice from Fig. 4 that, although the inference procedure successfully identifies excitatory and inhibitory synapses, the corresponding distributions are centered around values of different module, while the excitatory and inhibitory synaptic efficacies were chosen in the simulation to have equal absolute values. It would be of course interesting to give a theoretical account of such asymmetry, which would also allow one to remap the inferred values onto quantitatively reliable estimates of the real synaptic efficacies.
For excitatory synapses, an intuitive argument can hint at a strategy of computation. Starting from Eq. (2), the probability 〈 S S i | j = 1 〉 that neuron i will fire at time t ( S i ( ) = 1) t upon receiving a spike from neuron j ( S j ( t -δ ij ) = 1) is roughly equal to the probability that the membrane potential V ( ) of the post-synaptic neuron is at a distance less than t the synaptic efficacy J ij > 0 from the firing threshold θ ; such probability is the integral, between θ -J ij and θ , of the probability density p V ( ) of the membrane potential; assuming that the rest of the network, and possibly external sources, provide a noisy input (of infinitesimal variance σ 2 ) such that the diffusion approximation holds [12], the threshold is an absorbing barrier for the stochastic process V i ( ), t and this implies that p θ ( ) = 0. Therefore, in stationary conditions, expanding p V ( ) close to θ (indeed, consistently with the diffusion approximation, J ij glyph[lessmuch] θ -V rest , being V rest the equilibrium value of the membrane potential absent any external input):
$$\langle S _ { i } | S _ { j } = 1 \rangle \ \simeq \ \int _ { \theta - J _ { i j } } ^ { \theta } p ( v ) \, d v \simeq \int _ { \theta - J _ { i j } } ^ { \theta } p ^ { \prime } ( \theta ) \, ( v - \theta ) \, d v = \\ - \frac { 2 \nu } { \sigma ^ { 2 } } \int _ { \theta - J _ { i j } } ^ { \theta } ( v - \theta ) \, d v = \frac { \nu J _ { i j } ^ { 2 } } { \sigma ^ { 2 } }$$
where we have used, for the stationary average firing rate, the equivalence ν = -( σ / 2 2) p ′ ( θ ). Inserting this result into Eq. (2), then we have J Inf ∼ J 2 True /dt ; thus (for J ij > 0) the relationship is quadratic and divergent with decreasing time-bin dt .
In Methods this rough derivation is refined to take into account afferent external spikes to neuron i in a single time-bin dt , that can make the neuron fire for V < θ -J ij , and even when J ij < 0 (inhibitory synapse); due to the fact that for the latter case neuron i will fire only when external spikes overcompensate the inhibitory synaptic event, the found scaling (Eq. (26)) is different: still quadratic in J ij , but constant in dt We remark that the found relationship for inhibitory synaptic couplings basically gives J Inf glyph[greaterorsimilar] -1 / 2 whenever J ij < -J ext where J ext is the synaptic efficacy of synapses from external neurons (see
Eq. (27)); thus the inference method loses sensitivity for | J ij | approaching J ext ; moreover, since the J Inf are estimated from the simulated data and one can actually have J Inf < -1 / 2 because of noise in the estimate, no values J True can be inferred in these cases.
In the following, we call J Estimated True the value obtained by inverting the relationship between J Inf and J True . Fig. 5 shows the histograms of J Estimated True for four values of the time-bin dt , for the same network of Fig. 4. Since J True = ± 0 54 mV in this network, . the expectation is to find two peaks, one for excitatory and one for inhibitory synapses, around these two values. This expectation is substantially confirmed by the shown results for excitatory synapses, with a slight gradual worsening for increasing time-bin dt ; such worsening can be understood by noting that the relationship Eq. (24), valid for excitatory synapses, is derived for dt → 0. We notice that, for instantaneous synaptic transmission, for J ij > 0 every incoming pre-synaptic spike from neuron j has a probability of making neuron i fire that does not vanish as dt → 0; this is not true for inhibitory synapses, for which what matters is the number of Poisson excitatory events in the time-bin dt (from other neurons) needed to overcompensate the effect of an inhibitory pre-synaptic spike in the same time-bin dt . The probability of having this number in dt vanishes with dt , thereby making the number of observed post-synaptic spikes following a pre-synaptic spike vanish with dt ; the correlation-based estimates J Inf are therefore affected by greater noise. The noise broadens the distribution of J Inf and populates a tail of un-physical values J Inf < -1 2 , as discussed above, which contributes to the grey bar in the shown histograms.
The above considerations highlight the existence of a trade off in the choice of the time-bin dt , if one wants to derive a reliable estimate of true excitatory and inhibitory synaptic values from the ones inferred through the Kinetic Ising Model. A too small time-bin dt does not allow to infer inhibitory synapses, while a too large dt introduces a systematic bias in the estimated synaptic efficacies.
To illustrate the predicted scaling of J Inf vs J with respect to the time-bin dt , we performed simulations with a uniform distribution of synaptic efficacies and three values of dt . In Fig. 6, left panel, we plot the inferred synaptic couplings against the real synaptic efficacies. All the predicted features are nicely matched: the quadratic dependence J Inf ∼ J 2 , the strong dependence on dt for excitatory synapses (divergence for dt → 0), and the approximate independence of the inhibitory couplings from dt . The above theoretical predictions are used to rescale the inferred synaptic couplings, which are compared to the true ones in the right panel of Fig. 6. The validity of the approximation is confirmed by the collapse of the three curves on the main diagonal.
## Inference on a bursting network with spatial structure
So far, the simulated networks were uniformly sparsely connected, with no spatial structure; moreover, the neural activity was stationary and asynchronous. As a step towards checking the robustness of the method in more complex situations, we simulated a network with the following spatial structure. The N = 1000 excitatory neurons are subdivided into P = 100
dt=0.6ms
Fig. 5 Histograms of J Estimated True : theoretical estimates of J True from the inferred coupling J Inf for the same network of Fig. 4. The four panels show the histograms of J Estimated True (darker bars) for existing synapses only, for different choices of the time-bin dt ; for excitatory and inhibitory synapses Eq. (24) and Eq. (26) are applied, respectively. The two peaks already seen in Fig. 4, corresponding to excitatory and inhibitory neurons, are now expected to center around the true values J True = ± 0 54 mV. The light-grey bar in each histogram represents the number of . inhibitory synapses for which the found value of J Inf fell below the lower bound of the physical range, -1 2 < J Inf < 0, outside of which the relationship in Eq. (26) has no inverse solution; the bar is arbitrarily placed just on the left of the minimum value for J Estimated True obtainable by inverting Eq. (26).
populations, organized on a circle; the probability c αβ ( α, β = 1 , . . . , P ) that a neuron in population β is pre-synaptic to a neuron in population α is c αβ = 0 134 . e -d α, β ( ) 135 -d α, P/ ( 2) 26 1 . , where d α, β ( ) = min( | α -β , P | -| α -β | ) is the distance along the circle. Therefore, each population is more connected to its immediate neighbors, and some populations receive
Fig. 6 Relationship between inferred couplings and synaptic efficacies for varying time-bin dt . Left panel: J Inf vs J True . Right panel: J Estimated True vs J True . In both cases only non-zero synapses were included. Results are for a network of N E = 25 excitatory neurons and N I = 25 inhibitory neurons, sparsely connected with probability c = 0 1, and firing at about 20 Hz; the synaptic . efficacies are uniformly distributed in -0 54 mV . < J ij < 0 54 mV; . δ ij = 3 ms.
more pre-synaptic synapses than others; the average connection probability is 5%. Besides, the excitatory synaptic efficacies have Gaussian distribution.
The neurons are furthermore endowed with short-term synaptic depression, implemented according to the model in [13]. Such mechanism, mediating a self-inhibiting effect on the collective firing of the network, can generate short-lived bursts of activity, followed by periods of quiescence [14], in response to random fluctuations of the overall activity, analogously to what is often observed in neuronal cultures [15]. We emphasize that the network capable of spontaneously generated bursts is close to the instability of the low-activity asynchronous state, thus making the activity, even when restricted to the inter-burst periods, more correlated than the one of networks used in the preceding sections: for this more excitable network, correlated fluctuations (even in the low state) constitute a kind of global component of the activity that makes the contribution of the single synaptic contact harder to detect in the correlation functions. Fig. 7 shows an illustrative time course of the non-stationary, bursting neural activity from simulations.
In Fig. 8 we show the results of the inference procedure carried out on the subset of the 50 most active neurons in the network; the time record used for inference is of about 11 hours, with dt = 1 ms.
We note that for stationary pre-synaptic activity ν j , synaptic short-term depression effectively reduces J ij to J ij 〈 r ij 〉 = J ij / (1 + uτ ν r j ) (see Eq. (7)) where r ij is the fraction of synaptic resources available and τ r is its recovery characteristic time scale. We corrected J Estimated True by this factor.
The obtained inference is good for synaptic delays (not shown), while J Estimated True are clearly off target, with a large overestimation of synaptic efficacies. Besides, the right
Fig. 7 Illustrative time course of neuronal activity in a bursting network with spatially nonuniform synaptic connectivity and synaptic short-term depression (see text for details). Top panel: sample time course of the average firing rate. Bottom panel: raster plot from a sub-set of 100 neurons, corresponding to the time record of the top panel. The network comprises N = 1000 excitatory neurons ( J ij are drawn from a Gaussian distribution with mean 0 846 mV . and standard deviation of 0 211 mV), . δ ij exponentially distributed from 1 ms to 15 ms, and synapses are endowed with short-term depression with recovery time τ r = 800 ms and synaptic release fraction u = 0 2 (see Methods for details); the external current is a train of Poisson spikes . with average rate ν ext = 0 6 kHz, and synaptic efficacy randomly sampled from a Gaussian . distribution of mean J ext = 0 9 mV and standard deviation 0 225 mV. . .
Fig. 8 Inference results for the Kinetic Ising Model with time delays for the network described in Fig. 7. The inference is carried out on the 50 (out of 1000) most active neurons in the network, with dt = 1 ms, over a recording time of about 11 hours. Delays inferred from the cross-correlation peaks vs actual delays (not shown) for existing synapses have R 2 = 0 976 with . the identity line. Left panel: scatter plot of J Estimated True vs J True , where J Estimated True is computed using Eq. (24) (purely excitatory network) and then rescaled J Estimated ij → J Estimated ij (1+ uτ ν r j ), where ν j is the average firing rate of the pre-synaptic neuron, in order to compensate for the average synaptic depression induced by short-term depression (see Eq. (7)); the dashed line marks the identity line. Right panel: histograms of J Estimated True for existing synapses (solid line) and unconnected pairs (dashed lines); we remind that the J True have Gaussian distribution with 〈 J True 〉 = 0 846 mV. .
panel of Fig. 8 shows a large overlap between the J Estimated True for existing and non-existing synapses. We remark that, with respect to the hypothesis underlying the approximate, static mean-field formulation we adopted, the bursting regimes violates all of them: correlated fluctuations, non-stationarities, and large deviations from the average firing rates. We notice also that all coherent dynamic components contribute to overestimate the synaptic couplings.
We therefore performed an inference restricted to the inter-burst periods, considering this time 2 hours of simulation only. Such restriction is shown in Fig. 9 to greatly improve the quality of the inference. We remark that, despite the restriction to the low activity component of the dynamics, performing inference on this network remains a non-trivial test with respect to the results reported in previous sections. Indeed, we have spatial structure in the synaptic connectivity, and the network expresses its high self-excitability also in the statistics of the fluctuations in the low activity state.
Fig. 9 Inference results for the Kinetic Ising Model with time delays for the network described in Fig. 7; all the details are as in Fig. 8, with the exception that the inference procedure is carried out considering only the inter-burst (low activity) periods during a recording time of 2 hours only (about 1 / 5 of the recording length used in Fig. 8). Delays inferred from the cross-correlation peaks vs actual delays for existing synapses have R 2 = 0 944 with the identity line. .
## Discussion
The main focus of our work was to extend existing methods for inferring synaptic couplings, based on the Kinetic Ising Model, in order to incorporate a distribution of interaction time scales in the neural network dynamics, in their relationship with the time-bin dt used to discretize the data. In doing this, we derived analytically the relationship between the inferred couplings of the Ising model and the true synaptic efficacies of the spiking neural network generating the data.
The impact of the choice of the time-bin on the quality of the inference has been considered in the past. In the context of maximum entropy estimates in equilibrium models, the authors of [16] argued that the needed constraints on dt doom the method to poor performance for large networks. Also in [17], extending an analysis performed in [18], the authors studied the influence of the network size and time-bin on the quality of the inference, comparing the independent pair and naive mean field approximations with the results of the computationally demanding maximization of the likelihood through Boltzmann learning; the reported result was a decrease of the quality when increasing either the network size or the time-bin. While in the cited papers the model neurons used in simulations did have a characteristic time scale (associated with the kernel of the conductance dynamics) the analysis of the choice of the time-bin in relation with such a time scale, which we addressed here, was outside their scope.
We first checked the results of an established inference method based on the Kinetic Ising Model, showing that acceptable results are obtained only if the time-bin dt matches the time scale for the interaction between neurons (spike transmission delay in our case).
Then we have shown that, for the more realistic case of a distribution of time scales across the network, any choices of the time-bin provide poor results (to the point that, for example, excitatory synapses can be mis-inferred as inhibitory ones).
Motivated by the above observations, we devised a two-step method to first estimate, for each sampled neurons pair, a characteristic time scale of their interaction (spike transmission delay), and then to use such estimated time scales as pair-dependent timelags in the modified Kinetic Ising Model we introduce. The method can cope with wide distributions of time scales, these latter are reliably estimated, and synaptic couplings are inferred with good quality.
Even if the newly proposed inference method is made largely independent from the choice of the time-bin dt , the numerical values of the resulting inferred couplings still depend on dt , with noticeable differences between excitatory and inhibitory synapses. To resolve this issue, we studied in detail the stochastic equation controlling the evolution of the neuron's membrane potential, thus deriving an analytical expression relating inferred couplings to true synaptic efficacies. Such a relation turns out to be always quadratic, J Inf ∼ J 2 True , and does depend critically on dt for excitatory synapses. This allowed us to rescale the couplings J Inf to obtain a quantitatively good match with the true synaptic efficacies J True .
The analytical relations we derived hold for the specific neuronal-synaptic model we considered, that is the leaky integrate-and-fire neuron with instantaneous synaptic transmission. Nonetheless, as long as the considered time-bin dt is long w.r.t. the synaptic transmission times, and the membrane potential dynamics in the proximity of the spike emission is well approximated by an integrator with a firing threshold, then we do expect those expressions to give reasonable results. Such conditions are probably not too unrealistic for biological neurons and the fast components of the synaptic transmission. However, we remark that for real neural data a quantitative assessment of the relationship between inferred connections and synaptic strengths is still not resolved, and will probably require an incremental refinement of the inference procedure based on increasingly realistic models.
An attempt was made in [19] to infer from spikes data the synaptic efficacies of nonleaky integrate-and-fire neurons (not the couplings of an Ising model); this was done by directly maximizing the likelihood of the paths travelled by the membrane potential of a post-synaptic neuron between subsequent spikes it emits. For the sake of analytical tractability approximations were adopted (to our understanding, the most consequential being the absence of a leakage term in the neuron dynamics). Though we did not perform a systematic comparison between the two methods, which depends on several factors (possibly including the more or less noisy firing regime of the network), we did check in some cases of interest, including the one showed in Fig. 6, and the method proposed in [19] provides worse performance (in particular, the distribution of inferred values for true null synapses has a much larger overlap with the distribution of inferred values for truly non-zero synapses).
Among the limitations of the present work, it is worth stressing that we have used integrated-and-fire models with instantaneous synaptic transmission; a natural extension would be to consider models with a more realistic synaptic dynamics. However, we believe we took a step towards making inference in realistic settings, by understanding the role of inherent time scales in neural network dynamics.
## Materials and Methods
## Spiking simulations
We simulated networks of sparsely connected leaky integrate-and-fire neurons; the i -th neuron is modelled as a single dynamic variable, the membrane potential V i ( ), that follows t the dynamics:
$$\frac { \text{d} V _ { i } ( t ) } { \text{d} t } = - \frac { V _ { i } ( t ) - V _ { r e s t } } { \tau _ { m } } + \sum _ { j, n } J _ { i j } \delta ( t - t _ { j n } - \delta _ { i j } ) + J _ { e x t } \sum _ { n } \delta ( t - \tilde { t } _ { i n } )$$
where τ m = 20 ms is the membrane's time constant and V rest = -70 mV is the membrane's resting potential; J ij measures the (instantaneous, being δ t ( ) the Dirac's delta function) post-synaptic potential induced by a spike of neuron j on V i ; t jn is the time of firing of the n -th spike by neuron j , that induces a jump J ij in V i after a delay δ ij . J ext = 0 9 mV . models the post-synaptic potential caused by spikes coming from neurons not belonging to the network and collectively firing at a frequency ν ext = 1 kHz; the time ˜ t in of arrival of the n -th external spike on neuron i is randomly chosen such that ˜ t in -˜ t in -1 are exponentially distributed with mean 1 /ν ext , so that the count of external spikes in a time window ∆ t will follow a Poisson distribution of mean ν ext ∆ . Neuron t i will emit a spike at time t whenever V i ( t -) ≥ θ = -52 mV; upon this event, its membrane potential is instantaneously brought to a reset potential which we take equal to the resting value, V i ( t + ) = V rest , where it will stay unaffected for a time τ refractory = 2 ms (such condition effectively bounds the firing rate of the neurons below 500 Hz).
The synaptic efficacies J ij are chosen randomly so that (on average) a fraction 1 -c of them is 0 (that is, neuron j and neuron i are not connected), whereas with probability c they are drawn from a (continuous) probability distribution that depends solely on the excitatory or inhibitory nature of the pre-synaptic neuron j . In this paper, we use three different distributions: a delta function (all the synaptic efficacies have equal values), a uniform distribution between two extremal values ( J min , J max ), and a Gaussian distribution with given mean J and variance σ 2 J . When J ij is not zero, a transmission delay δ ij is sampled from a probability distribution that, in the reported simulations, is either a delta function (all the synapses share a single delay δ ) or an exponential distribution of mean δ and an offset δ min . The distribution is truncated at a maximum value δ max : if δ ij > δ max , then the value is re-sampled ( δ max is chosen so that a re-sampling is triggered on average 5% of the times). Note that δ , in this case, does not coincide with the average delay.
For the results reported in Figs. 7 to 9, the dynamics of the single neuron, Eq. (5), is complemented by a synaptic dynamics implementing a form of synaptic short-term depression [13]; each synaptic efficacy J ij is replaced by J ij r ij , where r ij represents the instantaneous fraction of available synaptic resources, and evolves according to:
$$\dot { r } _ { i j } = \frac { 1 - r _ { i j } } { \tau _ { r } } - u \, r _ { i j } \, \sum _ { n } \delta ( t - t _ { j n } - \delta _ { i j } )$$
where τ r = 800 ms and u = 0 2; thus each pre-synaptic spike depletes synaptic resources . by a fraction u ; synaptic resources, in turn, recover toward 1 (full availability) with a time constant τ r . Under the assumption of constant pre-synaptic firing rate ν j :
$$\langle r _ { i j } \rangle _ { s s } = \frac { 1 } { 1 + u \, \tau _ { r } \, \nu _ { j } }$$
Short-term synaptic depression has been proposed as an activity-dependent network self-inhibition promoting oscillatory or bursting behavior [14].
All the simulations have been run using the simulator described in [20], that implements an event-driven simulation strategy that does not make use of an integration time-step (that would represent an effective lower-bound for admissible binarization dt ), allowing to record spike events with arbitrary temporal resolution (within the allowed numerical precision).
## Inverse Kinetic Ising Model
Following [5], we work with time-binned spike trains under the assumption that, for the chosen time-bin dt , there is (almost) never more than one spike per bin. We denote the spike train of neuron i by S i ( ), t where S i ( ) = 1 if neuron t i fired a spike in bin t , and S i ( ) = 0 otherwise (note that in [5] the convention is, instead, t S i = ± 1). Thus the data we work with is a N × T binary 'spike matrix', where N is the total number of neurons considered and T is the number of time-bins. This representation of the data lends itself to formulating the problem in terms of an Ising model, more specifically, given the noisy and evolving nature of spike trains, to a stochastic dynamic formulation of it [4,6,11]. At each time-step, for every neuron i in the network we compute the total 'field':
$$H _ { i } ( t ) = h _ { i } + \sum _ { j } J _ { i j } S _ { j } ( t )$$
where h i is an external field. Then we let S i sample its next value from the probability distribution:
$$P ( S _ { i } ( t + d t ) | \mathbf S ( t ) ) = \frac { \exp \left [ S _ { i } ( t + d t ) H _ { i } ( t ) \right ] } { 1 + \exp H _ { i } ( t ) }$$
that depends on the state of the system only at the previous time-step (Markovian dynamics). Thus we are able to compute the likelihood that the probabilistic model generated the binarized data:
$$L [ S, J, h ] \ = \ \sum _ { t = 1 } ^ { T - 1 } \sum _ { \substack { i = 1 \\ r } } ^ { N } \ln ( P ( S _ { i } ( t + d t ) | \mathbf S ( t ) \} ) ) =$$
$$= \sum _ { i t } ^ { \cdot \cdot \cdot \cdot } \left [ \begin{matrix} \cdot \cdot \cdot \\ S _ { i } ( t + d t ) \, H _ { i } ( t ) - \ln \left ( 1 + \exp ( H _ { i } ( t ) ) \right ) \end{matrix} \right ]$$
$$\begin{matrix} ( 1 0 ) \\ ( 1 1 ) \end{matrix}$$
and, in principle, maximize it to obtain the 'best' parameters J and h . The maximization can be performed iteratively using the gradient:
$$\frac { \partial L } { \partial h _ { i } } \ & = \ \langle S _ { i } ( t ) \rangle _ { t } - \frac { 1 } { 2 } \langle 1 + \tanh \frac { H _ { i } ( t ) } { 2 } \rangle _ { t } \\ \frac { \partial L } { \partial J _ { i j } } \ & = \ \langle S _ { i } ( t + d t ) S _ { j } ( t ) \rangle _ { t } - \frac { 1 } { 2 } \langle ( 1 + \tanh \frac { H _ { i } ( t ) } { 2 } ) \ S _ { j } ( t ) \rangle _ { t }$$
or, in order to avoid computationally expensive iterations and following again [5], it is possible to use the mean-field equations:
$$m _ { i } = \frac { 1 } { 2 } \left [ 1 + \tanh \left ( \frac { h _ { i } + \sum _ { j } J _ { i j } m _ { j } } { 2 } \right ) \right ]$$
where m i = 〈 S i 〉 , and then, assuming that the fluctuations around mean values δS i ≡ S i -m i are small, to write:
$$\frac { \partial L } { \partial J _ { i j } } \simeq D _ { i j } ( d t ) - m _ { i } \left ( 1 - m _ { i } \right ) \sum _ { k = 1 } ^ { N } J _ { i k } \, C _ { k j }$$
where:
$$D _ { i j } ( \tau ) = \langle \delta S _ { i } ( t + \tau ) \delta S _ { j } ( t ) \rangle _ { t } \equiv m _ { j } \left ( \langle S _ { i } ( t + \tau ) | S _ { j } ( t ) = 1 \rangle - m _ { i } \right ) \quad \ \ ( 1 4 )$$
is the delayed correlation matrix that, for τ = 0, gives the connected correlation matrix:
$$C _ { i j } \equiv \langle \delta S _ { i } ( t ) \delta S _ { j } ( t ) \rangle _ { t } \equiv D _ { i j } ( 0 )$$
whose diagonal elements are C jj = m j (1 -m j ). Putting ∂L/∂J ij = 0, Eq. (13) becomes a set of linear equations that, through a simple matrix inversion, gives the synaptic matrix J ; then, inverting the (non-linear) equations (12), also the local static fields h i are inferred.
## Inverse Kinetic Ising Model with time delays
We extend the Kinetic Ising Model to account for a distribution of delays δ ij by re-writing the local field as:
$$H _ { i } ( t ) = h _ { i } + \sum _ { j } J _ { i j } S _ { j } ( t - \delta _ { i j } )$$
This extension does not formally modify the expression for the likelihood of the model on the data, once the new form of H i is taken into account, whereas the mean-field approximation of the gradient now reads:
glyph[negationslash]
$$& \frac { \partial L } { \partial J _ { i j } } \simeq D _ { i j } ( \delta _ { i j } ) - m _ { i } \left ( 1 - m _ { i } \right ) \sum _ { k = 1 } ^ { N } J _ { i k } \, D _ { k j } ( \delta _ { i j } - \delta _ { i k } ) = \\ = \ & D _ { i j } ( \delta _ { i j } ) - m _ { i } \left ( 1 - m _ { i } \right ) J _ { i j } \, C _ { j j } - m _ { i } \left ( 1 - m _ { i } \right ) \sum _ { k \neq j } J _ { i k } \, D _ { k j } ( \delta _ { i j } - \delta _ { i k } ) \quad \quad ( 1 7 )$$
glyph[negationslash]
where in the second passage, in the sum over k , we singled out the diagonal term that in general will dominate the remainder of the sum, since D kj ( τ ), for k = , will be typically j much smaller than C jj at every τ , and even more so whenever τ is chosen far from δ kj , where the cross-correlation attains its peak value (see Fig. 3; here 'far' is roughly intended with respect to the width of the peak of the cross-correlation); this latter case is the most probable, in the sum in Eq. (17), since the converse would imply δ ij glyph[similarequal] δ kj + δ ik , an event that is unlikely as long as the the typical values of the delays are larger than the width of the han the D ij peak.
Putting ∂L/∂J ij = 0, and defining a set of N matrices M ( ) i , with i = 1 , . . . , N , whose elements are
$$M _ { k j } ^ { ( i ) } \equiv D _ { k j } ( \delta _ { i j } - \delta _ { i k } )$$
Eq. (17) can be written as N sets, one for each row of the matrix J , of N linear equations:
$$D _ { i j } ( \delta _ { i j } ) = m _ { i } \left ( 1 - m _ { i } \right ) \sum _ { k = 1 } ^ { N } J _ { i k } \, M _ { k j } ^ { ( i ) }$$
that allow, for each post-synaptic neuron i , by inverting the matrix M ( ) i , to infer the synaptic couplings with its pre-synaptic neurons. The local fields h i are found, as above, by then inverting Eq. (12).
## Relationship between synaptic efficacies and inferred couplings
In order to remap the inferred values onto quantitatively reliable estimates of the synaptic efficacies, we start by putting ∂L/∂J ij = 0 in Eq. (17) and noting that, by entirely
glyph[negationslash]
neglecting the typically small (see above) k = j terms in the sum, we can approximate the inferred J ij as:
$$\Lambda _ { J } J _ { I n f, i j } \simeq \frac { D _ { i j } } { C _ { j j } \, m _ { i } \, ( 1 - m _ { i } ) } = \frac { \langle S _ { i } | S _ { j } = 1 \rangle - m _ { i } } { m _ { i } \, ( 1 - m _ { i } ) \, ( 1 - m _ { j } ) }$$
where, in the second passage, we made use of Eq. (14). For reference, having binarized the spike trains with a time-bin dt , we can write the average 'magnetization' m i = ν dt i , where ν i is the average spike-frequency of neuron i . Therefore, we want to estimate the probability that neuron i will fire ( S i = 1) upon receiving a spike from neuron j ( i.e. , conditioned on the event S j = 1).
To start, we note that in Eq. (5) the current felt by neuron i in the network can be well approximated, over time scales of the order of the membrane's time constant τ m , with a Gaussian memoryless stochastic process, identified by a given infinitesimal mean µ and variance σ 2 . This diffusion approximation is warranted by the small size of the synaptic efficacies and the high frequency of the incoming spikes [12]. Under such approximation, Eq. (5) describes an Ornstein-Uhlenbeck process with an absorbing barrier [21], whose probability distribution p V, t ( ) obeys the boundary condition:
$$p ( \theta, \, t ) = 0$$
Moreover, it can be shown that the instantaneous emission rate of the neuron is given by:
$$\nu ( t ) = - \frac { \sigma ^ { 2 } } { 2 } \, \frac { \partial p ( V, \, t ) } { \partial V } \Big | _ { V = \theta }$$
glyph[negationslash]
Now when a spike from neuron j arrives at neuron i at a (random) time t + glyph[epsilon1] within the current time-bin ( t, t + dt ) (0 ≤ glyph[epsilon1] ≤ dt ), it produces a sudden jump J ij = 0 in the membrane potential V i ; glyph[epsilon1] is assumed to have uniform probability distribution between 0 and dt , since the only information is the arrival of the spike at some time during dt . In the same time-bin dt , the neuron also receives external spikes and recurrent spikes from other neurons in the network; in the following, we will assume that both the external and the internal contributions to the input current to the post-synaptic neuron are Poissonian trains; to simplify the discussion, we will assume that all the incoming spikes come through synapses with synaptic efficacy J ext and with a total frequency ν ext : such assumption is easily relaxed to account for pre-synaptic neurons (other than j ) having different synaptic efficacies J ik and firing frequencies. Finally, in the time interval glyph[epsilon1] preceding the arrival of the spike from pre-synaptic neuron j , we assume that the probability for neuron i to fire is the baseline probability ν glyph[epsilon1] i ; knowing that a pre-synaptic spike from neuron j is due at a given time, indeed, will in general offset this probability, but here we are neglecting this information, assuming the effects are small. The combined contribution of the baseline firing probability, the pre-synaptic spike S j = 1, and the external spikes, in the time-bin dt , can be estimated as follows.
When J ij > 0, then instantaneously the whole right tail ( θ -J ij < V < θ ) of p V, t ( ) will pass the threshold; in addition, we should consider the contribution from V < θ -J ij
deriving from realizations of the neuron i process that will cross the threshold upon receiving external (excitatory) spikes in the following ( t + glyph[epsilon1], t + dt ) interval. For k incoming external spikes, the whole interval of V between θ -J ij and θ -J ij -k J ext will contribute to the firing probability; an additional element to be taken into account is a drift term due to the leaky dynamics of V Eq. (5):
$$\text{due to the leaky dynamics of $V$ Eq. (5)$} \colon & \quad \begin{array} { c } \langle S _ { i } | S _ { j } = 1 \rangle = \frac { 1 } { d t } \int _ { 0 } ^ { d t } d \epsilon \left [ \nu _ { i } \, \epsilon + & ( 2 3 ) \\ & \sum _ { k \geq 0 } ^ { \infty } \text{Poisson} [ k | \nu _ { e x t } \, ( d t - \epsilon ) ] \, \langle \int _ { \theta _ { k } \, J _ { e x t } - J _ { i } - \mu _ { 0 } \, \bar { t } _ { k } \, ( d t - \epsilon ) } ^ { \theta } p ( v, t ) \, d v \rangle _ { \bar { t } _ { k } } \right ] \simeq \\ & \simeq & \ \frac { \nu _ { i } \, d t } { 2 } + \frac { 1 } { d t } \int _ { 0 } ^ { d t } d \epsilon \left [ \frac { 2 \, \nu _ { i } } { \sigma ^ { 2 } } \, \text{Poisson} [ 0 | \nu _ { e x t } \, ( d t - \epsilon ) ] \int _ { 0 } ^ { J _ { i j } } x \, d x + \\ & + \, \text{Poisson} [ 1 | \nu _ { e x t } \, ( d t - \epsilon ) ] \, \frac { 2 \, \nu _ { i } } { \sigma ^ { 2 } } \, \langle \int _ { 0 } ^ { J _ { i j } + J _ { e x t } + \mu _ { 0 } \, \bar { t } _ { 1 } \, ( d t - \epsilon ) } x \, d x \rangle _ { \bar { t } _ { 1 } } \right ] \\ \text{where, in the second passage, we made use of $Eq. (22)$ to approximate $p(V, t)$ close to} \end{array}$$
where, in the second passage, we made use of Eq. (22) to approximate p V, t ( ) close to the threshold θ , and neglected the Poisson terms for k > 1, that are order dt 2 or higher; µ 0 is a constant approximation of the leakage term in the V dynamics (Eq. (5)) close to the threshold, and 0 ≤ ˜ t k ≤ 1 is a random number representing the time of arrival of the k -th external spike in a time window (given that exactly k external spikes are due in the time window); it is easy to show that 〈 ˜ t k 〉 = k k +1 . Combining this expression with Eq. (20) and keeping only the leading terms in dt (averaging over ˜ t 1 has influence at order dt 2 or higher), we have:
$$J _ { I n f, i j } \simeq \frac { J _ { i j } ^ { 2 } } { \sigma ^ { 2 } \, d t } + \frac { J _ { i j } ^ { 2 } \left ( \nu _ { i } + \nu _ { j } \right ) + J _ { i j } \, J _ { e x t } \, \nu _ { e x t } } { \sigma ^ { 2 } } + O ( d t ) \quad \text{if } J _ { i j } > 0$$
To derive this expression we furthermore used the equality σ 2 = J 2 ext ν ext that holds when the input current is comprised of a single Poissonian train of impulses; as stated above, it is easy to relax this assumption and to generalize the result for the case of many trains with different frequencies ν ext,k and synaptic efficacies J ext,k .
Thus, for J ij > 0, the value of the inferred synaptic coupling J Inf,ij depends quadratically on the real value J ij , and critically on dt ( J Inf ∼ 1 /dt ); this latter dependence derives from the fact that the contribution to 〈 S S i | j = 1 〉 from the spike of pre-synaptic neuron j is order 1 (Poisson[0 | ν ext dt ] glyph[similarequal] 1 for dt → 0), whereas at the same time, in Eq. (20), the denominator vanishes with dt ( m i (1 -m i ) glyph[similarequal] ν dt i ).
When J ij < 0, the only chance for neuron i to fire after the arrival of a spike from neuron j at time t + glyph[epsilon1] is exclusively given by the probability that one or more external
spikes will compensate for the sudden negative jump J ij , making V i pass the threshold θ :
$$& \langle S _ { i } | S _ { j } = 1 \rangle = \frac { 1 } { d t } \int _ { 0 } ^ { d t } \mathrm d \epsilon \left [ \nu _ { i } \, \epsilon + \\ & \sum _ { k \geq 1 } ^ { \infty } \text{Poisson} ( k | \nu _ { e x t } \left ( d t - \epsilon \right ) ) \, \langle \int _ { \theta - k \, J _ { e x t } + | J _ { i j } | - \mu _ { 0 } \, \tilde { t } _ { k } \left ( d t - \epsilon \right ) } ^ { \theta } p ( v, \, t ) \, \mathrm d v \rangle _ { \tilde { t } _ { k } } \right ]$$
It is important to note that, if | J ij | > J ext , the term with k = 1 will vanish (a single external spike won't suffice to compensate for the negative jump J ij ); if | J ij | > 2 J ext , the second term too will disappear (not even two external spikes will be enough) and so on.
Now let's assume that | J ij | < J ext , so that all the terms in the sum are non-zero and the sum will be dominated, for small dt , by the first term only; then, reasoning as above for J ij > 0, we have:
$$J _ { I n f, i j } \simeq \frac { ( J _ { i j } ^ { 2 } + 2 \, J _ { i j } \, J _ { e x t } ) \, \nu _ { e x t } } { 2 \, \sigma ^ { 2 } } + O ( d t ) \quad \text{if } \, - J _ { e x t } \leq J _ { i j } < 0$$
Under the hypothesis | J ij | < J ext , then, the dependence of the inferred coupling on the real synaptic efficacy is again quadratic, as for the J ij > 0 case, but its leading term does not depend on dt .
If we assume, instead, J ext < J | ij | ≤ 2 J ext , the sum in Eq. (25) will be dominated by the term k = 2, since the first term vanishes, and thus we get:
$$J _ { I n f, i j } \simeq - \frac { 1 } { 2 } + O ( d t ) \quad \text{if } \, - \, 2 \, J _ { e x t } \leq J _ { i j } < - J _ { e x t }$$
Basically then, for | J ij | > J ext , the inferred J will be largely independent of the true value of J ij . This result can be intuitively understood by examining the extreme case J ij →-∞ ; in this case, the only surviving contribution to the probability of firing during dt (given the arrival of the inhibitory spike from j ), is the baseline probability ν glyph[epsilon1] i before the large, negative jump J ij in V i ; such term, integrated over a uniform distribution becomes ν dt/ i 2, which inserted into Eq. (20) gives J Inf,ij glyph[similarequal] -1 2. / Eq. (27) then shows that, for dt → 0, the limit case's behavior of J Inf is essentially attained already for J ij ≤ -J ext .
## Acknowledgments
We thank Y. Roudi for interesting discussions.
## References
- 1. Eggermont JJ (1990) The correlative brain. Springer.
- 2. Pernice V, Rotter S (2013) Reconstruction of sparse connectivity in neural networks from spike train covariances. Journal of Statistical Mechanics: Theory and Experiment 2013: P03008.
- 3. Schneidman E, Berry MJ, Segev R, Bialek W (2006) Weak pairwise correlations imply strongly correlated network states in a neural population. Nature 440: 10071012.
- 4. Marre O, El Boustani S, Fr´gnac Y, Destexhe A (2009) Prediction of spatiotemporal e patterns of neural activity from pairwise correlations. Physical Review Letters 102: 138101.
- 5. Hertz J, Roudi Y, Tyrcha J (2013) Ising models for inferring network structure from spike data. In: Quian Quiroga R, Panzeri S, editors, Principles of neural coding, CRC Press.
- 6. Roudi Y, Hertz J (2011) Mean field theory for nonequilibrium network reconstruction. Physical Review Letters 106: 048702.
- 7. Bialek W (2012) Biophysics: searching for principles. Princeton University Press.
- 8. Truccolo W, Eden UT, Fellows MR, Donoghue JP, Brown EN (2005) A point process framework for relating neural spiking activity to spiking history, neural ensemble, and extrinsic covariate effects. Journal of neurophysiology 93: 1074-1089.
- 9. Pillow JW, Shlens J, Paninski L, Sher A, Litke AM, et al. (2008) Spatio-temporal correlations and visual signalling in a complete neuronal population. Nature 454: 995-999.
- 10. Roudi Y, Dunn B, Hertz J (2015) Multi-neuronal activity and functional connectivity in cell assemblies. Current opinion in neurobiology 32: 38-44.
- 11. Glauber RJ (1963) Time-dependent statistics of the ising model. Journal of Mathematical Physics 4: 294-307.
- 12. Renart A, Brunel N, Wang XJ (2004) Mean-field theory of irregularly spiking neuronal populations and working memory in recurrent cortical networks. Computational Neuroscience: A comprehensive approach : 431-490.
- 13. Tsodyks MV, Markram H (1997) The neural code between neocortical pyramidal neurons depends on neurotransmitter release probability. Proceedings of the National Academy of Sciences 94: 719-723.
- 14. Holcman D, Tsodyks M (2006) The emergence of up and down states in cortical networks. PLoS Computational Biology 2: e23.
- 15. Eytan D, Marom S (2006) Dynamics and effective topology underlying synchronization in networks of cortical neurons. The Journal of Neuroscience 26: 8465-8476.
- 16. Roudi Y, Nirenberg S, Latham PE (2009) Pairwise maximum entropy models for studying large biological systems: when they can work and when they can't. PLoS Computational Biology 5: e1000380.
- 17. Roudi Y, Aurell E, Hertz JA (2009) Statistical physics of pairwise probability models. Frontiers in Computational Neuroscience 3.
- 18. Roudi Y, Tyrcha J, Hertz J (2009) Ising model for neural data: Model quality and approximate methods for extracting functional connectivity. Physical Review E 79: 051915.
- 19. Cocco S, Leibler S, Monasson R (2009) Neuronal couplings between retinal ganglion cells inferred by efficient inverse statistical physics methods. Proceedings of the National Academy of Sciences 106: 14058-14062.
- 20. Mattia M, Del Giudice P (2000) Efficient event-driven simulation of large networks of spiking neurons and dynamical synapses. Neural Computation 12: 2305-2329.
- 21. Cox DR, Miller HD (1965) The theory of stochastic processes. Chapman & Hall.
|
10.1371/journal.pone.0118412
|
[
"Cristiano Capone",
"Carla Filosa",
"Guido Gigante",
"Federico Ricci-Tersenghi",
"Paolo del Giudice"
] |
2014-08-05T16:10:45+00:00
|
2015-02-18T14:35:21+00:00
|
[
"q-bio.NC",
"cond-mat.dis-nn",
"q-bio.QM"
] |
Inferring Synaptic Structure in presence of Neural Interaction Time Scales
|
Biological networks display a variety of activity patterns reflecting a web
of interactions that is complex both in space and time. Yet inference methods
have mainly focused on reconstructing, from the network's activity, the spatial
structure, by assuming equilibrium conditions or, more recently, a
probabilistic dynamics with a single arbitrary time-step. Here we show that,
under this latter assumption, the inference procedure fails to reconstruct the
synaptic matrix of a network of integrate-and-fire neurons when the chosen time
scale of interaction does not closely match the synaptic delay or when no
single time scale for the interaction can be identified; such failure,
moreover, exposes a distinctive bias of the inference method that can lead to
infer as inhibitory the excitatory synapses with interaction time scales longer
than the model's time-step. We therefore introduce a new two-step method, that
first infers through cross-correlation profiles the delay-structure of the
network and then reconstructs the synaptic matrix, and successfully test it on
networks with different topologies and in different activity regimes. Although
step one is able to accurately recover the delay-structure of the network, thus
getting rid of any \textit{a priori} guess about the time scales of the
interaction, the inference method introduces nonetheless an arbitrary time
scale, the time-bin $dt$ used to binarize the spike trains. We therefore
analytically and numerically study how the choice of $dt$ affects the inference
in our network model, finding that the relationship between the inferred
couplings and the real synaptic efficacies, albeit being quadratic in both
cases, depends critically on $dt$ for the excitatory synapses only, whilst
being basically independent of it for the inhibitory ones.
|
1408.1016v1
|
## Stellar irradiated discs and implications on migration of embedded planets III: viscosity transitions
Bertram Bitsch 1 2 , , Alessandro Morbidelli 2 , Elena Lega , Katherine Kretke 2 3 , Aur´ elien Crida 2
- 1 Lund Observatory, Department of Astronomy and Theoretical Physics, Lund University, Box 43, 22100 Lund, Sweden
- 2 Laboratoire Lagrange, UMR7293, Universit´ e Nice Sophia-antipolis / CNRS / Observatoire de la Cˆte d'Azur, 06300 Nice, France o
- 3 Southwest Research Institute, Department of Space Studies, 1050 Walnut Street, Suite 300, Boulder, CO 80302, USA
Preprint online version: August 6, 2014
## ABSTRACT
Context. The migration strength and direction of embedded low-mass planets depends on the disc structure. In discs with an e ffi cient radiative transport, the migration can be directed outwards for planets with more than 3 -5 Earth masses. This is due to the entropy driven corotation torque, a process that extends the lifetimes of growing planetary embryos. However, smaller mass planets are still migrating inwards and might be lost to the central star.
Aims. We investigate the influence on the disc structure caused by a jump in the α parameter of the viscosity to model a dead-zone structure in the disc. We focus on ˙ M discs, which have a constant net mass flux. Using the resulting disc structure, we investigate the consequences for the formation of planetesimals and determine the regions of outward migration for proto-planets.
Methods. We performed numerical hydrosimulations of ˙ M discs in the r-z-plane. We used the explicit implicit hydrodynamical code / FARGOCA that includes a full tensor viscosity and stellar irradiation as well as a two-temperature solver that includes radiation transport in the flux-limited di ff usion approximation. The migration of embedded planets was studied by using torque formulae.
Results. Viscosity transitions inside the disc create transitions in density that stop inward migration for small planets through the so-called 'planet trap' mechanism. This mechanism also works for planets down to MP > 0 5 . MEarth , while in radiative discs with no viscosity transition the lowest mass with which inward migration can be avoided is 3 -5 Earth masses. Additionally, the viscosity transitions change the pressure gradient in the disc, which facilitates planetesimal formation via the streaming instability. However, a very steep transition in viscosity is needed to achieve in a pressure bump in the disc.
Conclusions. The transition in viscosity facilitates planetesimal formation and can stop the migration of small-mass planets ( MP > 0 5 . MEarth ), but still does not halt inward migration of smaller planets and planetesimals that are a ff ected by gas drag. A very steep, probably unrealistic viscosity gradient is needed to trap planets of smaller masses and halt gas-drag-driven planetesimal migration at a pressure bump.
Key words. accretion discs - planet formation - hydrodynamics - radiative transport - planet disc interactions
## 1. Introduction
Planet formation occurs in accretion discs around young stars. The evolution of accretion discs is still not understood completely, but it takes place on a Myr time-scale during which the accretion rate of gas onto the host star drops (Hartmann et al. 1998). The formation of gas giants has to take place during this time. In the most commonly favoured scenario (Pollack et al. 1996) a massive solid core of several Earth masses has to form first to accrete gas. However, the formation of cores via the accretion of embryos and planetesimals is not very e ffi cient and therefore can not explain the growth of proto-planetary cores (Levison et al. 2010). Recently, the idea of pebble accretion onto planetary embryos o ff ered a possible solution to this problem, because it occurs on a very short time-scale (Lambrechts & Johansen 2012; Morbidelli & Nesvorny 2012).
the planetary cores in the disc depends on the thermodynamics in the disc (Paardekooper & Mellema 2006; Kley & Crida 2008; Baruteau & Masset 2008; Kley et al. 2009). For a review on planet migration see Baruteau et al. (2014). The thermodynamics inside the discs are determined by viscous heating, radiative cooling, and stellar illumination (Bitsch et al. (2013), hereafter Paper I). Stellar irradiation maintains a flared disc structure in the outer parts of the disc (Chiang & Goldreich (1997); Paper I). A consequence of the disc structure on the migration of planets is described in Paper I, where we find that outward migration is possible in the regions of the disc where H r / decreases, which e ff ectively means that the regions of outward migration are smaller in a flared disc than a shadowed disc, where H r / decreases in the outer parts of the disc.
Regardless of how the cores may form, they will migrate through the disc because of interactions with the gas (Ward 1997). As long as discs had been assumed to be locally isothermal, cores were expected to migrate very fast towards the central star, which was a problem for understanding the formation of giant planets. It has been recently shown that the migration of
Bitsch et al. (2014) hereafter Paper II studied the evolution of discs where the mass flow towards the star ( ˙ ) is independent M of radius. In these accretion discs, the mass flux ˙ M is defined as
$$\underset { \cdot } { \text{isother} } - \dot { M } = 3 \pi \nu \Sigma _ { G } = \alpha H ^ { 2 } \Omega _ { K } \Sigma _ { G } \,,$$
where ν is the viscosity, following the α -viscosity approach (Shakura & Sunyaev 1973). H is the height of the disc, and Ω K is the Keplerian rotation frequency. Σ G denotes the gas surface density. We have found that during the evolution of the disc (decreasing ˙ ) M the regions of outward migration shrink until for
small ˙ M no outward migration is possible. Consequently, all planets migrate inwards, similarly to the isothermal disc scenario. This, again, may be problematic for understanding giant planet formation. Moreover, even when regions of outward migration exist, they typically concern only planets more massive than ≈ 5 Earth masses. Smaller planets migrate inwards as they grow, and therefore it is debated whether they can achieve this threshold mass quickly enough to be captured in the region of outward migration. Additionally, the inward migration of small-mass planets might be even faster than estimated in Paardekooper et al. (2011), see Lega et al. (2014). All this indicates that we still miss a mechanism for preventing inward migration of low-mass protoplanets.
The inward migration of small-mass planets can be stopped by a bump in vortensity (Masset et al. 2006), which is defined as ( ∇× v ) / Σ , where v is the velocity of the gas. The bump in vortensity (positive radial gradient) results in a positive contribution of the barotropic part of the corotation torque that acts on the planet which can overcompensate for the negative Lindblad torque and halt its inward migration.
Additionally, the formation of planetesimals and planetary embryos can be aided by a positive radial pressure gradient in a protoplanetary disc (Johansen & Youdin 2007). This stops the inward migration of aggregates that undergo gas drag (Brauer et al. 2008) and significantly facilitates planetesimal formation (Bai & Stone 2010a,b). This indicates that a bump in vortensity and pressure would be very useful for the formation of planetesimals and planetary embryos by the core accretion scenario, which raises the question of how such a vortensity bump and a pressure bump can appear in the disc.
The radial changes of the vortensity and pressure fields can be triggered by changes in viscosity, which can be caused by changes in the gas ionisation fraction or condensation at the ice line (see sect 2.2). In Paper II we only explored discs that have a unique value of α in time and space. In this paper, we investigate the e ff ect of viscosity transitions on the structure of the disc, with special attention devoted to the radial gradients of vortensity and pressure. Our aim is to determine the conditions for the following appealing scenario for giant planet formation:
At a pressure bump in the disc, particles can accumulate and form planetesimals and planetary embryos. These embryos would then migrate inwards, therefore we assume that a bump in vortensity exists at the same location in the disc, which can trap the planetary embryos (Masset et al. 2006; Morbidelli et al. 2008). After they are formed, these embryos could then grow via pebble accretion (Lambrechts & Johansen 2012; Morbidelli & Nesvorny 2012) until they reach a mass of ≈ 10 -20 MEarth and gas accretion onto the core can start (Pollack et al. 1996). The inward migration of cores of a few Earth masses can also be stopped by the entropy related corotation torque (Paardekooper & Mellema 2006; Kley & Crida 2008; Baruteau & Masset 2008; Kley et al. 2009), so that at this stage of evolution a bump in vortensity is not necessarily needed any more. As the planet starts to grow beyond ≈ 50 MEarth , it opens up a gap in the disc (Crida et al. 2006) and is then released into type-II-migration (Lin & Papaloizou 1986) and slowly moves towards the star as the disc is accreted. The final stopping point of the type-II migrating gas giant planet is then set by the inner edge of the disc (Crida & Morbidelli 2007) or the photo-evaporation radius of the disc (Alexander & Pascucci 2012), which explains the pile-up of large gas giants in the inner systems.
This paper is structured as follows: First, we give an overview of the methods used, especially of the viscosity (sec- tion 2). Then we compare in great detail a ˙ M = 1 × 10 -7 M /circledot / yr disc with and without transition in viscosity in section 3. In section 4 we investigate the changes of the disc structure as the disc evolves to smaller ˙ . M Di ff erent variations of viscosity transitions are then discussed in section 5. In section 6 we discuss which viscosity transition is actually needed to realise our planet formation scenario, and we discuss how realistic it might be. Finally, we give a summary in section 7.
## 2. Methods
## 2.1. General setup
The general simulation setup follows the descriptions of Papers I and II. The protoplanetary gas disc is treated as a threedimensional (3D) non-self gravitating gas whose motion is described by the Navier-Stokes equations. Without any perturbers in the disc, the disc has an axisymmetric structure. We therefore can compute all quantities in 2D, in the r -θ -plane, where θ is the colatitude of the spherical coordinates centred on the star, such that θ = 90 ◦ in the midplane. We used a 386 × 66 grid cells for all simulations. The dissipative e ff ects are described via the standard viscous stress tensor approach (e.g. Mihalas & Weibel Mihalas 1984). We also included the irradiation from the central star, which was described in detail in Paper I. For that purpose we modified and substantially extended the existing multi-dimensional hydrodynamical code FARGOCA, as presented in Lega et al. (2014) and Paper II.
Viscous and stellar heating generates thermal energy in the disc, which di ff uses radiatively through the disc and is emitted from its surfaces. To describe this process we used the fluxlimited di ff usion approximation (FLD, Levermore & Pomraning 1981), an approximation that allows the transition from the optically thick mid-plane to the thin regions near the disc's surface; more details of this are given in Paper I. For comparison purposes with Paper II, we used a star with M /star = 1 M /circledot and R /star = 3 R /circledot and T /star = 5600K, which corresponds to L ≈ 9 L /circledot . The central star mainly influences the parts of the disc that are dominated by stellar irradiation. For the inner disc, we do not assume that this plays a significant role.
We adopted the opacity law of Bell & Lin (1994), with κ/star = 3 5cm 2 . / g as in Paper II. Throughout all the simulations presented here, we used a metallicity of 1% in µ m size dust grains. We also apply the ˙ M boundary conditions described in the appendix of Paper II.
## 2.2. Prescription of viscosity transitions
Here we explore the e ff ects of putative viscosity transitions at the ice line on the formation of planetesimals via the streaming instability as well as their capability of halting the migration of small planets. More precisely, we are interested in the viscosity gradients (that change the gradients in surface density and pressure) necessary to support our proposed scenario of planet formation. We modelled the transition in viscosity in a simple toy model, instead of using full-scale M agnetoR otationalI nstability (MRI) simulations (Balbus & Hawley 1998). We made the general assumption that the α -parameter of our viscosity (Shakura & Sunyaev 1973) in the active layer ( α A ) is much larger than in the midplane regions of the disc, which corresponds to a dead zone like structure (Balbus & Hawley 1998).
Pierens & Nelson (2010) have studied the influence of a dead zone on the accretion of gas onto giant cores in isothermal discs. They modelled the dead zone by an α transition in the vertical
direction dependent on H in the disc. They found that the gas accretion process is not su ffi ciently altered by the presence of a dead zone, but that inward migration is somewhat reduced for giant planets. Here we instead focus on the migration of smallmass planets, and moreover, we link the transitions in α to the temperature and column density in the disc.
We took two main boundaries for the α transition into account. (i) An inner radial boundary that is dependent on the temperature (thermal ionisation), and (ii) a boundary that depends on the vertically integrated volume density (due to the absorption of X-ray and or cosmic rays). In the midplane regions of the disc, / we then explored the influence of an additional transition of viscosity at the ice line, where the di ff erent amount of condensed dust ice grains there can quench the MRI at di / ff erent degrees (Kretke & Lin 2007). The idea is not to give a priori quantitative estimates for these transition, but to understand the general behaviour of a disc that has viscosity transitions. Therefore the prescriptions also are slightly modified to test di ff erent parameters in section 5.
The general assumption is that for T > Tcrit = 1000K the disc is fully ionised and therefore MRI active. This would correspond to the inner parts of the disc, and maybe to some superheated parts in the atmosphere of the accretion disc. For T < Tcrit the MRI activity has to be triggered by di ff erent sources. At the upper and lower layers of the disc, ionization originates from cosmic rays, stellar X-rays, ultraviolet photons and radionuclides. All these ionization sources have di ff erent penetration depths into the disc and therefore ionize the disc down to different levels.
Ultraviolet photons will only penetrate to a mass column of 0 1 . g cm / 2 (Perez-Becker & Chiang 2011). X-rays can contribute down to Σ = 10 g cm / 2 (Bai & Goodman 2009), while cosmic rays can penetrate down to Σ = 100 g cm / 2 (Dzyurkevich et al. 2013). The di ff erent penetration depths for cosmic rays and Xrays suggests that an additional transition of α exists in the vertically integrated volume density.
Additionally, we imposed a transition of α at the ice line because there might be ice grains below the water condensation temperature that can then quench the MRI more e ff ectively than in the region where only silicate grains are available (Kretke & Lin 2007). For the di ff erent α values we used we defined the following relation:
$$\alpha _ { A } & = 1 0 \alpha _ { U } = 1 0 0 \alpha _ { D } \, & ( 2 ) \quad \text{viscosity} \quad \text{thon in } \mathfrak { r }$$
where we set α A = 0 0264. This indicates that . α U = 0 00264 ( . α undead), which is still a significant, but low, alpha factor. In the α D ( α dead) region, the α factor is even smaller.
A zero viscosity in the dead zone can lead to outbursts in the disc triggered by gravitationally instabilities in the disc (Martin & Lubow 2014). Inside the dead zone, however, radionuclides can provide an additional source of ionization, and hydrodynamical instabilities might create turbulence, for instance, the vertical shear instability (Nelson et al. 2013) or the baroclinic instability (Klahr & Bodenheimer 2003), which will provide some viscosity. Therefore we assumed that α D is not zero.
We set the following α regions in the disc:
$$2$$
$$\cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdot \\ \Sigma _ { P } ( z ) \leq 1 0 g / c m ^ { 2 } \quad \text{OR} \quad 1 0 0 0 K \leq T \quad \Rightarrow \quad \alpha _ { A } \\ 1 0 g / c m ^ { 2 } < \Sigma _ { P } ( z ) \leq 5 0 g / c m ^ { 2 } \quad \Rightarrow \quad \alpha _ { A } \to \alpha _ { U } \\ 5 0 g / c m ^ { 2 } < \Sigma _ { P } ( z ) ; 1 6 0 K \leq T < 2 2 0 K \quad \Rightarrow \quad \alpha _ { U } \to \alpha _ { D } \\ 5 0 g / c m ^ { 2 } < \Sigma _ { P } ( z ) \leq 9 0 g / c m ^ { 2 } ; T < 1 6 0 K \quad \Rightarrow \quad \alpha _ { U } \to \alpha _ { D }$$
$$9 0 g / c m ^ { 2 } < \Sigma _ { P } ( z ) ; T < 1 6 0 K \quad & \Rightarrow \quad \alpha _ { D } \\ 8 0 0 K \leq T < 1 0 0 0 K \quad & \Rightarrow \quad \alpha _ { A } \to \alpha _ { U } \, ( 3 ) \\ \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot$$
where → symbolizes a transition region for α . Σ P ( z ) is the vertically integrated volume density from the top of the disc down to a height z , where the desired quantity of Σ P is reached. This basically represents the penetration depth of cosmic and X-rays. We obtained the gas surface density by integrating both sides of the disc from top to midplane and from bottom to midplane. Therefore the gas surface density of the disc Σ G is 2 Σ P (0). Transitions for the di ff erent α regions need to be smoothed out, because this is closer to real MHD simulations than a step function (Dzyurkevich et al. 2013).
Note here that if either T > 1000K or Σ P ( z ) < 10g cm 2 / , the α value corresponds to α A , because the disc is either completely ionised by thermal ionisation or by X-rays and cosmic rays. The transition regions are smoothed in the following, linear way:
$$\frac { \lambda \cdot \cdot \cdot \cdot } { \delta } K \text{ the } \frac { \alpha _ { T } } { \alpha _ { A } } & = \alpha _ { A } - ( \alpha _ { A } - \alpha _ { U } ) \frac { \sum _ { P } - \sum _ { A } } { 4 0 g / c m ^ { 2 } } \quad \text{for} \quad \alpha _ { A } \to \alpha _ { U } \quad \text{(4)} \\ \quad \text{$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\col:$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\col.$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\col.:$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colots$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colot$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\colon$\col.}$$
$$\text{$\text{$row_use}$} \\ \text{old cor-} \\ \text{e-sinner-}$$
and likewise for the other transition regions. Parts of the disc that are in these transition regions either follow the gradient of Σ P or of T , depending on the steepness of these gradients (where the shallower gradient is favoured). Note here that the temperature range over which we imposed the α -transition ( α U → α D ) at the ice line is exactly the range over which the opacity changes in the opacity we used (Bell & Lin 1994). We therefore mark in our plots the ice line at 190K, which is the middle of that transition range. The viscosity is given by
$$\colon \cos \text{mic} \quad \nu = \alpha c _ { s } ^ { 2 } / \Omega _ { K }$$
where Ω K is the Keplerian frequency and cs is the sound speed. Here we used the midplane sound speed cs ( z = 0), which provides a better control over the viscosity, which varies with height only because of changes in the α parameter. Additionally, the vertical changes in cs are expected to be much smaller than the changes in α . The α map as a function of Σ P and T is shown in Fig. 1 (top). We also display the α map of an ˙ M = 1 × 10 -7 M /circledot / yr disc (bottom panel) in equilibrium state (constant mass flux through each radial section of the disc) with α A = 0 0264. The . viscosity used for α A in this work is generally ten times higher than in Paper II.
## 3. Viscosity transitions
In Paper II we focused on discs with constant ˙ . M These discs also featured a constant α parameter in the whole disc. As stated above, the α parameter of the viscosity now follows some transitions, so that α changes in the disc with r and z because the temperature and the volume density are also functions of r and z . The radial and vertical changes in α will lead to changes in volume density relative to the uniform α case, because ˙ M has to be independent of r in the steady state. We compare here a disc with the described transition in α with a disc that has constant α throughout the disc equal to the value of α A .
## 3.1. Disc structure
In Fig. 2 (top) the midplane temperature profiles for ˙ M = 1 × 10 -7 M /circledot / yr discs with and without a transition in α are shown. Up to ≈ 20AU the midplane temperature is smaller
Fig. 1. Top: α as a function of Σ P and T as specified by eq. 3. Bottom: α parameter in an ˙ M = 1 × 10 -7 M /circledot / yr disc that is discussed in section 3. Inside the black line α ≤ α U , inside the white line α = α D .
in the case with α transition than the case without α transition. Consequently, the H r / profile of the disc with α transition is smaller than the H r / profile of the disc without α transition (second panel in Fig. 2). Here we define H as the pressure scale height using the midplane sound speed cs ( z = 0) with H = cs ( z = 0) / Ω K . The bumps in the H r / and T profiles at ≈ 4AU are also caused by transitions in the opacity profile (see Paper II), which changes the heating and cooling properties of the disc.
In a 1D radial model of a disc with constant radial ˙ , M a reduction of α is compensated for by an increase of Σ G of the same magnitude, as Σ G ∝ ˙ M /ν . This would allow for the same amount of viscous heating in the 1D disc and hence the same temperature profile. However, in r -z -discs this is not the case. A reduction in α does not imply an increase of Σ G by the same factor, because the active layer of the disc can carry most of the accretion rate (see Fig. 5), which is not possible in 1D models. Indeed, the surface density profile in the inner part of the disc with α transition (third from top in Fig. 2) only shows an increase by a factor of 3 in contrast with the reduction in α by a factor of 100 in the dead zone.
In the outer parts of the disc, the surface density is lower for the disc with an α transition. This is caused by the higher H r / in the outer parts, which results in a higher viscosity and hence in a lower surface density for an equal ˙ M disc.
In the disc with an α transition, the surface density shows a significant local maximum at ≈ 4AU, which is exactly where we set the transition of α U to α D . This is because, as ˙ M is independent of r , a change in viscosity is compensated for by a change in Σ G , resulting in the high-density regions close to the location of the change in α . This local maximum in surface density at ≈ 4AU is very appealing, because it might function as a planet trap (by changing the disc vortensity) and stop inward migration of low-mass planets that migrate in the type-I-migration regime
Pressure in g/(cm s ) 2
Fig. 2. Midplane temperature (top), H r / (second from top), integrated surface density Σ G (third from top) and midplane pressure (bottom) for the ˙ M = 1 × 10 -7 M /circledot / yr disc with and without a α transition.
(Masset et al. 2006). In contrast, the disc without an α transition shows a dip in surface density around ≈ 4AU, which is caused by the transition in opacity at the ice line and the corresponding increase in viscosity caused by the increase in H r / (see Paper II).
The midplane pressure of the same discs is displayed in the bottom panel of Fig. 2. For the disc without an α transition, the pressure follows a power law without any derivations. No pressure bumps exist in this case. In the case with an α transition, however, small wiggles in the pressure are visible, especially at ≈ 4AU, where α U transitions to α D .
Despite the bumps in surface density, no real pressure bump exists in the disc because the gradient of pressure is not inverted. In 2D discs the pressure is defined as
$$P _ { 2 D } = R _ { g a, s } \Sigma _ { G } T / \mu \circ \Sigma _ { 0 } T _ { 0 } r ^ { - s - \beta } \,, \quad \ \ & ( 7 ) \quad \ \ 0. 4 \ \Big |$$
with Rgas being the gas constant and µ the mean molecular weight. Here s denotes the power-law index of the surface density profile and β the power-law index of the temperature profile. This indicates that when a jump in surface density can compensate for the gradient in T , a pressure bump exists. However, in 3D discs the pressure is related to the volume density ρ G as
$$P _ { 3 D } = R _ { g a s } \rho _ { G } T / \mu \infty \, \frac { \Sigma _ { 0 } } { H _ { 0 } } T _ { 0 } r ^ { - s - \beta - ( 1 + f ) } \,, & & ( 8 ) &$$
as in hydrostatic equilibrium ρ G = Σ G / ( √ 2 π H ), where H is the disc vertical thickness. Here f denotes the flaring index of the disc. By comparing P D 3 with P D 2 , we immediately see that for the same temperature profile a steeper positive radial surface density gradient is needed in 3D than in 2D to invert the radial pressure gradient, as long as the flaring index of the disc is not strongly negative, which generally is the case. In the models shown in Fig. 2, the local maximum of the surface density gradient is not large enough to result in a pressure bump.
In Fig. 3 the ( r z , ) temperature (top) and density (bottom) distribution are shown. In the inner parts close to midplane, the disc is heated by viscosity. With increasing distance to the star, this heating decreases as the density decreases. The upper layers of the disc are heated by the star, therefore they are hotter than the midplane region. The regions of the disc far away from the star with low density are heated by heat di ff usion downwards from the upper layers. Close to the α transition at ≈ 4AU, the temperature is vertically constant for a very wide vertical range (up to ≈ 0 3AU). .
The volume density ρ G near the midplane is nearly radially constant around the α transition at ≈ 4AU, as shown in Fig. 3 (bottom). In this region the density is vertically constant up to ≈ 0 1AU. To calculate the surface density, the density shown here . was vertically integrated, which resulted in the local maximum in the surface density (middle panel in Fig. 2). But, this is not enough to generate a pressure bump because T decreases with r .
The radial velocity of the disc without an α transition is displayed in the top panel of Fig. 4. A positive value indicates an outward motion, while a negative value indicates an inward motion. Here, we see an outflow in the midplane regions of the disc and an inflow at the top layers of the disc, which was also observed by Urpin (1984); Kley & Lin (1992); Takeuchi & Lin (2002). This flow is caused by the α prescription for the viscosity (Fromang et al. 2011).
The radial velocity of the disc with α -transition is displayed in the bottom panel of Fig. 4. We now observe several layers in the flow pattern. In the very top regions of the disc (where α = α A ) the flow is directed inwards. Farther down, the flow
Fig. 3. Temperature (top) and density (bottom) of the ˙ M = 1 × 10 -7 M /circledot / yr with α transition. The black and white lines mark the transitions in the α viscosity parameter as defined by eq. 3 and as shown in Fig. 1. The grey line marks the disc's pressure scale height H .
is directed outwards, while it then flows inwards again at the α U transition. The flow farther inside the disc near the midplane is again directed outwards. Both discs show similar inflow behaviour in the top layers.
In Fig. 5 the ˙ M rates for the di ff erent α -layers of the disc with α transition are displayed. The layers are defined by their α values as indicated in the figure. Additionally, the total ˙ M rate is displayed, which is the sum of all the di ff erent layers. While the total ˙ M rate remains approximately constant in radius, the di ff erent ˙ M rates of the di ff erent layers vary strongly. Between 2 -4AU, the ˙ M rate in the active layer is highest (in inward direction) and in the same region, the layer with α < α U has the largest outflow. At 4AU the flux changes and each layer seems to carry the same amount of ˙ . At 8AU, M α > α U , so that only two layers in the disc exist. At that distance the flow returns to the state of an unlayered disc, with outflow in midplane and inflow in the top layers. We recall that the accretion rate of the active layer only depends on the radial velocity of the gas, because its density is fixed to 10 g cm / 2 at each r by construction.
In a simple view of a dead zone, with zero viscosity, the material flows inwards in the top layers of the disc and then dumps material into the dead zone, which then grows in density. Eventually, the dead zone becomes so massive that the disc becomes gravitationally unstable and produces outbursts of accretion (Martin & Lubow 2014). But this view is not supported by the flow picture we found in our simulations. Just above the region of low viscosity, we found outflow and in the very top layers inflow. This suggests that the mass flux in the active layer can adjust to the dead zone so that the mass flux is still constant in r and at the same time does not dump that much material in the dead zone. This consideration is also supported by the fact that we did not find an increase of Σ G of a factor of 100 in
Fig. 4. Radial velocity of the disc without α transition (top) and with α transition (bottom). The black and white lines mark the transitions in the α viscosity parameter as defined by eq. 3 and as shown in Fig. 1. In the bottom panel, the parts of the disc above the red line are fully active with α A . They are overplotted in the top plot as well to show where these transitions would be. The grey line represents the pressure scale-height H . Positive velocities mark an outward motion, while negative velocities indicate an inward motion.
Fig. 5. ˙ M rates for the di ff erent layers of the ˙ M = 1 × 10 -7 M /circledot / yr disc with an α transition. Here, a positive ˙ M rate indicates accretion, while a negative ˙ M indicates decretion (motion away from the star).
the dead zone in our simulations, where α is reduced by a factor of 100. This basically suggests that the discs cannot become gravitationally unstable which disagrees with the hypothesis of Martin & Lubow (2014).
Fig. 6. ∆ v cs of the ˙ M = 1 × 10 -7 M /circledot / yr disc with and without an α transition. The black line marks the ∆ parameter in the standard MMSNdisc.
## 3.2. Planetesimal formation and embryo migration
The reduction of the gravitational force by the radially outwardspointing force of the pressure gradient causes a di ff erence ∆ v between the azimuthal mean gas flow and the Keplerian orbit. This is given by
$$\stackrel { \partial } { c } \, \text{and} \, \text{as} \\ \stackrel { \partial } { c } \, \text{above} \, \quad \frac { \Delta v } { c _ { s } } = \eta \frac { v _ { K } } { c _ { s } } = - \frac { 1 } { 2 } \frac { c _ { s } } { v _ { K } } \frac { d \ln ( P ) } { d \ln ( r ) } \,,$$
where vK = √ GM r / is the Keplerian velocity cs / vK = H r / (for the H r / profile, see the second panel from top in Fig. 2). η represents a measure of the gas pressure support (Nakagawa et al. 1986).
Because of the sub-Keplerian rotation of the gas, small solid particles drift towards the star. A local reduction in ∆ v cs facilitates particle clumping in the streaming instability (Johansen et al. 2007; Johansen & Youdin 2007; Bai & Stone 2010a,b). Reduced ∆ v cs would also help planetesimals to grow further by pebble accretion (Lambrechts & Johansen 2012).
The ∆ v cs parameter is displayed in Fig. 6. In addition to the ˙ M = 1 × 10 -7 M /circledot / yr discs with and without α transition, we display the ∆ v cs parameter in the standard Minimum Mass Solar Nebular (MMSN) case (Hayashi 1981). The MMSN model is constructed with simple power laws in surface density and temperature, which results in a power law for the ∆ v cs parameter as well. The increasing ∆ v cs with r means that it is much harder to form planetesimals via the streaming instability at larger distances.
The disc without transition in α has a maximum at r ≈ 3 5 . AU and a drop for larger r that is caused by the maximum in the H r / profile, which is originally caused by transitions in opacity (Paper II). That ∆ v cs has the same value at 2AU and at 5AU implies that the formation of planetesimals at these locations are equally likely. But at 5AU the disc is much colder than at 2AU, which allows for the condensation of ice grains, meaning that more solid material is available at 5AU. In the end, this might actually facilitate planetesimal formation more at 5AU than at 2AU.
In the case with an α transition ∆ v cs is much reduced at 4AU. This is the exact location of the α transition from α U to α D . This transition causes a wiggle in the midplane pressure (bottom panel in Fig. 2). The change of the pressure gradient then
significantly reduces ∆ v cs at that location, making planetesimals formation via the streaming instability much more likely there.
After they are formed, small planetesimal can migrate inwards. These planetesimals can then grow to planetary embryos by oligarchic growth (Kokubo & Ida 1998) or by pebble accretion (Lambrechts & Johansen 2012; Morbidelli & Nesvorny 2012). We now shift our attention to planet migration. To measure the migration rate, one should in principle extend our disc model to full 3D (with su ffi cient resolution in azimuth), introduce a planet in the disc and measure the torques. This would be very computationally expensive, however, in particular because for the small-mass planets that we are interested in here, a very high resolution would be required to resolve the planet's horseshoe region. To estimate the smallest mass needed for outward migration of embryos triggered by the entropy related corotation torque, we therefore used the formula by Paardekooper et al. (2011), which captures the effects of torque saturation due to viscosity e ff ects in contrast to Paardekooper et al. (2010), where the torques are fully unsaturated. The torque formula by Paardekooper et al. (2011) was observed to match with the torques in 3D simulations quite well (Bitsch & Kley 2011). However for small-mass planets, a new phenomenon has been observed that can result in a more negative torque that drives a faster inward migration than described by the formula of Paardekooper et al. (2011) (Lega et al. 2014). However, this phenomenon deserves more detailed studies. We therefore used the formula by Paardekooper et al. (2011) as an estimate for planet migration in gas discs. This was made in the same way as described in Papers I and II, so we do not state the torque formula explicitly. The total torque acting on an embedded planet is a composition of its Lindblad torque and its corotation torque:
$$\Gamma _ { t o t } & = \Gamma _ { L } + \Gamma _ { C } \. & ( 1 0 ) \ \underset { \text{smaller} \, t } { \max }$$
The Lindblad and corotation torque depend on the local radial gradients of entropy S ∝ r -ξ , with ξ = β -( γ -1 0) . s , and gas surface density Σ G ∝ r -s . β describes the radial gradient of the temperature profile, T ∝ r -β . Very approximately, for large s a large ξ caused by a large β will lead to outward migration, while a flat radial entropy gradient will lead to inward migration.
The migration maps for discs without (top) and with (bottom) an α transition are displayed in Fig. 7. In the case without an α transition, two regions of outward migration are visible, one around 2AU and one between 4 and 7AU. For these regions of outward migration a planetary mass of at least ≈ 17 MEarth is needed, which is very high. The reason for this lies in the relatively high viscosity with α = 0 0264 (Paper II). .
In the case of an α transition this changes dramatically. There are still two regions of outward migration, but they are much more narrow in radial distance than for the case without an α transition.
More precisely, now the region of outward migration is only a small band between 3 5 to 4 0AU. This is because farther away . . the viscosity is too low to keep the entropy-driven corotation torque unsaturated. The outer migration region is now due to the vortensity-driven corotation toque, which is positive and locally exceeds the Lindblad torque because of the positive radial gradient of the surface density. This is the planet-trap mechanism first studied in Masset et al. (2006). 1 The planet trap mechanism
1 In a disc without an α transition the outer radial boundary of the outward migration region is approximately the location where the aspect ratio of the disc has a minimum, i.e. at the transition between the part of the disc that is dominated by viscous heating and that dominated
Fig. 7. Migration map for discs without (top) and with (bottom) an α transition. The encircled regions in black mark regions in which outward migration is possible. The vertical red line marks the ice line at 190K in both plots. The two plots feature slightly di ff erent radial extensions and planetary masses.
also changes the smallest and largest masses for outward migration, relative to the entropy-driven corotation case. The smallest mass for outward migration is now ≈ 0 5 . MEarth , which is much smaller than before. On the other hand, the largest mass for undergoing outward migration is reduced to ≈ 23 MEarth .
Planetesimal formation for discs with and without the α transition, the streaming instability is more likely to operate in a region of outward migration. However, this may not be enough to explain the formation of a giant planet at the snowline (i.e. at the location of the α -transition). In fact, only objects more massive than 0 5 . MEarth avoid inward migration in this region. Thus, if the streaming instability and the subsequent pebble accretion (Lambrechts & Johansen 2012) do not form a 0 5 . MEarth object su ffi ciently fast, the object will have the time to migrate into the inner system. Therefore, a steeper viscosity gradient than assumed here may be needed to allow a giant planet to form at the snowline to produce a steeper surface density gradient (capable of stopping the inward migration of smaller-mass planets) and possibly even a pressure bump (preventing the inward migration of planetesimals).
## 4. Disc evolution
We follow here the prescription for the transition in α from eq. 3 and decrease the disc's ˙ M by decreasing the Σ G value to follow the evolution of the disc through di ff erent accretional stages (same approach as in Paper II). We expect that with the decrease
by stellar irradiation. Thus, changing the luminosity properties of the central star a ff ects the extension of the outward migration region. In a disc with an α transition the outward migration region is instead located near the maximum of H r / , and thus it is entirely in the part of the disc dominated by viscous heating. Therefore, the outward migration region becomes quite insensitive to the assumed stellar properties.
Fig. 8. Surface density (top) and midplane temperature (middle) for discs featuring di ff erent ˙ M rates with α transitions. The bottom plot features ∆ v cs of the streaming instability for the same discs.
of the surface density of the disc, the radial gas-flow pattern eventually returns to the pattern of an unlayered disc. This is because the vertical integrated surface density from the disc's surface down to the midplane, which cannot exceed Σ G / 2, will not exceed the 100g cm / 2 limit to allow a full transition from α U to α D (see eq. 3).
The surface density (top) and midplane temperature (middle) are displayed in Fig. 8. The ratio between the surface densities in the ˙ M = 1 × 10 -7 M /circledot / yr and the ˙ M = 1 × 10 -8 M /circledot / yr is higher than a factor of 10 (the ratio of the accretion rates) in the inner parts of the disc. This is because the disc with the smaller ˙ M has a larger vertically averaged α (because the column density of the
Fig. 9. Migration map for discs with an α transition. The discs feature ˙ M = 5 × 10 -8 M /circledot / yr (top) and ˙ M = 1 × 10 -8 M /circledot / yr (bottom). The regions encircled in black mark regions in which outward migration is possible. The vertical red line in the top plot marks the ice line at 190K. The bottom plot has a di ff erent colour scale.
gas is lower). Consequently, the full surface density is reduced more than proportionally to the accretion rate. The outer parts of both discs are fully active, so that the di ff erence in Σ G is ≈ 10, i.e. proportional to the change of the ˙ M rate. This means that the change in total surface density Σ G is not linear in the change of ˙ , because the change of viscosity is not linear in the first place. M
A local maximum in the surface density (positive surface density gradient) is visible only for the ˙ M = 1 × 10 -7 M /circledot / yr disc . For the ˙ M = 5 × 10 -8 M /circledot / yr disc a dip in the surface density profile at ≈ 2 -3AU is visible, which is at the outer edge of the α U layer. In fact, inside of ≈ 2 5AU, this reduction in . α causes the surface density to increase. For lower ˙ M values no such dips are visible, because the highest surface density of these discs at any radius is lower than the condition for reducing α A fully towards α U .
∆ v cs is displayed in the bottom panel of Fig. 8. The dip in ∆ v cs that was visible for the ˙ M = 1 × 10 -7 M /circledot / yr disc vanishes for smaller ˙ . M This is because the transition in α form α U to α D disappears, because the disc's surface density is decreased. The big bump of ∆ v cs at ≈ 2AU in the ˙ M = 5 × 10 -8 M /circledot / yr disc is caused by the very steep negative surface density gradient in this region, which results in a steeper negative pressure gradient (eq. 8), which increases ∆ v cs (eq. 9). In the later stages of the disc evolution, ∆ v cs becomes much smaller in the inner regions of the disc, which in principle facilitates planetesimal formation by the streaming instability. However, at such late stages the lifetime of the disc is quite short, so that there may not be enough time to form the cores and then accrete gas to form giant planets.
The migration maps for discs with ˙ M = 5 × 10 -8 M /circledot / yr (top) and ˙ M = 1 × 10 -8 M /circledot / yr (bottom) discs are displayed in Fig. 9. A large region of outward migration exists around ≈ 3 -4 5AU .
in the ˙ M = 5 × 10 -8 M /circledot / yr disc. This region is quite similar to the region of outward migration for the ˙ M = 1 × 10 -7 M /circledot / yr disc without an α transition shown in the top panel of Fig. 7. This is because the disc with ˙ M = 5 × 10 -8 M /circledot / yr has a surface density lower than 100 g cm / 2 beyond 3AU and therefore is close to being fully ionised. The region of outward migration here is related to the entropy-driven corotation torque and not to the vortensitydriven corotation torque.
The ˙ M = 1 × 10 -8 M /circledot / yr disc features for 2 < r < 4AU a region of strong inward migration, which is caused by the positive temperature gradient in the disc. In the very inner parts of the disc ( r < 1 0AU) a region of outward migration exists, because . of the strong negative temperature gradient, as shown in the top panel of Fig. 8. This migration map is also similar to the migration map of a disc without an α transition. This means that the region of outward migration moves inwards in time as the disc evolves to smaller ˙ . M
The time the disc spends at high ˙ M rates ( ˙ M > 5 × 10 -8 M /circledot / yr ) is expected to be very short (a few 100kyr, see Hartmann et al. (1998)). But in the evolution from ˙ M = 1 × 10 -8 M /circledot / yr to 5 × 10 -8 M /circledot / yr the lowest mass for outward planet migration increases from 2 5 to 10 Earth masses. This means . that the formation of a giant planet in this region can occur only if the growth rate of the core is fast enough. Otherwise the core eventually finds itself below the lower mass boundary of the outward migration region and starts to migrate towards the inner disc. Similarly, when the disc reaches ˙ M ≈ 1 × 10 -8 M /circledot / yr , the outward migration region disappears for all planetary masses. Thus, it is necessary that by this time, the core has reached a large enough mass to accrete a massive envelope and transit to a slow type-II migration mode. We now examine whether such a fast growth is possible. Considering pebble accretion, the growth rates of planetesimals can be estimated (eq.42 Lambrechts & Johansen (2012))
$$\Delta t _ { d } \approx 8 \times 1 0 ^ { 6 } \left ( \frac { \Delta v / c _ { s } } { 0. 0 5 } \right ) ^ { 3 } \left ( \frac { \rho _ { p } / \rho _ { G } } { 0. 0 1 } \right ) ^ { - 1 } \left ( \frac { M _ { 0 } } { 1 0 ^ { - 5 } M _ { E } } \right ) ^ { - 1 } \left ( \frac { r } { 5 A U } \right ) y r \,, ( 1 1 ) \quad 3 0 g / c m ^ { 2 }$$
where t d defines the growth timescale, ρ P the pebble density, and M 0 the initial seed mass that accretes the pebbles. Assuming that a fraction of 10% of solids is in pebbles and an initial seed mass of 10 -4 ME , the growth time-scale is about 600kyr, which is similar to the time-scale of the disc evolution according to the observations of Hartmann et al. (1998). This indicates that it may be possible for the cores to grow at the same rate as the outward migration region shifts upward in mass while the stellar accretion rate is reduced (see top panel in Fig. 9).
## 5. Different α transitions
We explore here di ff erent prescriptions for the change of α as a function of Σ G and T .
## 5.1. Change of the cosmic ray penetration depth
The penetration of cosmic and X-rays into the disc is of crucial importance for the drive of the MRI. It is still debated how deep cosmic rays penetrate the disc. We therefore changed the transition from α A to α D to model a reduced penetration depth of cosmic and x-rays in the following way:
$$^ { \prime } c m ^ { 2 } \prec \Sigma _ { \xi } < 3 0 a l c m ^ { 2 }$$
$$m ^ { 2 } < \Sigma _ { P } ; 1 6 0 K < T \leq 2 2 0 K \ \Rightarrow \ \alpha _ { U } \rightarrow \alpha _ { D }$$
$$1 0 g / c m ^ { 2 } < \Sigma _ { P } \leq 3 0 g / c m ^ { 2 } \quad & \Rightarrow \quad \alpha _ { A } \rightarrow \alpha _ { U } \\ 3 0 g / c m ^ { 2 } < \Sigma _ { P } ; 2 2 0 K < T \leq 8 0 0 K \quad & \Rightarrow \quad \alpha _ { U } \\ 3 0 g / c m ^ { 2 } < \Sigma _ { P } ; 1 6 0 K < T \leq 2 2 0 K \quad & \Rightarrow \quad \alpha _ { U } \rightarrow \alpha _ { D }$$
Fig. 10. Density (top) and migration map (bottom) of the ˙ M = 5 × 10 -8 M /circledot / yr disc with a lower penetration depth for cosmic rays. In the top plot, the black and white lines mark the transitions in the α viscosity parameter as defined by eq. 12, the grey line marks the disc's height H . In the bottom plot, the regions encircled in black enclose the regions of outward migration, while the vertical red line marks the ice line at 190K.
$$| y r \,, ( 1 1 ) \begin{array} { c c c } 3 0 g / c m ^ { 2 } < \Sigma _ { P } \leq 5 0 g / c m ^ { 2 } ; T < 1 6 0 K & \Rightarrow & \alpha _ { U } \rightarrow \alpha _ { D } \\ 5 0 g / c m ^ { 2 } < \Sigma _ { P } ; T < 1 6 0 K & \Rightarrow & \alpha _ { D } \,. & ( 1 2 ) \\ \text{density} \end{array}$$
This implies that α D can still be reached at lower surface density values, hence it can be reached at lower ˙ M values. Accordingly, we expect to observe the features related to the presence of a dead zone (surface density bump, planet trap, etc.) for discs with smaller ˙ M with this prescription of α (eq. 12) and not with that of eq. 3.
In Fig. 10 the ( r z , ) density structure of a ˙ M = 5 × 10 -8 M /circledot / yr disc with the α transition defined in eq. 12 is shown in the top panel. Clearly, the discs features a local maximum in density at ≈ 3 3AU. This then translates into a local maximum in the verti-. cally integrated surface density (not displayed). No real pressure bump (positive pressure gradient) is visible in this case either, only a flattening of the pressure gradient, because the negative gradient in T overcompensates for the positive gradient in ρ G (see eq. 8).
The migration map (bottom panel in Fig. 10) shows a very similar pattern as that of the ˙ M = 1 × 10 -7 M /circledot / yr disc with the α transition described by eq. 3 (Fig. 7). More precisely, the stopping point of inward migration is shifted slightly inwards to ≈ 3AU. This is expected because the surface density gradients, the α distribution, and the temperature gradients are similar. For smaller ˙ , M the prescription of eq. 12 is not enough to create a low viscosity region. This means that low ˙ M discs are always nearly fully ionised, and their structure is similar to that described above.
Fig. 11. Surface density for ˙ M = 1 × 10 -7 M /circledot / yr discs with different ratios of α /α A U (red, blue, and black) and one (light blue) with α U → α D for 185K < T < 195K.
## 5.2. Change of the value of α U
In this section the ratio between the di ff erent α is changed, that is, we modify eq. 2. These three independent modifications are
$$\alpha _ { A } \ & = \ 2 0 \alpha _ { U } = 1 0 0 \alpha _ { D } \, \\ \alpha _ { A } \ & = \ 5 0 \alpha _ { U } = 1 0 0 \alpha _ { D }$$
$$\alpha _ { A } \, = \, 1 0 0 \alpha _ { U } = 1 0 0 \alpha _ { D } \,, & & ( 1 3 ) \quad \text{fig.13}.$$
where we kept the transition parameters for temperature and surface density as in eq. 3. This means that the change of viscosity at the ice line is reduced, because we aim to keep the same total reduction of α from α A to α D . For the last test case, there is only one transition in α because the transition α U → α D does not exist any more.
In Fig. 11, the surface density for the four di ff erent α ratios for a disc with ˙ M = 1 × 10 -7 M /circledot / yr is displayed. In the outer parts of the disc, the surface density remains the same for all α cases because the outer parts of the disc are fully ionised and hence always have α A , which is the same in all simulations. In the inner parts of the disc, the surface density increases as α U becomes smaller, because all discs have the same ˙ . M
The local maximum in surface density (positive surface density gradient) at the snow line also deceases when α U decreases, because the contrast α U to α D decreases. In the case of α A = 50 α U the positive surface density gradient at the ice line has completely vanished and only a flat plateau is visible. This might be enough to stop inward migration.
The migration map for α A = 10 α U is discussed in section 3. We do not display the migration maps for α A = 20 α U and α A = 50 α U here because the migration map looks quite similar to that presented in Fig. 7 (bottom). The only di ff erence is that the region of outward migration is smaller (e.g. outward migration is still possible only between 3 and 10 MEarth ). In Fig. 12 the migration map for α A = 100 α U is displayed. Interestingly, this migration map still features a small region of outward migration, even though that there is only a plateau in the surface density and not a bump.
This is because the barotropic parts of the corotation torque still generate a positive (although weak) contribution, and the steep negative radial temperature gradient in the region produces a positive entropy related corotation torque. The viscosity, al-
Fig. 12. Migration map for the disc with ˙ M = 1 × 10 -7 M /circledot / yr that features α A = 100 α U . The regions encircled in black mark regions in which outward migration is possible. The vertical red line in the plot marks the ice line at 190K.
Fig. 13. Migration map for the disc with ˙ M = 1 × 10 -7 M /circledot / yr with steep gradient in the α transition. The encircled regions in black mark regions in which outward migration is possible. The vertical red line in the plot marks the ice line at 190K.
though low, is still high enough to prevent torque saturation (this is not the case in the example discussed in Appendix A).
## 5.3. Change of the transition α U to α D
The transition from α U to α D in eq. 3 was implemented across a range of 60K, because this corresponds to the temperature at which the opacity transition is smoothed at the ice line. However, 60K can be considered a broad temperature range just for the sublimation and condensation of ice grains. We now change this range to 10K. This is the only modification to eq. 3. The corresponding surface density is shown in Fig. 11 as the light-blue line.
A much steeper surface density gradient at the location of the ice line is clearly visible. This change of surface density will result in a planet trap, similar to Fig. 7, but with a higher intensity. This also traps planets with masses smaller than 0 5 . MEarth (Fig. 13). Note that an inverse pressure gradient now exists (Fig. 14). An inverse pressure gradient in a protoplanetary disc stops the inward migration of small bodies that undergo gas drag (Brauer et al. 2008), promotes the onset of the streaming instability (Johansen & Youdin 2007), and significantly facilitates planetesimals formation (Bai & Stone 2010a,b), as also discussed in section 3.2.
The maximum steepness a density gradient can have is constrained by the disc height H because a density bump would be unstable to the Rayleigh instability if its width were smaller than H (Yang & Menou 2010). Here the transition occurs on a length scale of 2 H .
Fig. 14. Midplane pressure for the ˙ M = 1 × 10 -7 M /circledot / yr discs with di ff erent transition ranges in the temperature for α U to α D .
Fig. 15. Pressure gradient at the α transition as a function of the viscosity gradient. The data are taken from simulations of the discs with ˙ M = 1 × 10 -7 M /circledot / yr and di ff erent α transitions.
## 6. Gradients at the α transition
We have presented here detailed simulations of discs with transitions in the α parameter that resulted in changes in the disc structure. If the α transition is large enough, small mass planets can be trapped. The smallest mass of trapped planets decreases with increasing value of the radial positive surface density gradient. At the same location of the disc, ∆ v cs of the streaming instability (eq. 9) is greatly reduced, which facilitates planetesimal formation. Again, an inversion of the pressure gradient is only visible if the disc features a very steep positive surface density gradient that can overcompensate for the negative temperature gradient (eq. 8).
We now investigate which changes in α , or more precisely, in viscosity are needed to create a pressure bump in the disc. In Fig. 15 the pressure gradient, d ln( P ) / d ln( r ), is displayed as a function of the viscosity gradient, d ln( ν / ) d ln( r ). In addition to the simulation data, we provide a rough fit through the data points to estimate the viscosity gradient necessary to cause a pressure bump in the disc.
The change in viscosity required to observe a pressure bump is quite large, d ln( ν / ) d ln( r ) ≈ -28. The reduction in viscosity needs to be this high because of the vertical structure of the disc. As stated above, in a 1D (radial) disc, a reduction in viscosity will be equally compensated for by an increase in surface density. If the vertical structure is taken into account, however, this is di ff erent because most of the accretion rate can be carried by the active layer (see Fig. 5), which results in a lower density increase in the midplane regions of the disc. Clearly, a pressure bump is much harder to achieve in a disc with a layered vertical structure than in a 1D disc structure.
The only simulation in which we observed a bump in the midplane pressure featured a jump of α by a factor of 10 over a temperature range of 10K, which corresponds to 0 13AU in . this simulation. We justified this jump in α by more particles being available beyond the ice line ( r > rice ) that can quench the MRI (Kretke & Lin 2007). It seems realistic that the temperature range over which water vapour condensates into ice grains is only 10K. But a change of α of a factor of 10 caused by more ice grains is doubtful. A smaller reduction of α seems more reasonable. However, if α is only changed by a factor of 2, no bump in the midplane pressure can be observed. This indicates the reduction in α needed to create a pressure bump is ≈ 5.
## 7. Summary
We have investigated the influence of viscosity transitions on the structure and migration rate of planets in accretion discs, which feature the same mass flow ˙ M through every radial section of the disc. The structure of the disc was calculated by using 2D hydrodynamical simulations that feature viscous and stellar heating as well as radiative cooling. The viscosity transitions were implemented by reducing the α parameter of the viscosity prescription. When the viscosity is reduced in parts of the disc, the disc's flow adapts to keep the vertically integrated ˙ M independent of radius.
For high ˙ M rates, the gas surface density of the disc Σ G is high enough to shield the inner parts of the disc from cosmic and X-rays, resulting in a region of reduced α , the dead zone. This region of reduced viscosity is denser than the active layers, which creates a local maximum in surface density. Because we imposed a radial α transition at the ice line as well, we created a planet trap at ≈ 4AU. This planet trap is very e ffi cient in trapping small-mass planets ( MP > 0 5 . MEarth ), in contrast to discs without a transition in α , where a much larger planetary mass is needed to stop inward migration by the positive corotation torque (Fig. 7).
As the disc evolves to lower ˙ M rates, the surface density decreases, so that the disc is unable to completely shield itself from cosmic and X-rays. This implies that as the disc evolves, the regions of low α shrink and the disc finally reaches a state without reduced α . In this case the only possibility to stop planet migration is the positive entropy related corotation torque (when it exists).
The migration of small solids (pebbles) is instead related to gas drag and is sensitive to how sub-Keplerian the gas-disc is. A pressure bump is needed to stop the inward migration of pebbles. To create a pressure bump a positive radial volume density gradient is needed that is strong enough to compensate for the negative temperature gradient. This can only be achieved if the viscosity transition is very steep. More precisely, our simulations show that a reduction of about a factor of 5 of α is needed over a temperature range of ≈ 10K.
Weexplained the reduction in α at the ice line by more grains that can block the MRI (Kretke & Lin 2007). This view is under debate and the magnitude of this change is not known.
Fig. A.1. Surface density for ˙ M = 1 × 10 -8 M /circledot / yr disc with α A = 0 003, which is . ≈ 10 lower than in Fig. 2. The ˙ M = 1 × 10 -7 M /circledot / yr disc with larger α A is overplotted for comparison.
Nevertheless, we can conclude here that quenching the MRI at the ice line by ice grains (Kretke & Lin 2007) can be an e ff ective trap for Earth-mass planets. This does not necessarily imply a pressure bump in the disc structure, however small dust and ice particles can be trapped there. It implies a significantly reduced pressure gradient, which in turn can significantly facilitate planetesimal formation by the streaming instability (Bai & Stone 2010a,b).
Acknowledgements. B. Bitsch and A. Morbidelli have been partly sponsored through the Helmholtz Alliance Planetary Evolution and Life . We thank the Agence Nationale pour la Recherche under grant ANR-13-BS05-0003-01 (MOJO). We also thank A. Johansen for useful discussions. In addition we thank an anonymous referee for his her / input. The computations were made done on the Mesocentre SIGAMM machine, hosted by the Observatoire de la Cˆ ote d'Azur.
## Appendix A: Interchange between Σ and ν
In Paper II we mainly focused on discs that have a ten times higher surface density and a ten times smaller α than the discs presented here, which gives the same ˙ M rate. Because the penetration of cosmic rays into the disc is dependent on the discs surface density, this is a crucial quantity for developing a low viscosity region through the shielding of cosmic rays. Here, we take a disc with ˙ M = 1 × 10 -8 M /circledot / yr , but with the viscosity parameter α A = 0 003, which is ten times lower than presented in . section 3.
In Fig. A.1 the surface density of the ˙ M = 1 × 10 -8 M /circledot / yr disc with α A = 0 003 is displayed. The prescription of the . α transition follows the one given in eq. 3. As in Fig. 2 (middle), a bump in the surface density profile is visible.
However, the migration map (Fig. A.2) does not show a positive torque that acts on the planet at any radial distance. The negative torque is smallest in the region close to the surface density bump at ≈ 2 5AU. This indicates that in principle the bump . in the surface density should work as a planet trap, but as the α parameter in that region of the disc is very small ( α D = 3 × 10 -5 ), the torques do saturate and cannot sustain outward migration any more (Masset et al. (2006)). This result shows that the parameter range of the α transitions that allow the formation of a planet trap is very narrow.
Fig. A.2. Migration map for ˙ M = 1 × 10 -8 M /circledot / yr disc with α A = 0 003, which is . ≈ 10 lower than in Fig. 9. The vertical blue line marks the ice line at T = 190K.
## References
Alexander, R. D. & Pascucci, I. 2012, MNRAS, 442, pp.82
Bai, X. & Goodman, J. 2009, ApJ, 701, pp.737 Bai, X. N. & Stone, J. M. 2010a, ApJ, 722, pp. 1437 Bai, X. N. & Stone, J. M. 2010b, ApJ, 722, L220 Balbus, S. A. & Hawley, J. F. 1998, Reviews of Modern Physics, 70, 1 Baruteau, C., Crida, A., Paardekooper, S. J., et al. 2014, in Protostars and Planets VI, arXiv:1312.4293 Baruteau, C. & Masset, F. 2008, ApJ, 672, 1054 Bell, K. R. & Lin, D. N. C. 1994, ApJ, 427, 987 Bitsch, B., Crida, A., Morbidelli, A., Kley, W., & Dobbs-Dixon, I. 2013, A&A, 549, id.A124 Bitsch, B. & Kley, W. 2011, A&A, 536, A77 Bitsch, B., Morbidelli, A., Lega, E., & Crida, A. 2014, astro-ph.EP Brauer, F., Henning, T., & Dullemond, C. P. 2008, A&A, 487, pp.L1 Chiang, E. I. & Goldreich, P. 1997, ApJ, 490, 368 Crida, A. & Morbidelli, A. 2007, MNRAS, 377, 1324 Crida, A., Morbidelli, A., & Masset, F. 2006, Icarus, 181, 587 Dzyurkevich, N., Turner, N., Henning, T., & Kley, W. 2013, ApJ, 765, id.114 Fromang, S., Lyra, W., & Masset, F. 2011, A&A, id.A107 Hartmann, L., Calvet, N., Gullbring, E., & D'Alessio, P. 1998, ApJ, 495, p.385 Hayashi, C. 1981, Progress of Theoretical Physics Supplement, 70, pp.35 Johansen, A., Oishi, J. S., Mac Low, M. M., et al. 2007, Nature, 448, pp. 1022 Johansen, A. & Youdin, A. 2007, ApJ, 662, pp. 627 Klahr, H. H. & Bodenheimer, P. 2003, ApJ, 582, 869 Kley, W., Bitsch, B., & Klahr, H. 2009, A&A, 506, 971 Kley, W. & Crida, A. 2008, A&A, 487, L9 Kley, W. & Lin, D. N. C. 1992, ApJ, 397, 600 Kokubo, E. & Ida, S. 1998, Icarus, 131, pp.171 Kretke, K. & Lin, D. N. C. 2007, ApJ, 664, L55 Lambrechts, M. & Johansen, A. 2012, A&A, 544, id.A32 Lega, E., Crida, A., Bitsch, B., & Morbidelli, A. 2014, astro-ph.EP, 1402.2834 Levermore, C. D. & Pomraning, G. C. 1981, ApJ, 248, 321 Levison, H. F., Thommes, E., & Duncan, M. J. 2010, AJ, 139, pp.1297 Lin, D. N. C. & Papaloizou, J. 1986, ApJ, 307, 395 Martin, R. G. & Lubow, S. H. 2014, MNRAS, 437, p.682 Masset, F. S., Morbidelli, A., Crida, A., & Ferreira, J. 2006, ApJ, 642, 478 Mihalas, D. & Weibel Mihalas, B. 1984, Foundations of radiation hydrodynamics (New York: Oxford University Press, 1984) Morbidelli, A., Crida, A., Masset, F., & Nelson, R. 2008, A&A, 478, 929 Morbidelli, A. & Nesvorny, D. 2012, A&A, 546, id.A18 Nakagawa, Y., Sekiya, M., & Hayashi, C. 1986, Icarus, 67, p. 375 Nelson, R. P., Gressel, O., & Umurhan, O. M. 2013, MNRAS, 435, p.2610 Paardekooper, S. J., Baruteau, C., Crida, A., & Kley, W. 2010, MNRAS, 401, 1950 + Paardekooper, S. J., Baruteau, C., & Kley, W. 2011, MNRAS, 410, 293 Paardekooper, S. J. & Mellema, G. 2006, A&A, 459, L17 Perez-Becker, D. & Chiang, E. 2011, ApJ, 727, id.2 Pierens, A. & Nelson, R. P. 2010, A&A, 520, id.A14 Pollack, J. B., Hubickyj, O., Bodenheimer, P., et al. 1996, Icarus, 124, 62 Shakura, N. I. & Sunyaev, R. A. 1973, A&A, 24, 337 Takeuchi, T. & Lin, D. N. C. 2002, ApJ, 581, pp.1344 Urpin, V. 1984, Soviet Astronomy, 280, p.50 Ward, W. R. 1997, Icarus, 126, 261
Yang, C.-C. & Menou, K. 2010, MNRAS, 402, pp. 2436
|
10.1051/0004-6361/201424015
|
[
"Bertram Bitsch",
"Alessandro Morbidelli",
"Elena Lega",
"Kathrine Kretke",
"Aurélien Crida"
] |
2014-08-05T16:12:03+00:00
|
2014-08-05T16:12:03+00:00
|
[
"astro-ph.EP"
] |
Stellar irradiated discs and implications on migration of embedded planets III: viscosity transitions
|
The migration strength and direction of embedded low-mass planets depends on
the disc structure. In discs with an efficient radiative transport, the
migration can be directed outwards for planets with more than 3-5 Earth masses.
This is due to the entropy driven corotation torque, a process that extends the
lifetimes of growing planetary embryos. We investigate the influence on the
disc structure caused by a jump in the alpha parameter of the viscosity to
model a dead-zone structure in the disc. We focus on M-dot discs, which have a
constant net mass flux. Using the resulting disc structure, we investigate the
consequences for the formation of planetesimals and determine the regions of
outward migration for proto-planets. We performed numerical hydrosimulations of
M-dot discs in the r-z-plane. We used the explicit/implicit hydrodynamical code
FARGOCA that includes a full tensor viscosity and stellar irradiation as well
as a two-temperature solver that includes radiation transport in the
flux-limited diffusion approximation. Viscosity transitions inside the disc
create transitions in density that stop inward migration for small planets
through the so-called "planet trap" mechanism. This mechanism also works for
planets down to M_P > 0.5M_E. Additionally, the viscosity transitions change
the pressure gradient in the disc, which facilitates planetesimal formation via
the streaming instability. However, a very steep transition in viscosity is
needed to achieve in a pressure bump in the disc. The transition in viscosity
facilitates planetesimal formation and can stop the migration of small-mass
planets (M_P>0.5M_E), but still does not halt inward migration of smaller
planets and planetesimals that are affected by gas drag. A very steep, probably
unrealistic viscosity gradient is needed to trap planets of smaller masses and
halt gas-drag-driven planetesimal migration at a pressure bump.
|
1408.1017v1
|
## The complexity of approximating a trembling hand perfect equilibrium of a multi-player game in strategic form
Kousha Etessami , Kristoffer Arnsfelt Hansen , Peter Bro Miltersen , and 1 2 2 Troels Bjerre Sørensen 3
- 2
- IT-University of Copenhagen. [email protected]
- 1 University of Edinburgh. [email protected] Aarhus University. { arnsfelt,bromille @cs.au.dk } 3
Abstract. We consider the task of computing an approximation of a trembling hand perfect equilibrium for an n -player game in strategic form, n ≥ 3. We show that this task is complete for the complexity class FIXP a . In particular, the task is polynomial time equivalent to the task of computing an approximation of a Nash equilibrium in strategic form games with three (or more) players.
## 1 Introduction
Arguably [17], the most important refinement of Nash equilibrium for finite games in strategic form (a.k.a. games in normal form, i.e., games given by their tables of payoffs) is Reinhard Selten's [15] notion of trembling hand perfection . The set of trembling hand perfect equilibria of a game is a non-empty subset of the Nash equilibria of that game. Also, many 'unreasonable' Nash equilibria of many games, e.g., those relying on 'empty threats' in equivalent extensive forms of those games, are not trembling hand perfect, thus motivating and justifying the notion. The importance of the notion is illustrated by the fact that Selten received the Nobel prize in economics together with Nash (and Harsanyi), 'for their pioneering analysis of equilibria in the theory of non-cooperative games'. In this paper, we study the computational complexity of finding trembling hand perfect equilibria of games given in strategic form.
The computational complexity of finding a Nash equilibrium of a game in strategic form is well-studied. When studying this computational task, we assume that the game given as input is represented as a table of integer (or rational) payoffs, with each payoff given in binary notation. The output is a strategy profile, i.e., a family of probability distributions over the strategies of each player, with each probability being a rational number with numerator and denominator given in binary notation. The computational task is therefore discrete and we are interested in the Turing machine complexity of solving it. Papadimitriou [13] showed that for the case of two players, the problem of computing an exact Nash equilibrium is in PPAD , a natural complexity class introduced in that paper, as a consequence of the Lemke-Howson algorithm [10] for solving this task. For the case of three or more players, there are games where no Nash equilibrium which uses only rational probabilities exists [12], and hence considering some relaxation of the notion of 'computing' a Nash equilibrium is necessary to stay within the discrete input/output framework outlined above. In particular, Papadimitriou showed that the problem of computing an /epsilon1 -Nash equilibrium, with /epsilon1 > 0 given as part of the input in binary notation, is also in PPAD , as a consequence of Scarf's algorithm [14] for solving this task. Here, an /epsilon1 -Nash equilibrium is a strategy profile where no player can increase its utility by more than /epsilon1 by deviating. In breakthrough papers, Daskalakis et al. [5] and Chen and Deng [4] showed that both tasks are also hard for PPAD , hence settling their complexity: Both are PPAD -complete. Subsequently, Etessami and Yannakakis [6] pointed out that for some games, /epsilon1 -Nash equilibria can be so remote from any exact Nash equilibrium (unless /epsilon1 is so small that its binary notation has encoding size exponential in the size of the game), that the former tells us little or nothing about the latter. For such games, the /epsilon1 -Nash relaxation is a bad proxy for Nash equilibrium, assuming the latter is what we are actually interested in computing. Motivated by this, they suggested a different relaxation: Compute a strategy profile with /lscript ∞ -distance at most δ from an exact Nash equilibrium, with δ > 0 again given as part of the input in binary notation. In other words, compute an actual Nash equilibrium to a desired number of bits of accuracy . They showed that this problem is complete for a natural complexity class FIXP a that they introduced in the same paper. Informally, FIXP a is the class of discrete search problems that can be reduced to approximating (within desired /lscript ∞ -distance) any one of the Brouwer fixed points of a function given by an algebraic circuit using gates: + , - ∗ , , /, max min. (We will formally define , FIXP a later.)
In this paper, we want to similarly understand the case of trembling hand perfect equilibrium. For the case of two players, the problem of computing an exact trembling hand perfect equilibrium is PPAD -complete. This follows from a number of known exact pivoting algorithms for computing refinements of this notion [18,11,16]. For the case of three or more players, we are not aware of any natural analogue of the notion of /epsilon1 -Nash equilibrium as an approximate proxy for a trembling hand perfect equilibrium. 4 Thus, we only discuss in this paper the approximation notion of Etessami and Yannakakis. The main result of the present paper is the following:
Theorem 1. The following computational task is FIXP a -complete for any n ≥ 3 : Given an integer payoff table for an n -player game Γ , and a rational δ > 0 , with all numbers given in standard binary notation, compute (the binary representation of) a strategy profile x ′ with rational probabilities having /lscript ∞ distance at most δ to a trembling hand perfect equilibrium of Γ .
4 The already studied notion of an /epsilon1 -perfect equilibrium ( /epsilon1 -PE), which we discuss later, does not qualify as such an analogue: For some three-player games, every /epsilon1 -PE uses irrational probabilities, and thus 'computing' an (exact) /epsilon1 -PE is just as problematic as computing an exact NE. Indeed, the notion of a /epsilon1 -PE is used as a technical step towards the definition of trembling hand perfect equilibrium, rather than as a natural 'numerical relaxation' of this notion.
As an immediate corollary of our main theorem, and the results of Etessami and Yannakakis, we have that approximating a Nash equilibrium and approximating a trembling hand perfect equilibrium are polynomial time equivalent tasks . In particular, there is a polynomial time algorithm that finds an approximation to a trembling hand perfect equilibrium of a given game, using access to any oracle solving the corresponding approximation problem for the case of Nash equilibrium. To put this result in perspective, we note that Nash equilibrium and trembling hand perfect equilibrium are computationally quite different in other respects: if instead of finding an equilibrium, we want to verify that a given strategy profile is such an equilibrium, the case of Nash equilibrium is trivial, while the case of trembling hand perfect equilibrium for games with 3 (or more) players is NP -hard [8]. This might lead one to believe that approximating a trembling hand perfect equilibrium for games with 3 or more players is likely to be harder than approximating a Nash equilibrium, but we show that this is not the case.
## 1.1 About the proof
Informally (for formal definitions, see below), FIXP (resp., FIXP a ) is defined as the complexity class of search problems that can be cast as exactly computing (resp., approximating) a Brouwer fixed point of functions represented by circuits over basis { + , ∗ -, , /, max min , } with rational constants. It was established in [6] that computing (resp., approximating) an actual Nash Equilibrium (NE) for a finite n -player game is FIXP -complete (resp., FIXP a -complete), already for n = 3. Since trembling hand perfect equilibria constitute a refinement of Nash Equilibria, to show that approximating a trembling hand perfect equilibrium is FIXP a -complete, we merely have to show that this task is in FIXP a .
An /epsilon1 -perfect equilibrium ( -PE for short) is defined to be a fully mixed strat/epsilon1 egy profile, x , where every strategy j of every player i that is played with probability x i,j > /epsilon1 must be a best response to the other player's strategies x -i . Then, a trembling hand perfect equilibrium (PE for short) is defined to be a limit point of a sequence of /epsilon1 -PEs, for /epsilon1 > 0, /epsilon1 → 0. Here, by limit point we mean, as usual, any point to which a subsequence of the sequence converges. Such a point must exist, by the Bolzano-Weierstrass theorem.
In rough outline, our proof that approximating a PE is in FIXP a has the following structure:
- 1. We first define (in section 3) for any n -player game Γ , a map, F /epsilon1 Γ , parameterized by a parameter /epsilon1 > 0, so that F /epsilon1 Γ defines a map from D /epsilon1 Γ to itself, where D /epsilon1 Γ denotes the space of fully mixed strategy strategy profiles x such that every player plays each strategy with probability at least /epsilon1 . Also, F /epsilon1 Γ ( x ) is described by a { + , - ∗ , , min max -circuit with , } /epsilon1 as one of its inputs. In particular, the Brouwer fixed point theorem applies to this map. We show that the circuit defining F /epsilon1 Γ can be computed in polynomial time from the input game instance Γ , and that every Brouwer fixed point of F /epsilon1 Γ is an /epsilon1 -PE of the original game Γ , making crucial use of, and modifying, a new fixed point characterization of NEs that was defined and used in [6].
- 2. We then show (in section 4) that if /epsilon1 ∗ > 0 is made sufficiently small as a function of the encoding size | Γ | of the game Γ , and of a parameter δ > 0, specifically if /epsilon1 ∗ ≤ δ 2 g ( | Γ | ) , where g is some polynomial, then any /epsilon1 ∗ -PE must be δ -close (in the l ∞ -norm) to an actual PE. This part of the proof relies on real algebraic geometry.
- 3. We then observe (in section 5) that for any desired δ , we can encode such a sufficiently small /epsilon1 ∗ > 0 as a circuit that is polynomially large in the encoding size of Γ and δ , simply by repeated squaring . We think of this as constructing a virtual infinitesimal and believe that this technique will have many other applications in the context of proving FIXP a membership using real algebraic geometry. Finally, plugging in the circuit for /epsilon1 ∗ for the input /epsilon1 in the circuit for F /epsilon1 Γ , we obtain a Brouwer function F /epsilon1 ∗ Γ ( x ), defined by a { + , - ∗ , , max min -circuit, such that any fixed point of , } F /epsilon1 ∗ Γ ( x ) is guaranteed to be a fully mixed strategy profile, x ∗ /epsilon1 ∗ , that is also within l ∞ distance δ of a PE, x ∗ , of Γ . The triangle inequality completes the proof.
## 2 Definitions and Preliminaries
## 2.1 Game-theoretic notions
We use Q + to denote the set of positive rational numbers. A finite n -player normal form game, Γ = ( N, S 〈 i 〉 i ∈ N , 〈 u i 〉 i ∈ N ), consists of a set N = { 1 , . . . , n } of n players indexed by their number, a set of n (disjoint) finite sets of pure strategies , S i , one for each player i ∈ N , and n rational-valued payoff functions u i : S → Q , from the product strategy space S = Π n i =1 S i to Q .
The elements of S , i.e., combinations of pure strategies, one for each player, are called pure strategy profiles . The assumption of rational values is for computational purposes. Each rational number r is represented as usual by its numerator and denominator in binary, and we use size r ( ) to denote the number of bits in the representation. The size | Γ | of the instance (game) Γ is the total number of bits needed to represent all the information in the game: the strategies of all the players and their payoffs for all s ∈ S .
A mixed strategy , x i , for a player i is a probability distribution on its set S i of pure strategies. Letting m i = | S i | , we view x i as a real-valued vector x i = ( x i, 1 , . . . , x i,m i ) ∈ [0 , 1] m i , where x i,j denotes the probability with which player i plays pure strategy j in the mixed strategy x i . Note that we must have x i ≥ 0 and ∑ m i i =1 x i,m i = 1. That is, a vector x i is a mixed strategy of player i iff it belongs to the unit simplex ∆ m i = { y ∈ R m i | y ≥ 0; ∑ m i j =1 y j = 1 . We } use the notation π i,j to identify the pure strategy j of player i , as well as its representation as a mixed strategy that assigns probability 1 to strategy j and probability 0 to the other strategies of player i .
A mixed strategy profile x = ( x , . . . , x 1 n ) is a combination of mixed strategies for all the players. That is, vector x is a mixed strategy profile iff it belongs to the product of the n unit simplexes for the n players, { x ∈ R m | x ≥ 0; ∑ m i j =1 x i,j = 1 for i = 1 , . . . , k } . We let D Γ denote the set of all mixed profiles for game Γ .
The profile is fully mixed if all the pure strategies of all players have nonzero probability. We use the notation x -i to denote the subvector of x induced by the pure strategies of all players except for player i . If y i is a mixed strategy of player i , we use ( y i ; x -i ) to denote the mixed profile where everyone plays the same strategy as x except for player i , who plays mixed strategy y i .
The payoff function of each player can be extended from pure strategy profiles to mixed profiles, and we will use U i to denote the expected payoff function for player i . Thus the (expected) payoff U i ( x ) of mixed profile x for player i is ∑ x 1 ,j 1 . . . x k,j k u i ( j 1 , . . . , j k ) where the sum is over all pure strategy profiles ( j 1 , . . . , j k ) ∈ S .
A Nash equilibrium (NE) is a (mixed) strategy profile x ∗ such that all i = 1 , . . . , n and every mixed strategy y i for player i , U i ( x ∗ ) ≥ U i ( y i ; x ∗ -i ). It is sufficient to check switches to pure strategies only, i.e., x ∗ is a NE iff U i ( x ∗ ) ≥ U i ( π i,j ; x ∗ -i ) for every pure strategy j ∈ S i , for each player i = 1 , . . . , n . Every finite game has at least one NE [12].
A mixed profile x is called a /epsilon1 -perfect equilibrium ( /epsilon1 -PE) if it is (a) fully mixed, i.e., x i,j > 0 for all i , and (b), for every player i and pure strategy j , if x i,j > /epsilon1 , then the pure strategy π i,j is a best response for player i to x -i . We call a mixed profile x ∗ , a trembling hand perfect equilibrium (PE) of Γ if it is a limit point of /epsilon1 -PEs of the game Γ . In other words, we call x a PE if there exists a sequence /epsilon1 k > 0, such that lim k →∞ /epsilon1 k = 0, and such that for all k there is a corresponding /epsilon1 k -PE, x /epsilon1 k of Γ , such that lim k →∞ x /epsilon1 k = x ∗ . Every finite game has at least one PE, and all PEs are NEs [15].
## 2.2 Complexity theoretic notions
A { + , - ∗ , , max min , } -circuit is a circuit with inputs x , x 1 2 , . . . , x n , as well as rational constants, and a finite number of (binary) computation gates taken from { + , - ∗ , , min max , with a subset of the computation gates labeled , } { o , o 1 2 , . . . , o m } and called output gates. 5
All circuits of this paper are { + , - ∗ , , min max -circuits, so we shall often just , } write 'circuit' for ' { + , - ∗ , , min max -circuit'. A circuit computes a continuous , } function from R n → R m (and Q n → Q m ) in the natural way. Abusing notation slightly, we shall often identify the circuit with the function it computes.
By a (total) multi-valued function , f , with domain A and co-domain B , we mean a function that maps each a ∈ A to a non-empty subset f ( a ) ⊆ B . We use f : A /dblarrowheadright B to denote such a function. Intuitively, when considering a multivalued function as a computational problem, we are interested in producing just one of the elements of f ( a ) on input a , so we refer to f ( a ) as the set of allowed outputs . A multi-valued function f : { 0 1 , } ∗ /dblarrowheadright R ∗ is said to be in FIXP if there is a polynomial time computable map, r , that maps each instance I ∈ { 0 1 , } ∗ of f to r I ( ) = 〈 1 k I , 1 d I , P I , C I , a I , b I 〉 , where
-k , d I I are positive integers and a , b I I ∈ Q d I .
5 Note that the gates { + , - ∗ , , min max , } are of course redundant: gates { + , ∗ , max } with rational constants are equally expressive.
- -P I is a convex polytope in R k I , given as a set of linear inequalities with rational coefficients.
- -C I is a circuit which maps P I to itself.
- -φ I : { 1 , . . . , d I } → { 1 , . . . , k I } is a finite function given by its table.
/negationslash
- -f ( I ) = { ( a y I i φ I ( ) i + b I i ) d I i =1 | y ∈ P I ∧ C I ( y ) = y } . Note that f ( I ) = ∅ , by Brouwer's fixed point theorem.
The above is in fact one of many equivalent characterizations of FIXP [6]. Informally, FIXP are those real vector multi-valued functions, with discrete inputs, that can be cast as Brouwer fixed point computations. A multi-valued function f : { 0 1 , } ∗ /dblarrowheadright R ∗ is said to be FIXP -complete if:
## 1. f ∈ FIXP , and
- 2. for all g ∈ FIXP , there is a polynomial time computable map, mapping instances I of g to 〈 y I , 1 k I , 1 d I , φ I , a I , b I 〉 , where y I is an instance of f , k I and d I are positive integers, φ I maps { 1 , . . . , d I } to { 1 , . . . , k I } , a I and b I are d I -tuples with rational entries, so that g I ( ) ⊇ { ( a z I i φ I ( ) i + b I i ) d I i =1 | z ∈ f ( y I ) } . In other words, for any allowed output z of f on input y I , the vector ( a z I i φ I ( ) i + b I i ) d I i =1 is an allowed output of g on input I .
Etessami and Yannakakis [6] showed that the multi-valued function which maps games in strategic form to their Nash equilibria is FIXP -complete. 6
Since the output of a FIXP function consists of real-valued vectors, and as there are circuits whose fixed points are all irrational, a FIXP function is not directly computable by a Turing machine, and the class is therefore not directly comparable with standard complexity classes of total search problems (such as PPAD , PLS , or TFNP ). This motivates the following definition of the discrete class FIXP a , also from [6]. A multi-valued function f : { 0 1 , } ∗ /dblarrowheadright { 0 1 , } ∗ (a.k.a. a totally defined discrete search problem) is said to be in FIXP a if there is a function f ′ ∈ FIXP , and polynomial time computable maps δ : { 0 1 , } ∗ → Q + and g : { 0 1 , } ∗ →{ 0 1 , } ∗ , such that for all instances I ,
$$f ( I ) \supseteq \{ \, g ( \langle I, y \rangle ) \, | \, y \in \mathbb { Q } ^ { * } \, \wedge \, \exists y ^ { \prime } \in f ^ { \prime } ( I ) \, \colon \, \| y - y ^ { \prime } \| _ { \infty } \leq \delta ( I ) \, \}.$$
Informally, FIXP a are those totally defined discrete search problems that reduce to approximating exact Brouwer fixed points. A multi-valued function f : { 0 1 , } ∗ /dblarrowheadright { 0 1 , } ∗ is said to be FIXP a -complete if:
## 1. f ∈ FIXP a , and
- 2. For all g ∈ FIXP a , there are polynomial time computable maps r 1 , r 2 : { 0 1 , } ∗ →{ 0 1 , } ∗ , such that g I ( ) ⊇ { r 2 ( 〈 I, z 〉 ) | z ∈ f ( r 1 ( I )) } .
Etessami and Yannakakis showed that the multi-valued function that maps pairs 〈 Γ, δ 〉 , where Γ is a strategic form game and δ > 0, to the set of rational δ -approximations (in /lscript ∞ -distance) of Nash equilibria of Γ , is FIXP a -complete.
6 To view the Nash equilibrium problem as a total multi-valued function, f Nash : { 0 1 , } ∗ /dblarrowheadright R ∗ , we can view all strings in { 0 1 , } ∗ as encoding some game, by viewing 'ill-formed' input strings as encoding a fixed trivial game.
## 3 Computing /epsilon1 -PEs in FIXP
Given a game Γ , let m = ∑ i ∈ N m i denote the total number of pure strategies of all players in Γ . For /epsilon1 > 0, let D /epsilon1 Γ ⊆ D Γ denote the polytope of fully mixed profiles of Γ such that furthermore every pure strategy is played with probability at least /epsilon1 > 0 (recall that D Γ is the polytope of all strategy profiles). In this section, we show the following theorem.
Theorem 2. There is a function, F /epsilon1 Γ ( x ) : D Γ → D /epsilon1 Γ , given by a circuit computable in polynomial time from Γ , with the circuit having both x and /epsilon1 > 0 as its inputs, such that for all fixed 0 < /epsilon1 < 1 /m , every Brouwer fixed point of the function F /epsilon1 Γ ( x ) is an /epsilon1 -PE of Γ . In particular, the problem of computing an /epsilon1 -perfect equilibrium for a finite n -player normal form game is in FIXP .
The rest of the section is devoted to the proof of Theorem 2. We will directly use, and somewhat modify, a construction developed and used in [6] (Lemmas 4.6 and 4.7, and definitions before them) which characterize the Nash Equilibria of a game as fixed points of a { + , - ∗ , , max min -circuit. In particular, compared , } to Nash's original functions [12], the use of division is avoided. The construction defined in [6] that we modify amounts to a concrete algebraic realization of certain geometric characterizations of Nash Equilibria that were described by Gul, Pierce, and Stachetti in [7].
Concretely, suppose we are given 0 < /epsilon1 < 1 /m . For each mixed strategy profile x , let v x ( ) be a vector which gives the expected payoff of each pure strategy of each player with respect to the profile x for the other players. That is, vector x is a vector of dimension m , whose entries are indexed by pairs ( i, j ) , i = 1 , . . . , n ; j = 1 , . . . , m i , and v x ( ) is also a vector of dimension m whose ( i, j )-entry is U i ( π i,j ; x -i ). Let h x ( ) = x + v x ( ). We can write h x ( ) as ( h 1 ( x , . . . , h ) n ( x )) where h i ( x ) is the subvector corresponding to the strategies of player i . For each player i , consider the function f i,x ( ) = t ∑ j ∈ S i max( h ij ( x ) -t, /epsilon1 ). Clearly, this is a continuous, piecewise linear function of t . The function is strictly decreasing as t ranges from -∞ (where f i,x ( ) = + t ∞ ) up to max j h ij ( x ) -/epsilon1 (where f i,x ( ) = t m i · /epsilon1 ). Since we have m i · /epsilon1 < 1, there is a unique value of t , call it t i , where f i,x ( t i ) = 1. The function F /epsilon1 Γ is defined as follows:
$$F _ { \Gamma } ^ { \epsilon } ( x ) _ { i j } = \max ( h _ { i j } ( x ) - t _ { i }, \epsilon )$$
for every i = 1 , . . . , n , and j ∈ S i . From our choice of t i , we have ∑ j ∈ S i F /epsilon1 Γ ( x ) ij = 1 for all i = 1 , . . . , n , thus for any mixed profile, x , we have F /epsilon1 Γ ( x ) ∈ D /epsilon1 Γ . So F /epsilon1 Γ maps D Γ to D /epsilon1 Γ , and since it is clearly also continuous, it has fixed points, by Brouwer's theorem.
Lemma 1. For 0 < /epsilon1 < 1 /m , every fixed point of the function F /epsilon1 Γ is an /epsilon1 -PE of Γ .
Proof. If x is a fixed point of F /epsilon1 Γ , then x ij = max( x ij + ( ) v x ij -t , /epsilon1 i ) for all i, j . Recall that v x ( ) ij = U i ( π i,j ; x -i ) is the expected payoff for player i of his j 'th pure strategy π i,j , with respect to strategies x -i of the other players.
Note that the equation x ij = max( x ij + U i ( π i,j ; x -i ) -t , /epsilon1 i ) implies that U i ( π i,j ; x -i ) = t i for all i, j such that x ij > /epsilon1 , and that U i ( π i,j ; x -i ) ≤ t i for all i, j such that x ij = . Consequently, by definition, /epsilon1 x constitutes an /epsilon1 -PE. /intersectionsq /unionsq
The following Lemma shows that we can implement the function F /epsilon1 Γ ( x ) by a circuit which has x and /epsilon1 as inputs. The proof exploits sorting networks.
Lemma 2. Given Γ , we can construct in polynomial time a { + , - ∗ , , max min , } -circuit that computes the function F /epsilon1 Γ ( x ) , where x and /epsilon1 are inputs to the circuit.
Proof. The circuit does the following.
Given a vector x ∈ D Γ , first compute y = h x ( ) = x + ( ). It is clear from v x the definition of v x ( ) that y can be computed using + , ∗ gates. For each player i , let y i be the corresponding subvector of y induced by the strategies of player i . Sort y i in decreasing order, and let z i be the resulting sorted vector, i.e. the components of z i = ( z i 1 , . . . , z im i ) are the same as the components of y i , but they are sorted: z i 1 ≥ z i 2 ≥ . . . ≥ z im i . To obtain z i , the circuit uses a polynomial sized sorting network, W i , for each i (see e.g. Knuth [9] for background on sorting networks). For each comparator gate of the sorting network we use a max and a min gate.
Using this, for each player i , we compute t i as the following expression:
$$\max \{ ( 1 / l ) \cdot ( ( \sum _ { j = 1 } ^ { l } z _ { i j } ) + ( m _ { i } - l ) \cdot \epsilon - 1 ) | l = 1, \cdots, m _ { i } \}$$
We will show below that this expression does indeed give the correct value of t i . Finally, we output x ′ ij = max( y ij -t , /epsilon1 i ) for each i = 1 , . . . , d ; j ∈ S i .
We now have to establish that t i = max (1 { /l ) (( · ∑ l j =1 z ij )+( m i -l ) · /epsilon1 -1) | l = 1 , · · · , m i } . Consider the function f i,x ( ) = t ∑ j ∈ S i max( z ij -t, /epsilon1 ) as t decreases from z i 1 -/epsilon1 where the function value is at its minimum of m /epsilon1 i , down until the function reaches the value 1. In the first interval from z i 1 -/epsilon1 to z i 2 -/epsilon1 the function is f i,x ( ) t = z i 1 -t + ( m i -1) · /epsilon1 ; in the second interval from z i 2 -/epsilon1 to z i 3 -/epsilon1 it is f i,x ( ) t = z i 1 + z i 2 -2 t + ( m i -2) · /epsilon1 , and so forth. In general, in the l -th interval, f i,x ( ) t = ∑ l j =1 ( z ij -t ) + ( m i -l ) · /epsilon1 = ∑ l j =1 z ij -lt + ( m i -l ) · /epsilon1 . If the function reaches the value 1 in the l 'th interval, then clearly t i = (( ∑ l j =1 z ij ) + ( m i -l ) · /epsilon1 -1) /l . In that case, furthermore for k < l , we have ∑ k j =1 ( z ij -t i ) + ( m i -k ) · /epsilon1 ≤ ∑ l j =1 ( z ij -t i ) + ( m i -l ) · /epsilon1 = 1, because in that case we know ( z ij -t i ) ≥ /epsilon1 for every j ∈ { 1 , . . . , l } . Therefore, in this case (( ∑ k j =1 z ij ) + ( m i -k ) · /epsilon1 -1) /k ≤ t i . On the other hand, if l < m i , then for k > l we have t i ≥ z ik -/epsilon1 , i.e., z ik -t i ≤ /epsilon1 , and thus for all k > l , k ≤ m i , we have ∑ k j =1 ( z ij -t i ) + ( m i -k ) · /epsilon1 ≤ ∑ l j =1 ( z ij -t i ) + ( m i -l ) · /epsilon1 = 1. Thus again (( ∑ k j =1 z ij ) + ( m i -k ) · /epsilon1 -1) /k ≤ t i . Therefore, t i = max (1 { /l ) · (( ∑ l j =1 z ij ) + ( m i -l ) · /epsilon1 -1) | l = 1 , · · · , m i } . /intersectionsq /unionsq
Lemma 1 and Lemma 2 together immediately imply Theorem 2.
## 4 Almost implies near
As outlined in the introduction, in this section, we want to exploit the 'uniform' function F /epsilon1 Γ ( x ) devised in the previous section for /epsilon1 -PEs, and construct a 'small enough' /epsilon1 ∗ > 0 such that any fixed point of F /epsilon1 ∗ Γ ( x ) is δ -close, for a given δ > 0, to an actual PE.
The following is a special case of the simple but powerful 'almost implies near' paradigm of Anderson [1].
Lemma 3 (Almost implies near). For any fixed strategic form game Γ , and any δ > 0 , there is an /epsilon1 > 0 , so that any /epsilon1 -PE of Γ has /lscript ∞ -distance at most δ to some PE of Γ .
Proof. Assume to the contrary that there is a game Γ and a δ > 0 so that for all /epsilon1 > 0, there is an /epsilon1 -PE x /epsilon1 of Γ so that there is no PE in a δ -neighborhood (with respect to l ∞ norm) of x /epsilon1 . Consider the sequence ( x 1 /n n ) ∈ N . Since this is a sequence in a compact space (namely, the space of mixed strategy profiles of Γ ), it has a limit point, x ∗ , which is a PE of Γ (since x /epsilon1 is a /epsilon1 -PE). But this contradicts the statement that there is no PE in a δ -neighborhood of any of the profiles x 1 /n . /intersectionsq /unionsq
A priori, we have no bound on /epsilon1 , but we next use the machinery of real algebraic geometry [2,3] to obtain a specific bound as a 'free lunch', just from the fact that Lemma 3 is true.
Lemma 4. There is a constant c , so that for all integers n, m, k, B ∈ N and δ ∈ Q + , the following holds. Let /epsilon1 ≤ min( δ, 1 /B ) n cm 3 . For any n -player game Γ , with a combined total of m pure strategies for all players, and with integer payoffs of absolute value at most B , any /epsilon1 -PE of Γ has /lscript ∞ -distance at most δ to some PE of Γ .
Proof. The proof involves constructing formulas in the first order theory of real numbers, which formalize the 'almost implies near' statement of Lemma 3, with δ being 'hardwired' as a constant and /epsilon1 being the only free variable. Then, we apply quantifier elimination to these formulas. This leads to a quantifier free statement to which we can apply standard theorems bounding the size of an instantiation of the free variable /epsilon1 making the formula true. We shall apply and refer to theorems in the monograph of Basu, Pollack and Roy [2,3]. Note that we specifically refer to theorems and page numbers of the online edition [3]; these are in general different from the printed edition [2].
First-order formula for an /epsilon1 -perfect equilibrium: Define R x i ( \ k ) as the polynomial expressing U i ( π i,k ; x -i ), that is, the expected payoff to player i when it uses pure strategy k , and the other players play according to their mixed strategy in the profile x . Thus,
/negationslash
$$R _ { i } ( x \, \bar { \ } k ) \coloneqq \sum _ { a - i } u _ { i } ( k ; a _ { - i } ) \prod _ { j \neq i } x _ { j, a _ { j } }.$$
Let EPS-PE( x, /epsilon1 ) be the quantifier-free first-order formula, with free variables x ∈ R m and /epsilon1 ∈ R , defined by the conjunction of the following formulas that together express that x is an /epsilon1 -perfect equilibrium:
$$x _ { i, j } > 0 \quad & \text{for $i=1\dots,n$, and $j=1,\dots,m_{i}$} \, \\ x _ { i, 1 } + \dots + x _ { i, m _ { i } } = 1 \quad & \text{for $i=1\dots,n$} \, \\ ( R _ { i } ( x \, \| \kappa \| ) \| \geq R _ { i } ( x \, \| \, l ) ) \vee ( x _ { i, k } \leq \epsilon ) \quad & \text{for $i=1\dots,n$, and $k,l=1,\dots,m_{i}$} \.$$
First-order formula for perfect equilibrium: Let PE( x ) denote the following firstorder formula with free variables x ∈ R m , expressing that x is a perfect equilibrium:
$$\forall \epsilon > 0 \, \exists y \in \mathbb { R } ^ { m } \, \colon \text{EPS-PE} ( y, \epsilon ) \wedge \| x - y \| ^ { 2 } < \epsilon \.$$
First-order formula for 'almost implies near' statement: Given a fixed δ > 0 let PE-bound ( ) denote the following first-order formula with free variable δ /epsilon1 /epsilon1 ∈ R , denoting that any /epsilon1 -perfect equilibrium of G is δ -close to a perfect equilibrium (in /lscript 2 -distance, and therefore also in /lscript ∞ -distance):
$$\forall x \in \mathbb { R } ^ { m } \, \exists y \in \mathbb { R } ^ { m } \, \colon ( \epsilon > 0 ) \wedge \left ( \neg \text{EPS-PE} ( x, \epsilon ) \vee \left ( \text{PE} ( y ) \wedge \| x - y \| ^ { 2 } < \delta ^ { 2 } \right ) \right ) \.$$
Suppose δ 2 = 2 -k and the payoffs have absolute value at most B = 2 . Then τ for this formula we have
- -The total degree of all involved polynomials is at most max(2 , n -1).
- -The bitsize of coefficients is at most max( k, τ ).
- -The number of free variables is 1.
- -Converting to prenex normal form, the formula has 4 blocks of quantifiers, of sizes m m , , 1, m , respectively.
We now apply quantifier elimination [3, Algorithm 14.6, page 555] to the formula PE-bound ( ), converting it into an equivalent quantifier free formula δ /epsilon1 PE-bound ( ) with a single free variable ′ δ /epsilon1 /epsilon1 . This is simply a Boolean formula whose atoms are sign conditions on various polynomials in /epsilon1 . The bounds given by Basu, Pollack and Roy in association to Algorithm 14.6 imply that for this formula:
- -The degree of all involved polynomials (which are univariate polynomials in /epsilon1 ) is max(2 , n -1) O m ( 3 ) = n O m ( 3 ) .
- -The bitsize of all coefficients is at most max( k, τ ) max(2 , n -1) O m ( 3 ) = max( k, τ ) n O m ( 3 ) .
By Lemma 3, we know that there exists an /epsilon1 > 0 so that the formula PE-bound ( ) is true. We now apply (as the involved polynomials are univari-′ δ /epsilon1 ate, simpler tools would also suffice) Theorem 13.14 of Basu, Pollack and Roy [3, Page 521] to the set of polynomials that are atoms of PE-bound ( ) and ′ δ /epsilon1 conclude that PE-bound ( ′ δ /epsilon1 ∗ ) is true for some /epsilon1 ∗ ≥ 2 -max( k,τ ) n O m ( 3 ) . By the semantics of the formula PE-bound ( ), we also have that PE-bound ( δ /epsilon1 δ /epsilon1 ′ ) is true for all /epsilon1 ′ ≤ /epsilon1 ∗ , and the statement of the lemma follows. /intersectionsq /unionsq
## 5 Proof of the main theorem
We now prove Theorem 1. Let Γ be the n -player game given as input. Let m be the combined total number of pure strategies for all players. Let B ∈ N be the largest absolute value of any payoff of Γ . By the definition of FIXP a , our task is the following. Given a parameter δ > 0, we must construct a polytope P , a circuit C : P → P , and a number δ ′ , so that δ ′ -approximations to fixed points of C can be efficiently transformed into δ -approximations of PEs of Γ . In fact, we shall let δ ′ = δ/ 2 and ensure that δ ′ -approximations to fixed points of C are δ -approximations of PEs of Γ . The polytope P is simply the polytope D Γ of all strategy profiles of Γ ; clearly we can output the inequalities defining this polytope in polynomial time. The circuit C is the following: We construct the circuit for the function F /epsilon1 Γ of Section 3. Then, we construct a circuit for the number /epsilon1 ∗ = min( δ/ 2 , B -1 ) 2 /ceilingleft cm 3 lg n /ceilingright ≤ min( δ/ 2 , B -1 ) n cm 3 , where c is the constant of Lemma 4: The circuit simply repeatedly squares the number min( δ/ 2 , B -1 ) (which is a rational constant) and thereby consists of exactly /ceilingleft cm 3 lg n /ceilingright multiplication gates, i.e., a polynomially bounded number. We then plug in the circuit for /epsilon1 ∗ for the parameter /epsilon1 in the circuit for F /epsilon1 Γ , obtaining the circuit C , which is obviously a circuit for F /epsilon1 ∗ Γ . Now, by Theorem 2, any fixed point of C on P is an /epsilon1 ∗ -PE of Γ . Therefore, by Lemma 4, any fixed point of C is a δ/ 2-approximation in /lscript ∞ -distance to a PE of Γ . Finally, by the triangle inequality, any δ ′ = δ/ 2approximation to a fixed point of C on P is a δ/ 2 + δ/ 2 = δ approximation of a PE of Γ . This completes the proof.
## 6 Conclusion
We have showed that the problem of approximating a trembling hand perfect equilibrium for a finite strategic form game is in FIXP a . We do not know if exactly computing a trembling hand perfect equilibrium is in FIXP , and we consider this an interesting open problem, although it should be noted that if one is interested exclusively in the Turing Machine complexity of the problem, FIXP a membership of the approximation version is arguably 'the real thing'. We also note that this makes our proof interesting as a case where membership in FIXP a is not established as a simple corollary of the exact problem being in the 'abstract class' FIXP , as was the case for all examples in the original paper of Etessami and Yannakakis.
## Acknowledgements
Hansen and Miltersen acknowledge support from the Danish National Research Foundation and The National Science Foundation of China (under the grant 61061130540) for the Sino-Danish Center for the Theory of Interactive Computation, within which their work was performed. They also acknowledge support from the Center for Research in Foundations of Electronic Markets (CFEM), supported by the Danish Strategic Research Council.
## References
- 1. R. M. Anderson. 'almost' implies 'near'. Transactions of the American Mathematical Society , 296(1):229-237, 1986.
- 2. S. Basu, R. Pollack, and M. Roy. Algorithms in Real Algebraic Geometry . Springer, second edition, 2008.
- 3. S. Basu, R. Pollack, and M. Roy. Algorithms in Real Algebraic Geometry . http://perso.univ-rennes1.fr/marie-francoise.roy/bpr-ed2-posted2.html, online edition, 2011.
- 4. X. Chen and X. Deng. Settling the complexity of two-player Nash equilibrium. In Proceedings of 47th Annual IEEE Symposium on Foundations of Computer Science (FOCS'06) , pages 261-272, 2006.
- 5. C. Daskalakis, P. W. Goldberg, and C. H. Papadimitriou. The complexity of computing a Nash equilibrium. In Procedings of the 38th Annual ACM Symposium on the Theory of Computing (STOC'06) , pages 71-78, 2006.
- 6. K. Etessami and M. Yannakakis. On the complexity of Nash equilibria and other fixed points. SIAM J. Comput. , 39(6):2531-2597, 2010.
- 7. F. Gul, D. Pearce, and E. Stacchetti. A bound on the proportion of pure strategy equilibria in generic games. Math. of Oper. Res. , 18:548-552, 1993.
- 8. K. A. Hansen, P. B. Miltersen, and T. B. Sørensen. The computational complexity of trembling hand perfection and other equilibrium refinements. In Algorithmic Game Theory - Third International Symposium, SAGT 2010 , volume 6386 of Lecture Notes in Computer Science , pages 198-209. Springer, 2010.
- 9. D. E. Knuth. The Art of Computer Programming, Volume III: Sorting and Searching . Addison-Wesley, 1973.
- 10. C.E. Lemke. Bimatrix equilibrium points and mathematical programming. Management Science , 11:681-689, 1965.
- 11. P. B. Miltersen and T. B. Sørensen. Computing a quasi-perfect equilibrium of a two-player game. Economic Theory , 42(1):175-192, 2010.
- 12. J. Nash. Non-cooperative games. Annals of Mathematics , 54:289-295, 1951.
- 13. C. H. Papadimitriou. On graph-theoretic lemmata and complexity classes. In Proceedings of the 31st Annual Symposium on Foundations of Computer Science , pages 794-801, St. Louis, MS, October 1990. IEEE Computer Society Press.
- 14. H. E. Scarf. The approximation of fixed points of a continuous mapping. SIAM Journal of Applied Mathematics , 15:1328-1343, 1967.
- 15. R. Selten. A reexamination of the perfectness concept for equilibrium points in extensive games. International Journal of Game Theory , 4:25-55, 1975.
- 16. T. B. Sørensen. Computing a proper equilibrium of a bimatrix game. In ACM Conference on Electronic Commerce, EC '12 , pages 916-928, 2012.
- 17. E. van Damme. Stability and Perfection of Nash Equlibria . Springer-Verlag, 2nd edition, 1991.
- 18. B. von Stengel, A. van den Elzen, and D. Talman. Computing normal form perfect equilibria for extensive two-person games. Econometrica , 70(2):693-715, 2002.
| null |
[
"Kousha Etessami",
"Kristoffer Arnsfelt Hansen",
"Peter Bro Miltersen",
"Troels Bjerre Sorensen"
] |
2014-08-05T16:12:21+00:00
|
2014-08-05T16:12:21+00:00
|
[
"cs.GT",
"cs.CC"
] |
The complexity of approximating a trembling hand perfect equilibrium of a multi-player game in strategic form
|
We consider the task of computing an approximation of a trembling hand
perfect equilibrium for an n-player game in strategic form, n >= 3. We show
that this task is complete for the complexity class FIXP_a. In particular, the
task is polynomial time equivalent to the task of computing an approximation of
a Nash equilibrium in strategic form games with three (or more) players.
|
1408.1018v4
|
## THE NUMBER OF RAMIFIED PRIMES IN NUMBER FIELDS OF SMALL DEGREE
## ROBERT J. LEMKE OLIVER AND FRANK THORNE
Abstract. In this paper we investigate the distribution of the number of primes which ramify in number fields of degree d ≤ 5. In analogy with the classical Erd˝s-Kac theorem, o we prove for S d -extensions that the number of such primes is normally distributed with mean and variance log log X .
## 1. Introduction and statement of results
Consider the following problem: How many number fields are there of absolute discriminant less than X ? By work of Minkowski and Hermite, this number is known to be finite, and so it is natural question to ask for an asymptotic estimate.
Hermite's methods suggest that it is natural to count number fields of each fixed degree separately. To this end, write
$$\Sigma \quad & ( 1. 1 ) \quad \mathcal { F } _ { d } ( X ) \coloneqq \{ K / \mathbb { Q } \colon \deg K = d, | D _ { K } | \leq X \}, \quad N _ { d } ( X ) \coloneqq \# \mathcal { F } _ { d } ( X ), \\ \quad \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-}" \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-}\text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-} \text{-}$$
where D K is the discriminant of K . It has been proved for each d ≤ 5 that N X d ( ) ∼ c X d for an appropriate constant c d ; the results are classical for d = 2 and due to Davenport and Heilbronn [DH71] ( d = 3), Cohen, Diaz y Diaz, and Olivier [CDyDO02, Coh03] and Bhargava [Bha05] ( d = 4), and Bhargava [Bha10] ( d = 5). For higher d this has been conjectured by Malle [Mal04] (see also Bhargava [Bha07]), and an upper bound roughly of the form X d /epsilon1 was proved by Ellenberg and Venkatesh [EV06].
In the cases where we have an asymptotic formula, one may also ask about the rate of convergence for these formulas. Write F S d ( X ) and N S d ( X ) for the analogues of (1.1) where only those fields whose Galois closure is an S d -extension of Q are enumerated. Then, for d ≤ 5, power saving error terms have been proved for N S d ( X ), by Belabas, Bhargava, and Pomerance [BBP10] ( d = 3 4) and Shankar and Tsimerman [ST14] ( , d = 5).
It is also possible to count number fields satisfying a given finite set of local conditions, e.g., number fields in which a fixed prime p is ramified. The error terms will depend on the local conditions, but this dependence can be explicitly controlled, and it is again possible to give power saving error terms. Such results are essentially known in complete generality, although they are not all stated explicitly enough in the literature. 1 In Section 2 we give explicit results of this nature which suffice to prove our main theorem.
It is typical that good estimates for the distribution of an arithmetic sequence in arithmetic progressions open the door to further distribution results. For example:
The first author was supported by an NSF Mathematical Sciences Postdoctoral Research Fellowship at Stanford University.
The second author is partially supported by the National Science Foundation under Grant No. DMS1201330.
1 As we note later, such results did appear after we completed the initial version of our paper. The strongest results to date are due to Ellenberg, Pierce, and Wood [EPW16].
- · Yang [Yan09], and independently Cho and Kim [CK15], used results on cubic fields to obtain the distribution of the low-lying zeroes of the associated Dedekind zeta functions. Yang also did the same for quartic fields, and Cho and Kim did the same for some particular families of higher degree number fields (conditionally on several unproved hypotheses).
- · In a second paper [CK14], Cho and Kim prove a central limit theorem for the partial sums of coefficients of Artin L -functions associated to S d -number fields with d ≤ 5. Their preprint, which we learned of after this paper was nearly complete, gives independent proofs of the results of Section 2 (with different values of α d and β d ).
- · An old theorem of Erd˝s obtained the mean value of the least quadratic nonresidue o to a varying modulus, or equivalently the mean value of the least non-split prime in a quadratic field. Generalizing this, Martin and Pollack [MP13] obtained the mean value of the least prime that does not split completely in K , averaged over all cubic fields K .
- · Zhao [Zha13] counted cubic extensions of F q ( ) with a power saving error term, and t the second author and Xiong [TX14] applied this to prove 2 that the number of F q -rational points on random trigonal covers of P 1 ( F q ) converges to a Gaussian distribution, as does the number of zeros of the associated zeta function in prescribed arcs. (See also the bibliography of [TX14] for further related works.)
In this note we give another application and prove an Erd˝s-Kac theorem for number fields. o For an integer n , write ω n ( ) for the number of prime factors of n (counted without multiplicity). The classical Erd˝s-Kac theorem states that, on average over 1 o ≤ n ≤ X , the distribution of ω n ( ) approaches a Gaussian with mean and variance log log X .
To formulate a number field analogue, write ω K ( ) := ω D ( K ) for a number field K/ Q ; we will study the distribution of the values ω K ( ). Related questions have been studied previously; for example, Belabas and Fouvry [BF99] proved that there are infinitely many S 3 -cubic fields whose discriminant is fundamental and has at most 7 prime factors. 3 Although probabilistic considerations suggest that there should be infinitely many cubic, quartic, and quintic fields of prime discriminant, it seems that this question is beyond the reach of existing methods.
However, as we will see, an Erd˝ os-Kac theorem for number fields is straightforward to prove:
Theorem 1.1. Let d = 2 3 4 , , , or 5 . For any z ∈ R , we have that
$$& \text{I theorem } 1. 1. \ L e t \, d = 2, 3, 4, \, \text{or} \, \text{s. } \, \text{$t^{\prime}$ or any $z\in \mathbb{K}$,\, we have that} \\ & ( 1. 2 ) \quad \lim _ { X \to \infty } \frac { \# \{ K \in \mathcal { F } _ { S _ { d } } ( X ) \, \colon \omega ( K ) - \log \log X \leq z \sqrt { \log \log X } \} } { N _ { S _ { d } } ( X ) } = \frac { 1 } { \sqrt { 2 \pi } } \int _ { - \infty } ^ { z } e ^ { - t ^ { 2 } / 2 } \, d t. \\ & \text{That is as $K$ numbers over $S$ number fields of decrease $d$ ( K )$ theomes nammallu distributed}$$
That is, as K ranges over S d number fields of degree d , ω K ( ) becomes normally distributed.
Remark. 1. The d = 2 case requires less deep input than the other cases, and, in particular, is already in the literature. A proof of a slightly more general result can be found, for example, in [KLO16].
2 Analogues of the results in Section 2, i.e., an extension of Zhao's results to allow for local conditions, have yet to be worked out in full detail in the function field setting. This is not likely to be especially difficult, but for the moment the results of [TX14] should be considered incomplete.
3 Similar results for S 4 -quartic and S 5 -quintic fields, albeit with a larger bound on the number of prime factors, should follow from the results described in Section 2.
- 2. One might also ask for a quantitative version of Theorem 1.1, i.e., for a version with an explicit error term. We could presumably do this with more care; the results would depend on the error terms in Section 2 and would likely be far from optimal.
/negationslash
Our result is stated only for S d -extensions, and, for the most part, we leave open the analogous question for other Galois groups. However, it can be proved for d = 3 or 5 that such extensions may be included in (1.2) as follows. If G d = D 4 is a transitive permutation group on d letters, d ≤ 5, then it is known that N G d ( X ) /lessmuch X 1 -δ for some fixed δ > 0. Such results have been proved by a variety of authors, and all of these cases may be proved by using the Selberg sieve (exactly following [ST14]) in conjunction with an argument along the lines of Section 3.2 of [Bha10]. In the A 5 case, the (easy) details of this argument, along with a stronger error term, will appear in forthcoming work of Bhargava, Cojocaru, and the second author [BCT]. In particular, for d = 3 or 5, these G d -fields make a negligible contribution to (1.2), and it follows that one may count all degree d extensions in Theorem 1.1 instead of just S d -extensions.
In the D 4 case, it was proved by by Cohen, Diaz y Diaz, and Olivier [CDyDO02, Coh03] that N D 4 ( X ) = c D 4 X + O X ( 3 4+ / /epsilon1 ). The proof involves a careful study of associated arithmetic Selmer groups, and yields an explicit representation for the associated Dirichlet series ∑ K | Disc( K ) | -s . However, the density ρ D 4 ( q ) of field discriminants divisible by q is not obviously multiplicative in q , so that more care would be needed when applying the probabilistic model we use in Section 3. We leave the question of whether an Erd˝s-Kac theorem o should hold for D 4 -extensions for later work.
For S 3 -sextic extensions (i.e., extensions Galois over Q with Galois group S 3 ), our methods should yield a proof of the analogue of Theorem 1.1. The required analogue of Theorem 2.1 was essentially proved in [TT14], except that the dependence of the error terms on q was not made explicit there. It would not be difficult to do so, but the details are a bit messy and so we omit them here.
Belabas has implemented an algorithm to quickly enumerate cubic fields (see [Bel97], with accompanying software available from his website), which allows us to compare our results for d = 3 to numerical data. The graph on the left gives the count of ω K ( ) for all cubic fields K with | Disc( K ) | ≤ 10 ; for comparison, the right graph gives the distribution of 8 ω n ( ) over n ≤ 10 . 8
The graphs are extremely similar, although the cubic discriminants have slightly fewer prime divisors on average. This does not seem too surprising, as, for example, there are no field discriminants divisible by 2 but not 4.
This paper is organized as follows. In Section 2, we recall the necessary results from the literature on the distribution of number fields. In fact, while essentially all the results we need are known, not all have appeared in the literature. We therefore also sketch proofs of these results. In Section 3, we prove Theorem 1.1 by adapting a proof of the classical Erd˝ os-Kac theorem due to Billingsley.
## Acknowledgments
We would like to thank Etienne Fouvry, Ken Ono, Ari Shnidman, Arul Shankar, and Jacob Tsimerman for helpful feedback.
## 2. Counting fields with power-saving error terms
In this section we give the estimates for counting number fields needed in the proof of our main result.
Deviating slightly from the notation in the introduction, for each d ≤ 5, write N X d ( ) for the counting function of degree d number fields K with | D K | < X , and whose Galois closure has Galois group S d or A d over Q . Additionally, write N X,q d ( ) for the count of such number fields whose discriminant is divisible by q . The main technical result we need is the following.
## Theorem 2.1. Let
$$\begin{matrix} \frac { - \dots \dots \dots \dots \dots \dots \dots } { } & \alpha _ { 2 } = \frac { 1 } { 2 }, & \alpha _ { 3 } = \frac { 1 } { 6 }, & \alpha _ { 4 } = \frac { 1 } { 2 4 0 }, & \alpha _ { 5 } = \frac { 1 } { 2 0 0 }, \end{matrix}$$
$$\beta _ { 2 } = - \frac { 1 } { 2 }, \quad \beta _ { 3 } = \frac { 2 } { 3 }, \quad \beta _ { 4 } = \frac { 9 } { 1 0 }, \quad \beta _ { 5 } = 1. \\ \dots \quad. \quad. \quad. \quad. \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots$$
Then, there are constants c d and multiplicative functions ρ d ( q ) , the latter given by
$$1 \text{hen, there are constants } c _ { d } \text{ and multiple caltive functions } \rho _ { d } ( q ), \text{ the latter given } l \\ ( 2. 3 ) & \rho _ { d } ( p ) \coloneqq \begin{cases} \begin{array} { c c } 1 - \frac { 1 } { 1 + p ^ { - 1 } } & \text{if } d = 2, \\ 1 - \frac { 1 } { 1 + p ^ { - 1 } + p ^ { - 2 } } & \text{if } d = 3, \\ 1 - \frac { 1 } { 1 + p ^ { - 1 } + 2 p ^ { - 2 } + p ^ { - 3 } } & \text{if } d = 4, \text{ and } \\ 1 - \frac { 1 } { 1 + p ^ { - 1 } + 2 p ^ { - 2 } + 2 p ^ { - 3 } + p ^ { - 4 } } & \text{if } d = 5, \end{array} \\ \text{such that for all squarefree} \, q \text{ we have} \\ ( 2. 4 ) & N _ { d } ( X, q ) = c _ { d } \rho _ { d } ( q ) X + O ( X ^ { 1 + \epsilon - \alpha _ { d } } q ^ { \beta _ { d } } ), \end{cases}$$
$$N _ { d } ( X, q ) = c _ { d } \rho _ { d } ( q ) X + O ( X ^ { 1 + \epsilon - \alpha _ { d } } q ^ { \beta _ { d } } ),$$
the implied constant being absolute.
The precise values of α d and β d given are not optimal, and for our purposes would be relevant only if we wanted to establish a quantitative rate of convergence to the Gaussian. For improved constants, together with a very detailed exposition of how they are obtained, we refer ahead to forthcoming work of Ellenberg, Pierce, and Wood [EPW16]. (Similar results were also obtained independently by Cho and Kim [CK14], and previously in part by Yang [Yan09].) Here we offer a quick proof which does not aim at optimal values of α d and β d .
When d = 2, the result follows by an elementary inclusion-exclusion argument.
For d = 3, this is a variation of the Davenport-Heilbronn theorem, essentially first proved in work of Belabas, Bhargava, and Pomerance [BBP10], and with the constants above proved in [TT13]. The counting functions N X,q 3 ( ) also have a negative secondary term of order X 5 6 / ,
as proved (independently) in [BST13] and [TT13], and with this secondary term included above we could take α d = 2 9 / , β d = 8 9. /
For d = 5 (we handle d = 4 last), this result was obtained for q = 1 in an elegant short paper of Shankar and Tsimerman [ST14] by means of the Selberg sieve. To extend their result to general q it is only necessary to adjust their final computation on p. 7, in which they 'sieve to fields.' We very briefly recap the situation at this point in their work, but our exposition is not to be taken independent of theirs. Let V Z denote the space of quadruples of integral skew-symmetric 5 × 5 matrices, equipped with an action of G Z = GL ( 4 Z ) × SL ( 5 Z ), and recall that the G Z -orbits of V Z parametrize quintic rings, with quintic fields corresponding to orbits of irreducible rings that are everywhere maximal. Combining equations (2) and (8) of [ST14], the number of orbits of discriminant up to X that are i) irreducible, but outside the 'main ball,' ii) reducible, and inside the main ball, or iii) corresponding to an order in a nonS 5 quintic field, is
$$O ( X ^ { 1 9 9 / 2 0 0 + \epsilon } ).$$
Thus, to obtain Theorem 2.1, it remains to count those points in the main ball which are everywhere maximal and satisfy the specified local conditions at q .
Let V ( ) i Z ,q denote the subset of x ∈ V ( ) i Z corresponding to quintic rings which are maximal at q and for which q | Disc( x ), and write k ′ q for its density. Then, analogously to p. 7 of [ST14] and with U p , W d , and k d as they are there, we find that the desired number of points in the main ball is
$$\{ \Omega \, \i i + \jmath \, \text{and} \, \i i \, \upsilon _ { p }, \, \nu _ { d }, \, \text{and} \, \sigma _ { d } \, \text{as every are there, we must use the same number or points} \\ \text{in the main ball is} \\ N _ { 1 2 } ^ { * } ( \cap _ { p } U _ { p } \cap V _ { z, q } ^ { ( i ) }, X ) & = \sum _ { d \in \mathbb { N }, \, ( d, q ) = 1 } \mu ( d ) N _ { 1 2 } ^ { * } ( W _ { d } \cap V _ { z, q } ^ { ( i ) }, X ) \\ & = \sum _ { d < T, \, ( d, q ) = 1 } \left ( c _ { i } \mu ( d ) k _ { d } k _ { q } ^ { \prime } X + O ( X ^ { 3 9 / 4 0 } d ^ { \ } e ^ { q 1 + \ } ) \right ) + \sum _ { d > T } O _ { \epsilon } ( X / d ^ { 2 - \epsilon } ) \\ & = \sum _ { d \in \mathbb { N }, \, ( d, q ) = 1 } c _ { i } \mu ( d ) k _ { d } k _ { q } ^ { \prime } X + O _ { \epsilon } ( X / T ^ { 1 - \epsilon } + X ^ { 3 9 / 4 0 } T ^ { 1 + \epsilon } q ^ { 1 + \epsilon } ) \\ & = c _ { i } k _ { q } ^ { \prime } \prod _ { p ( q } ( 1 - k _ { p } ) X + O _ { \epsilon } ( X / T ^ { 1 - \epsilon } + X ^ { 3 9 / 4 0 } T ^ { 1 + \epsilon } q ^ { 1 + \epsilon } ). \\ \text{In the second line we have used that } W _ { d } \cap V _ { z, q } \text{ consists of the union of } O ( d ^ { 7 8 + \epsilon } q ^ { 7 9 + \epsilon } ) \text{ translates} \\ \text{of } ( d q ) ^ { 2 } V _ { z }. \text{ with } ( T q ) ^ { 2 } \, < \, X ^ { 1 / 4 0 } \text{ so that we may apply } ( 4 ) \text{ of } \text{[ST14]. and choosing } T \, =$$
In the second line we have used that W d ∩ V Z ,q consists of the union of O d ( 78+ /epsilon1 q 79+ /epsilon1 ) translates of ( dq ) 2 V Z , with ( Tq ) 2 < X 1 40 / so that we may apply (4) of [ST14], and choosing T = X 1 80 / q -1 we obtain an error term of X 79 80+ / /epsilon1 q 1+ /epsilon1 . Combining this in the most simple-minded way with the previous error O X ( 199 200+ / /epsilon1 ), we obtain Theorem 2.1.
Remark. This corrects a typo in [ST14], where N ( -) should be replaced with N ∗ 12 ( -) on p. 7, as we did above. (In other words, we count elements of V Z without regard to reducibility but with the requirement that a 12 = 0; then (2) and (8) of [ST14] establish the stated bounds.)
/negationslash
For d = 4, it was proved by Bhargava [Bha04] that G Z := GL ( 3 Z ) × GL ( 2 Z )-orbits on V Z = (Sym 2 Z 3 ⊗ Z 2 ) ∗ are in bijection with isomorphism classes of pairs ( Q,R ), where Q is a quartic ring and R is a cubic resolvent of Q . Bhargava used this parameterization [Bha05] to obtain the asymptotic density of quartic fields, and with Belabas and Pomerance [BBP10] refined this with a power-saving error term.
An analogue of (4) of [ST14] is given in Theorem 4.11 of [BBP10]. For any G Z -invariant subset S ⊆ V Z , define N ∗ 11 ( S, X ) to be the average number of elements v ∈ S in a fundamental
/negationslash domain for G Z \ V , with a 11 = 0 and discriminant less than X , as in Theorem 4.11 of [BBP10] or Sections 2 and 3 of [ST14]. Theorem 4.11 then yields that for any translate L of mV Z , for m ≤ X 1 12 / , we have
$$N _ { 1 1 } ^ { * } ( L \cap V _ { Z } ^ { ( i ) }, X ) & = c _ { i } m ^ { - 1 2 } X + O \left ( m ^ { - 1 2 } X ^ { 2 3 / 2 4 } + m ^ { - 7 } X ^ { 1 1 / 1 2 } \right ) \\ & = c _ { i } m ^ { - 1 2 } X + O \left ( m ^ { - 7 } X ^ { 2 3 / 2 4 } \right ).$$
Imitating the d = 5 argument given above (alternatively, we could follow [BBP10]) we find the desired number of points in the main ball is
$$( 2. 5 ) & N _ { 1 1 } ^ { * } ( L \cap V _ { \mathbb { Z } } ^ { ( i ) }, X ) = c _ { i } m ^ { - 1 2 } X + O ( m ^ { - 1 2 } X ^ { 2 3 / 2 4 } + m ^ { - 7 } X ^ { 1 1 / 1 } \\ ( 2. 6 ) & = c _ { i } m ^ { - 1 2 } X + O ( m ^ { - 7 } X ^ { 2 3 / 2 4 } ). \\ & \text{Imitating the } d = 5 \, \argument \, \text{given above (alternatively, we could } f c \\ \text{find the desired number of noints in the main ball is}$$
$$\text{find the desired number of points in the main ball is} \stackrel { \sim } { N } _ { 1 1 } ^ { * } ( \cap _ { p } U _ { p } \cap V _ { Z, q } ^ { ( i ) } X ) & = \sum _ { d \in \mathbb { N }, \ ( d, q ) = 1 } \mu ( d ) N ( W _ { d } \cap V _ { Z, q } ^ { ( i ) } X ) \\ & = \sum _ { d < T, \ ( d, q ) = 1 } \left ( c _ { i } \mu ( d ) k _ { d } k _ { q } ^ { \prime } X + O ( X ^ { 2 3 / 2 4 } d ^ { 8 + \epsilon } q ^ { 9 + \epsilon } ) \right ) + \sum _ { d > T } O _ { \epsilon } ( X / d ^ { 2 - \epsilon } ) \\ & = \sum _ { d \in \mathbb { N }, \ ( d, q ) = 1 } c _ { i } \mu ( d ) k _ { d } k _ { q } ^ { \prime } X + O _ { \epsilon } ( X / T ^ { 1 - \epsilon } + X ^ { 2 3 / 2 4 } T ^ { 9 + \epsilon } q ^ { 9 + \epsilon } ) \\ & = c _ { \epsilon } k _ { q } ^ { \prime } \prod _ { p / q } ( 1 - k _ { p } ) X + O _ { \epsilon } ( X / T ^ { 1 - \epsilon } + X ^ { 2 3 / 2 4 } T ^ { 9 + \epsilon } q ^ { 9 + \epsilon } ). \\ \text{Similarly to above, we have used that } W _ { d } \cap V _ { Z, q } \text{ consists of the union of } O ( d ^ { 2 + \epsilon } q ^ { 2 3 + \epsilon } ) \text{translates of } ( d \lambda ) ^ { 2 } V _ { \epsilon } \text{ the bound outer for } d \ > T \text{ follows from Lemma 4 3 of } [ \text{RPP10} ]$$
Similarly to above, we have used that W d ∩ V Z ,q consists of the union of O d ( 22+ /epsilon1 q 23+ /epsilon1 ) translates of ( dq ) 2 V Z ; the bound quoted for d > T follows from Lemma 4.3 of [BBP10]. We then choose T = X 1 240 / q -9 10 / , observe that the condition ( Tq ) 2 < X 1 12 / required for (2.5) is easily satisfied, and obtain an error term /lessmuch X 239 240 / q 9 10 / . The main term is as in [Bha05].
Finally, Lemmas 4.9 and 4.10 of [BBP10] establish that up to an error of O X ( 11 12+ / /epsilon1 ) we may exchange the 'main ball' condition on a 11 for 'total irreducibility', which, by Theorem 4.1 of [BBP10]), is equivalent to the property that the associated quartic ring is an S 4 -quartic order. This completes the proof.
Remark. The tail estimate used to estimate the sum over d > T (Lemma 4.3 of [BBP10]) is proved for S 4 -quartic orders rather than for elements of V Z , and is therefore subject to the O X ( 11 12+ / /epsilon1 ) error term from Lemmas 4.9 and 4.10 of [BBP10] described above. Since any given order or element of V Z is counted at most O X ( /epsilon1 ) times (at most once for each d > T dividing the discriminant), this error term may be absorbed into the stated error term.
## 3. The Erd˝ os-Kac machinery
In order to prove Theorem 1.1, we adapt a proof of the classical Erd˝s-Kac theorem due o to Billingsley [Bil74]. We proceed via the method of moments, which in this case is made easier as we are content with establishing a qualitative result; in particular, the moment calculation need not be uniform. With not too much more effort, our results could be made uniform, and therefore quantitative, but in order to obtain an optimal result, we would likely need to proceed in a different direction.
Fix d = 2 3 4 , , , or 5, and, for each squarefree q , define ρ d ( q ) as in Theorem 2.1. Define a random variable R d,p to be 1 with probability ρ d ( p ) and 0 with probability 1 -ρ d ( p ), so that R d,p models the event that p ramifies in a number field of degree d . As a consequence of the
central limit theorem (e.g., via Lyapunov's criterion), we know that the quantity
$$R _ { d } ( Z ) & \coloneqq \sum _ { p \leq Z } R _ { d, p } \\ \mu & \cdot \cdot \cdot \cdot \cdot \cdot$$
becomes normally distributed as Z →∞ and, provided Z = X δ for some δ > 0, with mean and variance each µ X ( ) := log log X . Thus, for fixed k , we have that
$$& ( 3. 1 ) & & \mathbb { E } \left ( ( R _ { d } ( Z ) - \mu ( X ) ) ^ { k } \right ) = c _ { k } \mu ( X ) ^ { k / 2 } + o _ { k } ( \mu ( X ) ^ { k / 2 } ), \\ & \text{where} & & \quad \, \quad \, \quad \, \quad \,.$$
$$c _ { k } = \begin{cases} \frac { k! } { 2 ^ { k / 2 } ( k / 2 )! }, & \text{if $k$ is even, and} \\ 0, & \text{if $k$ is odd} \end{cases} \\ \therefore \, \text{how to load $C$ emission}$$
is the k -th moment of the standard Gaussian.
Following Billingsley's proof and the notation of Section 2, the key idea is to compare the k -th moment
$$M _ { d, k } ( X ) & \coloneqq \frac { 1 } { N _ { d } ( X ) } \sum _ { K \in \mathcal { F } _ { d } ( X ) } \left ( \omega ( K ) - \mu ( X ) \right ) ^ { k } \\ \intertext { m o n i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t od i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o D i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o d i n t o c o n }$$
of the quantity we are interested in to that of the model; here, F d ( X ) := { K/ Q : [ K : Q ] = d, Gal( K/ Q ) = ∼ S d or A , d and | D K | ≤ X } , so that N d ( X ) = # F d ( X ). In particular, as the moments of the Gaussian determine the distribution (for example, the moment generating function converges on all of C , and the distribution function can be recovered simply by Fourier inversion), our main theorem follows if we prove the analogue of (3.1) for each of these moments.
We let Z = X α / k β d 2 ( d +1) , write ω K Z ( ; ) := # { p | D K : p ≤ Z } , and for each number field K of degree d and each prime p , define
$$\delta _ { p } ( K ) \coloneqq \begin{cases} \ 1, & \text{if $p|D_{K}$, and} \\ 0, & \text{otherwise.} \end{cases}$$
Finally, for any k ′ ≤ k , define
$$\tilde { M } _ { d, k ^ { \prime } } ( X, Z ) \coloneqq \frac { 1 } { N _ { d } ( X ) } \sum _ { K \in \mathcal { F } _ { d } ( X ) } \omega ( K ; Z ) ^ { k ^ { \prime } }.$$
We now observe that
$$\begin{array} { r c l } \text{We now observe that} \\ \tilde { M } _ { d, k ^ { \prime } } ( X, Z ) & = & \frac { 1 } { N _ { d } ( X ) } \sum _ { K \in \mathcal { F } _ { d } ( X ) } \left ( \sum _ { p \leq Z } \delta _ { p } ( K ) \right ) ^ { k ^ { \prime } } \\ & = & \sum _ { p _ { 1 }, \dots, p _ { k ^ { \prime } } < Z } \frac { 1 } { N _ { d } ( X ) } \sum _ { K \in \mathcal { F } _ { d } ( X ) \colon } 1 \\ & = & \sum _ { p _ { 1 }, \dots, p _ { k ^ { \prime } } < Z } \left ( \rho _ { d } ( \text{cm} ( p _ { 1 }, \dots, p _ { k ^ { \prime } } ) ) + O \left ( X ^ { - \alpha _ { d } } Z ^ { k ^ { \prime } \beta _ { d } } \right ) \right ) \\ & = & \mathbb { E } ( R _ { d } ( Z ) ^ { k ^ { \prime } } ) + O \left ( X ^ { - \alpha _ { d } / 2 } \right ), \end{array}$$
by Theorem 2.1, the construction of the random variables R d,p , and the choice of Z . Letting M d,k ( X,Z ) := E (( ω K Z ( ; ) -µ X ( )) k ), this shows that
In particular, by (3.1), we see that M d,k ( X,Z ) = c k µ X ( ) k/ 2 + o k ( µ X ( ) k/ 2 ), where we recall that c k is the k -th moment of the standard Gaussian.
$$\mathcal { I } _ { d, k } ( X, Z ) \coloneqq \mathbb { E } ( ( \omega ( K ; Z ) - \mu ( X ) ) ^ { k } ), \, \text{this shows that} \\ M _ { d, k } ( X, Z ) \ \coloneqq \ \frac { 1 } { N _ { d } ( X ) } \sum _ { K \in \mathcal { F } _ { d } ( X ) } ( \omega ( K ; Z ) - \mu ( X ) ) ^ { k } \\ \quad = \ \sum _ { j = 0 } ^ { k } \binom { k } { j } \, ( - \mu ( X ) ) ^ { j } \tilde { M } _ { d, k - j } ( X, Z ) \\ \quad = \ \sum _ { j = 0 } ^ { k } \binom { k } { j } \, ( - \mu ( X ) ) ^ { j } \left ( \mathbb { E } ( R _ { d } ( Z ) ^ { k - j } ) + O ( X ^ { - \alpha _ { d } / 2 } ) \right ) \\ \quad = \ \mathbb { E } ( ( R _ { d } ( Z ) - \mu ( X ) ) ^ { k } ) + O _ { k } ( X ^ { - \alpha _ { d } / 2 } \mu ( X ) ^ { k } ). \\ \intertext { \text{particular, by } ( 3. 1 ), \, \text{we see that } M _ { d, k } ( X, Z ) = c _ { k } \mu ( X ) ^ { k / 2 } + o _ { k } ( \mu ( X ) ^ { k / 2 } ), \, \text{where we rec} } \, \text{at} \, c _ { k } \, \text{ is the $k$-th moment of the standard Gaussian}.$$
Finally, it remains to compare the truncation M d,k ( X,Z ) to the full moment M d,k ( X ). For this, we see that
$$\begin{smallmatrix} 1 \\ - \end{smallmatrix}$$
( ) where in advancing to the final line we employ the Cauchy-Schwarz inequality for those j that are odd. This completes the proof.
$$\text{this, we see that} \\ M _ { d, k } ( X ) \ & = \ \frac { 1 } { N _ { d } ( X ) } \sum _ { K \in \mathcal { F } _ { d } ( X ) } \left ( \omega ( K ) - \mu ( X ) \right ) ^ { k } \\ & = \ \frac { 1 } { N _ { d } ( X ) } \sum _ { K \in \mathcal { F } _ { d } ( X ) } \left ( \omega ( K ; Z ) - \mu ( X ) + O ( k ) \right ) ^ { k } \\ & = \ M _ { d, k } ( X, Z ) + O _ { k } \left ( \sum _ { j = 0 } ^ { k - 1 } \frac { 1 } { N _ { d } ( X ) } \sum _ { K \in \mathcal { F } _ { d } ( X ) } \left | \omega ( K ; Z ) - \mu ( X ) \right | ^ { j } \right ) \\ & = \ M _ { d, k } ( X, Z ) + o _ { k } \left ( \mu ( X ) ^ { k / 2 } \right ), \\ \text{where in advancing to the final line we employ the Cauchy-Schwarz inequality for those $j$}. \. \. \. \. \. \.$$
## References
[BBP10] Karim Belabas, Manjul Bhargava, and Carl Pomerance. Error estimates for the Davenport- Heilbronn theorems. Duke Math. J. , 153(1):173-210, 2010.
[BCT]
M. Bhargava, A.C. Cojocaru, and F. Thorne. The square sieve and the number of A 5 -quintic extensions of bounded discriminant. Work in preparation .
[Bel97]
K. Belabas. A fast algorithm to compute cubic fields. Math. Comp. , 66(219):1213-1237, 1997.
[BF99]
K. Belabas and E. Fouvry. Sur le 3-rang des corps quadratiques de discriminant premier ou presque premier. Duke Math. J. , 98(2):217-268, 1999.
[Bha04]
Manjul Bhargava. Higher composition laws. III. The parametrization of quartic rings. Ann. of Math. (2) , 159(3):1329-1360, 2004.
[Bha05]
Manjul Bhargava. The density of discriminants of quartic rings and fields. Ann. of Math. (2) , 162(2):1031-1063, 2005.
[Bha07]
Manjul Bhargava. Mass formulae for extensions of local fields, and conjectures on the density of number field discriminants. Int. Math. Res. Not. IMRN , (17):Art. ID rnm052, 20, 2007.
[Bha10]
Manjul Bhargava. The density of discriminants of quintic rings and fields. Ann. of Math. (2) , 172(3):1559-1591, 2010.
[Bil74]
Patrick Billingsley. The probability theory of additive arithmetic functions. Ann. Probability , 2:749-791, 1974.
| [BST13] | Manjul Bhargava, Arul Shankar, and Jacob Tsimerman. On the Davenport-Heilbronn theorems and second order terms. Invent. Math. , 193(2):439-499, 2013. |
|-----------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [CDyDO02] | Henri Cohen, Francisco Diaz y Diaz, and Michel Olivier. Enumerating quartic dihedral exten- sions of Q . Compositio Math. , 133(1):65-93, 2002. |
| [CK14] | Peter J. Cho and Henry H. Kim. Central limit theorem for Artin L -functions. Preprint , 2014. |
| [CK15] | Peter J. Cho and Henry H. Kim. Low lying zeros of Artin L -functions. Math. Z. , 279(3-4):669- 688, 2015. |
| [Coh03] | Henri Cohen. Enumerating quartic dihedral extensions of Q with signatures. Ann. Inst. Fourier (Grenoble) , 53(2):339-377, 2003. |
| [DH71] | H. Davenport and H. Heilbronn. On the density of discriminants of cubic fields. II. Proc. Roy. Soc. London Ser. A , 322(1551):405-420, 1971. |
| [EPW16] | J. Ellenberg, L. B. Pierce, and M. Matchett Wood. On /lscript -torsion in class groups of number fields. ArXiv e-prints , June 2016. |
| [EV06] | Jordan S. Ellenberg and Akshay Venkatesh. The number of extensions of a number field with fixed degree and bounded discriminant. Ann. of Math. (2) , 163(2):723-741, 2006. |
| [KLO16] | Zev Klagsbrun and Robert J. Lemke Oliver. The distribution of 2-Selmer ranks of quadratic twists of elliptic curves with partial two-torsion. Mathematika , 62(1):67-78, 2016. |
| [Mal04] | Gunter Malle. On the distribution of Galois groups. II. Experiment. Math. , 13(2):129-135, 2004. |
| [MP13] | Greg Martin and Paul Pollack. The average least character non-residue and further variations on a theme of Erd˝s. o J. Lond. Math. Soc. (2) , 87(1):22-42, 2013. |
| [ST14] | Arul Shankar and Jacob Tsimerman. Counting S 5 fields with a power saving error term. Forum of Math., Sigma , 2014. |
| [TT13] | Takashi Taniguchi and Frank Thorne. Secondary terms in counting functions for cubic fields. Duke Math. J. , 162(13):2451-2508, 2013. |
| [TT14] | Takashi Taniguchi and Frank Thorne. An error estimate for counting S 3 -sextic number fields. Int. J. Number Theory , 10(4):935-948, 2014. |
| [TX14] | Frank Thorne and Maosheng Xiong. Distribution of zeta zeroes for trigonal curves over a finite field. Preprint , 2014. |
| [Yan09] | Andrew Yang. Distribution problems associated to zeta functions and invariant theory . 2009. Thesis (Ph.D.)-Princeton University. |
| [Zha13] | Yongqiang Zhao. On sieve methods for varieties over finite fields. 2013. Thesis (Ph.D.)- University of Wisconsin-Madison. |
Department of Mathematics, Tufts University, 503 Boston Ave, Medford, MA 02155 E-mail address : robert.lemke [email protected]
Department of Mathematics, University of South Carolina, 1523 Greene St, Columbia, SC 29201
E-mail address : [email protected]
| null |
[
"Robert J. Lemke Oliver",
"Frank Thorne"
] |
2014-08-05T16:13:36+00:00
|
2016-09-02T21:07:21+00:00
|
[
"math.NT"
] |
The number of ramified primes in number fields of small degree
|
In this paper we investigate the distribution of the number of primes which
ramify in number fields of degree d <= 5. In analogy with the classical
Erdos-Kac theorem, we prove for S_d-extensions that the number of such primes
is normally distributed with mean and variance log log X.
|
1408.1021v1
|
## The Adaptive Priority Queue with Elimination and Combining
Irina Calciu, Hammurabi Mendes, and Maurice Herlihy Department of Computer Science Brown University 115 Waterman St., 4th floor Providence RI - USA { irina,hmendes,mph @cs.brown.edu }
August 6, 2014
## Abstract
Priority queues are fundamental abstract data structures, often used to manage limited resources in parallel programming. Several proposed parallel priority queue implementations are based on skiplists, harnessing the potential for parallelism of the add() operations. In addition, methods such as Flat Combining have been proposed to reduce contention by batching together multiple operations to be executed by a single thread. While this technique can decrease lock-switching overhead and the number of pointer changes required by the removeMin() operations in the priority queue, it can also create a sequential bottleneck and limit parallelism, especially for non-conflicting add() operations.
In this paper, we describe a novel priority queue design, harnessing the scalability of parallel insertions in conjunction with the efficiency of batched removals. Moreover, we present a new elimination algorithm suitable for a priority queue, which further increases concurrency on balanced workloads with similar numbers of add() and removeMin() operations. We implement and evaluate our design using a variety of techniques including locking, atomic operations, hardware transactional memory, as well as employing adaptive heuristics given the workload.
## 1 Introduction
A priority queue is a fundamental abstract data structure that stores a set of keys (or a set of key-value pairs), where keys represent priorities. It usually exports two main operations: add() , to insert a new item in the priority queue, and removeMin() , to remove the first item (the one with the highest priority). Parallel priority queues are often used in discrete event simulations and resource management, such as operating systems schedulers. Therefore, it is important to carefully design these data structures in order to limit contention and improve scalability. Prior work in concurrent priority queues exploited parallelism by using either a heap [6] or a skiplist [8] as the underlying data structures. In the skiplist-based implementation of Lotan and Shavit [8] each node has a 'deleted' flag, and processors contend to mark such 'deleted' flags concurrently, in the beginning of the list. When a thread logically deletes a node, it tries to remove it from the skiplist using the standard removal algorithm. A lock-free skiplist implementation is presented in [11].
However, these methods may incur limited scalability at high thread counts due to contention on shared memory accesses. Hendler et al. [3] introduced Flat Combining, a method for batching together multiple operations to be performed by only one thread, thus reducing the contention on the data structure. This idea has also been explored in subsequent work on delegation [9, 1], where a dedicated thread called a server performs work on behalf of other threads, called clients . Unfortunately, the server thread could become a sequential bottleneck. A method of combining delegation with elimination has been proposed to alleviate this problem for a stack data structure [2]. Elimination [4] is a method of matching concurrent inverse operations so that they don't access the shared data structure, thus significantly reducing contention and
increasing parallelism for otherwise sequential structures, such as stacks. An elimination algorithm has also been proposed in the context of a queue [10], where the authors introduce the notion of aging operations -operations that wait until they become suitable for elimination.
In this paper, we describe, to the best of our knowledge, the first elimination algorithm for a priority queue. Only add() operations with values smaller than the priority queue minimum value are allowed to eliminate. However, we use the idea of aging operations introduced in the queue algorithm to allow add() values that are small enough to participate in the elimination protocol, in the hope that they will soon become eligible for elimination. We implement the priority queue using a skiplist and we exploit the skiplist's capability for both operations-batching and disjoint-access parallelism. RemoveMin() requests can be batched and executed by a server thread using the combining/delegation paradigm. Add() requests with high keys will most likely not become eligible for elimination and need to be inserted in the skiplist, sometimes requiring expensive traversals towards the end of the data structure to do so. Therefore, these operations represent a bottleneck for the server and a missed opportunity for parallelism. To alleviate these issues, we split the underlying skiplist into two parts: a sequential part, managed by the server thread and a parallel part, where high-valued add() operations can insert their arguments in parallel. Our design reduces contention by performing batched sequential removeMin() and small-value add() operations, while also leveraging parallelism opportunities through elimination and parallel high-value add() operations. We show that our priority queue outperforms prior algorithms in high contention workloads on a SPARC Niagara II machine. Finally, we explore whether the use of hardware transactions could simplify our design and improve throughput. Unfortunately, machines that support hardware transactional memory (HTM) are only available for up to four cores (eight hardware threads), which is not enough to measure scalability of our design in high contention scenarios. Nevertheless, we showed that a transactional version of our algorithm is better than a non-transactional version on a Haswell four-core machine. We believe that these preliminary results will generalize to machines with more threads with support for HTM, once they become available. In summary, our main contributions are:
- · We propose the first elimination algorithm for a priority queue, consisting of (1) immediate elimination , where suitable add() and removeMin() operations exchange arguments; and (2) upcoming elimination , where add() operations with small keys, yet not suitable for elimination, wait some time until either they become suitable or time out.
- · We describe a scalable design for a priority queue based on our elimination algorithm and the delegation/combining paradigm introduced by prior work.
- · We augment our priority queue design with an adaptive component that allows it to perform combining and elimination efficiently, while also allowing add() operations not involved in the elimination to insert in parallel.
- · We analyze how hardware transactions could be used to simplify and improve our initial design and show performance results on a Haswell machine with transactional memory enabled.
## 2 Design
Our priority queue exports two operations: add() and removeMin() and is implemented using an underlying skiplist. The elements of the skiplist are buckets associated with keys. For a bucket b , the field b .key denotes the associated key. We split the skiplist in two distinct parts. The sequential part, in the beginning of the skiplist, is likely to serve forthcoming removeMin() operations of the priority queue ( PQ::removeMin() for short) as well as add( ) v operations of the priority queue ( PQ:add() for short) with v small enough (hence expected to be removed soon). The parallel part , which complements the sequential part, is likely to serve PQ::add( ) v operations where v is large enough (hence not expected to be removed soon). Either the sequential or the parallel part may become empty. Both lists are complete skiplists, with (dummy) head buckets called headSeq and headPar , respectively, with key -∞ . Both lists also contain (dummy) tail
buckets, with key + ∞ . We call the last non-dummy bucket of the sequential part lastSeq , which is the logical divisor between parts. Figure 1 shows the design.
Add (small keys)/Remove:
Add (bigger keys):
parallel server thread
Figure 1: Skiplist design. An elimination array is used for removeMin() s and add() s with small keys. A dedicated server thread collects the operations that do not eliminate and executes them on the sequential part of the skiplist. Concurrent threads operate on the parallel part, performing add() s with bigger keys. The dotted lines show pointers that would be established if the single skiplist was not divided in two parts.
When a thread performs a PQ::add( ) v , either (1) v > lastSeq key . , and the thread inserts the value concurrently in the parallel part of the skiplist, calling the SL::addPar() skiplist operation; or (2) v ≤ lastSeq.key , and the thread tries to perform elimination with a PQ::removeMin() using the elimination array. A PQ::add( ) v with v less than the smallest value in the priority queue can immediately eliminate with a PQ::removeMin() , if one is available. A PQ::add( ) v operation with v bigger than minValue (the current minimal key) but smaller than lastSeq key . lingers in an elimination array for some time, waiting to become eligible for elimination or timeout. A server thread executes sequentially all operations that fail to eliminate.
This mechanism describes the first elimination algorithm for a priority queue, well integrated with delegation/combining, presented in more detail in Section 2.2. Specifically: (1) The scheme harnesses the parallelism of the priority queue add() operations, letting the ones with keys physically distant and large enough (bigger than lastSeq.key ) execute in parallel. (2) At the same time, we batch concurrent priority queue add() with small keys and removeMin() operations that timed out in the elimination array, serving such requests quickly through the server thread - this latter operation simply consumes elements from the sequential part by navigating through elements in its bottom level, merely decreasing counters and moving pointers in the most common situation. While detaching a sequential part is non-negligible cost-wise, a sequential part has the potential to serve multiple removals.
## 2.1 Concurrent Skiplist
Our underlying skiplist is operated by the server thread in the sequential part and by concurrently inserting threads with bigger keys in the parallel part.
Sequential part. The server calls the skiplist function SL::moveHead() to extract a new sequential part from the parallel part if some PQ::removeMin() operation was requested and the sequential part was empty. Conversely, it calls the skiplist function SL::chopHead() to relink the sequential and the parallel parts, forming a completely parallel skiplist, if no PQ::removeMin() operations are being requested for some time. In SL::moveHead() , we initially determine the elements to be moved to the sequential part. If no elements are found, the server clears the sequential part, otherwise separating the sequential part from the rest of the list, which becomes the parallel part. The number of elements that SL::moveHead() tries to detach to the sequential part adaptively varies between 8 and 65,536. Our policy is simple: if more than
N insertions (e.g. N = 1000) occurred in the sequential part since the last SL::moveHead() , we halve the number of elements moved; otherwise, if less than M insertions (e.g. M = 100) were made, we double this number. After SL::moveHead() executes, a pointer called currSeq indicates the first bucket in the sequential part, and another called lastSeq indicates the final bucket. The server uses SL::addSeq() and SL::removeSeq() within the sequential part to remove elements or insert elements with small keys (i.e., belonging to the sequential part) that failed to eliminate. Buckets are not deleted at this time; they are deleted lazily when the whole sequential part gets consumed. A new sequential part can be created by calling SL::moveHead() again.
Parallel part. The skiplist function SL::addPar() inserts elements into the parallel part, and is called by concurrent threads performing PQ::add() . While these insertions are concurrent, the skiplist still relies on a Single-Writer Multi-Readers lock with writer preference for the following purpose. Multiple SL::addPar() operations acquire the lock for reading (executing concurrently), while SL::moveHead() and SL::chopHead() operations acquire the lock for writing. This way, we avoid that SL::addPar() operates on buckets that are currently being moved to the sequential part by SL::moveHead() , or interferes with SL::chopHead() . Despite the lock, SL::addPar() is not mutually exclusive with the head-moving operations ( SL::moveHead() and SL::chopHead() ). Only the pointer updates (for new buckets) or the counter increment (for existing buckets) must be done in the parallel part (and not have been moved to the sequential part) after we determine the locations of these changes. Hence, in the SL::addPar() operation, we first try to get a clean SL::find() : a SL::find() followed by acquiring the lock for reading, with no intervening head-moving operations. We can tell whether no head-moving operation took place since our lock operations always increases a timestamp variable, checked in the critical section. After a clean SL::find() , now holding the lock, if a bucket corresponding to the key is found, we insert the element in the bucket (incrementing a counter). Otherwise, a new bucket is created, and inserted level by level using CAS() operations. If a CAS() fails in a certain level, we release the lock and retry a clean SL::find() .
Our algorithm differs from the traditional concurrent skiplist insertion algorithms in two ways: (1) we hold a lock to avoid head-moving operations to take place after a clean SL::find() ; and (2) if the new bucket is moved out of the parallel section while we insert the element in the upper levels, we stop SL::addPar() , leaving this element with a capped level. This bucket is likely to be soon consumed by a SL::removeSeq() operation, resulting from a PQ::removeMin() operation.
## 2.2 Elimination and Combining
Elimination allows matching operations to complete without accessing the shared data structure, thus increasing parallelism and scalability. In a priority queue, any SL::removeMin() operation can be eliminated, but only SL::add() operations with values smaller or equal to the current minimum value can be so. If the priority queue is empty, any SL::add() value can be eliminated. We used an elimination array similar to the one in the stack elimination algorithm [4]. Each slot uses 64 bits to pack together a 32-bit value that represents either an opcode or a value to be inserted in the priority queue and a stamp that is unique for each operation. The opcodes are: EMPTY, REMREQ, TAKEN and INPROG. These are special values that cannot be used in the priority queue. All other values are admissible. In our implementation, each thread has a local count of how many operations it performed. This count is combined with the thread ID to obtain a unique stamp for each operation. Overflow was not an issue in our experiments, but if it becomes a problem a different algorithm for associating unique stamps to each operation could be used. The unique stamp is used to ensure linearizability, as explained in Section 3. All slots are initially empty, marked with the special value EMPTY, and the stamp value is zero.
A PQ::removeMin() thread loops through the elimination array until it finds a request to eliminate with or it finds an empty slot in the array, as described in Algorithm 1. If it finds a value in the slot, then it must ensure that the stamp is positive, otherwise the value was posted as a response to another thread. The value it finds must be smaller than the current priority queue minimum value. Then, the PQ::removeMin() thread can CAS the slot, which contains both the value and the stamp, and replace it with an indicator that the value was taken (TAKEN, with stamp zero). The thread returns the value found. If instead, the PQ::remove() thread finds an empty slot, it posts a remove request (REMREQ), with a unique stamp
generated as above. The thread waits until the slot is changed by another thread, having a value with stamp zero. The PQ::removeMin() thread can then return that value.
glyph[negationslash]
| Algorithm 1 PQ::removeMin() | Algorithm 1 PQ::removeMin() |
|-------------------------------|-------------------------------------------------------------------------|
| 1: while true do | 1: while true do |
| 2: | pos ← ( id +1)% ELIM SIZE; (value, stamp) ← elim[pos] |
| 3: | if IsValue(value) and (stamp > 0) and (value ≤ skiplist.minValue)) then |
| 4: | if CAS(elim[pos], (value, stamp), (TAKEN, 0)) then |
| 5: | return value |
| 6: | if value = EMPTY then |
| 7: | if CAS(elim[pos], (value, stamp), (REMREQ, uniqueStamp())) then |
| 8: | repeat |
| 9: | (value, stamp) ← elim[pos] |
| 10: | until value = REMREQ and value = INPROG |
| 11: | elim[pos] ← (EMPTY, 0); return value |
| 12: | inc(pos) |
glyph[negationslash]
A PQ::add() thread initially tries to use SL::addPar() to add its value in parallel. A failed attempt to add in parallel indicates that the value should try to eliminate or should be inserted in the sequential part. The PQ::add() thread tries to eliminate by checking through the elimination array for REMREQ indicators. If it finds a remove request, and its value is smaller than the priority queue minValue , it can CAS its value with stamp zero, effectively handing it to another thread. If multiple such attempts fail, the thread changes its behavior: it still tries to perform elimination as above, but as soon as an empty slot is found, it uses a CAS to insert its own value and the current stamp in the slot, waiting for another thread to match the operation (and change the opcode to TAKEN) returning the corresponding value.
The PQ::add() and PQ::removeMin() threads that post a request in an empty slot of the elimination array wait for a matching thread to perform elimination. However, elimination could fail because no matching thread shows up or because the PQ::add() value is never smaller than the priority queue minValue . To ensure that all threads make progress, we use a dedicated server thread that collects add and remove requests that fail to eliminate. The server thread executes the operations sequentially on the skiplist, calling SL::addSeq() and SL::removeSeq() operations. To ensure linearizability, the server marks a slot that contains an operation it is about to execute as in progress (INPROG). Subsequently, it executes the sequential skiplist operation and writes back the response in the elimination slot for the other thread to find it. A state machine showing the possible transitions of a slot in the elimination array is shown in Figure 2, and the algorithm is described in Algorithm 2.
| Algorithm 2 Server::execute() | Algorithm 2 Server::execute() |
|---------------------------------|---------------------------------------------------|
| 1: | while true do |
| 2: | for i : 1 → ELIM SIZE do |
| 3: | (value, stamp) ← elim[i] |
| 4: | if value = REMREQ then |
| 5: | if CAS(elim[i], (value, stamp), (INPROG, 0)) then |
| 6: | min ← skiplist.removeSeq(); elim[i] ← (min, 0) |
| 7: | if IsValue(value) and (stamp > 0) then |
| 8: | if CAS(elim[i], (value, stamp), (INPROG, 0)) then |
| 9: | skiplist.addSeq(value); elim[i] ← (TAKEN, 0) |
Figure 2: Transitions of a slot in the elimination array.
## 3 Linearizability
Our design provides a linearizable priority queue algorithm. Some operations have multiple possible linearization points by design, requiring careful analysis and implementation.
Skiplist. A successful SL::addPar( ) v (respectively, SL::addSeq( ) v ) usually linearizes when it inserts the element in the bottom level of the skip list with a CAS (respectively, with a store), or when the bucket for key v has its counter incremented with a CAS (respectively, with a store). However, a thread inserting a minimal bucket, whenever v < minValue , is required to update minValue . When the sequential part is not empty, only the server can update minValue (without synchronization). When the sequential part is empty, a parallel add with minimal value needs to update minValue . The adding thread loops until a CAS decreasing minValue succeeds or another thread inserts a bucket with key smaller than v . Note that no head-moving operation is taking place (the SL::addPar() threads hold the lock ). Threads that succeed changing minValue linearize their operation at the point of the successful CAS .
The head-moving operations SL::moveHead() and SL::chopHead() execute while holding the lock for writing, which effectively linearizes the operation at the lock.release() instant because: (1) no SL::addPar() is running; (2) no SL::addSeq() or SL::removeSeq() are running, as the server thread is the single thread performing those operations. Head-moving operations do not change minValue , in fact they preclude any changes to it. During these operations, however, threads may still perform elimination, which we discuss next.
Elimination. A unique stamp is used in each request posted in the array entries to avoid the 'ABA' problem. Each elimination slot is a 64-bit value that contains 32 bits for the posted value (for PQ::add() ) or a special opcode (for PQ::removeMin() ) and 32 bits for the unique stamp. In our implementation, the unique stamp is obtained by combining the thread id with the number of operations performed by each thread. Each thread, either adding or removing, that finds the inverse operation in the elimination array must verify that the exchanged value is smaller than minValue . If so, the thread can CAS the elimination slot, exchanging arguments with the waiting thread. It is possible that the priority queue minimum value is changed by a concurrent PQ::add() . In that case, the linearization point for both threads engaged in elimination is at the point where the value was observed to be smaller than the priority queue minimum. See Fig. 3.
The thread performing the CAS first reads the stamp of the thread that posted the request in the array and verifies that it is allowed to eliminate. Only then it performs a CAS on both the value and the stamp, guaranteeing that the thread waiting did not change in the meantime. Because both threads were running at the time of the verification, they can be linearized at that point. Without the unique stamp, the eliminating
op to slot
op on skiplist result
Figure 3: Concurrent execution of an op thread posting its request to an empty slot, and an inv thread, executing a matching operation. The operation by the inv thread could begin anytime before the Read and finish any time after the CAS . The linearization point is marked with a red X.
thread could perform a CAS on an identical request (i.e., identical operation and value) posted in the array by a different thread. The CAS would incorrectly succeed, but the operations would not be linearizable because the new thread was not executing while the suitable minimum was observed.
The linearizability of the combining operation results from the linearizability of the skiplist. The threads post their operation in the elimination array and wait for the server to process it. The server first marks the operation as in progress by CAS ing INPROG into the slot. Then it performs the sequential operation on the skiplist and writes the results back in the slot, releasing the waiting thread. The waiting thread observes the new value and returns it. The linearization point of the operation happens during the sequential operation on the skiplist, as discussed above. See Fig. 4.
Figure 4: Concurrent execution of a client thread and the server thread. The client posts its operation op to an empty slot and waits for the server to collect the operation and execute it sequentially on the skiplist. The linearization point occurs in the sequential operation and is marked with a red X.
Start
Post
(op, stamp)
Observe
Return
Time
Add/Remove
(op)
Thread to slot
result
## Remove/Add 4 Evaluation
(inv)
Thread
x
Read (op, stamp) Check val <= minValue CAS (op, stamp) In this section, we discuss results on a Sun SPARC T5240, which contains two UltraSPARC T2 Plus chips with 8 cores each, running at 1.165 GHz. Each core has 8 hardware strands, for a total of 64 hardware threads per chip. A core has a 8KB L1 data cache and shares an 4MB L2 data cache with the other cores on a chip. Each experiment was performed five times and we report the median. Variance was very low for all experiments. Each test was run for ten seconds to measure throughput. We used the same benchmark as flat combining [3]. A thread randomly flips a coin with probability p to be an PQ::add() and 1 -p to be a PQ::removeMin() . We started a run after inserting 2000 elements in the priority queue for stable state results.
x
Our priority queue algorithm ( pqe ) uses combining and elimination, and leverages the parallelism of PQ::add() . We performed experiments to compare against previous priority queues using combining methods, such as flat combining skiplist ( fcskiplist ) and flat combining pairing heap ( fcpairheap ). We also compared against previous priority queues using skiplists with parallel operations, such as a lock free skiplist ( lfskiplist ) and a lazy skiplist ( lazyskiplist ). The flat combining methods are very fast at performing PQ::removeMin() operations, which then get combined and executed together. However, performing the PQ::add() operations sequentially is a bottleneck for these methods. Conversely, the lfskiplist and
lazyskiplist algorithms are very fast at performing the parallel adds, but get significantly slowed down by having PQ::removeMin() operations in the mix, due to the synchronization overhead involved. Our pqe design tries to address these limitations through our dual (sequential and parallel parts), adaptive implementation that can be beneficial in the different scenarios.
Figure 5: Priority queue performance with add() s, 50% removeMin() s.
Figure 7: add() work breakdown.
50% Figure 6: Priority queue performance with 80% add() s, 20% removeMin() s.
Figure 8: removeMin() work breakdown.
We considered different percentages of PQ::add() and PQ::removeMin() in our tests. When the operations are roughly the same number, pqe can fully take advantage of both elimination and parallel adds, so it has peak performance. Figure 5 shows how for 50% PQ::add() and 50% PQ::removeMin() , pqe is much more scalable and can be up to 2.3 times faster than all other methods. When there are more PQ::add() than PQ::removeMin() , as in Figure 6 with 80% PQ::add() and 20% PQ::removeMin() , pqe behavior approaches the other methods, but it is still 70% faster than all other methods at high thread counts. In this specific case there is only little potential for elimination, but having parallel insertion operations makes our algorithm outperform the flat combining methods. The lazyskiplist algorithm also performs better than other methods, as it also takes advantage of parallel insertions. However, pqe uses the limited elimination and the combining methods to reduce contention, making it faster than the lazyskiplist . For more PQ::removeMin() operations than PQ::add() operations, the pqe 's potential for elimination and parallel adds are both limited, thus other methods can be faster. Pqe is designed for high contention scenarios, in which elimination and combining thrive. Therefore, it can incur a penalty at lower thread counts, where there is not enough contention to justify the overhead of the indirection caused by the elimination array and the server thread.
To better understand when each of the optimizations used is more beneficial, we analyzed the breakdown of the PQ::add() and PQ::removeMin() operations for different PQ::add() percentages. When we have 80% PQ::add() , most of them are likely to be inserted in parallel (75%), with a smaller percentage being able to eliminate and an even smaller percentage being executed by the server, as shown in Fig. 7. In the
same scenario, 75% of removeMin() operations eliminate, while the rest gets execute by the server, as seen in Fig. 8. For balanced workloads (50% -50%), most operations eliminate and a few PQ::add() operations are inserted in parallel. When the workload is dominated by PQ::removeMin() , most PQ::add() eliminate, but most PQ::removeMin() are still left to be executed by the server thread, thus introducing a sequential bottleneck. Eventually the priority queue would become empty, not being able to satisfy PQ::removeMin() requests with an actual value anymore. In this case, any add() operation can eliminate, allowing full parallelism. We do not present results for this case because it is an unlikely scenario that unrealistically favors elimination.
## 4.1 Evaluating the Overhead of PQ::moveHead() and PQ::chopHead()
Maintaining separate skiplists for the sequential and the parallel part of the priority queue is beneficial for the overall throughput, but adds some overhead, which we quantify in this section. The number of elements that become part of the sequential skiplist changes dynamically based on the observed mix of operations. This adaptive behavior helps reduce the number of moveHead() and chopHead() operations required. Table 1 shows the percentage of the number of head-moving operations out of the total number of PQ::removeMin() operations for different mixes of PQ::add() and PQ::removeMin() operations. The head-moving operations are rarely called due to the priority queue's adaptive behavior.
Table 1: The number of head-moving operations as a percentage of the total number of PQ::removeMin() operations, considering different add() and removeMin() mixes.
| Add() percentages | % moveHead() | % chopHead() |
|---------------------|----------------|----------------|
| 80 | 0.24% | 0.03% |
| 50 | 0.32% | 0.01% |
| 20 | 0.00% | 0.00% |
## 5 Hardware Transactions
Transactional memory [5] is an optimistic mechanism to synchronize threads accessing shared data. Threads are allowed to execute critical sections speculatively in parallel, but, if there is a data conflict, one of them has to roll back and retry its critical section. Recently, IBM and Intel added HTM instructions to their processors [12, 7]. In our priority queue implementation, we used Intel's Transactional Synchronization Extensions (TSX) [7] to simplify the implementation and reduce the overhead caused by the synchronization necessary to manage a sequential and a parallel skiplist. We evaluate our results on an Intel Haswell four core processor, Core i7-4770, with hardware transactions enabled (restricted transactional memory - RTM), running at 3.4GHz. There are 8GB of RAM shared across the machine and each core has a 32KB L1 cache. Hyperthreading was enabled on our machine so we collected results using all 8 hardware threads. Hyperthreading causes resource sharing between the hyperthreads, including L1 cache sharing, when running with more than 4 threads, thus it can negatively impact results, especially for hardware transactions. We did not notice a hyperthreading effect in our experiments. We used the GCC 4.8 compiler with support for RTM and optimizations enabled (-O3).
## 5.1 Skiplist
The Single-Writer-Multi-Readers lock used to synchronize the sequential and the parallel skiplists complicates the priority queue design and adds overhead. In this section, we explore an alternative design using
hardware transactions. However, the naive approach of making all operations transactional causes too many aborts. Instead, the server increments a timestamp whenever a head-moving operation SL::moveHead() or SL::chopHead() -starts or finishes. A SL::addPar() operation first reads the timestamp and executes a nontransactional SL::find() and then starts a transaction for the actual insertion, adding the server's timestamp to its read set and aborting if it is different from the initially recorded value. Moreover, if the timestamp changes after starting the transaction, indicating a head-moving operation, the transaction will be aborted due to the timestamp conflict. If the timestamp is valid, SL::find() must have recorded the predecessors and successors of the new bucket at each level i in preds[i] and succs[i] , respectively. If a bucket already exists, the counter is incremented inside the transaction and the operation completes. If the bucket doesn't exist, the operation proceeds to check if preds[i] points to succs[i] for all levels 0 ≤ i ≤ MaxLvl. If so, the pointers have not changed before starting the transaction and the new bucket can be correctly inserted between preds[i] and succs[i] . Otherwise, we commit the (innocuous) transaction, yet restart the operation.
Figures 9 and 10 compare the performance of the lock-based implementation and the implementation based on hardware transactions for two different percentages of PQ::add() s and PQ::removeMin() s. When fewer PQ::removeMin() operations are present, the timestamp changes less frequently and the PQ::add() transactions are aborted fewer times, which increases performance in the 80%-20% insertion-removal mix. In the 50%-50% mix, we obtain results comparable to the pqe algorithm using the lock-based approach, albeit with a much simpler implementation.
Figure 9: Priority queue performance when we use Figure 10: a transaction-based dual skiplist; 80% add() s, 20% a transaction-based dual skiplist;
removeMin() s.
Priority queue performance when we use 50% add() s, 50% removeMin() s.
## 5.2 Evaluating the Overhead of Aborted Transactions
The impact of aborted transactions is reported in Tables 2 and 3. As the number of threads increases, the number of transactions per successful operation also increases, as does the percentage of operations that need more than 10 retries to succeed. Note that the innocuous transactions that find inconsistent pointers, changed between the SL::find() and the start of the transaction are not included in the measurement. After 10 retries, threads give up on retrying the transactional path and the server executes the operations on their behalf, either in the sequential part, using sequential operations, or in the parallel part, using CAS() for the pointer changes, but without holding the readers lock. The server does not need to acquire the readers lock because no other thread will try to acquire the writer lock.
The number of transactions per successful operation is at most 3 92, but 3 22 in the 50% . . -50% case. The percentage of operations that get executed by the server (after aborting 10 times) is at most 10% of the total number of operations, but between 1 73% and 2 01% for the 50% . . -50% case.
| Working Threads | Transactions per successful operation | Fallbacks per total operations |
|-------------------|-----------------------------------------|----------------------------------|
| 1 | 1.01 | 0.00% |
| 2 | 2.34 | 0.51% |
| 3 | 3.21 | 1.73% |
| 4 | 3.31 | 2.12% |
| 5 | 3.46 | 2.74% |
| 3.46 | | 6 2.67% |
| 7 | 3.61 | 3.25% |
| Add percentage | Transactions per successful operation | Fallbacks per total operations |
|------------------|-----------------------------------------|----------------------------------|
| 100 | 1.32 | 0.00% |
| 80 | 1.77 | 0.01% |
| 60 | 2.37 | 0.29% |
| 50 | 3.22 | 2.01% |
| 40 | 3.64 | 5.24% |
| 20 | 3.92 | 10.34% |
| 0 | 1.09 | 0.00% |
Table 2: Transaction stats for varying # of threads, with 50% PQ::add() s and 50% PQ::removeMin() s
Table 3: Transaction stats for varying mixes, with 1 server thread and 3 working threads.
Table 4: Statistics on the overhead of aborted transactions.
## 6 Conclusion
In this paper, we describe a technique to implement a scalable, linearizable priority queue based on a skiplist, divided into a sequential and a parallel part. Our scheme simultaneously enables parallel PQ::add() operations as well as sequential batched PQ::removeMin() operations. The sequential part is beneficial for batched removals, which are performed by a special server thread . While detaching the sequential part from the parallel part is non-negligible cost-wise, the sequential part has the potential to serve multiple subsequent removals at a small constant cost. The parallel part is beneficial for concurrent insertions of elements with bigger keys (smaller priority), not likely to be removed soon. In other words, we integrate the flat combining/delegation paradigm introduced in prior work with disjoint-access parallelism.
In addition, we present a novel priority queue elimination algorithm, where PQ::add() operations with keys smaller than the priority queue minimum can eliminate with PQ::removeMin() operations. We permit PQ::add() operations, with keys small enough, to linger in the elimination array, waiting to become eligible for elimination. If the elimination is not possible, the operation is delegated to the server thread. Batched removals (combining) by the server thread is well-integrated with both: (1) parallelism of add() operations with bigger keys; and (2) the elimination algorithm, that possibly delegates failed elimination attempts (of elements with smaller keys) to the server thread in a natural manner. Our priority queue integrates delegation, combining, and elimination, while still leveraging the parallelism potential of insertions.
## References
- [1] Irina Calciu, Dave Dice, Tim Harris, Maurice Herlihy, Alex Kogan, Virendra J. Marathe, and Mark Moir. Message passing or shared memory: Evaluating the delegation abstraction for multicores. In OPODIS , pages 83-97, 2013.
- [2] Irina Calciu, Justin Gottschlich, and Maurice Herlihy. Using delegation and elimination to implement a scalable numa-friendly stack. In 5th USENIX Workshop on Hot Topics in Parallelism , 2013.
- [3] Danny Hendler, Itai Incze, Nir Shavit, and Moran Tzafrir. Flat combining and the synch-ronizationparallelism tradeoff. In Proceedings of the 22nd ACM symposium on Parallelism in algorithms and architectures , SPAA '10, pages 355-364, New York, NY, USA, 2010. ACM.
- [4] Danny Hendler, Nir Shavit, and Lena Yerushalmi. A scalable lock-free stack algorithm. J. Parallel Distrib. Comput. , 70(1):1-12, January 2010.
- [5] Maurice Herlihy and J. Eliot B. Moss. Transactional memory: Architectural support for lock-free data structures. SIGARCH Comput. Archit. News , 21(2):289-300, May 1993.
- [6] Galen Hunt, M. Michael, S. Parthasarathy, and M. Scott. An efficient algorithm for concurrent priority queue heaps. Information Processing Letters , 60(3):151 - 157, 1996.
- [7] Intel Corporation. Transactional Synchronization in Haswell. Retrieved from http://software.intel.com/en-us/blogs/2012/02/07/ transactional-synchronization-in-haswell/ , 8 September 2012.
- [8] I. Lotan and N. Shavit. Skiplist-based concurrent priority queues. In Proc. of the 14th International Parallel and Distributed Processing Symposium (IPDPS) , pages 263-268, 2000.
- [9] Zviad Metreveli, Nickolai Zeldovich, and M. Frans Kaashoek. Cphash: a cache-partitioned hash table. In Proceedings of the 17th ACM SIGPLAN symposium on Principles and Practice of Parallel Programming , PPoPP '12, pages 319-320, New York, NY, USA, 2012. ACM.
- [10] Mark Moir, Daniel Nussbaum, Ori Shalev, and Nir Shavit. Using elimination to implement scalable and lock-free fifo queues. In Proceedings of the seventeenth annual ACM symposium on Parallelism in algorithms and architectures , SPAA '05, pages 253-262, New York, NY, USA, 2005. ACM.
- [11] H. Sundell and P. Tsigas. Fast and lock-free concurrent priority queues for multi-thread systems. In IEEE International Symposium on Parallel and Distributed Processing , page 11 pp., april 2003.
- [12] Amy Wang, Matthew Gaudet, Peng Wu, Jos´ e Nelson Amaral, Martin Ohmacht, Christopher Barton, Raul Silvera, and Maged Michael. Evaluation of blue gene/q hardware support for transactional memories. In Proceedings of the 21st international conference on Parallel architectures and compilation techniques , PACT '12, pages 127-136, New York, NY, USA, 2012. ACM.
## A Algorithms for the Concurrent Skiplist
In this section, we present the algorithms for the concurrent skiplist described in Sec. 2.1. The skiplist contains a Single-Writer-Multi-Readers lock with writer preference, called simply lock . In terms of notation, lock.acquireR() acquires the lock for reads, and lock.acquireW() acquires the lock for writes. The SL::removeSeq() skiplist procedure is described in Alg. 3.
glyph[negationslash]
```
lock.acquirev, acquires the lock for reaus, au
SL::removeSeq() skiplist procedure is described i
Algorithm 3 SL::removeSeq()
1: if minValue = MaxInt then
2: return MaxInt
3: if currSeq = -1 then
4: moveHead()
5: key <- currSeq.key
6: currSeq.counter <- currSeq.counter - 1
7: if currSeq.counter = 0 then
8: while currSeq \ne lastSeq do
9: currSeq <- currSeq.next[0]
10: if currSeq.counter > 0 then
11: minValue = currSeq.key
12: return key
13:
```
The variable lock.timestmap contains the timestamp associated with the lock (and hence with the head-moving operations). Algorithm 4 returns a pair of elements ( b, r ): b is a bucket found using the skiplist SL::find() operation, and r is a boolean defined as follows. If a head-moving operation happened anywhere between Lines 1 and 4, the timestamp moved and r will be false.
## Algorithm 4 cleanFind( v , preds, succs)
```
Algorithm 4 cleanFind(v, preds, succs)
1: t <= lock.timestamp
2: b <= find(headPar, v, preds, succs)
3: lock.acquireR()
4: if t < lock.timestamp then
5: lock.release()
6: return (\_, false)
7: return (b, true)
```
The SL::addPar() skiplist procedure is described in Alg. 5. It uses the clean find protocol above. It performs a clean find, followed by mutable operations (either increasing a counter or inserting a bucket), executed with lock acquired for reading.
The SL::moveHead() skiplist procedure is described in Alg. 6. Line 19 creates the sequential part starting from where the parallel part used to be, and the operations starting at Line 21 separate the skiplist in two parts. Note how SL::find() is used to locate the pointers that will change in order to separate the skiplist.
Finally, the SL::chopHead() skiplist procedure is described in Alg. 7. Note that all the SL::find() operations are executed outside the critical section. These operations identify the pointers that will change in order to relink the skiplist.
Algorithm 5 SL::addPar( v )
glyph[negationslash]
```
Algorithm 5 SL::addPar(v)
1: if v < lastSeq.key then
2: return false
3: (b, r) <- cleanFind(v, preds, succs)
4: if r = false then
5: restart at line 3
6: if b \== 1 then
7: Atomically increment b.counter
8: lock.release()
9: return true
10: b <- newNode(v)
11: for i: 1 - 1 - p.topLevel do
12: b.next[i] <- succs[i]
13: if not CAS(preds[0].next[0]: succs[0] - 1 - b) then
14: lock.release()
15: restart at line 3
16: repeat
17: m <- minValue
18: until m < v or CAS(minValue: m - 1 - v)
19: for i: 1 - 1 - p.topLevel do
20: b.next[i] <- succs[i]
21: if CAS(preds[i].next[i]: succs[i] - 1 - b) then
22: continue
23: lock.release()
24: repeat
25: (b, r) <- cleanFind(v, preds, succs)
26: until r = true
27: if b = 1 - then
28: lock.release()
29: return true
30: return true
B Algorithms for Elimination and C.
In this section, we present the algorithms for the elimination a
described in Section 2? The priority name removal ie shown
```
## B Algorithms for Elimination and Combining
In this section, we present the algorithms for the elimination and combining strategies for our priority queue, described in Section 2.2. The priority queue removal is shown in Alg. 1.
The priority queue insertion algorithm is shown in Alg. 8. If the value being inserted is not suitable for the parallel part ( PQ::addPar() returns false), the request is posted in the elimination array, until eliminated with a suitable PQ::removeMin() or consumed by the server thread. Details are discussed in Sec. 2.2.
## C Implementing Combining and Elimination with Transactions
In this section, we describe our experience using Intel TSX to simplify combining and elimination. Adapting the elimination algorithm to use transactions was straightforward, by just replacing the pessimistic synchronization with transactions. We note that a unique stamp as described in Section 2.2 is not necessary for linearizability of elimination if the operations are performed inside hardware transactions. If a thread finds a matching operation and ensures in a transaction that the value is smaller than the minimum, then elimination is safe. If a change in the matching operation had occurred, the transaction would have aborted. We retry each transaction N times (e.g. N = 3 in our implementation). If a thread's transaction is aborted too many times during elimination, the thread moves on to other slots without retrying the failed slot in
## Algorithm 6 SL::moveHead()
glyph[negationslash]
```
Algorithm 6 SL::moveHead()
1: n is determined dynamically (see text)
2: lock.acquireW()
3: currSeq <- 1
4: pred <- headPar
5: curr <- headPar.next[0]
6: i = 0
7: while i < n and curr \ne tail do
8: i <- i + curr.counter
9: if currSeq = 1 then
10: currSeq <- curr; minValue <- curr.key
11: pred <- curr; curr <- curr.next[0]
12: if i = 0 then
13: for i : MaxLvl -> 0 do
14: headPar[i], headSeq[i] <- tail
15: lastSeq <- headPar, minValue <- MaxLvl
16: lock.release()
17: return false
18: lastSeq <- pred
19: for i : MaxLvl -> 0 do
20: headSeq[i] <- headPar[i]
21: find(headSeq, lastSeq + 1, preds, succs)
22: for i : MaxLvl -> 0 do
23: preds[i].next[i] <- tail
24: headPar.next[i] <- succs[i]
25: lock.release()
26: return true
Algorithm 7 SL::chopHead()
1: if currSeq = 1 then
2: return false
- e
-
```
## Algorithm 7 SL::chopHead()
- 1: if currSeq = ⊥ then
glyph[negationslash]
- 2: return false 3: find(headSeq, lastSeq.key + 1, preds, ⊥ ) 4: find(headSeq, currSeq.key, ⊥ , succs) 5: lock.acquireW() 6: for i : MaxLvl → 0 do preds[i].next[i] ← headPar.next[i] 7: lastSeq ← headPar, currSeq ← ⊥ 8: for i : MaxLvl → 0 do 9: headPar.next[i] ← succs[i] if succs[i] = tail lock.release() return true
- 10:
- 11:
a fallback path. However, if the transaction fails while trying to insert an PQ::add() or PQ::removeMin() operation in an empty slot to be collected by the server thread, the original pessimistic algorithm is used as a software fallback path in order to guarantee forward progress. Unfortunately, the unique stamp needs to be used to ensure linearizability of the operations executed on the fallback path.
Using transactions in the server thread implementation required including SL::addSeq() and SL::removeSeq() inside a transaction, which in turn caused too many aborts. Therefore, we designed an alternative combining algorithm that executes these operations outside the critical section. The complete algorithm is presented
## Algorithm 8 PQ::add(inValue)
```
Algorithm 8 PQ::add(inValue)
1: if inValue < skiplist.minValue then
2: rep <- MAX_ELIM_MIN
3: else
4: if skiplist.addPar(inValue) then
5: return true
6: rep = MAX_ELIM
7: while rep > 0 do
8: pos <- (id+1)% ELIM_SIZE; (value, stamp) <- elim[pos]
9: if value = REMREQ and (inValue < skiplist.minValue)) then
10: if CAS(elim[pos], (value, stamp), (inValue, 0)) then
11: return true
12: rep <- rep -1; inc(pos)
13: if skiplist.addPar(inValue) then
14: return true
15: while true do
16: (value, stamp) <- elim[pos]
17: if value = REMREQ and (inValue < skiplist.minValue)) then
18: if CAS(elim[pos], (value, stamp), (inValue, 0)) then
19: return true
20: if value = EMPTY then
21: if CAS(elim[pos], (value, stamp), (inValue, uniqueStamp())) then
22: repeat
23: (value, stamp) <- elim[pos]
24: until value = TAKEN
25: elim[pos] <- (EMPTY, 0); return true
26: inc(pos)
in Algorithm 9. It is based on the observation that, as long as there is a sequential part
the SL: :removeSeq() and the SL: :addSeq() operations can be executed lazily. The ser
keblistie* minValue to return a value to a remove reset and only execute the sequential
```
in Algorithm 9. It is based on the observation that, as long as there is a sequential part in the skiplist, the SL::removeSeq() and the SL::addSeq() operations can be executed lazily. The server can use the skiplist's minValue to return a value to a remove request and only execute the sequential operation after, without the remove thread waiting for it. Note that the skiplist's minValue could, in the meantime, return a value that is outdated. However, this value is always smaller or equal to the actual minimum in the skiplist, because it can only lag behind one sequential remove. This function is used by the PQ::add() operations to determine if they can eliminate or not. Therefore, estimating a minimum smaller than the actual minimum can affect performance, but will not impact correctness of our algorithm. Moreover, the server performs the PQ::removeMin() operation immediately after writing the minimum, thus cleaning up the sequential part and updating the minimum estimate. The PQ::add() case is similar too. If there is a sequential part to the skiplist, the server can update the skiplist lazily, after it releases the waiting thread. There is one difference. If the value inserted is smaller than minValue , then this needs to be updated before releasing the waiting thread.
Using these changes in the combining algorithm allowed a straightforward implementation using hardware transactions. However, our experiments indicated that certain particularities of the best-effort HTM design make it unsuitable for this scenario. First of all, because of its best-effort nature, a fallback is necessary in order to make progress. Therefore, the algorithm might be simplified on the common case, but it is still as complex as the fallback. Moreover, changes are often needed to adapt algorithms for an implementation using hardware transactions. Because these changes involve decreasing the sizes of the critical sections and decreasing the number of potential conflicts, these changes could be beneficial to the original algorithm too. Finally, it seems that communications paradigms, such as elimination and combining, are best implemented using pessimistic methods. Intel TSX has no means of implementing non-transactional operations inside
transactions (also called escape actions) and no polite spinning mechanism to allow a thread to wait for a change that is going to be performed in a transaction. The spinning thread could often abort the thread that it is waiting for. We used the PAUSE instruction in the spinning thread to alleviate this issue, but better hardware support for implementing communication paradigms using hardware transactions is necessary. For our elimination and combining algorithms, we concluded that pessimistic synchronization works better.
glyph[negationslash]
glyph[negationslash]
| 1: | true do |
|------|--------------------------------------------------------------|
| 2: | for i : 1 → ELIM SIZE do |
| 3: | (value, stamp) ← elim[i] |
| 4: | if value = REMREQ then |
| 5: | if skiplist.currSeq = ⊥ then |
| 6: | skiplist.moveHead() |
| 7: | if skiplist.currSeq = ⊥ then |
| 8: | if CAS(elim[i], (value, stamp), (skiplist.minValue, 0)) then |
| 9: | skiplist.removeSeq() |
| 10: | else |
| 11: | if CAS(elim[i], (value, stamp), (INPROG, 0)) then |
| 12: | min ← skiplist.removeSeq(); elim[i] ← (min, 0) |
| 13: | if IsValue(value) and (stamp > 0) then |
| 14: | if skiplist.currSeq = ⊥ then |
| 15: | if CAS(elim[i], (value, stamp), (INPROG, 0)) then |
| 16: | if value < skiplist.minValue then |
| 17: | skiplist.minValue ← value |
| 18: | elim[i] ← (TAKEN, 0); skiplist.addSeq(value) |
| 19: | else |
| 20: | if CAS(elim[i], (value, stamp), (INPROG, 0)) then |
| 21: | skiplist.addSeq(value); elim[i] ← (TAKEN, 0) |
| null |
[
"Irina Calciu",
"Hammurabi Mendes",
"Maurice Herlihy"
] |
2014-08-05T16:16:56+00:00
|
2014-08-05T16:16:56+00:00
|
[
"cs.DC"
] |
The Adaptive Priority Queue with Elimination and Combining
|
Priority queues are fundamental abstract data structures, often used to
manage limited resources in parallel programming. Several proposed parallel
priority queue implementations are based on skiplists, harnessing the potential
for parallelism of the add() operations. In addition, methods such as Flat
Combining have been proposed to reduce contention by batching together multiple
operations to be executed by a single thread. While this technique can decrease
lock-switching overhead and the number of pointer changes required by the
removeMin() operations in the priority queue, it can also create a sequential
bottleneck and limit parallelism, especially for non-conflicting add()
operations.
In this paper, we describe a novel priority queue design, harnessing the
scalability of parallel insertions in conjunction with the efficiency of
batched removals. Moreover, we present a new elimination algorithm suitable for
a priority queue, which further increases concurrency on balanced workloads
with similar numbers of add() and removeMin() operations. We implement and
evaluate our design using a variety of techniques including locking, atomic
operations, hardware transactional memory, as well as employing adaptive
heuristics given the workload.
|
1408.1022v2
|
## A Note on Kuhn's Theorem with Ambiguity Averse Players
Gaurab Aryal and Ronald Stauber ∗
June 17, 2014
## Abstract
Kuhn's Theorem shows that extensive games with perfect recall can equivalently be analyzed using mixed or behavioral strategies, as long as players are expected utility maximizers. This note constructs an example that illustrates the limits of Kuhn's Theorem in an environment with ambiguity averse players who use a maxmin decision rule and full Bayesian updating.
Keywords: Extensive games; Ambiguity; Maxmin; Dynamic consistency
JEL Classification: C72; D81
∗ G. Aryal: The University of Chicago, Department of Economics, 1126 E. 59th Street, Chicago, IL 60637, Email: [email protected] ; R. Stauber (corresponding author): Research School of Economics, Australian National University, ACT 0200, Australia, Telephone: (+61) 2 6125 7138, Email: [email protected]
## 1 Introduction
A classic result in game theory, Kuhn's Theorem (Kuhn, 1953) shows that for each mixed strategy in an extensive game with perfect recall, there exists an outcome-equivalent behavioral strategy, and vice versa. 1 In light of the fact that expected utility maximizers are dynamically consistent based on Bayesian updating, the principal implication of Kuhn's Theorem is that the optimality of a mixed strategy in the strategic form of an extensive game is equivalent to the conditional optimality of its outcome-equivalent behavioral strategy, at all information sets that are reached with positive probability according to a player's beliefs about the strategies of his opponents, as pointed out by Brandenburger (2007).
This equivalence between mixed and behavioral strategies does not extend to games where players are ambiguity averse in the sense of Gilboa and Schmeidler (1989), so that beliefs are represented by sets of probability measures, and a maxmin decision rule is used to define optimality. Such set-valued beliefs of a player in an extensive game may arise in a number of settings, which could include Bayesian games where players have ambiguous information about opponents' types as in Kajii and Ui (2005), games where equilibria are defined by ambiguous beliefs as in Lo (1996), games where players are allowed to use ambiguous randomization devices as in Riedel and Sass (2014), or games where players' beliefs are defined through ambiguous trembles of their opponents' unambiguous strategies as in Aryal and Stauber (2014). If we consider an extensive game with complete (but not necessarily perfect) information, so that each player i only faces uncertainty regarding his opponents' strategies, and this uncertainty is represented by set-valued beliefs, the standard Kuhn's Theorem still yields the existence of (unambiguous) behavioral and mixed strategies β i and µ i for this player that are outcome-equivalent for every element of his (set-valued) beliefs. However, if this player uses full Bayesian updating to derive beliefs at each information set, combined with a maxmin decision rule, optimality of the mixed strategy µ i in the strategic form need not correspond to the conditional optimality of the outcome-equivalent behavioral strategy β i at information sets that are reached according to any element of the player's ambiguous beliefs. This is a straightforward consequence of the well-known fact that maxmin expected utility is not dynamically consistent with full
1 See Maschler et al. (2013) for a thorough exposition of this result. Two strategies are outcome-equivalent if they induce the same distribution over terminal histories, irrespective of the strategies used by a player's opponents.
## Bayesian updating. 2
In an interesting paper, Epstein and Schneider (2003) show that dynamic consistency can be restored in dynamic choice settings with maxmin preferences and full Bayesian updating, as long as the overall uncertainty a decision-maker faces can be captured by a set of priors that has a 'rectangularity' property. Roughly, rectangularity means that the initial set of priors can be constructed by recursively combining all its corresponding conditional and marginal probabilities at each stage of the information filtration. Sass (2013) shows that the characterization of Epstein and Schneider (2003) can be applied in extensive games with perfect information to recover a version of Kuhn's Theorem, which implies that in environments with perfect information, the dynamic consistency associated to the standard Kuhn's Theorem can be recovered through the use of mixed strategies that possess a type of rectangularity property. The principal contribution of this note is to consider whether this approach can be extended to derive restrictions on ex-ante ambiguous beliefs/strategies, such as the rectangularity property identified by Epstein and Schneider (2003), that yield dynamic consistency in extensive games with imperfect information. We show by means of an example that such restrictions are not possible, if any two players' ambiguous beliefs about a third opponent's strategies are required to be consistent-in the sense that these beliefs are the same-and each player's beliefs regarding the strategies of his opponents are independent across players.
## 2 Analysis
We first present an example that shows that the dynamic consistency associated to standard expected utility maximization may not hold in extensive games with ambiguity averse players, which implies that an analysis based on the extensive form of a game may yield different conclusions than one based on its strategic form. We then show how dynamic consistency might be recovered in this example, using the rectangularity property identified by Epstein and Schneider (2003). Finally, we construct an extension of the example which shows that this approach to achieving dynamic consistency cannot be extended to all extensive games with perfect recall. Our examples are all standard extensive games with imperfect information and perfect recall, as defined in Maschler et al. (2013), for example.
2 See, for example, Epstein and Schneider (2003) and Epstein and Le Breton (1993).
Figure 1: A two-player example.
Consider the game described in Figure 1, which is a slight modification of an example from Aryal and Stauber (2014). Since each player only moves at one information set, the two players' sets of mixed and behavioral strategies are identical, so Kuhn's Theorem holds trivially. However, if we interpret mixed strategies as strategies in the strategic form associated to the extensive game, choosing a mixed strategy by player 2 implicitly requires that this player commits to a randomization over actions before the start of the game, and hence before he knows whether his information set is reached or not, following player 1's actions. If 2's beliefs about 1's strategies are unambiguous, and if 2's information set is reached with strictly positive probability according to these beliefs, the ex ante optimality of a (mixed) strategy of player 2 is equivalent to the conditional optimality of the respective (behavioral) strategy at 2's information set. This is a consequence of the dynamic consistency of expected utility maximization.
Now assume that player 2 is ambiguity averse, and that his (ambiguous) beliefs about player 1's strategy are given by an ε -contamination of the strategy that assigns probability 1 to action R . 3 Denoting the strategy assigning probability 1 to R by β ∗ 1 = ( l ∗ , r ∗ , o ∗ ) = (0 1 0), and the set of all distributions over , , { L, R, O } by ∆ , an 1 ε -contamination of β ∗ 1 for some small ε > 0 is defined by the set of distributions
$$\beta _ { 1 } ^ { \varepsilon } \coloneqq ( 1 - \varepsilon ) \beta _ { 1 } ^ { * } + \varepsilon \Delta _ { 1 }.$$
The projection of the set β ε 1 onto the l -r plane is illustrated by the gray-shaded triangle in Figure 2. 4 Denote a strategy of player 2 by β 2 = ( m, 1 -m ), where m is the probability
3 See Aryal and Stauber (2014) for an interpretation of such beliefs.
4 The origin of the coordinate system corresponds to player 1 playing O with probability 1.
r
Figure 2: An ε -contamination.
assigned to action M , and assume that player 2 chooses a value of m in the strategic form of the game, and hence is able to commit to m before knowing whether his information set is reached or not. If 2 follows a maxmin decision rule given β ε 1 , he then chooses m to solve
$$\max _ { m \in [ 0, 1 ] } \min _ { ( l, r, o ) \in \beta _ { 1 } ^ { \epsilon } } \{ o ( - 1 ) + l [ 1 0 1 ( 1 - m ) ] + r [ 1 0 1 m + 1 0 0 ( 1 - m ) ] \}.$$
Clearly, for every m ∈ [0 , 1], the minimum in this problem is attained at ( l, r, o ) = (0 1 , -ε, ε ), and hence 2's optimal choice is to set m = 1.
In order to consider 2's optimal strategy conditional on reaching his information set, we derive his conditional beliefs at the information set using full Bayesian updating of all distributions in β ε 1 . 5 In Figure 2, the resulting conditional beliefs over the information set { L, R } are represented by the thick black line connecting the points (0 , 1) and ( ε, 1 -ε ). Letting δ denote the probability assigned to history R , 2's optimal choice of a (behavioral) strategy must solve
$$\max _ { m \in [ 0, 1 ] } \min _ { \delta \in [ 1 - \varepsilon, 1 ] } \{ \delta [ 1 0 1 m + 1 0 0 ( 1 - m ) ] + ( 1 - \delta ) ( 1 - m ) 1 0 1 \} \\ = \max _ { m \in [ 0, 1 ] } \min _ { \delta \in [ 1 - \varepsilon, 1 ] } \{ \delta ( 1 0 2 m - 1 ) + 1 0 1 ( 1 - m ) \} \\ = \max _ { m \in [ 0, 1 ] } \begin{cases} 1 0 0 + m, & \text{ if } m \leq \frac { 1 } { 1 0 2 }, \\ 1 0 0 + \varepsilon + ( 1 - 1 0 2 \varepsilon ) m, & \text{ if } m > \frac { 1 } { 1 0 2 }. \end{cases} \underset { \text{$\bullet$ that in this particular case $2^{\prime}$ is information set is reached with positive probability record}$$
5 Note that in this particular case, 2's information set is reached with positive probability according to every element of β ε 1 .
It follows that as long as 1 -102 ε < 0, or equivalently, ε > 1 102 , the optimal conditional strategy of player 2 is to set m = 1 102 , so his optimal choice is not dynamically consistent.
If we interpret 2's beliefs as defined by the ε -contamination as a set of priors corresponding to a dynamic decision problem for player 2, then this set of priors does not satisfy the rectangularity condition of Epstein and Schneider (2003) that guarantees dynamic consistency, so the dynamic inconsistency is not unexpected. We now explain the implications of rectangularity for player 2's dynamic decision problem, referring to Epstein and Schneider (2003) for a general definition. 2's information structure is given by a filtration ( F 0 2 , F 1 2 ), where F 0 2 = { L, R, O } , F 1 2 = {{ L, R } { , O }} , and all information is revealed after 2 chooses an action. The restriction of every prior p 0 ∈ ∆ to 1 F 1 2 defines a marginal, or one-stepahead , distribution p 0 + over F 1 2 , such that p 0 can be decomposed in terms of p 0 + and its conditional given F 1 2 , denoted by p 1 , using the standard form
$$p ^ { 0 } = \int _ { \mathcal { F } _ { 2 } ^ { 0 } } p ^ { 1 } d p _ { + } ^ { 0 }.$$
A set of priors P ∈ ∆ 1 is rectangular if it is defined by all the compositions of all its conditionals and one-step-ahead distributions, so for every one-step-ahead distribution p 0 + corresponding to some p 0 ∈ P , and every conditional q 1 corresponding to some q 0 ∈ P , the composition ∫ F 0 2 q 1 dp 0 + is also an element of P . Epstein and Schneider (2003) note that any set of conditionals P 1 combined with any set of one-step-ahead distributions P 0 + define an induced rectangular set of priors given by
$$\mathcal { P } ( P ^ { 1 }, P _ { + } ^ { 0 } ) \coloneqq \left \{ \int _ { \mathcal { F } _ { 2 } ^ { 0 } } q ^ { 1 } d p _ { + } ^ { 0 } \, \Big | \ q ^ { 1 } \in P ^ { 1 }, \ p _ { + } ^ { 0 } \in P _ { + } ^ { 0 } \right \}.$$
In our example, player 2's beliefs are defined by a set of priors P ∈ ∆ such that 1 P = β ε 1 . If P 1 and P 0 + denotes the sets of all conditionals and one-step-ahead distributions associated to elements of P , then P ( P , P 1 0 + ) is the smallest rectangular set of priors containing P , as noted by Epstein and Schneider (2003). The set of induced one-step-ahead distributions is given by all distributions that assign a probability between 0 and ε to the event { O } , and the remaining probability to the event { L, R } . Combining these distributions with the previously described conditionals according to the formula for P ( P , P 1 0 + ), yields the smallest rectangular set of priors containing P = β ε 1 . In Figure 3, the rectangular set P ( P , P 1 0 + ) is given by the union of the gray and light gray shaded triangles. As explained in Epstein and Schneider (2003), the two lines with slope -1 form the boundary of the
r
Figure 3: The smallest rectangular set of priors containing β ε 1 .
region where the probability of { O } lies between 0 and ε , and the vertical axis together with the dashed line form the boundary of the region that induces the conditionals over { L, R } represented by the thick black line. Combining the associated one-step-ahead and conditional distributions yields the resulting rectangular set.
If we now consider player 2's problem resulting from (ambiguous) beliefs induced by the rectangular set P ( P , P 1 0 + ), we get the same optimal choice of m at 2's information set, since the set of conditional beliefs resulting from full Bayesian updating does not change. However, the ex ante problem of choosing a mixed strategy in the strategic form now becomes max m ∈ [0 1] , min p + ∈ [0 ,ε , δ ] ∈ [1 -ε, 1] { p + ( -1) + (1 -p + ) { δ [101 m +100(1 -m )] + (1 -δ )(1 -m )101 }} , which clearly has the same solution as the conditional problem, and hence dynamic consistency is restored.
The previous example suggests that it might be possible to restore dynamic consistency in extensive games by restricting every player's beliefs about his opponents' strategies in a way that generates rectangular priors for the filtration defining the player's associated decision problem. We show next, by constructing an appropriate example, that this approach cannot be used to ensure dynamic consistency in general games.
Consider the three player game described in Figure 4. This game is an extension of the two player game we analyzed previously, and since player 2's payoffs are assumed
Figure 4: A three-player example.
to be independent of player 3's actions, our previous analysis and the result of Epstein and Schneider (2003), which shows the equivalence of dynamic consistency and rectangular priors, imply that in order to guarantee dynamic consistency for player 2, we need to assume that 2's beliefs about 1's actions are given by a rectangular prior. Since 2's information is given by the same filtration as in the previous game, any rectangular set of priors must be of the form described by the gray area in Figure 5. If player 3's (ambiguous) beliefs about player 1's actions are required to be consistent with player 2's beliefs, then the rectangularity requirements that we need to ensure dynamic consistency for player 2, imply that 3's beliefs about 1's actions must be rectangular relative to 2's information filtration. We thus assume that players 2 and 3 have identical priors over 1's actions, given by a rectangular set such as the one described in Figure 5. Furthermore, assuming that 3's beliefs about 2's actions are independent of his beliefs about 1's actions, we can represent such (ambiguous) beliefs by an interval [ a, b ] ⊆ [0 , 1], where any n ∈ [ a, b ] denotes the probability 2 assigns to action N .
To analyze dynamic consistency for player 3, note that his information filtration is given by F 0 3 = { LM,LN,RM,RN,O } and F 1 3 = {{ LM,LN,RM , RN,O } { }} . Since 3's payoffs arising from the terminal histories LM LN , and RM are assumed to be identical, we can combine these histories into a single 'state' for player 3, which we will denote by Z . We can then rewrite the filtration as F 0 3 = { Z, RN,O } and F 1 3 = {{ Z , } { RN,O }} , and represent any prior in a two-dimensional 'Machina triangle' where the horizontal axis represents the
Figure 5: A rectangular set of priors for player 2.
probability of state Z , denoted by z , the vertical axis represents the probability of state RN , denoted by rn , and the origin is equivalent to state O . Given a strategy of player 1 denoted by ( l, r, o ), and an independent strategy of player 2 denoted by n , the induced prior over 3's state space assigns probability o to state O , probability rn to state RN , and probability l + (1 r -n ) to state Z . Since the origins of the triangles representing the prior ( l, r, o ) for players 2 and the corresponding prior ( l + (1 r -n , rn, o ) ) ≡ ( z, rn, o ) for player 3 both coincide with probability 1 assigned to state O , we can represent both priors on the same triangle. As the probability of O stays constant, a 'shift' from ( l, r, o ) to ( z, rn, o ) is captured by a movement along the line of slope -1 that corresponds to the probability o , as described by a move from point E to point F in Figure 6.
If both 2's and 3's beliefs about 1's actions are captured by the same arbitrary set of priors that is rectangular with respect to 2's information filtration, such as the one described in Figure 5, and 3's beliefs about 2's actions are captured by a set [ a, b ] representing probabilities n assigned to action N , we can represent the induced set of priors for player 3's filtration by shifting all points of the rectangular set spanned by the points { A,B,C,D } in Figure 5, analogously to the shift from Figure 6. As long as 0 < a < b < 1, this would imply that each point ( l, r, o ) from the initial rectangular set will yield a set of points along the line of slope -1 corresponding to o , such that the corresponding set of ' rn -coordinates' lie in the interval [ ra, rb ]. The full initial set spanned by { A,B,C,D } in the l -r plane then yields a set of priors for player 3 in the z rn -plane, which is represented in Figure 7 by the
Figure 6: Changing priors.
gray shaded area spanned by { A , B , C , D ′ ′ ′ ′ } . Note that the northwest boundary A D ′ ′ of this induced set of priors corresponds to the northwest boundary AD of the initial set of priors, transformed using the larger value n = ; similarly, the southeast boundary b B C ′ ′ is induced by the initial southeast boundary BC , transformed using the smaller value n = a . 6 Hence, if r A and r D , and l A and l D denote the r -and l -coordinates, respectively, of the points A and D in Figure 5, then the rn -coordinates of the points A ′ and D ′ in Figure 7 will be given by r A b and r D b , and the z -coordinates of A ′ and D ′ are z A ′ = l A + r A (1 -b ) and z D ′ = l D + r D (1 -b ). Since r A > r D and l A > l D , this implies that both the rn and the z coordinates of A ′ exceed those of D ′ . A similar argument shows that the rn and the z coordinates of B ′ exceed those of C ′ . Since the lines connecting A ′ and B ′ , and C ′ and D ′ , must both have slope -1, this implies that the set spanned by { A , B , C , D ′ ′ ′ ′ } cannot be rectangular relative to player 3's filtration, because such a rectangular set would have to be bounded by two vertical lines representing constant values of z , and by two lines that
6 Note that any point ( αl +(1 -α l , αr ) ′ +(1 -α r ) ′ ) on the line connecting two elements ( l, r ) and ( l ′ , r ′ ) of the initial rectangular set, if transformed using some value of n , will lie on the line connecting the respective transformations of ( l, r ) and ( l ′ , r ′ ), if based on the same value of n , since
( αl +(1 -α l ) ′ +[ αr +(1 -α r ) ′ ](1 -n , ) [ αr +(1 -α r ) ′ ] n ) = ( α l [ + (1 r -n )]+(1 -α )[ l ′ + r ′ (1 -n )] , αrn +(1 -α r n . ) ′ )
This implies that the boundaries A D ′ ′ and B C ′ ′ must always be straight lines. Furthermore, since the initial points A and D yield identical conditionals over { L, R } , the points A ′ and D ′ will yield identical conditionals over { Z, RN } , and similarly for B ′ and C ′ .
Figure 7: Induced set of priors for player 3.
bound the set of all priors that map to some interval of conditional beliefs over { RN,O } (such as the thick black line on the rn -axis in Figure 7). The failure of rectangularity for player 3's information filtration implies that there exist payoffs for player 3 under which dynamic consistency does not hold.
## 3 Conclusion
Our analysis shows that if any two players' ambiguous beliefs about a third player in a game are required to be consistent, and each player's beliefs regarding the strategies of his opponents are assumed to be independent across players, then the dynamic consistency implicit in Kuhn's Theorem cannot be achieved in every extensive game with imperfect information and ambiguity averse players, even if additional rectangularity assumptions are introduced. The obvious implication is that an analysis of such games must rely on the extensive form and on a consistent planning assumption in the spirit of Strotz (1955-1956). Alternatively, if an analyst wants to use the strategic form of a game and retain dynamic consistency, he may do so by using ambiguous beliefs that are rectangular, but then he must either allow for inconsistencies of beliefs across players, or restrict his analysis to extensive games forms for which rectangular beliefs are guaranteed to yield dynamic consistency. A recent paper by Mouraviev, Riedel, and Sass (2014) identifies restrictions that characterize a corresponding class of games.
## References
Aryal, G., Stauber, R., 2014. Trembles in extensive games with ambiguity averse players. Economic Theory (forthcoming), -.
Brandenburger, A., 2007. A note on Kuhn's theorem. In: van Benthem, J., Gabbay, D., L¨we, B. o (Eds.), Interactive Logic. Proceedings of the 7th Augustus de Morgan Workshop, London. Texts in Logic and Games 1. Amsterdam University Press, pp. 71-88.
Epstein, L. G., Le Breton, M., 1993. Dynamically consistent beliefs must be Bayesian. Journal of Economic Theory 61 (1), 1-22.
Epstein, L. G., Schneider, M., 2003. Recursive multiple-priors. Journal of Economic Theory 113 (1), 1-31.
Gilboa, I., Schmeidler, D., 1989. Maxmin expected utility with non-unique prior. Journal of Mathematical Economics 18 (2), 141-153.
Kajii, A., Ui, T., 2005. Incomplete information games with multiple priors. Japanese Economic Review 56 (3), 332-351.
Kuhn, H. W., 1953. Extensive games and the problem of information. Contributions to the Theory of Games II, 193-216.
Lo, K. C., 1996. Equilibrium in beliefs under uncertainty. Journal of Economic Theory 71 (2), 443-484.
Maschler, M., Solan, E., Zamir, S., 2013. Game Theory. Cambridge University Press.
Mouraviev, I., Riedel, F., Sass, L., 2014. Kuhn's theorem for extensive form Ellsberg games, Working paper, Bielefeld University.
Riedel, F., Sass, L., 2014. Ellsberg games. Theory and Decision 76 (4), 469-509.
Sass, L., 2013. Kuhn's theorem for extensive form Ellsberg games, Working Paper 478, Institute of Mathematical Economics, Bielefeld University.
Strotz, R. H., 1955-1956. Myopia and inconsistency in dynamic utility maximization. Review of Economics Studies 23 (3), 165-180.
|
10.1016/j.econlet.2014.08.018
|
[
"Gaurab Aryal",
"Ronald Stauber"
] |
2014-08-05T16:25:08+00:00
|
2014-08-06T00:34:52+00:00
|
[
"q-fin.EC",
"cs.GT"
] |
A Note on Kuhn's Theorem with Ambiguity Averse Players
|
Kuhn's Theorem shows that extensive games with perfect recall can
equivalently be analyzed using mixed or behavioral strategies, as long as
players are expected utility maximizers. This note constructs an example that
illustrate the limits of Kuhn's Theorem in an environment with ambiguity averse
players who use maxmin decision rule and full Bayesian updating.
|
1408.1023v6
|
## DTKI: a new formalized PKI with verifiable trusted parties
Jiangshan Yu , Vincent Cheval , and Mark Ryan ∗ † ∗
∗ University of Birmingham, United Kingdom † LORIA, CNRS, France
F
Abstract -The security of public key validation protocols for web-based applications has recently attracted attention because of weaknesses in the certificate authority model, and consequent attacks.
Recent proposals using public logs have succeeded in making certificate management more transparent and verifiable. However, those proposals involve a fixed set of authorities. This means an oligopoly is created. Another problem with current log-based system is their heavy reliance on trusted parties that monitor the logs.
We propose a distributed transparent key infrastructure (DTKI), which greatly reduces the oligopoly of service providers and allows verification of the behaviour of trusted parties. In addition, this paper formalises the public log data structure and provides a formal analysis of the security that DTKI guarantees.
Index Terms -PKI, SSL, TLS, key distribution, certificate, transparency, trust, formal verification.
## 1 INTRODUCTION
The security of web-based applications such as e-commerce and web-mail depends on the ability of a user's browser to obtain authentic copies of the public keys for the application website. For example, suppose a user wishes to log in to her bank account through her web browser. The web session will be secured by the public key of the bank. If the user's web browser accepts an inauthentic public key for the bank, then the traffic (including log-in credentials) can be intercepted and manipulated by an attacker.
The authenticity of keys is assured at present by certificate authorities (CAs). In the given example, the browser is presented with a public key certificate for the bank, which is intended to be unforgeable evidence that the given public key is the correct one for the bank. The certificate is digitally signed by a CA. The user's browser is preconfigured to accept certificates from certain known CAs. A typical installation of Firefox has about 100 root certificates in its database.
Unfortunately, numerous problems with the current CA model have been identified. Firstly, CAs must be assumed to be trustworthy. If a CA is dishonest or compromised, it
Corresponding Author: Jiangshan Yu. Email: [email protected]. This paper is published at The Computer Journal, Vol.59 No.11, pp. 16951713, 2016
may issue certificates asserting the authenticity of fake keys; those keys could be created by an attacker or by the CA itself. Secondly, the assumption of honesty does not scale up very well. As already mentioned, a browser typically has hundreds of CAs registered in it, and the user cannot be expected to have evaluated the trustworthiness and security of all of them. This fact has been exploited by attackers [1], [2], [3], [4], [5], [6]. In 2011, two CAs were compromised: Comodo [7] and DigiNotar [8]. In both cases, certificates for high-profile sites were illegitimately obtained, and in the second case, reportedly used in a man in the middle (MITM) attack [9].
## Proposed solutions
Several interesting solutions have been proposed to address these problems. For a comprehensive survey, see [10].
Key pinning mitigates the problem of untrustworthy CAs, by defining in the client browser the parameters concerning the set of CAs that are considered entitled to certify the key for a given domain [11], [12]. However, scalability is a challenge for key pinning.
Crowd-sourcing techniques have been proposed in order to detect untrustworthy CAs, by enabling a browser to obtain warnings if the received certificates are different from those that other people are being offered [13], [14], [15], [16], [17], [18], [19], [20]. Crowd-sourcing techniques have solved some CA-based problems. However, the technique cannot distinguish between attacks and authentic certificate updates, and may also suffer from an initial unavailability period.
Solutions for revocation management of certificates have also been proposed; they mostly involve periodically pushing revocation lists to browsers, in order to remove the need for on-the-fly revocation checking [21], [22]. However, these solutions create a window during which the browser's revocation lists are out of date until the next push.
More recently, solutions involving public append-only logs have been proposed. We consider the leading proposals here.
Public log-based systems : Sovereign Keys (SK) [23] aims to get rid of browser certificate warnings, by allowing domain owners to establish a long term ('sovereign') key and by providing a mechanism by which a browser can hard-fail if it doesn't succeed in establishing security via that key. The sovereign key is used to cross-sign operational TLS [24], [25] keys, and it is stored in an append-only log on a 'time-line server', which is abundantly mirrored. However, in SK, internet users and domain owners have to trust mirrors of time-line servers, as SK does not enable mirrors to provide efficient verifiable proofs that the received certificate is indeed included in the append-only log.
Certificate transparency (CT) [26] is a technique proposed by Google that aims to efficiently detect fake public key certificates issued by corrupted certificate authorities, by making certificate issuance transparent. They improved the idea of SK by using append-only Merkle tree to organise the append-only log. This enables the log maintainer to provide two types of verifiable cryptographic proofs: (a) a proof that the log contains a given certificate, and (b) a proof that a snapshot of the log is an extension of another snapshot ( i.e. , only appends have taken place between the two snapshots). The time and size for proof generation and verification are logarithmic in the number of certificates recorded in the log. Domain owners can obtain the proof that their certificates are recorded in the log, and provide the proof together with the certificate to their clients, so the clients can get a guarantee that the received certificate is recorded in the log.
Accountable key infrastructure (AKI) [27] also uses public logs to make certificate management more transparent. By using a data structure that is based on lexicographic ordering rather than chronological ordering, they solve the problem of key revocations in the log. In addition, AKI uses the 'checks-and-balances' idea that allows parties to monitor each other's misbehaviour. So AKI limits the requirement to trust any party. Moreover, AKI prevents attacks that use fake certificates rather than merely detecting such attacks (as in CT). However, as a result, AKI needs a strong assumption namely, CAs, public log maintainers, and validators do not collude together - and heavily relies on third parties called validators to ensure that the log is maintained without improper modifications.
Certificate issuance and revocation transparency (CIRT) [28] is a proposal for managing certificates for end-to-end encrypted email. It proposes an idea to address the revocation problem left open by CT, and the trusted party problem of AKI. It collects ideas from both CT and AKI to provide transparent key revocation, and reduces reliance on trusted parties by designing the monitoring role so that it can be distributed among user browsers. However, CIRT can only detect attacks that use fake certificates; it cannot prevent them. In addition, since CIRT was proposed for email applications, it does not support the multiplicity of log maintainers that would be required for web certificates.
Attack Resilient Public-Key Infrastructure (ARPKI) [29] is an improvement on AKI. In ARPKI, a client can designate n service providers (e.g. CAs and log maintainers), and only needs to contact one CA to register her certificate. Each of the designated service providers will monitor the behaviour of other designated service providers. As a result, ARPKI prevents attacks even when n -1 service providers are colluding together, whereas in AKI, an adversary who successfully compromises two out of three designated service providers can successfully launch attacks [29]. In addition, the security property of ARPKI is proved by using a protocol verification tool called Tamarin prover [30]. The weakness of ARPKI is that all n designated service providers have to be involved in all the processes (i.e. certificate registration, confirmation, and update), which would cause considerable extra latencies and the delay of client connections.
In public log-based systems, efforts have been made to integrate revocation management with the certificate auditing. CT introduced revocation transparency (RT) [31] to deal with certificate revocation management; and in AKI and ARPKI, the public log only stores currently valid certificates (revoked certificates are purged from the log). However, the revocation checking processes in both RT and A(RP)KI are linear in the number of issued certificates making it inefficient. CIRT allows efficient proofs of non-revocation, but it does not scale to multiple logs which are required for web certificates.
## Remaining problems
Afoundational issue is the problem of oligopoly . The presentday certificate authority model requires that the set of global certificate authorities is fixed and known to every browser, which implies an oligopoly. Currently, the majority of CAs in browsers are organisations based in the USA, and it is hard to become a browser-accepted CA because of the strong trust assumption that it implies. This means that a Russian bank operating in Russia and serving Russian citizens living in Russia has to use an American CA for their public key. This cannot be considered satisfactory in the presence of mutual distrust between nations regarding cybersecurity and citizen surveillance, and also trade sanctions which may prevent the USA offering services (such as CA services) to certain other countries.
None of the previously discussed public log-based systems address this issue. In each of those solutions, the set of log maintainers (and where applicable, time-line servers, validators, etc.) is assumed to be known by the browsers, and this puts a high threshold on the requirements to become a log maintainer (or validator, etc.). Moreover, none of them solve the problem that a multiplicity of log maintainers reduces the usefulness of transparency, since a domain owner has to check each log maintainer to see if it has mis-issued certificates. This can't work if there is a large number of log maintainers operating in different geographical regions, each one of which has to be checked by every domain owner.
A second issue is the requirement of trusted parties. Currently, all existing proposals have to rely on some sort of trusted parties or at least assume that not all parties are colluding together. However, a strong adversary (e.g. a government agency) might be able to control all service providers (used by a given client) in a system.
A third foundational issue of a different nature is that of analysis and correctness. SK, CT, AKI and CIRT are
large and complex protocols involving sophisticated data structures, but none of them have been subjected to rigorous analysis. It is well-known that security protocols are notoriously difficult to get right, and the only way to avoid this is with systematic verification. For example, attacks on AKI and CIRT have been identified in [29] and in the appendix of our technical report [32], respectively. The flaws may be easily fixed, but only once they have been identified. It is therefore imperative to verify this kind of complex protocol. ARPKI is the first formally verified log-based PKI system. However, they used several abstractions during modelling in the Tamarin prover. For example, they represent the underlying log structure (a Merkle tree) as a list. However, in systems like CIRT and this paper with more complex data structures, it is important to have a formalised data structure and its properties to prove the security claim. The formalisation of complex data structures and their properties in the log-based PKI systems is a remaining problem.
The last problem is the management of certificate revocation. As explained previously, existing solutions for managing certificate revocation (e.g. CRL, OCSP, RT) are still unsatisfactory.
## This paper
We propose a new public log-based architecture for managing web certificates, called Distributed Transparent Key Infrastructure (DTKI), with the following contributions.
- · We identify anti-oligopoly as an important property for web certificate management which has hitherto not received attention.
- · Compared to its predecessors, DTKI is the first system to have all desired features - it minimises the presence of oligopoly, prevents attacks that use fake certificates, provides a way to manage certificate revocation, verifies output from trusted parties, and is secure even if all service providers (e.g. CAs and log maintainers) collude together (see Section 5 for our security statement). A comparison of the properties of different log-based systems can be found in Section 6.
- · We provide formal machine-checked verification of its core security property using the Tamarin prover. In addition, we formalise the data structures needed for transparent public logs, and provide rigorous proofs of their properties.
## 2 OVERVIEW OF DTKI
Distributed Transparent Key Infrastructure (DTKI) is an infrastructure for managing keys and certificates on the web in a way which is transparent , minimises oligopoly , and allows verification of the behaviour of trusted parties. In DTKI, we mainly have the following agents:
Certificate log maintainers (CLM): A CLM maintains a database of all valid and invalid (e.g. expired or revoked) certificates for a particular set of domains for which it is responsible. It commits to digests of its log, and provides efficient proofs of presence and absence of certificates in the log. CLMs behave transparently and their actions can be verified.
A mapping log maintainer (MLM): To minimise oligopoly, DTKI does not fix the set of certificate logs. The MLM maintains association between certificate logs and the domains they are responsible for. It also commits to digests of the log, and provides efficient proof of current association, and behaves transparently. Clients of the MLM are not required to blindly trust the MLM, because they can efficiently verify the obtained associations.
The MLM has a strategic role of determining the authorised CLMs, and the mapping log to be maintained rarely changes; therefore it can be easily governed by an international panel. In practice, ICANN is a possible party to be given the responsibility to run the MLM.
Users and their browsers: They query the MLM, and obtain and verify the proofs about the mapping of top-level domains (TLDs) to CLMs. They query CLMs and obtain and verify proofs about certificates.
Mirrors: Mirrors are servers that maintain a full copy of the mapping log and certificate logs respectively downloaded from the MLM and corresponding CLMs, and the corresponding digest of the log signed by the log maintainer. In other words, mirrors are distributed copies of logs. Anyone (e.g. ISPs, CLMs, CAs, domain owners) can be a mirror. Unlike in SK, mirrors are not required to be trusted in DTKI, because they give a proof for every association that they send to their clients. The proof is associated to the digest of the MLM.
Certificate authorities (CA): They check the identity of domain owners, and create certificates for the domain owners' keys. However, in contrast with today's CAs, the ability of CAs in DTKI is limited since the issuance of a certificate from a CA is not enough to convince web browsers to accept the certificate (proof of presence in the relevant CLM is also needed).
In DTKI, each domain owner has two types of certificate, namely TLS certificate and master certificate. Domain owners can have different TLS certificates but can only have one master certificate. A TLS certificate contains the public key of a domain server for a TLS connection, whereas the master certificate contains a public key, called 'master verification key'. The corresponding secret key of the master certificate is called 'master signing key'. Similar to the 'sovereign key' in SK [23], the master signing key is only used to validate a TLS certificate (of the same subject) by issuing a signature on it. This limits the ability of certificate authorities since without having a valid signature (issued by using the master signing key), the TLS certificate will not be accepted. Hence, the TLS secret key is the one for daily use; and the master signing key is rarely used. It will only be used for validating a new certificate, or revoking an existing certificate. We assume that domain owners can take care of their master signing key, as a master signing key can be kept offline, and is rarely used.
After a domain owner obtains a master certificate or a TLS certificate from a CA, he needs to make a registration request to the corresponding CLM to publish the certificate into the log. To do so, the domain owner signs the certificate using the master signing key, and submits the signed certificate to a CLM determined (typically based on the top-level domain) by the MLM. The CLM checks the signature, and
accepts the certificate by adding it to the certificate log if the signature is valid. The process of revoking a certificate is handled similarly to the process of registering a certificate in the log.
When establishing a secure connection with a domain server, the browser receives a corresponding certificate and proofs from a mirror of the MLM and a CLM, and verifies the certificate, the proof that the certificate is valid and recorded in the certificate log, and proof that this certificate log is authorised to manage certificates for the domain. Users and their browsers should only accept a certificate if the certificate is issued by a CA, and validated by the domain owner, and current in the certificate log.
Fake master certificates or TLS certificates can be easily detected by the domain owner, because the attacker will have had to insert such fake certificates into the log (in order to be accepted by browsers), and is thus visible to the domain owner.
Rather than relying solely on trusted monitors to verify the healthiness of logs and the relations between logs, DTKI uses a crowdsourcing-like way to ensure the integrity of the log and the relations between mapping log and a certificate log, and between certificate logs. In particular, the monitoring work in DTKI can be broken into independent little pieces, and thus can be done by distributing the pieces to users' browsers. In this way, users' browsers can perform randomly-chosen pieces of the monitoring role in the background (e.g. once a day). Thus, web users can collectively monitor the integrity of the logs. We envisage parameters in browsers allowing users to control how that works.
To avoid the case that attackers create a 'bubble' (i.e. an isolated environment) around a victim, we share the same assumption as other existing protocols (e.g. CT and CIRT) - we assume that gossip protocols [33] are used to disseminate digests of the log. So, users of logs can detect if a log maintainer shows different versions of the log to different sets of users. Since log maintainers sign and timestamp their digests, a log maintainer that issues inconsistent digests can be held accountable.
## 3 THE PUBLIC LOG
DTKI uses append-only logs to record all requests processed by the log maintainer, and allows log maintainers to efficiently generate some proofs that can be efficiently verified. These proofs mainly include that some data (e.g. a certificate or a revocation request) has or has not been added to the log; and that a log is extended from a previous version.
So, the log maintainer's behaviour is transparent to the public, and the public is not required to blindly trust log maintainers. Public log data structures have been widely studied [34], [35], [36], [37], [23], [26], [28]. To the best of our knowledge, no single data structure can provide all proofs required by DTKI. We adopt and extend the idea of CIRT log structure [28] which makes use of two data structures to provide all the kinds of proofs needed for DTKI.
This section presents the intuition of two abstract data structures encapsulating the desired properties, then introduces how to use the data structures to construct our public logs in a concrete manner by extending the CIRT data structure. The formalisation of our abstract data structures,
| Function | Output |
|----------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Chronological Data Structure | Chronological Data Structure |
| digest | given input a sequence S of data, it outputs the digest of sequence S of data organised |
| VerifPoP c | by using chronological data structure given input ( digest ( S ) , d, p ) , it outputs a boolean value indicating the verification result of the proof p that some data d is included in a set S |
| VerifPoE c | given input (( dg ′ , N ′ ) , ( dg, N ) , p ) , it outputs a boolean value indicating the verification result of the proof p that a sequence of data represented by its digest dg and size N is extended from another sequence of data represented by digest dg ′ and size N ′ |
| Ordered Data Structure | Ordered Data Structure |
| digest O | given input a sequence S of data, it outputs the digest of sequence S of data organised |
| VerifPoP O (resp. VerifPoAbs O ) | by using ordered data structure given input ( digest o ( S ) , d, p ) , it outputs a boolean value indicating the verification result of the proof p that some data d is (resp. is not) included in a set S |
| VerifPoAdd O (resp. VerifPoD O ) | given input ( d, dg, dg ′ , p ) , it outputs a boolean value indicating the verification result of the proof p that dg ′ is the digest obtained after adding data d into (resp. deleting data d from) the sequence of data represented by digest dg ′ ′ |
| VerifPoM O | given input ( d, d , dg, dg , p ) , it outputs a boolean value indicating the verification result of the proof p that dg ′ is the digest obtained after replacing d with d ′ in the sequence of data represented by dg |
Table 1: Some functions supported by the data structures, of size N . The full list of operations and functions supported by the data structures, and the detailed properties of the data structures, are formalised in our technical report.
log structures, and their properties, and our detailed implementation, are presented in our technical report [32]. We also present some examples of the data structures there.
## 3.1 Data structures
Our log makes use of two data structures, namely chronological data structure and ordered data structure, to provide all the proofs required by DTKI. We use the notion of digest to represent a unique set of data, such that the size of a digest is a constant. For example, a digest could be the hash value of a set of data.
A chronological data structure is an append-only data structure, i.e. only the operation of adding some data is allowed. With a chronological data structure, for a given sequence S of data of size N and with digest dg , we have d ∈ S for some data d , if and only if there exists a proof p of size O (log( N )) , called the proof of presence of d in S , such that p can be efficiently verified by using VerifPoP c (see Table 1); and for all sequence S ′ with digest dg ′ and size N < N ′ , we have that S ′ is a prefix of S , if and only if there exists a proof p ′ of size O (log( N )) , called the proof
of extension of S from S ′ , such that p ′ can be efficiently verified by using VerifPoE c (see Table 1).
In this way, to verify that some data is included in a sequence of data stored in a chronological data structure (of size N ), the verifier only needs to download the corresponding digest, and the corresponding proof of presence (with size O log N ( ( )) ). The verification of proof of extension is similarly efficient. Possible implementations are appendonly Merkle tree [34] and append-only skip list, as proposed in [26] and [36], respectively.
With the append-only property, the chronological data structure enables one to prove that a version of the data structure is an extension of a previous version. This is useful for our public log since it enables users to verify the history of a log maintainer's behaviours.
Unfortunately, the chronological data structure does not provide all desired features. For example, it is very inefficient to verify that some data (e.g. a revocation request) is not in the chronological data structure (the cost is O N ( ) , where N is the size of the data structure). To provide missing features, we need to use the ordered data structure .
An ordered data structure is a data structure allowing one to insert, delete, and modify stored data. In addition, with an ordered data structure, for a given sequence S of data of size N and with digest dg , we have d ∈ S (resp. d / S ∈ ) for some data d , if and only if there exists a proof p of size O (log( N )) , called the proof of presence (resp. absence) of d in (resp. not in) S , such that p can be efficiently verified by using VerifPoP O (resp. VerifPoAbs O ) (see Table 1).
Possible implementations of ordered data structure are Merkle tree which is organised as a binary search tree (as proposed in [28]), and authenticated dictionaries [35].
With an ordered data structure, however, the size of proof that the current version of the data is extended from a previous version is O N ( ) . As the chronological data structure and the ordered data structure have complementary properties, we use the combination of them to organise our log.
## 3.2 Mapping log
To minimise oligopoly, DTKI uses multiple certificate logs, and does not fix the set of certificate logs and the mapping between domains and certificate logs. A mapping log is used to record associations between domain names and certificate log maintainers, and can provide efficient proofs regarding the current association. It would be rather inefficient to explicitly associate each domain name to a certificate log, due to the large number of domains. To efficiently manage the association, we use a class of simple regular expressions to present a group of domain names, and record the associations between regular expressions and certificate logs in the mapping log. For example, the mapping might include (* \ .org, Clog 1 ) and ([a-h].* \ .com, Clog ) to mean that certifi1 cate log maintainer Clog 1 deals with domains ending .org and domains starting with letters from a to h ending .com . In our technical report [32], we have formally defined some constraints on the regular expressions we use, the relations between them, and how to use random verification to verify that no overlap between regular expressions exists.
Intuitively, as presented in Figure 1, the mapping log is organised by using a chronological data structure, and stores received requests 1 together with the request time, and four digests of different ordered data structures representing the status of the log. Each entry is of the form
$$\mathfrak { h } ( r e q, t, d g ^ { s }, d g ^ { b l }, d g ^ { r }, d g ^ { i } )$$
In the formula, req is the request received by the mapping log at time t ; dg s 2 stores information about CLMs (e.g. the certificate of the CLM, and the current digest of the certificate log clog ); dg bl stores the identity of blacklisted certificate log maintainers; dg r stores the mapping from a regular expression to the identity of CLMs, and dg i stores the mapping from the identity of CLMs to a set of regular expressions.
In more detail, each entry of the mapping log contains digests after processing the request req (received by the mapping log maintainer at time t ) on the digest stored in the previous record. Each of the notations is explained as follows:
- · req can be add ( rgx, id ) , del ( rgx, id ) , new ( cert ) , mod ( cert, sign sk ( cert ′ ) , sign sk ′ ( n, dg, t )) , bl ( id ) , and end , respectively corresponding to a request to add a mapping ( rgx, id ) of regular expression rgx and identity id of a clog , to delete a mapping ( rgx, id ) , to add a certificate cert of a new clog , to change the certificate of a clog from cert to cert ′ , to blacklist id of an existing clog , and to close the update request; where sk and sk ′ are signing keys associated to the certificate cert and cert ′ , respectively; cert and cert ′ share the same subject, and n and dg are the size and the digest of the corresponding clog at time t , respectively;
- · dg s is the digest of an ordered data structure storing the identity information of the form ( cert, sign sk ( n, dg, t )) for the currently active certificate logs, where cert is the certificate for the signing key sk of the certificate log, and n and dg are respectively the size and digest of the certificate log at time t . Data are ordered by the domain name in cert .
- · dg bl is the digest of an ordered data structure storing the domain names of blacklisted certificate logs. Data are ordered by the stored domain names.
- · dg r is the digest of an ordered data structure storing elements of the form ( rgx, id ) , which represents the mapping from regular expression rgx to the identity id of a clog , data are ordered by rgx ;
- · dg i is the digest of an ordered data structure storing elements of the form ( id, dg irgx ) , which represents the mapping from identity id of a clog to a digest dg irgx of ordered data structure storing a set of regular expressions, data are ordered by id .
The requests are used for modifying mappings or the existing set of certificate log maintainers. When a request
- 1. The request includes adding, removing, and modifying a certificate log and/or a mapping.
- 2. We simplified the description here: we should say the ordered data structure represented by dg s stores the information, rather than the digest dg s stores it. We will use this simplification through the paper.
del ( rgx, id ) has been processed, the maintainer of certificate log with identity id needs to remove all certificates whose subject is an instance of regular expression rgx ; when a request add ( rgx, id ) has been processed, the maintainer of certificate log with identity id needs to download all certificates whose subject is an instance of rgx from the previous authorised log maintainer, and adds them into his log. These requests require certificate logs to synchronise with the mapping log; see Section 3.4.
Figure 1: A figure representation of the format of each record in the mapping log.
## 3.3 Certificate logs
The mapping log determines which certificate log is used for a domain. The certificates for the domain are stored in that certificate log.
A certificate log mainly stores certificates for domains according to the mappings presented in the mapping log. In particular, a certificate log is also organised by using a chronological data structure, and each entry of the log is of the form
$$\mathbf h ( r e q, N, d g ^ { r g x } )$$
where req is the received request and is processed at the time such that the mapping log is of size N dg ; rgx represents an ordered data structure storing a set of mappings from regular expressions to the information associated to the corresponding domains, such that the domain name is an instance of the regular expression. The stored information of a domain includes the identity and the master certificate of the domain, and two digests dg a and dg rv each presents an ordered data structure storing a set of active TLS certificates and a set of expired or revoked TLS certificates, respectively.
Elements in a record (as shown in 2) of a certificate log are detailed as follows.
- · req can be reg sign ( sk ( cert, t, 'reg' )) , rev sign ( sk ( cert, t, 'rev' )) , upadd h ( ( id , h ) ) , and updel h ( ( id , h ) ) , corresponding to a request to register and revoke a certificate cert at an agreed time t such that ( cert, t, 'reg' ) or is additionally signed by the master key sk , and update the certificate log by adding and by deleting certificates
- of identity id according to the changes of mlog , respectively. 'reg' and 'rev' are constant, and h is some value and we will explain it later.
- · N is the size of mlog at the time req is processed;
- · dg rgx is the digest of an ordered data structure storing a set of elements of the form ( rgx, dg id ) , represents the status of the certificate log after processing the request req , and stores all the regular expressions rgx that the certificate log is associated to. dg id is the digest of an ordered data structure storing a set of elements of the form ( h ( id , ) h ( cert, dg a , dg rv )) . It represents all domains associated to rgx . id is an instance of rgx and is the subject of master certificate cert . dg a and dg rv are digests of two ordered data structures each of which respectively stores a set of active and revoked TLS certificates. In addition, data in the structure represented by dg rgx and dg id are ordered by rgx and h id ( ) , respectively; data in the structure represented by dg a and dg rv are ordered by the subject of TLS certificates.
Note that requests upadd h ( ( id , h ) ) and updel h ( ( id , h ) ) are made according to the mapping log. Even though these modifications are not requested by domain owners, it is important to record them in the certificate log to ensure the transparency of the log maintainer's behaviour. Request upadd h ( ( id , h ) ) states that the certificate log maintainer is authorised to manage certificates for the domain name id from now on, and the current status of certificates for id is represented by h , where h = h ( cert, dg a , dg rv ) for some certificate cert and some digest dg a and dg rv representing the active and revoked certificates of id . h is the value obtained from the certificate log that is previously authorised to manage certificates for domain id . Similarly, request updel h ( ( id , h ) ) indicates that the certificate log cannot manage certificates for domain id any more according to the request in the mapping log.
Figure 2: A figure representation of the format of each record in the certificate log.
## 3.4 Synchronising the mapping log and certificate logs
The mapping log periodically (e.g. every day) publishes a signature sign sk ( t, dg, N ) , called signed Mlog time-stamp , on a time t indicating the publishing time, and the digest dg and size N of the mapping log. Mirrors of the mapping log need to download this signed data, and update their copy of the mapping log when it is updated. A signed Mlog time-stamp is only valid during the issue period (e.g. the day of issue). Note that mirrors can provide the same set of proofs as the mapping log maintainer, because the mirror has the copy of the entire mapping log; but mirrors are not required to be trusted, they do not need to sign anything, and a mirror which changed the log by itself will not be able to convince other users to accept it since the mirror cannot forge the signed Mlog time-stamp .
When a mapping log maintainer needs to update the mapping log, he requests all certificate log maintainers to perform the required update, and expects to receive the digest and size of all certificate logs once they are updated. After the mapping log maintainer receives these confirmations from all certificate log maintainers, he publishes the series of update requests in the mapping log, and appends an extra constant request end after them in the log to indicate that the update is done.
Log maintainers only answer requests according to their newly updated log if the mapping log maintainer has published the update requests in the mapping log. If in the log update period, some user sends requests to the mapping log maintainer or certificate log maintainers, then they give answers to the user according to their log before the update started.
We say that the mapping log and certificate logs are synchronised , if certificate logs have completed the log update according to the request in the mapping log. Note that a mis-behaving certificate log maintainer (e.g. one recorded fake certificates in his log, or did not correctly update his log according to the request of the mapping log) can be terminated by the mapping log maintainer by putting the certificate log maintainer's identity into the blacklist, which is organised as an ordered data structure represented by dg bl (as presented in 3.2).
## 4 DISTRIBUTED TRANSPARENT KEY INFRASTRUCTURE
Distributed transparent key infrastructure (DTKI) contains three main phases, namely certificate publication, certificate verification, and log verification. In the certificate publication phase, domain owners can upload new certificates and revoke existing certificates in the certificate log they are assigned to; in the certificate verification phase, one can verify the validity of a certificate; and in the log verification phase, one can verify whether a log behaves correctly.
Wepresent DTKI using the scenario that a TLS user Alice wants to securely communicate with a domain owner Bob who maintains the domain example.com .
## 4.1 Certificate insertion and revocation
To publish or revoke certificates in the certificate log, the domain owner Bob needs to know which certificate log is currently authorised to record certificates for his domain. This can be done by communicating with a mirror of the mapping log. We detail the protocol for requesting the mapping for Bob's domain.
## 4.1.1 Request mappings
Upon receiving the request, the mirror locates the certificate of the authorised CLM, and generates the proofs that
- a) the CLM is authorised for the domain; and
- b) the certificate is the current valid certificate for the CLM.
Loosely speaking, proof a) is the proof that the mapping from regular expression rgx to identity id is present in the digest dg r (as presented in the mapping log structure), such that example.com is an instance of rgx , and id is the identity of the CLM; proof b) is the proof that the certificate with subject id is present in dg s ; additionally, a proof that both dg s and dg r are present in the latest record of the mapping log is needed. All proofs should be linked to the latest digest signed by the MLM. If Bob has previously observed a version of the mlog, then a proof that the current mlog is an extension of the version that Bob observed will also be provided.
Bob accepts the response if all proofs are valid. He then stores the verified data in his cache for future connection until the signed digest is expired.
In more detail, after a mirror receives a request from Bob, the mirror obtains the data of the latest element of its copy of the mapping log, denoted h = h ( req, t, dg s , dg bl , dg r , dg i ) , and generates the proof of its presence in the digest (denoted dg mlog ) of its log of size N . Then, it generates the proof of presence of the element ( cert, sign sk ( n, dg, t )) in the digest dg s for some sign sk ( n, dg, t ) , proving that the certificate log maintainer whose cert belongs to is still active. Moreover, it generates the proof of presence of some element ( rgx, id ) in the digest dg r where id is the subject of cert and example.com is an instance of the regular expression rgx , proving that id is authorised to store the certificates of example.com . The mirror then sends to Bob the hash h , the signature sign sk ( n, dg, t ) , the regular expression rgx , the three generated proofs of presence, and the latest signed Mlog time-stamp containing the time t mlog , and digest dg mlog and size N mlog of the mapping log.
Bob first verifies the received signed Mlog time-stamp with the public key of the mapping log maintainer embedded in the browser, and verifies whether t Mlog is valid or not. Then Bob checks that example.com is an instance of rgx , and verifies the three different proofs of presence. If all checks hold, then Bob sends the signed Mlog time-stamp containing ( t ′ Mlog , dg ′ mlog , N ′ mlog ) that he stored during a previous connection, and expects to receive a proof of extension of ( dg ′ mlog , N ′ mlog ) into ( dg mlog , N mlog ) . If the received proof of extension is valid, then Bob stores the current signed Mlog time-stamp , and believes that the certificate log with identity id , certificate cert , and size that should be no smaller than n , is currently authorised for managing certificates for his domain.
## 4.1.2 Insert and revoke certificates
The first time Bob wants to publish a certificate for his domain, he needs to generate a pair of master signing key,
denoted sk m , and verification key. The latter is sent to a certificate authority, which verifies Bob's identity and issues a master certificate cert m for Bob. After Bob receives his master certificate, he checks the correctness of the information in the certificate. The TLS certificate can be obtained in the same way.
To publish the master certificate, Bob signs the certificate together with the current time t by using the master signing key sk m , and sends it together with the request AddReq to the authorised certificate log maintainer whose signing key is denoted sk clog . The certificate log maintainer checks whether there exists a valid master certificate for example.com ; if there is one, then the log maintainer aborts the conversation. Otherwise, the log maintainer verifies the validity of time t and the signature.
If they are all valid, the log maintainer updates the log, generates the proof of presence that the master certificate for Bob is included in the log, and sends the signed proof and the updated digest of the log back to Bob. If the signature and the proof are valid, and the size of the log is no smaller than what the mirror says, then Bob accepts and stores the response as an evidence of successful certificate publication. If Bob has previously observed a version of the clog, then a proof that the current clog is an extension of the version that Bob observed is also required.
Figure 3 presents the detailed process to publish the master certificate cert m . After a log maintainer receives and verifies the request from Bob, the log maintainer updates the log, generates the proof of presence of ( h ( id , ) h ( cert m , dg a , dg rv )) in dg id , ( rgx, dg id ) in dg rgx , and h reg ( ( sign sk m ( cert m , t, 'reg' )) , N mlog , dg rgx ) is the last element in the data structure represented by dg clog , where id is the subject of cert m and an instance of rgx ; reg ( sign sk m ( cert m , t, 'reg' )) is the register request to adding cert m into the certificate log with digest dg clog at time t . The log maintainer then issues a signature on ( dg clog , N, h ) , where N is the size of the certificate log, and h = h (( rgx, dg id ) , dg rgx , P ) , where P is the sequence of the generated proofs, and sends the signature σ 2 together with ( dg clog , N, rgx, dg id , dg rgx , dg a , dg rv , P ) to Bob. If the signature and the proof are valid, and N is no smaller than the size n contained in the signed Mlog time-stamp that Bob received from the mirror, then Bob stores the signed ( dg clog , N, h ) , sends the previous stored ( dg ′ clog , N ′ ) to the certificate log maintainer, and expects to receive a proof of extension of ( dg ′ clog , N ′ ) into ( dg clog , N ) . If the received proof of extension is valid, then Bob believes that he has successfully published the new certificate.
Note that it is important to send ( dg ′ clog , N ′ ) after receiving ( dg clog , N ) , because otherwise the log maintainer could learn the digest that Bob has, then give a pair ( dg ′′ clog , N ′′ ) of digest and size of the log such that N < N ′ ′′ < N . This may open a window to attackers who wants to convince Bob to use a certificate which was valid in dg ′′ clog but revoked in dg clog .
In addition, if Bob has run the request mapping protocol more than once, and has obtained a digest that is different from his local copy of the corresponding certificate log, then he should ask the CLM to prove that one of the digests is an extension of the other.
The process of adding a TLS certificate is similar to the process of adding a master certificate, but the log maintainer needs to verify that the TLS certificate is signed by the valid master signing key corresponding to the master certificate in the log.
To revoke a (master or TLS) certificate, the domain owner can perform a process similar to the process of adding a new certificate. For a revocation request with sign sk m ( cert, t ) , the log maintainer needs to check that sign sk m ( cert, t ′ ) is already in the log and t > t ′ . This ensures that the same master key is used for the revocation.
## 4.2 Certificate verification
Figure 3: The protocol presenting how domain owner Bob communicates with certificate log (clog) maintainer to publish a master certificate cert m .
When Alice wants to securely communicate with example.com , she sends the connection request to Bob, and expects to receive a master certificate cert m and a signed TLS certificate sign sk m ( cert, t ) from him. To verify the received certificates, Alice checks whether the certificates are expired. If both of them are still in the validity time period, Alice requests (as described in 4.1.1) the corresponding mapping from a mirror to find out the authorised certificate log for example.com , and communicates with the (mirror of) authorised certificate log maintainer to verify the received certificate.
Note that this verification requests extra communication round trips, but it gives a higher security guarantee. An alternative way is that Bob provides both certificates and proofs, and Alice verifies the received proofs directly.
Figure 4 presents the detailed process of verifying a certificate. After Alice learns the identity of the authorised certificate log, she sends the verification request V erifReq
with her local time t A and the received certificate to the certificate log maintainer. The time t A is used to prevent replay attacks, and will later be used for accountability. The certificate log maintainer checks whether t A is in an acceptable time range (e.g. t A is in the same day as his local time). If it is, then he locates the corresponding ( rgx, dg id ) in dg rgx in the latest record of his log such that example.com is an instance of regular expression rgx , locates ( h ( id , ) h ( cert m , dg a , dg rv )) in dg id and cert in dg a , then generates the proof of presence of cert in dg a , ( h ( id , ) h ( cert m , dg a , dg rv )) in dg id , ( rgx, dg id ) in dg rgx , and h ( req, N mlog , dg rgx ) is the latest record in the digest dg clog of the log with size N . Then, the certificate log maintainer signs ( dg clog , N, t A , h ) , where h = h ( m ) such that m = ( dg a , dg rv , rgx, dg id , req, N mlog , dg rgx , P ) , and P is the set of proofs, and sends ( dg clog , N, σ ) to Alice.
Alice should verify that N mlog is the same as her local copy of the size of mapping log. If the received N mlog is greater than the copy, then it means that the mapping log is changed (it rarely happens) and Alice should run the request mapping protocol again. If N mlog is smaller, then it means the CLM has misbehaved. Alice then verifies the signature and proofs, and sends the previously stored dg ′ clog with the size N ′ to the log maintainer, and expects to receive the proof of extension of ( dg ′ clog , N ′ ) into ( dg clog , N ) . If they are all valid, then Alice replaces the corresponding cache by the signed ( dg clog , N, t A , h ) and believes that the certificate is an authentic one.
In order to preserve privacy of Alice's browsing history, instead of asking Alice to query all proofs from the log maintainer, Alice can send the request to Bob who will redirect the request to the log maintainer, and redirect the received proofs from the log maintainer to Alice.
With DTKI, Alice is able to verify whether Bob's domain has a certificate by querying the proof of absence of certificates for example.com in the corresponding certificate log. This is useful to prevent TLS stripping attacks, where an attacker can maliciously convert an HTTPS connection into an HTTP connection.
## 4.3 Log verification
Users of the system need to verify that the mapping log maintainer and certificate log maintainers did update their log correctly according to the requests they have received, and certificate log maintainers did follow the latest mappings specified in the mapping log.
These checks can be easily done by a trusted monitor. However, since we aim to provide a TTP-free system, DTKI uses a crowdsourcing-like method, based on random checking, to monitor the correctness of the public log. The basic idea of random checking is that each user randomly selects a record in the log, and verifies whether the request and data in this record have been correctly managed. If all records are verified, the entire log is verified. Users only need to run the random checking periodically (e.g. once a day). The full version (with formalisation) of random checking can be found in our technical report. We give a flavour here by providing an example. Example 1 presents the random checking process to verify the correct behaviour of the mapping log.
Figure 4: The protocol for verifying a certificate with the corresponding certificate log maintainer.
Example 1. Suppose verifier has randomly selected the k th record of the mapping log, and the record has the form h add ( ( rgx, id , t ) k , dg s k , dg bl k , dg r k , dg i k ) . The verifier must check that all digests in this record are updated from the ( k -1) th record by adding a new mapping ( rgx, id ) in the mapping log at time t k .
Let the label of the ( k -1) th record be h ( req k -1 , t k -1 , dg s k -1 , dg bl k -1 , dg r k -1 , dg i k -1 ) , then to verify the correctness of this record, the verifier should run the following process:
- · verify that dg s k = dg s k -1 and dg bl k = dg bl k -1 ; and
- · verify that ( id, dg irgx k ) is the result of replacing ( id, dg irgx k -1 ) in dg i k -1 by ( id, dg irgx k ) by using VerifPoM O ; and
- · verify that dg r k is the result of adding ( rgx, id ) into dg r k -1 by using VerifPoAdd O , and id is an instance of rgx ; and
- · verify that dg irgx k is the result of adding rgx into dg irgx k -1 by using VerifPoAdd O .
Note that all proofs required in the above are given by the log maintainer. If the above tests succeed, then the mapping log maintainer has behaved correctly for this record.
## 4.4 Performance Evaluation
In this section, we measure the cost of different protocols in DTKI.
Assumptions : We assume that the size of a certificate log is 10 8 (the total number of registered domain names currently is 2 71 . × 10 8 [38], though only a fraction of them
have certificates). In addition, we assume that the number of stored regular expressions, the number of certificate logs, and the size of the mapping log are 1000 each. (In fact, if we assume a different number or size (e.g. 100 or 10000) for them, it makes almost no difference to the conclusion). Moreover, in the certificate log, we assume that the size of the set of data represented by dg rgx is 10, by dg id is 10 5 , by dg a is 10, and by dg rv is 100. These assumptions are based on the fact that dg rgx represents the set of regular expressions maintained by a certificate log; the dg id represents the set of domains which is an instance of a regular expression; and dg a and dg rv represent the set of currently valid certificates and the revoked certificates, respectively. Furthermore, we assume that the size of a certificate is 1.5 KB, the size of a signature is 256 bytes, the length of a regular expression and an identity is 20 bytes each, and the size of a digest is 32 bytes.
Space : Based on these assumptions, the approximate size of the transmitted data in the protocol for publishing a certificate is 4 KB, for requesting a mapping is 3 KB, and for verifying a certificate is 5 KB. Since the protocols for publishing a certificate and requesting a mapping are run occasionally, we mainly focus on the cost of the protocol for verifying a certificate, which is required to be run between a log server and a web browser in each secure connection.
By using Wireshark, we 3 measure that the size of data for establishing an HTTPS protocol to log-in to the internet bank of HSBC, Bank of America, and Citibank are 647.1 KB, 419.9 KB, and 697.5 KB, respectively. If we consider the average size ( ≈ 588 KB) of data for these three HTTPS connections, and the average size ( ≈ 6 KB) of data for their corresponding TLS establishment connections, we have that in each connection, DTKI incurs 83 % overhead on the cost of the TLS protocol. However, since the total overhead of an HTTPS connection is around 588 KB, so the cost of DTKI only adds 0.9 % overhead to each HTTPS connection, which we consider acceptable.
Time : Our implementation uses a SHA-256 hash value as the digest of a log and a 2048 bit RSA signature scheme. The time to compute a hash 4 is ≈ 0 01 . millisecond (ms) per 1KB of input, and the time to verify a 2048 bit RSA signature is 0.48 ms. The approximate verification time on the user side needed in the protocol for verifying certificates is 0.5 ms.
Hence, on the user side, the computational cost on the protocol for verifying certificates incurs 83 % on the size of data for establishing a TLS protocol, and 0.9 % on the size of data for establishing an HTTPS protocol; the verification time on the protocol for verifying certificates is 1.25 % of the time for establishing a TLS session (which is approximately 40 ms measured with Wireshark on the TLS connection to HSBC bank).
## 5 SECURITY ANALYSIS
We consider an adversary who can compromise the private key of all infrastructure servers in DTKI. In other words,
3. We use a MacBook Air 1.8 GHz Intel Core i5, 8 GB 1600 MHz DDR3.
4. SHA-256 on 64 byte size block.
the adversary can collude with all log servers and certificate authorities to launch attacks.
Main result : Our security analysis shows that
- · if the distributed random checking has verified all required tests, and domain owners have successfully verified their initial master certificates, then DTKI can prevent attacks from the adversary; and
- · if the distributed random checking has not completed all required tests, or domain owners have not successfully verified their initial master certificates, then an adversary can launch attacks, but the attacks will be detected afterwards.
We provide all source codes and files required to understand and reproduce our security analysis at [39]. In particular, these include the complete DTKI models and the verified proofs.
## 5.1 Formal analysis
We analyse the main security properties of the DTKI protocol using the TAMARIN prover [30]. The TAMARIN prover is a symbolic analysis tool that can prove properties of security protocols for an unbounded number of instances and supports reasoning about protocols with mutable global state, which makes it suitable for our log-based protocol. Protocols are specified using multiset rewriting rules, and properties are expressed in a guarded fragment of first order logic that allows quantification over timepoints.
TAMARIN is capable of automatic verification in many cases, and it also supports interactive verification by manual traversal of the proof tree. If the tool terminates without finding a proof, it returns a counter-example. Counterexamples are given as so-called dependency graphs, which are partially ordered sets of rule instances that represent a set of executions that violate the property. Counterexamples can be used to refine the model, and give feedback to the implementer and designer.
Modeling aspects : We used several abstractions during modeling. We model our log as lists, similar to the abstraction used in [29], [40]. We also assume that the random checking is verified.
We model the protocol roles D (domain server), M (mapping log maintainer), C (certificate log maintainer), and CA (certificate authority) by a set of rewrite rules. Each rewrite rule typically models receiving a message, taking an appropriate action, and sending a response message. Our modeling approach is similar to the one used in most TAMARIN models. Our modeling of the roles directly corresponds to the protocol descriptions in the previous sections. TAMARIN provides built-in support for a Dolev-Yao style network attacker, i.e., one who is in full control of the network. We additionally specify rules that enable the attacker to compromise service providers, namely the mapping log maintainer, certificate log maintainers and CAs, learn their secrets, and modify public logs.
Our final DTKI model (available from [39]) consists of 959 lines for the base model and five main property specifications, examples of which we will give below.
Proof goals : We state several proof goals for our model, exactly as specified in TAMARIN's syntax. Since TAMARIN's property specification language is a fragment of first-order logic, it contains logical connectives ( | , & , ==> not , , ...) and quantifiers ( All Ex , ). In TAMARIN, proof goals are marked as lemma . The # -prefix is used to denote timepoints, and ' E @ #i ' expresses that the event E occurs at timepoint i .
The first goal is a check for executability that ensures that our model allows for the successful transmission of a message. It is encoded in the following way.
```
//
//
// /
} /*
}
}
/*
```
"
The property holds if the TAMARIN model exhibits a behaviour in which a domain server received a message without the attacker compromising any service providers. This property mainly serves as a sanity check on the model. If it did not hold, it would mean our model does not model the normal (honest) message flow, which could indicate a flaw in the model. TAMARIN automatically proves this property in several minutes and generates the expected trace in the form of a graphical representation of the rule instantiations and the message flow.
We additionally proved several other sanity-checking properties to minimize the risk of modeling errors.
The second example goal is a secrecy property with respect to a classical attacker, and expresses that when no service provider is compromised, the attacker cannot learn the message exchanged between a user and a domain server. Note that K(m) is a special event that denotes that the attacker knows m at this time.
```
Note that K(m) is a special event that denotes that the
```
```
==>
( /* then the adversary cannot know m */
not (Ex #i5. K(m) @ #i5)
)
"
```
"
TAMARIN proves this property automatically (in 575 steps).
The above result implies that if a domain server D, whose domain name is Did such that Did is an instance of regular expression rgx, receives a message that was sent by a user, and the attacker did not compromise server providers, then the attacker will not learn the message.
The next two properties encode the unique security guarantees provided by our protocol, in the case that even all service providers are compromised.
The first main property we prove is that when all service providers (i.e. CAs, the MLM, and CLMs) are compromised, and the domain owner has successfully verified his master certificate in the log, then the attacker cannot learn the message exchanged between a user and a domain owner. It is proven automatically by TAMARIN in 5369 steps.
lemma
```
lemma message_secrecy_compromise_all_domain_verified_master_cert:
"
All D Did m rgx ltpkD stpkD #1.
```
/* The user has sent an encrypted message aenc{m}stpkD to domain server D whose identity is Did and TLS key is stpk, and the user received from D a confirmation h(m) of receipt. */
```
received from D a confirmation h(m) of receipt. */
(Com_Done(D, Did, m, rgx, ltpkD, stpkD) @ #i1
/* and at an earlier time, the domain server has verified his
master certificate */
& Ex #i2.
DerivedMasterCert(D, Did, rgx, ltpkD) @ #i2
& #i2 < #i1
)
==>
( /* then the adversary cannot know m */
this
not (Ex #i3. K(m) @ #i3)
)
==>
```
"
The property states that if a domain server D receives a message that was sent by a user, and at an earlier time, the domain server has verified his master certificate, then even if the attacker can compromise all server providers, the attacker cannot learn the message.
The final property states that when all service providers can be compromised, and a domain owner has not verified his/her master certificate, and the attacker learns the message exchanged between a user and the domain owner, then afterwards the domain owner can detect this attack by checking the log. It is also verified by TAMARIN within a few minutes.
lemma detect\_bad\_records\_in\_the\_log\_when\_master\_cert\_not\_verified:
```
stpkD lemma detect_bad_records_in_the_log_when_master_cert_not_verified:
LS key
mation "
All D Did m rgx ltpkD flag stpkD #i1 #i2 #i3.
/* The user has sent an encrypted message aenc(m)stpkD to domain
server D whose identity is Did and TLS key is stpk, and the user
received from D a confirmation h(m) of receipt. */
(Com_Done(D, Did, m, rgx, ltpkD, stpkD) @ #i1
/* and the adversary knows m */
& K(m) @ #i2
/* and we afterwards check the log */
& CheckedLog(D, Did, rgx, ltpkD, flag, stpkD) @ #i3
```
==>
( /* then we can detect a fake record in the log */ (flag = 'bad')
)
"
## 6 COMPARISON
As mentioned previously, DTKI builds upon a wealth of ideas from SK [23], CT [26], CIRT [28], and AKI [27]. Figure 5 shows the dimensions along which DTKI aims to improve on those systems.
Compared with CT, DTKI supports revocation by enabling log providers to offer proofs of absence and currency of certificates. In CT, there is no mechanism for revocation. CT has proposed additional data structures to hold revoked certificates, and those data structures support proofs of their contents [41]. However, there is no mechanism to ensure that the data structures are maintained correctly in time.
Compared to CIRT, DTKI extends the log structure of CIRT to make it suitable for multiple log maintainers, and provides a stronger security guarantee as it prevents attacks rather than merely detecting them. In addition, the presence of the mapping log maintainer and multiple certificate log maintainers create some extra monitoring work. DTKI solves it by using a detailed crowd-sourcing verification system to distribute the monitoring work to all users' browsers.
Compared to AKI and ARPKI, in DTKI the log providers can give proof that the log is maintained append-only from one step to the next. The data structure in A(RP)KI does not allow this, and therefore they cannot give a verifiable guarantee to the clients that no data is removed from the log.
DTKI improves the support that CT and A(RP)KI have for multiple log providers. In CT and AKI, domain owners wishing to check if there exists a log provider that has registered a certificate for him has to check all the log providers, and therefore the full set of log providers has to be fixed and well-known. This prevents new log providers being flexibly created, creating an oligopoly. In contrast, DTKI requires the browsers only to have the MLM public key built-in, minimising the oligopoly element.
In DTKI, trusted monitors are optional, as it uses crowdsourced verification. More precisely, a trusted monitor's verification work can be done probabilistically in small pieces by users' browsers.
Unlike the mentioned previous work, DTKI allows the possibility that all service providers (i.e. the MLM, CLMs, and mirrors) to collude together, and can still prevent attacks. In contrast, SK and CT can only detect attacks, and to prevent attacks, A(RP)KI requires that not all service providers collude together. Similar to A(RP)KI, DTKI also assumes that the domain is initially registered by an honest party to prevent attacks, otherwise A(RP)KI and DTKI can only detect attacks.
## 7 DISCUSSION
Responding to incorrect proofs : How should the browser (and the user) respond if a received proof (e.g., a proof of presence in the log) is incorrect? Such situations should be handled in the background by the software in the browser that verifies proofs, and be sent to domain owners for further investigation. The browser can also present errors to the user in the same way as the current state of the art. So, the user interface will remain the same. For example, a user might be shown two options, i.e. either to continue anyway, or not to trust the certificate and abort this connection. Another possible way is to hard fail if the verification has not been successful, as suggested by Google certificate transparency. However, this might be an obstacle for deploying DTKI in early stages.
Coverage of random checking : As mentioned previously, several aspects of the logs are verified by user's browsers performing randomly-chosen checks. The number of things to be checked depends on the size of the mapping log and certificate logs. The size of the mapping log mainly depends on the number of certificate logs and the mapping from regular expressions to certificate logs; and the size of certificate logs mainly depends on the number of domain servers that have a TLS certificate. Currently, there are 2 71 . × 10 8 domains [38] (though not every domain has a certificate), and 3 × 10 9 internet users [42]. The probability of a given domain not being checked on a given day (or week) is (1 -1 2 71 . × 10 8 ) 3 × 10 9 ≈ 1 56 . × 10 -5 (resp. ((1 -1 2 71 . × 10 8 ) 3 × 10 9 ) 7 ≈ 2 25 . × 10 -34 ). Thus, the expected number of unchecked domains per day (resp. per week) is 4 23 . × 10 3 (resp. 6 10 . × 10 -26 ).
Accountability of mis-behaving parties: The main goal of new certificate management schemes such as CT, CIRT, AKI, ARPKI and DTKI is to address the problem of mis-issued certificates, and to make the mis-behaving (trusted) parties accountable.
In DTKI, a domain owner can readily check for rogue certificates for his domain. First, he queries a mirror of the mapping log maintainer to find which certificate log maintainers (CLM) are allowed to log certificates for the domain (section 4). Then he examines the certificates for his domain that have been recorded by those CLMs. The responses he obtains from the mirror and the CLMs are accompanied by proofs. If he detects a mis-issued certificate, he requests revocation in the CLM. If that is refused, he can complain to the top-level domain, who in turn can request the MLM to change the CLM for his domain (after that, the offending CLM will no longer be consulted by browsers). This request should not be refused because the MLM is governed by an international panel. The intervening step, of complaining to the top-level domain, reflects the way domain names are actually managed in practice. Different top-level domains have different terms and conditions, and domain owners take them into account when purchasing domain names. In DTKI, log maintainers are held accountable because they sign and time-stamp their outputs. If a certificate log maintainer issues an inconsistent digest, this fact will be detected and the log maintainer can be blamed and blacklisted. If the mapping log misbehaved, then its governing panel must meet and resolve the situation.
In certificate transparency, this process is not as smooth. Firstly, the domain owner doesn't get proof that the list of issued certificates is complete; he needs to rely on monitors and auditors. Next, the process for raising complaints with log maintainers who refuse revocation requests is less clear
| | SK [23] | CT [26] | AKI [27] | ARPKI [29] | DTKI |
|-----------------------------------------------------------------------------------------------------------------|---------------------|-----------------------|----------------------------|----------------------------|-----------------------------------------------|
| Terminology | | | | | |
| Log provider | Time-line server | Log | Integrity log server (ILS) | Integrity log server (ILS) | Certificate/Mapping log maintainer (CLM, MLM) |
| Log extension | - | Log consistency | - | - | Log extension |
| Trusted party | Mirror | Auditor & monitor | Validator | Validator (optional) | - |
| Whether answers to queries rely on trusted parties or are accompanied by a proof | | | | | |
| Certificate-in-log query: Certificate-current-in-log query: Subject-absent-from-log query: Log extension query: | Rely Rely Rely Rely | Proof Rely Rely Proof | Proof Proof Proof Rely | Proof Proof Proof Rely | Proof Proof Proof Proof |
| Non-necessity of trusted monitors | | | | | |
| The role of trusted monitors can be distributed to browsers | No | No | No + | No + | Yes |
| Trust assumptions | | | | | |
| Not all service providers collude together | Yes | Yes | Yes | Yes | No |
| Domain is initially registered by an honest party | No | No | Yes* | Yes* | Yes* |
| Security guarantee | | | | | |
| Attacks detection or prevention | Detection | Detection | Prevention | Prevention | Prevention |
| Oligopoly issues | | | | | |
| Log providers required to be built into browser (oligopoly) | Yes | Yes | Yes | Yes | Only MLM |
| Monitors required to be built into browser (oligopoly and trust non-agility) | Yes | No | Yes | Yes † | No |
+ The system limits the trust in each server by letting them to monitor each other's behaviour.
* Without the assumption, the security guarantee is detection rather than prevention.
† The trusted party is optional, if there is a trusted party, then the trusted party is required to be built into browser.
Figure 5: Comparison of log-based approaches to certificate management. Terminology helps compare the terminology used in the papers. How queries rely on trusted parties shows whether responses to browser queries come with proof of correctness or rely on the honesty of trusted parties. Necessity of trusted parties shows whether the TP role can be performed by browsers. Trust assumptions shows the assumption for the claimed security guarantee. Oligopoly issues shows the entities that browsers need to know about.
(indeed, the RFC [26] says that the question of what domain owners should do if they see an incorrect log entry is beyond scope of their document). In CT, a domain owner has no ability to dissociate himself from a log maintainer and use a different one.
AKI addresses this problem by saying that a log maintainer that refuses to unregister an entry will eventually lose credibility through a process managed by validators, and will be subsequently ignored. The details of this credibility management are not very clear, but it does not seem to offer an easy way for domain owners to control which log maintainers are relied on for their domain.
will only want to launch attacks that would not be caught. (Such an adversary model has been described by Franklin and Yung [43], Canetti and Ostrovsky [44], Hazay and Lindell [45], and Ryan [28]). In DTKI, if a CA attempts to publish a fake master certificate for some domain, it will have to leave evidence of its misbehaviour in the log, and the misbehaviour will eventually be detected by the genuine domain owner.
Master certificate concerns : One concern is that a CA might publish fake master certificates for domains that the CA doesn't own and are not yet registered. However, this problem is not likely to occur: CAs are businesses, they cannot afford the bad press from negative public opinion and they cannot afford the loss of reputation. Hence, they
Another concern is the assumption that the domain owners can securely handle their master keys. In practice, the domain owners might have problems looking after their master keys due to lack of awareness of good practices. This problem arises in any web PKI: it is assumed that domain owners can securely handle their TLS keys. Our system adds one more key (the master key) to that requirement. A possible practical solution for domain owners is to use a trustworthy service to handle TLS keys (and the master key); the details are beyond the scope of the paper.
Avoidance of oligopoly : As we mentioned in the introduction, the predecessors (SK, CT, CIRT, AKI, ARPKI) of DTKI do not solve a foundational issue, namely oligopoly . These proposals require that all browser vendors agree on a fixed list of log maintainers and/or validators, and build it into their browsers. This means there will be a large barrier to create a new log maintainer.
CT has some support for multiple logs, but it doesn't have any method to allocate different domains to different logs. In CT, when a domain owner wants to check whether mis-issued certificates are recorded in logs, he needs to contact all existing logs, and download all certificates in each of the logs, because there is no way to prove to the domain owner that no certificates for his domain is in the log, or to prove that the log maintainer has showed all certificates in the log for his domain to him. Thus, to be able to detect fake certificates, CT has to keep a very small number of log maintainers. This prevents new log providers being flexibly created, creating an oligopoly.
In contrast to its predecessors, DTKI does not have a fixed set of certificate log maintainers (CLMs) to manage certificates for domain owners, and it allows operations of adding or removing a certificate log maintainer by updating the mapping log. In DTKI, the public log of the MLM is the only thing that browsers need to know.
The MLM may be thought to represent a monopoly; to the extent that it does, it is likely to be a much weaker monopoly than the oligopoly of CAs or log maintainers. CAs and log maintainers offer commercial services and compete with each other, by offering different levels of service at different price points in different markets. The MLM should not offer any commercial services; it should perform a purely administrative role, and is not required to be trusted because it behaves fully transparently and does not manage any certificates for web domains. In addition, the MLM is expected to be operated by an international panel with a lot of members.
In practice, we expect ICANN to be the MLM, as it is responsible for coordinating name-spaces of the Internet, and is governed by a Governmental Advisory Committee containing representatives from 111 states. However, there might be concerns here, including the concern that ICANN might not be interested in being the MLM, due to the fact that the service won't generate any revenue. Our solution does not address political issues around making decisions of whether to add or remove some CLMs or not.
## 8 CONCLUSIONS AND FUTURE WORK
Sovereign keys (SK), certificate transparency (CT), accountable key infrastructure (AKI), certificate issuance and revocation transparency (CIRT), and attack resilient PKI (ARPKI) are recent proposals to make public key certificate authorities more transparent and verifiable, by using public logs. CT is currently being implemented in servers and browsers. Google is building a certificate transparency log containing all the current known certificates, and is integrating verification of proofs from the log into the Chrome web browser.
Unfortunately, as it currently stands, CT risks creating an oligopoly of log maintainers (as discussed in section 7), of which Google itself will be a principal one. Therefore, adoption of CT risks investing more power about the way the internet is run in a company that arguably already has too much power.
In this paper we proposed DTKI - a transparent public key validation system using an improved construction of public logs. DTKI can prevent attacks based on mis-issued certificates, and minimises undesirable oligopoly situations by using the mapping log. In addition, we formalised the public log structure and its implementation; such formalisation work was missing in the previous systems (i.e. SK, CT, A(RP)KI, and CIRT). Since devising new security protocols is notoriously error-prone, we provide a formalisation of DTKI, and formally proved its security properties by using Tamarin prover.
## ACKNOWLEDGMENTS
This work was supported in part by The University of Birmingham under EPSRC Grant EP/H005501/1. We thank anonymous reviewers for their feedback.
## REFERENCES
- [1] Eckersley, P. (2011). Iranian hackers obtain fraudulent HTTPS certificates: How close to a web security meltdown did we get? Electronic Frontier Foundation.
- [2] Leyden, J. (2012). Trustwave admits crafting SSL snooping certificate: Allowing bosses to spy on staff was wrong, says security biz. The Register.
- [3] MS01-017. Erroneous Verisign-issued digital certificates pose spoofing hazard. Microsoft Support.
- [4] Roberts, P. (2011). Phony SSL certificates issued for Google, Yahoo, Skype, others. Threat Post.
- [5] Sterling, T. (2011). Second firm warns of concern after Dutch hack. Yahoo! News.
- [6] Falliere, N., Murchu, L. O., and Chien, E. (2011). W32.stuxnet dossier. Symantec Corporation.
- [7] Appelbaum, J. (2011). Detecting certificate authority compromises and web browser collusion. Tor Blog.
- [8] DigiNotar public report (2012). Black tulip report of the investigation into the DigiNotar certificate authority breach. Fox-IT.
- [9] Arthur, C. (2011). Rogue web certificate could have been used to attack Iran dissidents. The Guardian.
- [10] Clark, J. and van Oorschot, P. C. (2013) SoK: SSL and HTTPS:revisiting past challenges and evaluating certificate trust model enhancements. Proceedings of the IEEE Symposium on Security and Privacy 2013 , Berkeley, CA, USA, 19-22 May, pp. 511-525. IEEE Computer Society.
- [11] Langley, A. (2011). Public-key pinning. ImperialViolet.
- [12] Marlinspike, M. and Perrin, T. (2012). Trust assertions for certificate keys (TACK). Internet draft.
- [13] Wendlandt, D., Andersen, D. G., and Perrig, A. (2008) Perspectives: improving SSH-style host authentication with multi-path probing. Proceedings of the USENIX Annual Technical Conference , Boston, MA, USA, 22-27 June, pp. 321-334. USENIX Association, CA, USA.
- [14] Eckersley, P. and Burns, J. (2010) Is the SSLiverse a safe place? Chaos Communication Congress.
- [15] Alicherry, M. and Keromytis, A. D. (2009) Doublecheck: Multipath verification against man-in-the-middle attacks. Proceedings of the IEEE Symposium on Computers and Communications (ISCC) 2009 , Sousse, Tunisia, 5-8 July, pp. 557-563. IEEE Computer Society.
- [16] TR-12-015 (2012). Revisiting SSL: A large-scale study of the internet's most trusted protocol. The International Computer Science Institute, UC Berkeley.
- [17] The EFF SSL Observatory. https://www.eff.org/observatory.
- [18] Certificate patrol. http://patrol.psyced.org.
- [19] Soghoian, C. and Stamm, S. (2011). Certified lies: Detecting and defeating government interception attacks against SSL. Proceedings of the Financial Cryptography 2011 , Gros Islet, St. Lucia, 28 February - 4 March, pp. 250-259. Springer, Berlin.
- [20] Marlinspike, M. (2011) SSL and the future of authenticity. Black Hat, USA.
- [21] Rivest, R. L. (1998) Can we eliminate certificate revocation lists? Proceedings of the Financial Cryptography 1998 , Anguilla, British West Indies, 23-25 February, pp. 178-183. Springer, Berlin.
- [22] Langley, A. (2012). Revocation checking and Chrome's CRL. ImperialViolet.
- [23] Eckersley, P. (2012). Sovereign key cryptography for internet domains. Internet Draft.
- [24] Dierks, T. and Rescorla, E. (2008). The transport layer security (TLS) protocol version 1.2. RFC 5246 (Proposed Standard). Updated by RFCs 5746, 5878, 6176.
- [25] Turner, S. and Polk, T. (2011). Prohibiting secure sockets layer (SSL) version 2.0. RFC 6176 (Proposed Standard).
- [26] Laurie, B., Langley, A., and Kasper, E. (2013). Certificate Transparency. RFC 6962 (Experimental).
- [27] Kim, T. H.-J., Huang, L.-S., Perrig, A., Jackson, C., and Gligor, V. (2013) Accountable key infrastructure (AKI): A proposal for a public-key validation infrastructure. Proceedings of the 22nd International World Wide Web Conference , Rio de Janeiro, Brazil, 13-17 May, pp. 679-690. ACM International World Wide Web Conferences Steering Committee.
- [28] Ryan, M. (2014) Enhanced certificate transparency and end-toend encrypted mail. Proceedings of the 21st Annual Network and Distributed System Security Symposium, NDSS 2014 , San Diego, California, USA, 23-26 February, The Internet Society.
- [29] Basin, D., Cremers, C., Kim, T. H.-J., Perrig, A., Sasse, R., and Szalachowski, P. (2014) ARPKI: Attack resilient public-key infrastructure. Proceedings of the 2014 ACM SIGSAC Conference on Computer and Communications Security , Scottsdale, AZ, USA, 3-7 November, pp. 382-393. ACM.
- [30] Meier, S., Schmidt, B., Cremers, C., and Basin, D. (2013) The TAMARIN Prover for the Symbolic Analysis of Security Protocols. Proceedings of the Computer Aided Verification 2013 , Saint Petersburg, Russia, 13-19 July, pp. 696-701. Springer, Berlin.
- [31] Laurie, B. and Kasper, E. (2012) Revocation transparency. Google Research.
- [32] Yu, J., Cheval, V., and Ryan, M. (2015) Technical Report of DTKI. www.jiangshanyu.com/doc/paper/DTKI-report.pdf.
- [33] Jelasity, M., Voulgaris, S., Guerraoui, R., Kermarrec, A.-M., and Van Steen, M. (2007) Gossip-based peer sampling. ACM Transactions on Computer Systems , 25 , 8.
- [34] Merkle, R. C. (1987) A digital signature based on a conventional encryption function. Proceedings of CRYPTO 1987 , Santa Barbara, California, USA, 16-20 August, pp. 369-378. Springer, Berlin.
- [35] Anagnostopoulos, A., Goodrich, M. T., and Tamassia, R. (2001) Persistent authenticated dictionaries and their applications. Proceedings of Information Security Conference 2001 , Malaga, Spain, 1-3 October, pp. 379-393. Springer-Verlag, Berlin.
- [36] CS.CR/0302010 (2003) Authenticated append-only skip lists. CoRR.
- [37] Yumerefendi, A. R. and Chase, J. S. (2007) Strong accountability for network storage. ACM Transactions on Storage , 3 , 11.
- [38] Domain name industry brief. Verisign. http://www.verisign. com/en US/innovation/dnib/index.xhtml (2015)
- [39] Yu, J. (2015). Tamarin models for the DTKI protocol. http://www. jiangshanyu.com/doc/paper/DTKI-proof.zip.
- [40] Yu, J., Ryan, M., and Cremers, C. (2015) How to detect unauthorised usage of a key. IACR Cryptology ePrint Archive, 486.
- [41] Laurie, B. and Kasper, E. (2012). Revocation transparency. Google Research.
- [42] Internet users. Internet Usage Statistics. http://www. internetlivestats.com/internet-users/ (2015)
- [43] Franklin, M. K. and Yung, M. (1992) Communication complexity of secure computation (extended abstract). Proceedings of the STOC 1992 , Victoria, British Columbia, Canada, 4-6 May, pp. 699-710. ACM.
- [44] Canetti, R. and Ostrovsky, R. (1999) Secure computation with honest-looking parties: What if nobody is truly honest? (extended abstract). Proceedings of STOC 1999 , Atlanta, Georgia, USA, 1-4 May, pp. 255-264. ACM.
- [45] Hazay, C. and Lindell, Y. (2008) Efficient protocols for set intersection and pattern matching with security against malicious and covert adversaries. TCC 2008 , New York, USA, 19-21 March, pp. 155-175. Springer, Berlin.
## APPENDIX
## IMPLEMENTATION OF DATA STRUCTURES
This section shows the implementation of the chronological data structure and ordered data structure. We give some examples to show how the proofs could be done. Full details can be found in our technical report. We consider a secure hash function (e.g. SHA256), denoted h .
## Chronological data structure
The chronological data structure is implemented based on Merkle tree structure that we call ChronTree .
A ChronTree T is a binary tree whose nodes are labelled by bitstrings such that:
- · every non-leaf node in T has two children, and is labelled with h ( t glyph[lscript] , t r ) where t glyph[lscript] (resp. t r ) is the label of its left child (resp. right child); and
- · the subtree rooted by the left child of a node is perfect, and its height is greater than or equal to the height of the subtree rooted by the right child.
Here, a subtree is 'perfect' if its every non-leaf node has two children and all its leaves have the same depth.
Note that a ChronTree is a not necessarily a balanced tree. The two trees in Figure 6 are examples of ChronTrees where the data stored are the bitstrings denoted d , . . . , d 1 6 .
Figure 6: Example of two ChronTrees, T a and T b .
$$\begin{array} { c } \omega & & \\ \overbrace { w w w. } & \overbrace { \overbrace { \overbrace { d _ { 4 }, \, h ( d _ { 4 }, \, h ( d _ { 2 }, \, h _ { 1 }, \, h _ { 3 } ), \, h _ { 5 } ) } } & \overbrace { \overbrace { d _ { 8 }, \, h _ { 5 } } } & \overbrace { \overbrace { d _ { 8 }, \, h ( d _ { 8 }, \, h ( d _ { 9 } ) } } & \overbrace { \overbrace { d _ { 1 2 }, \, h ( d _ { 1 2 }, \, h _ { 1 1 }, \, h _ { 1 1 } ) } } \\ \intertext { t h o r i s e d } \text{Geogle } & \overbrace { \overbrace { d _ { 1 }, \, h _ { 1 } } } & \overbrace { \overbrace { d _ { 3 }, \, h _ { 3 } } } & \overbrace { \overbrace { \overbrace { d _ { 7 }, \, h _ { 7 } } } & \overbrace { \overbrace { d _ { 9 }, \, h _ { 9 } } } & \overbrace { \overbrace { d _ { 1 1 }, \, h _ { 1 1 } } } & \overbrace { \overbrace { d _ { 1 1 }, \, h _ { 1 1 } ) } \end{array}$$
Figure 7: An example of a LexTree T c , where h i = h ( d , i null , null ) for all i = { 1 3 5 7 9 11 , , , , , }
Given a ChronTree T with k leaves, we denote by S ( T ) = [ d , . . . , d 1 k ] the sequence of bitstrings stored in T . Note that a ChronTree is completely defined by the sequence of data stored in the leaves. Moreover, we say that the size of a ChronTree is the number its leaves.
Given a bitstring d and a ChronTree T , the proof of presence of d in T exists if there is a leaf n 1 in T labelled by d ; and is defined as ( w, b , . . . , b [ 1 k ]) such that:
- · w is the position in { glyph[lscript], r } ∗ of n 1 (that is, the sequence of left or right choices which lead from the root to n 1 ), and | w | = k ; and
- · if n , . . . , n 1 k +1 is the path from n 1 to the root, then for all i ∈ { 1 , . . . , k } , b i is the label of the sibling node of n i .
Intuitively, a proof of presence of d in T contains the minimum amount of information necessary to recompute the label of the root of T from the leaf containing d .
Example 2. Consider the ChronTree T b of Figure 6. The proof of presence of d 3 in T b is the tuple ( w,seq ) where:
- · w = glyph[lscript] · r · glyph[lscript]
- · seq = [ d , 4 h ( d , d 1 2 ) , h ( d , d 5 6 )]
Note that the size of the proof of presence is logarithmic in the size of the tree; even if the tree grows considerably, the size of the proof does not increase much.
Let T and T ′ be ChronTrees of size N and N ′ , respectively, such that N ′ ≤ N , S ( T ) = [ d , . . . , d 1 N ′ , . . . , d N ] , and S ( T ′ ) = [ d , . . . , d 1 N ′ ] for some bitstrings d 1 , . . . , d N ′ , . . . , d N . Let m be the smallest position of the bit 1 in the binary representation of N ′ ; and let ( d, w ) be the ( m +1) th node in the path of the node labelled by d N ′ to the root in T , where d is a bitstring and w ∈ { glyph[lscript], r } ∗ indicates the position. At last, let ( w,seq ′ ) be the proof of presence of d in T . The proof of extension of T ′ into T is defined as the sequence seq of bitstrings such that
- · if N ′ = 2 k for some k , then seq = seq ′ ; otherwise
- · seq = d :: seq ′ , where :: is the concatenation operation.
Example 3. The proof of extensions of T a into T b (Figure 6) is the sequence seq = [ d , d 3 4 , h ( d , d 1 2 ) , h ( d , d 5 6 )] .
While a proof of presence is the minimal amount of information necessary to recompute the hash value of a ChronTree from the leaf containing some particular data, the proof of extension is the minimal amount of information necessary to recompute the hash value of ChronTree T from the hash value of a ChronTree T ′ where T is an extension of T ′ . Intuitively, the proof of extension of a ChronTree T ′ into a ChronTree T is the proof of presence in T of the last inserted data of T ′ , i.e. d N ′ when S ( T ′ ) = [ d , . . . , d 1 N ′ ] . With this proof and the sizes of both trees, we can reconstruct the label of the root T but also the label of the root of T ′ as means to verify the proof of extension. Note that when N ′ = 2 k for some k , it implies that the tree T ′ is perfect and so the label of the root of T ′ is also a label of a node in T . Therefore, to reconstruct the label of the root of T , we only need a fragment of the proof of presence of d N ′ in T .
Example 4. Coming back to Example 3, consider the bitstrings h b = h h h ( ( ( d , d 1 2 ) , h ( d , d 3 4 )) , h ( d , d 5 6 )) and h a = h h ( ( d , d 1 2 ) , d 3 ) . seq proves the extension of h a of size 3 into h b of size 6 . Figure 6 is the graphical representation of the verification of seq given h a and h b . In particular, ( glyph[lscript] · r · glyph[lscript], [ d , 4 h ( d , d 1 2 ) , h ( d , d 5 6 )]) proves the presence of d 3 in h b and ( r, [ h ( d , d 1 2 )]) proves the presence of d 3 in h a .
## Ordered data structure
The ordered data structure is implemented as the combination of a binary search tree and a Merkle tree. The idea is that we can regroup all the information about a subject into a single node of the binary search tree, and while being able to efficiently generate and verify the proof of presence. We consider a total order on bitstrings denoted ≤ . This order could be the lexicographic order in the ASCII representations but it could be any other total order on bitstrings. The implementation is called LexTree .
A LexTree T is a binary search tree over pairs of bitstrings
- · for all two pairs ( d, h ) and ( d , h ′ ′ ) of bitstrings in T , ( d, h ) occurs in a node left of the occurrence of ( d , h ′ ′ ) if and only if d ≤ d ′ lexicographically;
- · for all nodes n ∈ T , n is labelled with the pair ( d, h ( d, h , h glyph[lscript] r )) where d is some bistring and ( d , h glyph[lscript] glyph[lscript] ) (resp. ( d , h r r ) ) is the label of its left child (resp. right child) if it exists; or the constant null otherwise.
Note that contrary to a ChronTree, the same set of data can be represented by different LexTrees depending on how the tree is balanced. To avoid this situation, we assume that there is a pre-agreed way for balancing trees.
Example 5. The tree in Figure 7 is an example of LexTree where d 1 ≤ d 2 ≤ . . . ≤ d 12 .
Example 6. Consider the LexTree T of Figure 7. The proof of presence of d 8 in T is the tuple ( h , h glyph[lscript] r , seq d , seq h ) where:
- · h glyph[lscript] = h 7 and h r = h 9 ; and
- · seq d = [ d 10 , d 6 ]
- · seq h = [ h ( d 12 , h 11 , null ) , h ( d , 4 h ( d , 2 h 1 , h 3 ) , h 5 )]
Like in ChronTrees, verifying the proof of presence of some data d in a LexTree T consists of reconstructing the hash value of the root of T .
Example 7. Consider the T c of Figure 7. Consider some data d such that d 7 ≤ d ≤ d 8 . The proof of absence of d in T c is the tuple ( null , null , seq d , seq h ) where:
- · seq d = [ d , d 7 8 , d 10 , d 6 ]
- · seq h = [ h 9 , h ( d 12 , h 11 , null ) , h ( d , 4 h ( d , 2 h 1 , h 3 ) , h 5 )]
|
10.1093/comjnl/bxw039
|
[
"Jiangshan Yu",
"Vincent Cheval",
"Mark Ryan"
] |
2014-08-05T16:32:12+00:00
|
2017-03-16T15:07:19+00:00
|
[
"cs.CR"
] |
DTKI: a new formalized PKI with no trusted parties
|
The security of public key validation protocols for web-based applications
has recently attracted attention because of weaknesses in the certificate
authority model, and consequent attacks.
Recent proposals using public logs have succeeded in making certificate
management more transparent and verifiable. However, those proposals involve a
fixed set of authorities. This means an oligopoly is created. Another problem
with current log-based system is their heavy reliance on trusted parties that
monitor the logs.
We propose a distributed transparent key infrastructure (DTKI), which greatly
reduces the oligopoly of service providers and allows verification of the
behaviour of trusted parties. In addition, this paper formalises the public log
data structure and provides a formal analysis of the security that DTKI
guarantees.
|
1408.1025v2
|
## Stable Throughput Region of Cognitive-Relay Networks with Imperfect Sensing and Finite Relaying Buffer
Ahmed M. Alaa
Abstract -In this letter, we obtain the stable throughput region for a cognitive relaying scheme with a finite relaying buffer and imperfect sensing. The analysis investigates the effect of the secondary user's finite relaying capabilities under different scenarios of primary, secondary and relaying links outages. Furthermore, we demonstrate the effect of miss detection and false alarm probabilities on the achievable throughput for the primary and secondary users.
## I. INTRODUCTION
The paradigm of cognitive radio promises a more efficient utilization of the scarce spectrum occupied by the nonpermanently active primary (licensed) users [7]. In a conventional cognitive spectrum sharing scheme, a secondary user (SU) senses the spectrum to check the activity of a primary user (PU). If the PU is idle, the SU opportunistically utilizes its allocated frequency bands for its own data transmission. Moreover, the SU can offer cooperative relaying for the PU packets if the PU link suffers from a high outage probability [1]. In such scheme, the SU receives the packets that are unsuccessfully decoded by the primary destination [8], increasing the availability of the spectrum for its own data transmission and increasing the probability of successful PU transmission.
In a practical system, the SU will sense the existence of the PU in an imperfect manner. This means that the SU operation would perturb the achievable rate of a PU with exclusive spectrum access. In [2], the stable throughput region was characterized for a shared channel where the SU randomly accesses the channel while relaying unsuccessfully transmitted PU data. However, the relaying queue was assumed to be an M/M/1 queue with infinite buffering capability. In [1], cooperative relaying was incorporated with the original notion for cognitive radio where the SU is allowed to access the channel only when the PU is idle. Again, the SU was assumed to buffer any amount of primary data. Similar analyses incorporated the physical layer parameters to determine the outage probabilities in [3] and [4]. A generalization for the cognitive scheme was studied in [5], where multiple secondary users exists, but all are equipped with unlimited relaying buffers.
In this work, we focus on the impact of the size of the relaying queue on the stable throughput region of cognitiverelay networks. The impact of relaying size is found to depend on the outage probabilities of primary and secondary links.
Fig. 1. Queue model for cognitive-relay scheme with finite relaying buffer.
Moreover, we characterize the impact of the false alarm and detection probabilities on the stable throughput region.
The rest of the paper is organized as follows: the system model is presented in section II. In section III, we derive the stable-throughput region. Numerical results are presented in section IV. Finally, we draw our conclusion in section V.
## II. SYSTEM MODEL
The system under study is a conventional cognitive system where the SU accesses the channel only when the PU is idle [1]. The queuing model of such system is demonstrated in Fig. 1. The primary transmitter S p is equipped with a primary queue Q p which buffers primary data arriving at a rate of λ p and departing at a rate of µ p . Similarly, the secondary transmitter S s is equipped with a secondary queue Q s with parameters λ s and µ s . Besides, the secondary users employs a seperate relaying queue Q ps that holds the relayed data arriving and leaving with rates of λ ps and µ ps respectively. The relaying queue has a limited length of m .
The system is assumed to be time slotted, with the slot duration accomodating one packet. All the arrival and departure rates specified are in packets/slot. The primary and secondary destinations D p and D s are assumed to send immediate acknowledgments (ACK) on a perfect channel upon successful reception. While relaying the PU data, the SU emulates a primary destination and sends an ACK to the PU allowing it to drop the packet of interest from Q p . The physical links S p -D p , S p -S s , S s -D p , and S s -D s are characterized by the outage probabilities P p,p out , P p,s out , P s,p out and P s,s out , respectively.
The SU senses the PU activity (current queue length) with a probability of detection P d and a false alarm probability of P f . We assume that simultaneous transmissions from both PU and SU leads to a collision that corrupts both packets. The PU
accesses the channel whenever its queue is not empty, whereas the SU accesses the channel only when the primary user has no packets in its queue.
## III. STABLE THROUGHPUT REGION
Our objective is to characterize the stable throughput region of the presented cognitive-relay scheme taking into consideration. Since the mean service rates at S s and S p depend on each others queue size, these queues are called interacting, and consequently the rates of the individual departure processes cannot be computed directly. In order to bypass this problem, we utilize the idea of stochastic dominance, which has been employed before in [6] to analyze interacting queues. In this setting, the SU is assumed to be always backlogged. In other words, we construct a dominant system where the SU transmits dummy packets whenever its queue is empty. Thus, all collision events in the original system occurs in the dominant system.
In the stochastic dominance approach, we first construct an appropriate dominant system, which is a modification of the original system, that ensures that the queue sizes in the dominant system are, at all times, at least as large as those of the original system. Thus, the stability region of the new system inner bounds that of the original system. Furthermore, in the new system the queues are decoupled and not interacting anymore, thereby permitting the characterization of the stability region.
We start by characterizing the service rate µ p (probability of successful transmission) perceived by the primary queue
/negationslash
$$\mu _ { p } = P _ { d } ( ( 1 - P _ { o u t } ^ { p, p } ) + P _ { o u t } ^ { p, p } ( 1 - P _ { o u t } ^ { p, s } ). P ( Q _ { p s } \neq m ) ), \ ( 1 ) \quad ( 1 - P ^ { p, } )$$
which corresponds to the probability that SU detects the PU and the PU channel is either not in outage, or outaged but with a PU to SU relaying in a non-full buffer. By defining the loading factor for the M/M/1/m relaying queue as ρ ps = λ ps µ ps , the probability of no overflow for the relaying buffer is
/negationslash
$$P ( Q _ { p s } \neq m ) = 1 - \frac { 1 - \rho _ { p s } } { 1 - \rho _ { p s } ^ { m + 1 } } \rho _ { p s } ^ { m }.$$
Thus, the primary user's service rate is given by
$$\mu _ { p } = P _ { d } ( ( 1 - P _ { o u t } ^ { p, p } ) + P _ { o u t } ^ { p, p } ( 1 - P _ { o u t } ^ { p, s } ) ( 1 - \frac { 1 - \rho _ { p s } } { 1 - \rho _ { p s } ^ { m + 1 } } \rho _ { p s } ^ { m } ). \ ( 2 ) \quad \text{and}$$
The arrival rate of the relayed packets depend on the outage of the primary channel, the probability of relaying buffer overflow and the probability of an empty PU buffer. The relayed data arrival rate can be given by:
/negationslash
$$\lambda _ { p s } = P ( Q _ { p s } \neq m ). P ( Q _ { p } \neq 0 ). P _ { d }. ( 1 - P _ { o u t } ^ { p, s } ). P _ { o u t } ^ { p, p },$$
/negationslash which reduces to
$$\lambda _ { p s } = ( 1 - \frac { 1 - \rho _ { p s } } { 1 - \rho _ { p s } ^ { m + 1 } } \rho _ { p s } ^ { m } ). \frac { \lambda _ { p } } { \mu _ { p } }. P _ { d }. ( 1 - P _ { o u t } ^ { p, s } ). P _ { o u t } ^ { p, p }. \quad ( 3 ) \quad \text{through}$$
Fig. 2. Stable throughput region with high outage on the primary link.
Note that to obtain the value of λ ps , we have to solve a transcendental equation. This results from the usage of an M/M/1/m queue rather than an infinite relaying queue. The relaying queue service rate depends on the probability of an empty primary queue, and the outage and false alarm probabilities as
$$\mu _ { p s } = ( 1 - \frac { \lambda _ { p } } { \mu _ { p } } ). P _ { d }. ( 1 - P _ { o u t } ^ { s, p } ). ( 1 - P _ { f } ). \quad ( 4 )$$
Similarly, the secondary user's service rate can be formulated as
$$\intertext { n e } _ { f \text{ the } } \mu _ { s } = ( \frac { 1 - \rho _ { p s } } { 1 - \rho _ { p s } ^ { m + 1 } } ). ( 1 - \frac { \lambda _ { p } } { \mu _ { p } } ). P _ { d }. ( 1 - P _ { o u t } ^ { s, s } ). ( 1 - P _ { f } ). \quad ( 5 )$$
The stable throughput region is obtained by applying the Loynes' theorem. For the case of the relaying queue, it is always stable because of its finite length. Based on Loynes' theorem, we have that λ <µ p p and λ <µ s s . By defining the following auxiliary parameters ψ = (1 -P f ) (1 . -P s,p out ) , η = (1 -P p,p out .P s,p out ) , and φ = (1 -P p,s out ) .P p,p out (1 -P p,p out .P p,s out ) . The stable throughput region is obtained by scanning ρ ps over the range [0, 1] and finding the resulting pair of ( λ p , λ s ) by solving the following transcendental equations:
$$\overline { \mu _ { p s } }, \\ \lambda _ { s } < & \frac { 1 - \frac { \phi \lambda _ { p } } { \psi ( 1 - \frac { \lambda _ { p } } { \mu _ { p } } ) } } { 1 - ( \frac { \phi \lambda _ { p } } { \psi ( 1 - \frac { \lambda _ { p } } { \mu _ { p } } ) } ) ^ { m + 1 } }. ( 1 - \frac { \lambda _ { p } } { \mu _ { p } } ). ( 1 - P _ { f } ). ( 1 - P _ { o u t } ^ { s, s } ),$$
$$\lambda _ { p } < P _ { d } ( ( 1 - P _ { o u t } ^ { p, p } ) + P _ { o u t } ^ { p, p } ( 1 - P _ { o u t } ^ { p, s } ) ( 1 - \frac { 1 - \rho _ { p s } } { 1 - \rho _ { p s } ^ { m + 1 } } \rho _ { p s } ^ { m } )$$
and
$$\text{ge} \quad \rho _ { p s } = \frac { \lambda _ { p } } { \mu _ { p } - \lambda _ { p } } \frac { ( 1 - \frac { 1 - \rho _ { p s } } { 1 - \rho _ { p s } ^ { m + 1 } } \rho _ { p s } ^ { m } ). P _ { d }. ( 1 - P _ { o u t } ^ { p, s } ). P _ { o u t } ^ { p, p } } { ( 1 - P _ { o u t } ^ { s, p } ). ( 1 - P _ { f } ) }$$
## IV. NUMERICAL RESULTS
## V. IMPACT OF FINITE RELAYING BUFFER
An interesting fact is that a very small relaying buffer is sufficient to achieve the same throughput of an M/M/1 relaying queue. In Fig. 2, we plot the achievable rates with m = 1, 2 and 5. It is shown that all these buffer lengths, despite of being relatively small, are sufficient to expand the PU stable throughput region in a similar manner to the M/M/1 queue.
Fig. 3. Impact of imperfect sensing on the stable throughput region.
## VI. IMPACT OF IMPERFECT SENSING
Fig. 3 depicts the impact of the imperfect sensing on the stable throughput boundaries for both cooperative and noncooperative cases. We note that the imperfect probability of detection causes deterioration for the achievable primary user rate. This is due to the fact that miss detection causes sure collisions between primary ans secondary users. The deterioration is nearly proportional to the value of P d . On the other hand, the false alarm probability causes loss of both primary and secondary data rates. An interesting fact is that in case of no-relaying the false alarm probability does not affect the achievable rate for the PU. This is because when no cooperation, a false alarm will not cause a transmission opportunity loss to the PU but will cause a loss to the SU. However, when relaying is applied, a false alarm will prevent the SU from forwarding the relayed data and neither the PU nor the SU utilize the channel causing a degradation in both rates. Thus, we recommend using a Neyman-pearson test for sensing, where the false alarm probability is set to a certain value dictated by the desired secondary rate and maximizing the achievable PU rate.
## VII. CONCLUSION
In this letter, we obtained the stable throughput region for a cognitive relaying scheme with a finite relaying buffer and imperfect sensing. The analysis investigated the impact of the secondary user's finite relaying capabilities under different scenarios of primary, secondary and relaying links outages. Furthermore, we demonstrated the effect of miss detection and false alarm probabilities on the achievable throughput for the primary and secondary users. It was found that detection probability affects the PU achievable rates only, while false alarm probability can affect the SU rates if the SU offers relaying to the PU. Besides, it was shown that a relaying buffer that is as small as one packet sized can achieve almost the same rates of an infinite M/M/1 queuing buffer.
## REFERENCES
- [1] O. Simeone, Y. Bar-Ness, and U. Spagnolini: 'Stable Throughput of Cognitive Radios With and Without Relaying Capability,' IEEE Trans.Commun. , 2007, 55 , pp. 2351-2360.
- [2] S. Kompellay, G. D. Nguyeny, J. E. Wieselthierz and A .Ephremides: 'Stable Throughput Tradeoffs in Cognitive Shared Channels with Cooperative Relaying,' in Proc. IEEE INFOCOM 2011, pp. 1961-1969, Shanghai, China, Nov. 2011.
- [3] M. Elsaadany, M. Abdallah, T. Khattab, M. Khairy and M. Hasna: 'Cognitive Relaying in Wireless Sensor Networks: Performance Analysis and Optimization,' in Proc. IEEE GLOBECOM 2010 , pp. 411-3415, Miami, Florida, May 2010.
- [4] A. K. Sadek, K. J. Ray Liu, and A. Ephremides: 'Cognitive Multiple Access Via Cooperation: Protocol Design and Performance Analysis,' IEEE Trans. Inf. Theory , 2007, 53 , pp. 3677-3696.
- [5] J. Jose, S. Vishwanath,and L. Ying: 'Queue-Architecture and Stability Analysis in Cooperative Relay Networks,' IEEE J. Sel. Areas Commun , 2012, 30 , pp. 1754-1761.
- [6] R. Rao and A. Ephermides: 'On the Stability of Interacting Queues in a Multiple-Access System,' IEEE Trans. Inf. Theory , 1988, 34 , pp. 918-930.
- [7] H. -Po Shiang and M. Van Der Schaar. 'Delay-sensitive resource management in multi-hop cognitive radio networks.' New Frontiers in Dynamic Spectrum Access Networks , 2008. DySPAN 2008. 3rd IEEE Symposium on, pp. 1-12. IEEE, 2008.
- [8] M. Van Der Schaar, D. S. Turaga, and R. Wong. 'Classification-based system for cross-layer optimized wireless video transmission.' IEEE Transactions on Multimedia, 2006, 5 , pp. 1082-1095.
| null |
[
"Ahmed M. Alaa"
] |
2014-08-05T16:37:44+00:00
|
2018-03-18T09:53:24+00:00
|
[
"cs.NI"
] |
Stable Throughput Region of Cognitive-Relay Networks with Imperfect Sensing and Finite Relaying Buffer
|
In this letter, we obtain the stable throughput region for a cognitive
relaying scheme with a finite relaying buffer and imperfect sensing. The
analysis investigates the effect of the secondary user's finite relaying
capabilities under different scenarios of primary, secondary and relaying links
outages. Furthermore, we demonstrate the effect of miss detection and false
alarm probabilities on the achievable throughput for the primary and secondary
users.
|
1408.1026v1
|
## Chapter 1 Pauli's Principle in Probe Microscopy
SP Jarvis , AM Sweetman , L Kantorovich , E McGlynn , and P Moriarty † † ‡ glyph[sharp] †
It appears to be one of the few places in physics where there is a rule which can be stated very simply, but for which no one has found a simple and easy explanation. The explanation is deep down in relativistic quantum mechanics. This probably means that we do not have a complete understanding of the fundamental principle involved. RP Feynman, The Feynman Lectures on Physics , Vol III, Chapter 4 (1964)
## Abstract
Exceptionally clear images of intramolecular structure can be attained in dynamic force microscopy through the combination of a passivated tip apex and operation in what has become known as the 'Pauli exclusion regime' of the tip-sample interaction. We discuss, from an experimentalist's perspective, a number of aspects of the exclusion principle which underpin this ability to achieve submolecular resolution. Our particular focus is on the origins, history, and interpretation of Pauli's principle in the context of interatomic and intermolecular interactions.
Keywords: dynamic force microscopy;non-contact atomic force microscopy; NC-AFM; Pauli exclusion principle; submolecular resolution; intramolecular; single molecule imaging
From Imaging and Manipulation of Adsorbates using Dynamic Force Microscopy , Vol V of Advances in Atom and Single Molecule Machines , Springer-Verlag; http://www.springer.com/ series/10425 . To be published late 2014.
† School of Physics and Astronomy, University of Nottingham, Nottingham NG7 2RD, UK
- ‡ Department of Physics, Kings College London, The Strand, London WC2R 2LS, UK
glyph[sharp] School of Physical Sciences, Dublin City University, Glasnevin, Dublin 9, Ireland
## 1.1 Intramolecular resolution via Pauli exclusion
In 2009 the results of a pioneering dynamic force microscopy (DFM ) experiment by Leo 1 Gross and co-workers at IBM Z¨rich were published[1] and revolutionised the field of u scanning probe microscopy. Gross et al. captured arguably the clearest real space images of a molecule achieved up to that point, resolving the 'textbook' structure of the molecular architecture. Two important experimental protocols enabled Gross et al. - and, subsequently, a number of other groups[2, 3, 4, 5, 6, 7, 8] (see Fig. 1.1 for examples) - to attain this exceptionally high resolution. First, the apex of the probe was functionalised (by picking up a molecule) to render it inert. This enabled the scanning probe to be placed extremely close to the adsorbed molecule of interest - so close that the second experimental protocol, namely the exploitation of electron repulsion via the Pauli exclusion principle 2 , played a key role in the imaging mechanism.
It is this second protocol which is the primary focus of this chapter. We'll discuss just how Pauli exclusion is exploited in state-of-the-art scanning probe microscopy, what pitfalls there might be in interpreting features in DFM images as arising directly from chemical bonds, and to what extent scanning probe measurements of tip-sample interactions provide deeper experimental insights into the exclusion principle itself. We should also stress right from the outset that although we concentrate on dynamic force microscopy throughout this chapter, prior to Gross et al. 's 2009 paper, Temirov, Tautz and co-workers had achieved unprecedented spatial resolution using a technique for which they coined the term scanning tunnelling hydrogen microscopy (STHM)[9, 10, 11, 12]. Both STHM and the type of DFM imaging introduced by Gross et al. [1] exploit Pauli exclusion as a means to acquire exceptionally high resolution. Before covering the exploitation of the exclusion principle in scanning probe microscopy, we'll consider a number of aspects of the fascinating history of Pauli's Ausschließungsregel [13] and outline some of the rich physics underpinning the principle.
## 1.2 A potted history of Pauli's exclusion principle
Michela Massimi has written an authoritative and engaging history of the Pauli exclusion principle (PEP)[13], which impressively combines clear explanations of the quantum and statistical physics underlying the PEP with engaging discussions of both the history and the philosophical ramifications of the principle. As Massimi points out in the preface to her book, her research on the origin and validation of the exclusion principle took almost ten years. For those readers interested in a comprehensive account of the 'evolution' of the PEP we therefore strongly recommend Masimi's book. Here we will limit ourselves to providing a brief summary of those aspects of the PEP which are of key significance for (ultra)high resolution scanning probe microscopy.
The origins of the exclusion principle lie, like so many aspects of quantum physics, in the interpretation of spectroscopic data. In particular, a series of so-called anomalies in the spectra of alkali and alkaline earth metals, and, arguably more importantly, the response of atomic spectra to the application of a magnetic field, i.e. the ('anomalous') Zeeman effect, became a major challenge to the Bohr-Sommerfeld theory of the electronic structure of atoms in the early 1920s. It was only with the introduction of what came to be
1 Although the term non-contact atomic force microscopy (NC-AFM) is widespread - to the extent that the major conference in the field is the annual International NC-AFM meeting - it is arguably something of a misnomer to label the technique 'non-contact' when it is now commonplace to operate in a regime where the probe is in contact with the sample. We will therefore use the term dynamic force microscopy throughout this chapter.
2 We shall return, in Sections 1.4 and 1.5, to a detailed discussion of whether or not it is appropriate to describe the effects of Pauli exclusion as a repulsive force.
Fig. 1.1 Imaging bonds via the Pauli exclusion principle (A) . Combination of schematic illustration and experimental data to demonstrate experimental protocol used to acquire submolecular resolution. The apex of the probe used in a dynamic force microscope is passivated (in this case with a CO molecule) and scanned across a pentacene molecule at a height where Pauli exclusion plays a key role in determining the tip-sample interaction. (B) Experimental frequency shift image for a pentacene molecule. [A and B taken from Gross et al.[1]. c American Assocation for the Advancement of Science (2009)]. © (C) Dynamic force microscope image of four 8-hydroxyquinoline molecules. Both intra- and intermolecular features are observed. (See Section 1.7). (D) Schematic diagram of molecular arrangement shown in (C) with the expected positions of hydrogen-bonds drawn as lines between the molecules. [C and D taken from Zhang et al.[3]. c American Assocation for the Advancement of Science (2012)]. © (E) High resolution image of a chain of oligo-(E)-1,1-bi(indenylidene) with associated structural model. Taken from Riss et al. [8]. c American © Chemical Society (2014). (F) DFM image of two different conformers of dibenzo[a,h]thianthrene on a NaCl/Cu(111) substrate with (lower panel) structural models of both conformers. Taken from Pavlicek et al.[6]. c American Physical Society (2012). © (G) Structural model of a naphthalenetetracarboxylic diimide (NTCDI) molecule and a DFM image of a hydrogen-bonded assembly of NTCDI molecules. From Sweetman et al.[4]. c Nature Publishing Group (2014). ©
known as electron spin - but which Pauli initially called simply the electron Zweideutigkeit ('twofoldness') - that the spectroscopic data could be reconciled with the theoretical predictions. The introduction of electron Zweideutigkeit [14] was followed very closely by Pauli's statement of the exclusion principle[15] (or, as it was known at the time, the exclusion rule). Pauli subsequently won the Nobel prize in 1945 for his discovery of the exclusion principle.
It is worth quoting directly from Pauli's Nobel lecture, given on Dec. 13 1946, as this provides key insights into the original formulation of the principle 'straight from the horse's mouth', as it were:
On the basis of my earlier results on the classification of spectral terms in a strong magnetic field the general formulation of the exclusion principle became clear to me. The fundamental idea can be stated in the following way:
The complicated numbers of electrons in closed subgroups are reduced to the simple number one if the division of the groups by giving the values of the four quantum numbers of an electron is carried so far that every degeneracy is removed. An entirely non-degenerate energy level is already closed , if it is occupied by a single electron; states in contradiction with this postulate have to be excluded.
Or, if we couch this in the lexicon of modern quantum mechanics, no two electrons can have the same values of n , l , m l , and m s (i.e. the principal, orbital angular momentum, magnetic, and spin quantum numbers). More succinctly, no two electrons can occupy the same quantum state. (The Pauli exclusion principle of course holds for all fermions (half-integer spin particles), not just electrons. We'll return to this point very soon).
Pauli's Zweideutigkeit is now of course known as particle spin but the inferred connection with the classical concept of a spinning object is unfortunately misleading. Indeed, Pauli himself switched from being firmly opposed to any connection between his Zweideutigkeit and spin, to a somewhat grudging acceptance of a link, and then, as his Nobel lecture highlights, back to a significant degree of scepticism about the value of any classical analogy:
On the other hand, my earlier doubts as well as the cautious expression 'classically non-describable two-valuedness' experienced a certain verification during later developments, since Bohr was able to show on the basis of wave mechanics that the electron spin ... must therefore be considered as an essentially quantum-mechanical property of the electron.
## 1.2.1 Particle statistics and the quantum identity crisis
Following hot on the heels of Pauli's publication of the exclusion principle, first Fermi[16, 17] and then Dirac[18] explored the quantum statistics of an ideal gas of particles which was subject to the exclusion principle. Dirac coined the term fermion to describe a particle subject to the Fermi-Dirac statistics he and Fermi derived; a fermion is therefore a particle which obeys the Pauli exclusion principle (and concomitantly is of half-integer spin). At the very heart of quantum statistics - and, indeed, of classical statistical mechanics - lies the issue of the distinguishability of particles 3 . A simple back-of-the-envelope argument based on the (in)distinguishability of particles can provide a helpful insight into the origin of the exclusion principle[19].
Before we introduce that back-of-the-envelope approach, however, it is first important to define just what it is we mean by indistinguishable particles. This, despite first appearances, is a far from trivial question to address and has been the subject of quite considerable debate and interest for many decades. De Muynck[20], Berry and Robbins[21], Ginsberg et al. [22] (see also Fleischhauer[23] for a very readable overview of Ginsberg
3 Long before the advent of quantum mechanics, the effect of considering indistinguishable vs distinguishable particles on the partition function for a system was known as the Gibbs paradox in classical thermodynamics/statistical mechanics
et al's work), Omar[24], and Dieks and co-workers[25, 26], amongst many others, have considered and explored the important issue of how indistinguishability and quantum statistics are intrinsically coupled. We shall not delve into the detailed arguments - be they physical, philosophical, or semantic in scope - and instead restrict ourselves to the following relatively simple, although certainly not 'universal', definitions. (It is also important to note that the condition for antisymmetry and the exclusion principle are not equivalent statements).
First, we draw a distinction between identical and indistinguishable particles. Identical particles are those which have the same intrinsic (or 'internal') properties (and the same values associated with those intrinsic properties), i.e. mass, charge, spin. So two electrons are identical to each other. And two protons, or two neutrons, are similarly identical to each other. But electrons are clearly not identical to protons, nor to neutrons. (We apologise for labouring the point to this extent but the terms 'identical' and 'indistiguishable' are often used interchangeably - including in many textbooks - and this has led to quite some confusion at times).
If we have a collection of identical particles then they are indistinguishable if we cannot separate them on the basis of their 'external' properties such as position or momentum. But classically it is possible to distinguish between identical particles (at least in principle): we can effectively 'label' individual identical particles on the basis of their positions or trajectories and distinguish them accordingly 4 . Quantum mechanically, however, the standard argument is that due to delocalisation we lose this ability to label particles on the basis of their trajectories and they then become indistinguishable.
But to what extent is this true? Are quantum particles indeed indistinguishable? One can find undergraduate-level descriptions of quantum statistics[30] which claim that quantum particles can in fact be distinguished on the basis of what might be called a 'Rayleigh criterion' for wavepackets: if two particles are separated by a distance greater than their de Broglie wavelength (i.e. such that the wavefunction overlap is minimal) then they are distinguishable on the basis of their respective positions. Versteegh and Dieks[27] invoke similar arguments about the spatial extent of wavepackets enabling identical quantum particles to be distinguished.
However, whether this is a valid condition for distinguishability is far from clear-cut. In his commentary on Ginsberg et al. 's work[22], Fleicschhauer[23] states the following:
In the quantum world, particles of the same kind are indistinguishable: the wavefunction that describes them is a superposition of every single particle of that kind occupying every allowed state. Strictly speaking, this means that we can't talk, for instance, about an electron on Earth without mentioning all the electrons on the Moon in the same breath.
Why might Fleicschhauer say this? 5 The answer is, from one perspective at least, rather straight-forward. The universal superposition to which Fleicschhauer refers arises because in reality we never have perfect confinement of particles: there is no such thing as the infinite potential well beloved of introductory quantum physics courses and there is therefore some finite (albeit extremely small) probability for tunnelling. Thus, in this
4 In a thought-provoking paper, Versteegh and Dieks[27] discuss the importance of the distinguishability of identical particles and what these means for classical thermodynamics and statistical mechanics, including the Gibbs paradox. We note, however, that there is a very important omission in the list of papers cited by Versteegh and Dieks, namely a paper by Edwin Jaynes[28] who makes the point, following a similar analysis by Pauli, that the classical thermodynamic definition of entropy as the integration of d Q T / over a reversible path is only introduced in the context of constant particle number. This means that there is always (ultimately, see Ehrenfest and Trkal[29]) an arbitrary integration function (not an integration constant, but a function of N ) that can be used to yield the desired extensivity of the entropy.
5 It is perhaps worth noting at this point that the 'interconnectedness' to which Fleicschhauer alludes in this quote, and its relevance (or not) to the Pauli exclusion principle, was the subject of a great deal of sometimes ill-tempered online debate following the BBC's broadcast of a popular science lecture on quantum mechanics by Brian Cox, which included a discussion of the PEP. Jon Butterworth's post for The Guardian[31] is a short, clear and entertaining discussion of the furore and the physics surrounding Cox's lecture.
sense an electron on the Earth is indeed indistinguishable from an electron on the Moon (or on Alpha Centauri).
But what really matters, of course, are the effects that this type of 'coupling' might have on experimental measurements. And for electrons separated by centimetres, let alone light years, those effects are, to put it mildly, utterly negligible. If we consider a 'double well' system for an electron on Earth and an electron on Alpha Centauri, the energy level splitting is unimaginably tiny (and beyond anything we could ever begin to hope to measure), and the time-scale for evolution of the quantum state exceeds the age of the universe.
So in any practical sense , position can indeed be used to distinguish quantum particles. This is why we can treat electrons in well-separated atoms as being distinguishable. In principle, the electrons are indeed described by a single multi-particle ('universal') wavefunction and are thus indistinguishable. In practice, however, the spatial extent of the particle wavepacket is such that we can treat electrons in atoms separated by distances much greater than their equilibrium bond length as distinguishable. Only when those atoms are brought together so that there is appreciable overlap of electronic wavefuctions, as in chemical bond formation or, as we shall discuss below, a dynamic force microscopy experiment, can one state that the electrons on each atom become indistinguishable.
Following this lengthy 'detour' on the topic of distinguishability vs indistinguishability, we are now finally at the point where we can return to a consideration of that back-of-an-envelope argument for the PEP which was mentioned above.
## 1.3 Statistics, symmetry, and spin
Let's take a system where identical quantum particles can't be distinguished from another. As the particles are indistinguishable then when we compute the probability density for the system, i.e. | Ψ | 2 , we must get the same answer regardless of how we arrange the particles, i.e. their spatial positions have no influence on the probability density. We'll consider a very simple system with just two particles whose positions are r 1 and r 2 and whose single particle wavefunctions are ψ 1 and ψ 2 respectively. If we cannot distinguish Particle 1 from Particle 2 then it's clear that
$$| \bar { \Psi } ( { \mathbf r } _ { 1 }, { \mathbf r } _ { 2 } ) | ^ { 2 } = | \bar { \Psi } ( { \mathbf r } _ { 2 }, { \mathbf r } _ { 1 } ) | ^ { 2 }$$
This means one of two things. Either
$$\Psi ( { \mathbf r } _ { 1 }, { \mathbf r } _ { 2 } ) = \Psi ( { \mathbf r } _ { 2 }, { \mathbf r } _ { 1 } )$$
or
$$\Psi ( { \mathbf r } _ { 1 }, { \mathbf r } _ { 2 } ) = - \Psi ( { \mathbf r } _ { 2 }, { \mathbf r } _ { 1 } )$$
To meet the condition imposed by Eqn. 1.2, we must have the following two-particle wavefunction:
$$\Psi ( { \mathbf r } _ { 1 }, { \mathbf r } _ { 2 } ) = \frac { 1 } { \sqrt { 2 } } \left ( \psi _ { 1 } ( { \mathbf r } _ { 1 } ) \psi _ { 2 } ( { \mathbf r } _ { 2 } ) + \psi _ { 2 } ( { \mathbf r } _ { 1 } ) \psi _ { 1 } ( { \mathbf r } _ { 2 } ) \right ) & & ( 1. 4 )$$
Or to satisfy Eqn. 1.3 we need the following:
$$\Psi ( { \mathbf r } _ { 1 }, { \mathbf r } _ { 2 } ) = \frac { 1 } { \sqrt { 2 } } \left ( \psi _ { 1 } ( { \mathbf r } _ { 1 } ) \psi _ { 2 } ( { \mathbf r } _ { 2 } ) - \psi _ { 2 } ( { \mathbf r } _ { 1 } ) \psi _ { 1 } ( { \mathbf r } _ { 2 } ) \right ) & & ( 1. 5 )$$
Eqn. 1.4 represents what is called the symmetric case, while Egn. 1.5 is termed the antisymmetric case 6 . The antisymmetric equation leads us to a simple, but exceptionally important, result - a result that is at the very core of how the universe behaves because it is ultimately responsible for the stability of matter[32, 33, 34]. Note what happens when we make ψ 1 = ψ 2 in Eqn. 1.5 (or, in other words, we put both particles in the same quantum state): the two-particle wavefunction, Ψ , vanishes . This is the essence of the Pauli exclusion principle: in the antisymmetric case, no two particles can exist in the same quantum state . (We should also stress that the exclusion principle is 7 not equivalent to the statement that fermions have antisymmetric wave functions. Rather, the exclusion principle follows from the antisymmetric character of fermions).
A rather remarkable observation is that only antisymmetric and symmetric wavefunctions are found in nature for fundamental particles, i.e. we only have bosons (symmetric state) and fermions (anti-symmetry). No other particles have been found that fall outside these symmetry classes 8 As Omar[24] points out in a comprehensive and very readable review of the ramifications of indistinguishability in quantum mechanics, this existence of only symmetric and antisymmetric states 9 is best described as a postulate (the 'symmetrization postulate'). And, disconcertingly, it's a postulate that apparently can't be deduced from the framework of quantum mechanics (either the non-relativistic or relativistic 'breeds' of the theory). In other words, we simply have to accept that only bosons and fermions exist (or, at least, we have no good experimental evidence to date for fundamental particles arising from other rather more exotic statistics/symmetries such as parastatistics (see Omar[24])). In this sense, we have progressed very little since Pauli voiced his misgivings about the origin of the exclusion principle almost seventy years ago:
I was unable to give a logical reason for the Exclusion Principle or to deduce it from more general assumptions... in the beginning I hoped that the new quantum mechanics would also rigorously deduce the Exclusion Principle.
## 1.3.1 Putting a spin on the story
All known fundamental particles are either bosons or fermions. (Within the Standard Model, fermions are 'matter' particles, whereas bosons are generally force 'carriers' 10 . Again, we are not including quasiparticles in the discussion.). All bosons have integer spin while fermions have half-integer spin. Clearly there must be a strong connection between spin and symmetry. Indeed, this is known as the spin-statistics theorem and holds not just for individual particles but composites of fundamental particles.
This link between spin, statistics, and the exclusion principle, however, very much appears not to be something that can be deduced from non-relativistic quantum mechanics.
6 The use of the terms symmetric and antisymmetric follows from Eqn. 1.2 (where Ψ is a symmetric function with respect to the exchange of coordinates) and Eqn 1.3 (where Ψ is an antisymmetric function). Note also that the factor of 1 √ 2 in Eqn. 1.4 and Eqn. 1.5 arises from normalisation of the wavefunction
7 We are neglecting explicit consideration of the spin contribution here - see Section 1.3.1. Moreover, we are making drastic simplifications regarding the treatment of many electron systems in order to put across the 'essence' of the exclusion principle. For example, equations 1.4 and 1.5 are approximations because, in reality, there are many more contributing terms (as in the Configuration Interaction method of quantum chemistry. See Kantorovich[35] for a summary.)
8 Note, however, that the key principle underlying the concept of supersymmetry is that bosons can be converted into fermions and vice versa. Supersymmetry therefore introduces a bosonic partner for every fermion (and, again, vice versa). To the chagrin of (some of) the particle physics community, however, any evidence for supersymmetry remains frustratingly elusive. Moreover, we are omitting any discussion of quasiparticles here. The results of measurements of two-dimensional systems exhibiting the fractional quantum Hall effect have been interpreted in terms of anyons[36], quasiparticles with mixed symmetry. 9
...for the total
wavefunction. Again, see Section 1.3.1.
10 ...although the Higgs boson is an honourable exception.
This is the origin of the statement from Feynman quoted at the start of this chapter the link between spin and the exclusion principle is 'deep down' in relativistic quantum mechanics. More recently, Bartalucci et al. [37] have put it like this:
Although the principle has been spectacularly confirmed by the number and accuracy of its predictions, its foundation lies deep in the structure of quantum field theory and has defied all attempts to produce a simple proof...
This means that within the non-relativistic quantum framework the spin-statisticssymmetry link is generally accepted as a dictum, although alternative non-relativistic approaches have certainly been explored[21]. (Duck and Sudarshan[38] detail a proof of the spin-statistics theorem which can be 'recast' in non-relativistic quantum field theory, but only if an aspect of their proof which stems from relativistic quantum theory (via Lorentz invariance) can be invoked as a postulate).
Notwithstanding its essential relativistic origin, the spin contribution can be incorporated into the particle wavefunction in non-relativistic quantum mechanics in a straightforward fashion via the introduction of the spin orbital. A spin orbital is a product of a spatial wavefunction (such as those described in the preceding section) and a spin function, which we can represent as χ ( ↑ ) or χ ( ↓ ) for the spin-up and spin-down states respectively. So, if we use x as a variable which incorporates both the spatial and spin coordinates, and we switch to using φ to represent only the spatial part (so that we can, as per convention, use ψ to represent the wavefunction), we have the following for the spin-up state of an electron:
$$\psi ( { \mathbf x } _ { 1 } ) = \phi ( { \mathbf r } _ { 1 } ) \chi ( \uparrow )$$
We therefore now have two options for ensuring antisymmetry in a two electron (or multielectron) system: either the spatial part or the spin part can lead to an antisymmetric total wavefunction, Ψ ( x 1 , x 2 ). In other words, if two electrons have opposite spin states then there is no constraint on the spatial wavefunction. But this is nothing more than the statement of the Pauli exclusion principle given earlier: no two electrons can exist in the same quantum state.
## 1.4 The origin of Pauli repulsion: A Gedankenexperiment
At short interatomic or intermolecular separations, Pauli repulsion 11 is much stronger than any electrostatic interaction, increasing very rapidly with decreasing distance between atoms or molecules. Recall, for example, that the Pauli repulsion term in the Lennard-Jones potential is modelled not with a 1 r dependence, as one would expect for a classical electrostatic interaction (between point charges), but with a 1 r 12 function. This 1 r 12 dependence is, of course, purely empirical in the Lennard-Jones (L-J) potential - it has no grounding in theory - but, nonetheless, the exceptionally high sensitivity of the repulsive interaction to small changes in interatomic/intermolecular separation is captured well by the functional form.
Of course, and as Baerends[40] discusses in a clear overview of Pauli repulsion effects in adsorption, we are dealing not with point charges and a pure Coulombic interaction but with a screened Coulomb potential and delocalised electron 'clouds'. The overlap of the electron clouds at short separations leads in a classical model, and perhaps counterintuitively, to an attractive electrostatic interaction. It is only when the inter-atomic separation becomes so small that nuclear repulsion dominates that the overall electrostatic force becomes repulsive.
11 We focus throughout this chapter only on fermions. For bosons, and as discussed by Mullin and Blaylock[39], an effective attractive force is often invoked.
Thus, and as we hope is abundantly clear from previous sections, we cannot expect to understand electron repulsion due to Pauli exclusion in the context of classical electrostatics. The fundamental origin of the repulsion comes from, as we've seen, the physical impossibility of 'squeezing' two fermions into the same quantum state. But the central question is this: just how does the exclusion principle translate into a physically measurable interaction? We'll see in the following section how dynamic force microscopy allows us to directly probe the exclusion-derived repulsion between the electron density of two atoms or molecules. Before we consider the results of the real-world experiment, however, it's very helpful to think about a 'stripped-down' system involving the overlap of two single particle wavefunctions (see Section 1.3)[41, 42, 43]. This ' Gedankenexperiment ', if you will, provides compelling insights into the origin of Pauli repulsion.
First, recall that the kinetic energy operator is -glyph[planckover2pi1] 2 2 m ∇ 2 . The curvature of a wavefunction therefore determines its kinetic energy (via the Laplacian, ∇ 2 ). Wilson and Goddard's approach[41] to elucidating the origin of Pauli repulsion was to compare the kinetic energy (KE) of a Hartree product of the wavefunctions for two same-spin electrons with the KE of an antisymmetrized product (see Fig. 1.2). A Hartree product is simply the following:
$$\Psi _ { \text{Hart} } ( { \mathbf r } _ { 1 }, { \mathbf r } _ { 2 } ) = \psi ( { \mathbf r } _ { 1 } ) \psi ( { \mathbf r } _ { 2 } )$$
As should be clear from Section 1.3, the multiparticle wavefunction Ψ Hart is not antisymmetric (nor does it take into account indistinguishability of the particles) and is therefore in general not appropriate to use to describe fermions. However, we can take the Hartree product as a representation of the system when the Pauli exclusion principle is 'suppressed' and determine the resulting kinetic energy.
In order to incorporate Pauli exclusion we have to consider a multi-particle wavefunction which is appropriately anti-symmetrized . Slater introduced an elegant method of enforcing this antisymmetry requirement via the determinant approach which now bears his name[45]. Wilson and Goddard[41] focussed on the orthogonality of orbitals which is generally imposed in approaches which treat the multiparticle wavefunction in terms of (a sum of) Slater determinants (see Fig. 1.2, taken from the PhD thesis of Julius Su[44]). We note, however, that orthogonality is a constraint on the multiparticle wavefunction that is not strictly necessary [35] and, as discussed by Beylkin et al.[46] leads to ever-increasing levels of computational expense as the size of a system grows.
Nonetheless, to ensure antisymmetry (i.e. the requirement of Eqn. 1.5), wavefunction slope and curvature must necessarily increase and thus the overall picture emerging from Fig. 1.2 is correct (even if one doesn't invoke orthogonality as the root cause of the increase in wavefunction curvature). This change in curvature results in a corresponding increase in kinetic energy. A complementary explanation from a Fourier analysis perspective, as noted in the following section, is that the increase in curvature of the wavefunction necessitates the introduction of higher spatial frequency contributions, i.e. higher momen-
Fig. 1.2 The effective repulsion due to Pauli exclusion stems from the change in the curvature of the wavefunction due to the requirement for antisymmetrization in fermion systems. One approach to visualising this is to consider the orthogonalization of orbitals (which is placed as a constraint on Slater determinant approaches to constructing a multi-particle wavefunction). Higher wavefunction curvature leads to a higher kinetic energy. Equivalently, higher curvature is accounted for in Fourier space by higher spatial frequency (momentum) components. Figure taken from the PhD thesis of Julian Su[44]. c Julian © Su (2007).
tum components). It is this increase in KE (or momentum) which is responsible for the majority of Pauli repulsion.
There are two important assumptions built into this description of Pauli exclusion, however. First, we have adopted a 'pairwise' approach to considering electron-electron interactions when, in reality, Pauli exclusion is an n -body, rather than a two-body problem. The second, and related, issue is that the modification of the wavefunction due to orthogonalisation will mean that the electron density will be distributed differently, affecting electron-electron interactions and giving rise to the effect known as correlation. Interactions between same-spin electrons go by the name Fermi correlation , whereas those between opposite-spin electrons are known as Coulomb correlation 12 . Nonetheless, the dominating contribution to Pauli repulsion is the pure quantum-mechanical component arising from wavefunction antisymmetry.
## 1.5 Is there a Pauli exclusion force ?
Having spent much of the chapter up to this point using the term 'Pauli repulsion', it might seem a little perverse for us to now pose the question as to whether there is a Pauli exclusion force or not (particularly as the experimental technique we're considering is dynamic force microscopy). Notwithstanding the use of ' Pauli repulsion ' or ' Pauli exclusion force ' in the DFM literature - and, more broadly, throughout very many areas of science (spanning, for example, particle physics, single molecule imaging and spectroscopy, astrophysics 13 , and cosmology) - a number of authors have made the claim that Pauli exclusion does not produce a force in the traditional sense. Mullin and Blaylock[39], in particular, present a set of arguments as to why they are of the opinion that couching the effects of Pauli exclusion in terms of a repulsive force , or exchange force , can be rather misleading. Indeed, they go so far as to argue - and we quote directly from their paper - that 'there is no real force due to Fermi/Bose symmetries' , citing, amongst others, Griffiths' description of the effects of Pauli exclusion[47]:
We call it an exchange force but it is not really a force at all - no physical agency is pushing on the particles; rather it is purely a geometric consequence of the symmetrization requirement.
What does Griffiths (and, by extension, Mullin and Blaycock) mean by this?
To back up their assertion that Pauli 'repulsion' is not a force in the traditional sense, Mullin and Blaycock's consider a number of 'archetypal' physicochemical phenomena where the exclusion principle plays a key role. Arguably the most instructive of these is their discussion of the changes in momentum in a classical gas as compared to a Fermi gas. We encourage the reader to follow the detail of the analysis in Section II of their paper (under the sub-section entitled Virial Expansion ) and restrict ourselves here simply to highlighting the central point they make.
Consider first a classical ideal gas in a container. Pressure, P , arises from the combined impacts of each atom of that gas on the walls of the container and is given by the force per unit area. Force, in turn, is the rate of change of momentum. The mean force, ¯ , F which each individual molecule of the gas contributes is ¯ F = ∆p/∆t , where ∆p is the momentum change on striking the wall. (This is twice the atomic momentum because the sign of the momentum flips on collision). ∆t is the time required for an atom to cross the container, i.e. ∆t = mL/p ¯ where L is the width of the container and m is the atomic mass. The key point in the classical case is this: if we make the volume of an ideal gas smaller or
12 The combined contributions of the exclusion principle and electron correlation produce the exchangecorrelation contribution to the functional in density functional theory.
13 The Pauli exclusion principle prevents the collapse of white dwarf and neutron stars. See Neutron Stars 1: Equation of State and Structure , P. H¨ ansel, AY Potekhin, and DG Yakovlev, Springer (New York, 2007).
we introduce repulsive interactions (with no change in temperature), the pressure of the gas will rise because of a decreased ∆t due to a change in (the effective) L arising from collisions, but ¯ p remains the same . (Recall that for a classical gas the root mean square momentum, p rms is √ 3 mk T B )
Compare this to what happens for a Fermi gas subject to the exclusion principle. The effect of the exclusion principle is to modify the momentum distribution . Mullin and Blaylock argue that this is subtly different to what happens for the classical gas when repulsive interactions are introduced. Classically, the repulsive forces raise the pressure of the gas because the collisions and deflections of the atoms change the atomic transit time. Quantum-mechanically, the momentum distribution is 'intrisically' modified because of the higher curvature of the wavefunction which results from the exclusion principle. Position and momentum are conjugate variables and are thus two sides of the same coin - Fourier transformation allows us to switch between the two (entirely equivalent) representations. The higher wavefunction curvature demanded by Pauli exclusion is entirely equivalent to stating that higher spatial frequency components are required in reciprocal (i.e. momentum) space 14 . It is this intrinsic symmetry-driven modification of the momentum distribution which raises the pressure of the Fermi gas.
It is worth lifting another couple of quotes from Mullin and Blaylock's paper to highlight just how strongly opposed they are to equating Pauli exclusion with a repulsive force:
The idea of an effective repulsion between fermions ignores the real physics and gives a very poor analogy with classical repulsive gases...we offer the following guiding principle regarding statistical symmetries: 'May the force be not with you'.
Is this degree of anti-force scepticism justified, however?
## 1.6 Beyond Gedanken: Exploiting exclusion in force microscopy
At this point, the pragmatic scanning probe microscopist could quite reasonably take issue with the preceding arguments because the primary experimental observable in a dynamic force microscopy experiment is the frequency shift of the probe. And this, via the SaderJarvis formalism[48], for example, can be converted directly to a tip-sample force . The effects of Pauli exclusion are directly measurable in DFM because they shift the resonant frequency of the probe-cantilever system, and this ultimately can be interpreted as a change in the tip-sample force. Notwithstanding the arguments put forward by Mullin and Blaylock[39], and Griffiths[47], amongst others, if Pauli exclusion isn't giving rise to a force then it certainly very much looks like it in a DFM experiment.
The resolution of this apparent conflict may lie, as Moll et al. have discussed in a recent paper focussed on the interpretation of submolecular resolution DFM images[49], in the virial theorem. Slater showed in the 1930s that the virial theorem can be applied to a molecule[50], assuming that the nuclei are fixed in place by external forces. The total electron energy, E , is related to the electronic kinetic energy, T , and potential energy, V , as follows:
$$T = - E - r ( \frac { d E } { d r } )$$
$$V = 2 E + r ( \frac { d E } { d r } )$$
The electronic kinetic energy and potential energy are thus coupled via the virial theorem. Moll et al. [49] claim that, despite the Pauli exclusion force being non-conservative
14 This, of course, is the fundamental origin of the Heisenberg uncertainty principle.
in character, if it is assumed that we have a diatomic (or dimolecular) system with one degree of freedom - as is the case for the tip-sample system in DFM - the Pauli energy and the increase in electronic kinetic energy can be related as follows:
$$E _ { \text{Pauli} } ( z ) = \frac { 1 } { z } \int _ { z } ^ { \infty } \Delta E _ { k i n } ( z ^ { \prime } ) d z ^ { \prime } & & ( 1. 1 0 )$$
where z is the interatomic/intermolecular separation. The issue of extracting accurate measures of non-conservative forces from the frequency shift observable in DFM, however, continues to attract considerable debate and discussion. For example, the Sader-Jarvis inversion technique[48] widely applied to extract forces from frequency-shift-vs-separation curves must, as John Sader and his co-authors themselves highlight[51], be applied with great care under conditions were there is a significant contribution from non-conservative forces.
Although the authors cited in the previous section propose reasons for drawing a distinction between a traditional force and the effects arising from Pauli exclusion, the increase in kinetic energy and momentum resulting from the requirement for wavefunction antisymmetry nonetheless ultimately result in an interaction which is measured as a repulsive force in a DFM experiment. That is, the connection between the change in kinetic energy and the total energy of the tip-sample system appears to result in a measurable, and positive (i.e. repulsive), contribution to the frequency shift due to the Pauli exclusion principle. What is important to realise from the previous sections, however, is that Pauli exclusion really is not comparable to other types of interparticle interaction. In this sense it is a phenomenon which is distinct from the four fundamental forces, i.e. strong, weak, electromagnetic (in particular), and, if the graviton exists, gravity.
## 1.6.1 Intramolecular Imaging
Although DFM's 'sibling' technique, scanning tunnelling microscopy (STM), has long been capable of submolecular resolution imaging, in the sense that molecular orbital density can be probed (see an earlier volume of this Springer series on Atom and Single Molecule Machines [52]), only DFM is capable of resolving the chemical framework or atomic structure of a molecule. This is because STM probes orbital density only within a specific energy window (set by the potential difference between the tip and sample) and in conventional tunnelling microscopy therefore only the frontier molecular orbitals are accessible 15 . The spatial distribution of the frontier orbital density generally does not map onto the atomic positions, and indeed often bears very little relationship to the 'ball-and-stick' models of molecules so familiar to chemists and physicists.
As Giessibl has highlighted[53], however, DFM is not restricted to probing the frontier orbital density and is sensitive to the total charge density. This is because intramolecular forces depend on the total electron density, rather than the density of states within a certain energy window[54]. The sensitivity of DFM to the total electron density is particularly pronounced when in the Pauli exclusion regime of imaging, i.e. at very small tip-sample separations. Fig. 1.1 at the start of this chapter shows very clearly that, unlike STM, DFM in this Pauli exclusion regime produces images which are remarkably similar to the ball-and-stick structural models of molecules.
On the basis of Fig. 1.3 (and related theoretical and experimental data), Moll et al.[49] argue that there is a close connection between the charge density of a molecule and the increase in electron kinetic energy due to Pauli exclusion. This assumes that (a) the arguments regarding wavefunction curvature outlined in Sections 1.4 and 1.5 provide an
15 In the scanning tunnelling hydrogen microscopy (STHM)[9, 10, 11] variant of STM mentioned earlier, this constraint can be circumvented.
Fig. 1.3 Comparison of (a) Experimental frequency shift image and (b) a simulated frequency shift image for a 3,4,9,10-perylenetetracarboxylic dianhydride (PTCDA) molecule calculated on the basis of the Pauli exclusion-derived change in electron kinetic energy estimated using Eqn.1.12. (c) Charge density of a PTCDA molecule at a given tip-sample separation. Compare with (d) , the change in kinetic energy at the same tip-sample separation. Figure adapted from Moll et al.[49]. c Institute of Physics Publishing © (2012).
accurate model of electron-electron interactions at the tip-sample junction, and (b) the dominant effect is the change in kinetic energy, and that this can be 'deconvolved' from the overall response of the electron density as a function of the tip-sample separation. They approximate the complicated relationship between the increase in kinetic energy and the separation of two atoms with different nuclear charges (see Eqn 6 of their paper) as follows:
$$\Delta E _ { \text{kin} } ( z ) = A \rho _ { s } ( z ) ^ { B } & & ( 1. 1 1 )$$
where z is the interatomic/intermolecular separation, ρ z ( ) is the sample charge density at separation z , and A and B are two tunable parameters. As can be seen in Fig. 1.3, this simple power law model, which involves no explicit consideration of the probe, provides good agreement with experimental frequency shift images of a 3,4,9,10-perylenetetracarboxylic dianhydride (PTCDA) molecule. We also include in Fig. 1.3, again from Moll et al. 's paper, a comparison of the charge density of the PTCDA molecule with the increase in kinetic energy calculated using the simple model of Equation 1.11. There is again apparently good agreement, adding support to the idea that DFM is sensitive to the total charge density of the system.
What is not included in the model used to generate the simulated images in Fig. 1.3 - although Moll and co-workers deal with this point elsewhere[55] - is the relaxation or bending of the CO molecule at the tip apex as it is moved across the underlying PTCDA molecule. It turns out that this is an extremely important contribution to the observation of intramolecular and intermolecular contrast in DFM images and we'll return to it in the final section.
## 1.6.2 Density depletion
The modification of the curvature and spatial extent of the tip-sample wavefunction due to Pauli exclusion produces extensive modification to the total electron density of the system. A key aspect of this is the generation of regions of density depletion. Baerends[40] discusses the importance of density depletion in the context of the Ag-O bond where a
substantial degree of Pauli exclusion-derived depletion around the centre of the bond is observed.
As a more recent example in the context of DFM, a number of the authors of this chapter have explored the importance of density depletion in the interpretation of images taken in the Pauli exclusion regime. The molecular system we used is that shown in Fig. 1.1(G) - a hydrogen-bonded assembly of naphthalenetetracarboxylic diimide (NTCDI) molecules on a passivated silicon surface. Fig. 1.4 shows a comparison of the total electron density for an NTCDI assembly vs the density difference at a number of different z positions of the tip above a C-C bond (Fig. 1.4(a)-(c)) and above an intermolecular region where hydrogen-bonding is expected[4]. Pauli exclusion results in strong tip-induced electron depletion above both the intermolecular and intramolecular bond regions.
The most important insight to be derived from this analysis of density depletion is that, as is always the case in any type of scanning probe experiment (and as is wellunderstood across the SPM community), the influence of the tip on the imaging process must always be carefully considered. Tip-sample interactions and convolution have been an issue for scanning tunnelling microscopy since its invention, of course, but with the advent of DFM imaging in the so-called 'Pauli regime' the probe can certainly no longer be treated as just a perturbation of the electronic structure. The tip-sample separation for the type of high resolution images shown in Fig. 1.1 is such that the repulsive Pauli component makes a strong contribution - the tip interacts heavily with the underlying molecule adsorbed on the sample surface. In this sense, the sample-tip apex system should be considered as one large molecule.
In the following, and final, section of this chapter we'll see just how important a role the tip can play in generating high resolution DFM images.
## 1.7 But do we really see bonds?
A key 'ingredient' in attaining intramolecular contrast in DFM is the passivation of the tip apex. Gross et al. [1] first showed that CO was a particularly appropriate molecule to use for imaging submolecular structure. (In the same paper, and in subsequent work[56], they compared the imaging capabilities of CO with those of other species adsorbed at the tip apex). Although deliberate functionalisation with CO is certainly not necessary to obtain intra- (and inter-)molecular contrast[4], carbon monoxide remains the molecule of choice at present for high resolution DFM.
It turns out that CO is very far from a rigid probe, however, and the tilting of the molecule at the tip apex plays an essential role in the imaging process. The flexibility of CO has been studied in some detail by a number of groups [57, 58, 55, 59] but it is a very recent paper[5], available only at the condensed matter arXiv at the time of writing, on which we would like to focus here. This paper provides particularly telling insights into the extent to which the probe itself contributes to the structure seen in molecular and submolecular images.
Hapala et al. [5] use an exceptionally simple, but remarkably powerful, model to simulate DFM (and scanning tunnelling hydrogen microscopy[9, 10, 11]) images acquired either with a CO probe or any other type of tip apex. They represent the tip-sample geometry as shown in Fig. 1.5 and account for interactions between the probe and sample molecule using analytical Lennard-Jones potentials. It is very important to note that no account is taken of intra- or intermolecular charge density in this model: the approach adopted by Hapala et al. uses only the coordinates of the atoms within the molecule under study - electron density due to bonding between those atoms is not incorporated in their model. In other words, the force-field does not rely on the electron density of the system. Although this might at first glance appear to be a rather crude approach (as compared to, for example, modelling the system using an ab initio method such as density
Fig. 1.4 Total electronic density (TED) and electron density difference (EDD) calculated for an NTCDI assembly plotted 100pm above the molecular plane for a variety of different tip heights. At each tip height in a simulated F z ( ) curve, the EDD was obtained by first calculating the TED for (i) the isolated surface, and (ii) the isolated NTCDI tip. These two densities were then summed together and subtracted from the relaxed total density for the full system. The remaining quantity is the EDD. This quantifies the fraction of charge which is redistributed due to the interaction of the DFM tip and the NTCDI molecule. The TED (left) and EDD (right) are shown for an oxygen-down NTCDI probe molecule at (a)-(c) the C-C location on an NTCDI molecule, and (d)-(f) at the intermolecular H-bond location for the different tip heights specified in each figure. Figure from Sweetman et al.[4]. c Nature Publishing Group (2014). ©
functional theory), it is nonetheless the case that their 'stripped-down' model accurately reproduces the experimental data. This is the acid test of any theory or simulation.
Fig. 1.5 Schematic model of the tip-sample geometry used by Hapala et al.[5] in their analysis of the origin of intra- and intermolecular contrast in DFM images. The final metal atom at the tip apex and the 'probe particle' are shaded in gold and cyan respectively with the underlying molecular layer represented by the standard space-filling model. The coloured vectors show the various forces on the tip: F Tip,R (green) is the radial force; F Tip,xy (red) is the lateral force; and F urf S (yellow) is the force due to the sample molecules. ( T i and T t refer to tunnelling processes not of interest in this chapter.) Taken from Hapala et al .[5].
Fig. 1.6 shows a comparison between experimental DFM images and the output of Hapala et al. 's simulations for two systems comprising assemblies of 8-hydroxyquinoline tetramers and NTCDI molecules (as discussed above in the context of Fig. 1.4), respectively. For both of these systems, intermolecular interactions are mediated by hydrogen bonding. Note, however, how the sharp intra- and intermolecular features in the simulated image of Fig. 1.6(a) agree extremely well with those in the experimental data shown in Fig. 1.6(b), despite the absence of any intra- or intermolecular charge density in the model. Fig. 1.6(c) and 6(d) similarly show a comparison between the 'flexible tip' model and DFM images of a hydrogen-bonded NTCDI assembly[4] taken by a number of the authors of this chapter. Again, intramolecular and intermolecular features are observed in the simulated image, despite the absence of any charge density due to covalent- or hydrogen-bonding.
It therefore would appear that the flexibility of the probe molecule plays a major role in the imaging of intra- and intermolecular structure. But we've seen in previous sections that there's also a close correspondence between images simulated on the basis of an increase in electron kinetic energy due to Pauli exclusion and the experimental frequency shift data[49]. Moreover, the intensity of intramolecular bonds as observed by DFM is related to the Pauling bond order[55], i.e. the charge density. Similarly, the DFM images of de Oteyza et al. [2] clearly show a pronounced difference between single, double, and triple bonds. The key issue is therefore the extent to which the response of the tip to interatomic and/or intermolecular charge density is a 'first order' vs 'second order' contribution to the imaging mechanism, as compared to the flexibility of the probe. This is currently a very active area of debate.
In order to explore the influence of tip relaxation on the DFM images of NTCDI shown in Figs. 4 and Fig. 1.6, we (i.e. SJ, AS, LK, PJM, and co-workers[4]) generated simulated images using a variant of DFT where both the atomic geometry and the electronic structure of the system were 'frozen'. Despite the lack of probe relaxation, a weak feature at the expected position of the hydrogen-bond was observed. Nonetheless, another question remains: to what extent might convolution of the tip's electron density with molecular charge density at the edge of a molecule account for the observation of 'intermolecular' features? In the supplementary information file associated with their paper, Hapala et al. [5] suggest that this convolution effect could be as strong as the interaction of the probe with any charge density due to an intermolecular bond. This is an exceptionally important issue which needs to be addressed in a timely fashion by the scanning probe microscopy community.
## 1.8 Conclusions
The history of the development of the Pauli exclusion principle provides fascinating insights into just how problematic it is to associate purely quantum mechanical concepts with classical 'real world' analogies. In this sense, it's a shame that Pauli's Zweideutigkeit term did not gain wider acceptance as it's a less misleading, albeit rather more prosaic, description than 'spin'. Similarly, when we describe the Pauli exclusion principle as giving rise to a repulsive force we should bear in mind that the origin of the repulsion detected in dynamic force microscopy is not at all adequately explained via classical analogies. The interaction arises from the modification of the electrons' momentum distribution due to the increased curvature of the wavefunction imposed by the requirement for antisymmetrization in fermion systems. Classical analogies will clearly fail. Understanding the fundamental origin of the increased wavefunction curvature is ultimately, as Feynman puts it in the quote at the start of this chapter, ' deep down in relativistic quantum mechanics '.
Fig. 1.6 (a) (b) , : Comparison of a simulated DFM image of a hydrogen-bonded assembly of 8hydroxyquinoline molecules (from Hapala et al.[5]) with the corresponding experimental DFM image taken from Zhang et al.[3]. (c) Series of simulated frequency shift images at different tip-sample separations, again from Hapala et al.[5], of NTCDI molecules using a (top row) unrelaxed, and (bottom row) relaxed tip. (d) Experimental frequency shift image for comparison. (From Sweetman et al.[4]).
Dynamic force microscopy provides us with direct access to the effects of Pauli exclusion on an atom-by-atom and/or molecule-by-molecule basis, and with resolution comparable to the spatial extent of a single chemical bond. This is a remarkable capability. At the time of writing it has been only five years since Gross et al. [1] pioneered the exploitation of Pauli exclusion in force microscopy. As this variant of scanning probe microscopy is therefore in its infancy, there is potentially immense scope for detailed insights into the effects of the exclusion principle in a variety of atomic and molecular systems. However, every probe microscope image - indeed, every image (regardless of the technique used to generate that image) - is, at some level, a convolution of the properties of the sample and those of the imaging system. In the Pauli exclusion regime of dynamic force microscopy this convolution can be exceptionally strong. We therefore need to temper our enthusiasm for the acquisition of ultrahigh resolution images with caution regarding the interpretation of the data, as the examples included in this chapter clearly show.
## 1.9 Acknowledgments
We are very grateful for financial support from the UK Engineering and Physical Sciences Research Council in the form of a fellowship (EP/G007837), from the Leverhulme Trust (through grant F/00114/BI), from the European Commission's ICT-FET programme via the Atomic Scale and Single Molecule Logic gate Technologies (AtMol) project (www.atmol.eu), Contract No. 270028, and the ACRITAS Marie Curie Initial Training Network (www.acritas.eu). We are also very grateful for the support of the University of Nottingham High Performance Computing Facility. PJM thanks Christian Joachim (CNRS Toulouse) and Leo Gross (IBM Zurich) for helpful discussions and e-mail exchanges on the role of the exclusion principle in probe microscopy.
## References
- [1] L Gross, F Mohn, N Moll, P Liljeroth, and G Meyer. Science , 325:1110, 2009.
- [2] DG de Oteyza, P Gorman, Y-C Chen, S Wickenburg, A Riss, DJ Mowbray, G Etkin, Z Pedramrazi, H-Z Tsai, A Rubio, MF Crommie, and FR Fischer. Science , 340:1434, 2013.
- [3] J Zhang, P Chen, B Yuan, W Ji, Z Cheng, and X Qiu. Science , 342:611, 2013.
- [4] A Sweetman, SP Jarvis, H Sang, I Lekkas, P Rahe, Y Wang, J Wang, N Champness, L Kantorovich, and P Moriarty. Nature Comm. , 5:3931, 2014.
- [5] P Hapala, G Kichin, CF Wagner, S Tautz, R Temirov, and P Jelinek. Cond Matt arXiv; , arXiv:1406.3562, 2014.
- [6] N Pavlicek, B Fleury, M Neu, J Niedenfuehr, C Herranz-Lancho, M Ruben, and J Repp. Phys. Rev. Lett. , 108:086101, 2012.
- [7] R Pawlak, S Kawai, S Fremy, T Glatzel, and E Meyer. ACS Nano , 5:6349, 2011.
- [8] A Riss, S Wickenburg, P Gorman, LZ Tan, H-Z Tsai, DG de Oteyza, Y-C Chen, AJ Bradley, MM Ugeda, G Etkin, SG Louie, FR Fischer, and MF Crommie. Nano Lett. , 14:2251, 2014.
- [9] R. Temirov, S. Soubatch, O. Neucheva, A. C. Lassise, and F. S. Tautz. New J Phys , 10:053012, 2008.
- [10] C. Weiss, C. Wagner, C. Kleimann, M. Rohlfing, F. S. Tautz, and R. Temirov. Phys. Rev. Lett. , 105:086103, 2010.
- [11] C Weiss, C Wagner, R Temirov, and FS Tautz. J. Am. Chem. Soc. , 132:11864, 2010.
- [12] G Kichin, C Weiss, C Wagner, FS Tautz, and R Temirov. J. Am. Chem. Soc. , 133: 16847, 2011.
- [13] Michela Massimi. Pauli's exclusion principle: The origin and validation of a scientific principle . Cambridge University Press, Cambridge, 2005.
- [14] W. Pauli. Zeitschrift fur Physik , 31:373, 1925.
- [15] W. Pauli. Zeitschrift fur Physik , 31:765, 1925.
- [16] E. Fermi. Rendiconti Lincei , 3:145, 1926.
- [17] A. Zannoni. On the quantization of the monoatomic ideal gas. arXiv:condmat/9912229 [cond-mat.stat-mech]. Translation of Fermi's 1926 paper.
- [18] PAM Dirac. Proc. Royal Soc. A , 112:661, 1926.
- [19] PCWDavies and DS Betts. Quantum mechanics . Stanley Thornes (Publishers) Ltd., Cheltenham, England, 1994.
- [20] WM De Muynck. Int. J. Theor. Phys. , 14:327, 1975.
- [21] MV Berry and JM Robbins. J. Phys. A , 33:L207, 2000.
- [22] NS Ginsberg, SR Garner, and LV Hau. Nature , 445:623, 2007.
- [23] M Fleischhauer. Nature , 445:605, 2007.
- [24] Y Omar. Contemp. Phys. , 46:437, 2005.
- [25] D Dieks and MAM Versteegh. Found. Phys. , 38:923, 2008.
- [26] D Dieks and A Lubberdink. Found. Phys. , 41:1051-1064, 2011.
- [27] MAM Versteegh and D Dieks. Am. J. Phys. , 79:741, 2011.
- [28] E Jaynes. The gibbs paradox. In G. J. Erickson C. R. Smith and P. O. Neudorfer, editors, Maximum Entropy and Bayesian Methods , 1992.
- [29] P Ehrenfest and V Trkal. Deduction of the dissociation-equilibrium from the theory of quanta and a calculation of the chemical constant based on this. In Collected Scientific Papers , page 414. North-Holland, Amsterdam, 1959.
- [30] JW Rohlf. Modern physics:From α to Z 0 . John Wiley and Sons, Inc., New Jersey, U.S., 1994.
- [31] Jon Butterworth. On Pauli and the interconnectedness of all things . http: //www.theguardian.com/science/life-and-physics/2012/feb/28/1 , 2012. [Online; accessed 07-July-2014].
- [32] EH Lieb and WE Thirring. Phys. Rev. Lett. , 35:687-689, 1975.
- [33] EH Lieb. Rev. Mod. Phys. , 48:553-569, 1976.
- [34] EH Lieb. Bulletin Am. Math. Soc. , 22:1-49, 1990.
- [35] Lev Kantorovich. Quantum Theory of the Solid State: An Introduction . Kluwer Academic Publishers, Dordrecht, Holland, 2004.
- [36] F Wilczek. Phys. Rev. Lett. , 49:957, 1982.
- [37] S Bartalucci, M. Bragadireanu, M. Cargnelli, M Catitti, C Curceanu (Petrascu), S Di Matteo, J.-P. Egger, C Guaraldo, M Iliescu, T Ishiwatari, M Laubenstein, J Marton, E Milotti, D Pietreanu, T Ponta, D.L. Sirghi, F Sirghi, L Sperandio, E Widmann, and J Zmeskal. Phys. Lett. B , 641:18, 2006.
- [38] Ian Duck and EGC Sudarshan. Am. J. Phys. , 66:284, 1998.
- [39] WJ Mullin and G Blaylock. Am. J. Phys. , 71:1223, 2003.
- [40] EJ Baerends. In Cluster models for surface and bulk phenomena , volume 283, pages 189-207, 1992.
- [41] CW Wilson and WA Goddard. Theor. Chim. Acta , 26:195, 1972.
- [42] CW Wilson and WA Goddard. Theor. Chim. Acta , 26:211, 1972.
- [43] CW Wilson and WA Goddard. Chem. Phys. Lett. , 5:45-49, 1970.
- [44] Julius Su. An electron force field for simulating large scale excited electron dynamics . http://thesis.library.caltech.edu/1598/ , 2007. [Online; accessed 07-July2014].
- [45] JC Slater. Phys. Rev. , 34:1293, 1929.
- [46] G Beylkin, MJ Mohlenkamp, and F P´ erez. J. Math. Phys. , 49:032107, 2008.
- [47] D. J. Griffiths. Introduction to Quantum Mechanics . Prentice-Hall, Englewood Cliffs, NJ, US.
- [48] JE Sader and SP Jarvis. Accurate formulas for interaction force and energy in frequency modulation force spectroscopy. Appl. Phys. Lett. , 84(10):1801 - 1803, 2004.
- [49] N Moll, L Gross, F Mohn, A Curioni, and G Meyer. New J. Phys. , 14:083023, 2012.
- [50] JC Slater. The virial and molecular structure. J. Chem. Phys. , 1:687-691, 1933.
- [51] JE Sader, T Uchihashi, MJ Higgins, A Farrell, Y Nakayama, and SP Jarvis. Nanotechnology , 16:S94, 2005.
- [52] P Moriarty. Submolecular resolution imaging of C 60 :From orbital density to bond order. In Advances in atom and single molecule machines Vol III:Imaging and amnipulating molecular orbitals , page 195. Springer-Verlag, 2013.
- [53] FJ Giessibl. High-speed force sensor for force microscopy and profilometry utilizing a quartz tuning fork. Appl.Phys. Lett. , 73:3956, 1998.
- [54] RP Feynman. Forces in molecules. Phys. Rev. , 56:340-343, 1939.
- [55] L Gross, F Mohn, N Moll, B Schuler, A Criado, E Guitin, D Pea, A Gourdon, and G Meyer. Science , 337:1326, 2012.
- [56] F Mohn, B Schuler, L Gross, and G Meyer. Appl. Phys. Lett. , 102:073109, 2013.
- [57] Z Sun, MP Boneschanscher, I Swart, D Vanmaekelbergh, and P Liljeroth. Phys. Rev. Lett. , 106:046104, 2011.
- [58] J Welker and FJ Giessibl. Science , 336:444, 2012.
- [59] AJ Weymouth, T Hofmann, and FJ Giessibl. Science , 343:1120, 2014.
| null |
[
"Samuel Jarvis",
"Adam Sweetman",
"Lev Kantorovich",
"Enda McGlynn",
"Philip Moriarty"
] |
2014-08-05T16:39:00+00:00
|
2014-08-05T16:39:00+00:00
|
[
"cond-mat.mes-hall"
] |
Pauli's Principle in Probe Microscopy
|
Exceptionally clear images of intramolecular structure can be attained in
dynamic force microscopy through the combination of a passivated tip apex and
operation in what has become known as the "Pauli exclusion regime" of the
tip-sample interaction. We discuss, from an experimentalist's perspective, a
number of aspects of the exclusion principle which underpin this ability to
achieve submolecular resolution. Our particular focus is on the origins,
history, and interpretation of Pauli's principle in the context of interatomic
and intermolecular interactions.
|
1408.1027v3
|
## Bayesian Nonparametric Modeling for Multivariate Ordinal Regression
Maria DeYoreo ∗ Department of Statistical Science, Duke University and Athanasios Kottas † Department of Applied Mathematics and Statistics,
University of California, Santa Cruz
September 21, 2016
## Abstract
Univariate or multivariate ordinal responses are often assumed to arise from a latent continuous parametric distribution, with covariate effects which enter linearly. We introduce a Bayesian nonparametric modeling approach for univariate and multivariate ordinal regression, which is based on mixture modeling for the joint distribution of latent responses and covariates. The modeling framework enables highly flexible inference for ordinal regression relationships, avoiding assumptions of linearity or additivity in the covariate effects. In standard parametric ordinal regression models, computational challenges arise from identifiability constraints and estimation of parameters requiring nonstandard inferential techniques. A key feature of the nonparametric model is that it achieves inferential flexibility, while avoiding these difficulties. In particular, we establish full support of the nonparametric mixture model under fixed cut-off points that relate through discretization the latent continuous responses with the ordinal responses. The practical utility of the modeling approach is illustrated through application to two data sets from econometrics, an example involving regression relationships for ozone concentration, and a multirater agreement problem. Supplementary materials related to computation are available online.
Keywords: Dirichlet process mixture model; Kullback-Leibler condition; Markov chain Monte Carlo; polychoric correlations.
∗ M. DeYoreo ([email protected]) is Postdoctoral Researcher, Department of Statistical Science, Duke University. This research is part of the Ph.D. dissertation of M. DeYoreo, completed at University of California, Santa Cruz.
† A. Kottas ([email protected]) is Professor of Statistics, Department of Applied Mathematics and Statistics, University of California, Santa Cruz.
## 1 Introduction
Estimating regression relationships for univariate or multivariate ordinal responses is a key problem in many application areas. Correlated ordinal data arise frequently in the social sciences, for instance, survey respondents often assign ratings on ordinal scales (such as 'agree', 'neutral', or 'disagree') to a set of questions, and the responses given by a single rater are correlated. Ordinal data is also commonly encountered in econometrics, as rating agencies (such as Standard and Poor's) use an ordinal rating scale. A natural way to model data of this type is to envision each ordinal variable as representing a discretized version of an underlying latent continuous random variable. In particular, the commonly used (multivariate) ordinal probit model results when a (multivariate) normal distribution is assumed for the latent variable(s).
Under the probit model for a single ordinal response Y with C categories, and covariate vector x , Pr( Y ≤ m | x ) = Φ( γ m -x β T ), for m = 1 , . . . , C . Here, -∞ = γ 0 < γ 1 < · · · < γ C -1 < γ C = ∞ are cut-off points for the response categories, where typically γ 1 = 0 for identifiability. Albert and Chib (1993) have shown that posterior simulation is greatly simplified by augmenting the model with latent variables. In particular, assume that the ordinal response Y arises from a latent continuous response Z , such that Y = m if and only if Z ∈ ( γ m -1 , γ m ], for m = 1 , . . . , C , and Z | β ∼ N( x β T , 1), which yields Pr( Y = m | x ) = ∫ γ m γ m -1 N( z | x β T , 1)d z .
The multivariate probit model for binary or ordinal responses generalizes the probit model to accommodate correlated ordinal responses Y = ( Y , ..., Y 1 k ), where Y j ∈ { 1 , ..., C j } for j = 1 , ..., k , using a multivariate normal distribution for the underlying latent variables Z = ( Z , ..., Z 1 k ). To obtain an identifiable model, restrictions must be imposed on the covariance matrix Σ of the multivariate normal distribution for Z . One way to handle this is to restrict the covariance matrix to be a correlation matrix, which complicates Bayesian inference since there does not exist a conjugate prior for correlation matrices. Chib and Greenberg (1998) discuss inference under this model, using a random walk Metropolis algorithm to sample the correlation matrix. Liu (2001) and Imai and van Dyk (2005) use parameter expansion with data augmentation (Liu and Wu, 1999) to expand the parameter space such that unrestricted covariance matrices may be sampled. Talhouk et al. (2012)
work with a sparse correlation matrix arising from conditional independence assumptions, and use a parameter expansion strategy to expand the correlation matrix into a covariance matrix, which is updated with a Gibbs sampling step. Webb and Forster (2008) reparameterize Σ in a way that enables fixing its diagonal elements without sacrificing closed-form full conditional distributions. Lawrence et al. (2008) use a parameter expansion technique, in which the parameter space includes unrestricted covariance matrices, which are then normalized to correlation matrices. In addition to the challenges arising from working with correlation matrices, the setting with multivariate ordinal responses requires estimation for the cut-off points, which are typically highly correlated with the latent responses.
The assumption of normality on the latent variables is restrictive, especially for data which contains a large proportion of observations at high or low ordinal levels, and relatively few observations at moderate levels. As a consequence of the normal distribution shape, there are certain limitations on the effect that each covariate can have on the marginal probability response curves. In particular, Pr( Y j = 1 | x ) and Pr( Y j = C j | x ) are monotonically increasing or decreasing as a function of covariate x , and they must have the opposite type of monotonicity. The direction of monotonicity changes exactly once in moving from category 1 to C j (referred to as the single-crossing property). In addition, the relative effect of covariates r and s , i.e., the ratio of ∂ Pr( Y j = m | x ) /∂x r to ∂ Pr( Y j = m | x ) /∂x s , is equal to the ratio of the r -th and s -th regression coefficients for the j -th response, which does not depend on m or x . That is, the relative effect of one covariate to another is the same for every ordinal level and any covariate value. We refer to Boes and Winkelmann (2006) for further discussion of such properties.
Work on Bayesian nonparametric modeling for ordinal regression is relatively limited, particularly in the multivariate setting. In the special case of binary regression, there is only one probability response curve to be modeled, and this problem has received significant attention. Existing semiparametric approaches involve relaxing the normality assumption for the latent response (e.g., Newton et al., 1996; Basu and Mukhopadhyay, 2000), while others have targeted the linearity assumption for the response function (e.g., Mukhopadyay and Gelfand, 1997; Walker and Mallick, 1997; Choudhuri et al., 2007). For a univariate ordinal response, Chib and Greenberg (2010) assume that the latent response arises from
scale mixtures of normals, and the covariate effects to be additive upon transformation by cubic splines. This allows nonlinearities to be obtained in the marginal regression curves, albeit under the restrictive assumption of additive covariate effects. Gill and Casella (2009) extend the ordinal probit model by introducing subject-specific random effects terms modeled with a Dirichlet process (DP) prior.
Chen and Dey (2000) model the latent variables for correlated ordinal responses with scale mixtures of normal distributions, with means linear on the covariates. Related, Bao and Hanson (2015) assume the latent variables arise from mixtures of linear multivariate probit regressions, mixing on the regression coefficients and covariance matrices in the multivariate normal distribution. The latter modeling approach is an extension of the work in Kottas et al. (2005), where, in the context of multivariate ordinal data without covariates, the latent response distribution was modeled with a DP mixture of multivariate normals. In the absence of covariates, this model is sufficiently flexible to uncover essentially any pattern in a contingency table while using fixed cut-offs. This represents a significant advantage relative to the parametric models discussed, since the estimation of cut-offs requires nonstandard inferential techniques, such as hybrid Markov chain Monte Carlo (MCMC) samplers (Johnson and Albert, 1999) and reparamaterization to achieve transformed cut-offs which do not have an order restriction (Chen and Dey, 2000).
The preceding discussion reflects the fact that there are significant challenges involved in fitting multivariate probit models, and a large amount of research has been dedicated to providing new inferential techniques in this setting. While the simplicity of its model structure and interpretability of its parameters make the probit model appealing to practitioners, the assumptions of linear covariate effects, and normality on the latent variables are restrictive. Hence, from both the methodological and practical point of view, it is important to explore more flexible modeling and inference techniques, especially for the setting with multivariate ordinal responses. Semiparametric methods for binary regression are more common, since in this case there is a single regression function to be modeled. When taken to the setting involving a single ordinal response with C ≥ 3 classifications, it becomes much harder to incorporate flexible priors for each of the C -1 probability response curves. And, semiparametric prior specifications appear daunting in the multi-
variate ordinal regression setting where, in addition to general regression relationships, it is desirable to achieve flexible dependence structure between the ordinal responses.
Motivated by these limitations of parametric and semiparametric prior probability models, our goal is to develop a Bayesian nonparametric regression model for univariate and multivariate ordinal responses, which enables flexible inference for both the conditional response distribution and for regression relationships. We focus on problems where the covariates can be treated as random, and model the joint distribution of the covariates and latent responses with a DP mixture of multivariate normals to induce flexible regression relationships, as well as general dependence structure in the multivariate response distribution. In many fields - such as biometry, economics, and the environmental and social sciences - the assumption of random covariates is appropriate, indeed we argue necessary, as the covariates are not fixed prior to data collection.
In addition to the substantial distributional flexibility, an appealing aspect of the nonparametric modeling approach taken is that the cut-offs may be fixed, and the mixture kernel covariance matrix left unrestricted. Regarding the latter, we show that all parameters of the normal mixture kernel are identifiable provided each ordinal response comprises more than two classifications. For the former, we demonstrate that, with fixed cut-offs, our model can approximate arbitrarily well any set of probabilities on the ordinal outcomes. To this end, we prove that the induced prior on the space of mixed ordinal-continuous distributions assigns positive probability to all Kullback-Leibler (KL) neighborhoods of all densities in this space that satisfy certain regularity conditions.
We are primarily interested in regression relationships, and demonstrate that we can obtain inference for a variety of nonlinear as well as more standard shapes for ordinal regression curves, using data examples chosen from fields for which the methodology is particularly relevant. The joint response-covariate modeling approach yields more flexible inference for regression relationships than structured normal mixture models for the response distribution, such as the ones in Chen and Dey (2000) and Bao and Hanson (2015). Moreover, as a consequence of the joint modeling framework, inferences for the conditional covariate distribution given specific ordinal response categories, or inverse inferences, are also available. These relationships may be of direct interest in certain applications. Also
of interest are the associations between the ordinal variables. These are described by the correlations between the latent variables in the ordinal probit model, termed polychoric correlations in the social sciences (e.g., Olsson, 1979). Using a data set of ratings assigned to essays by multiple raters, we apply our model to determine regions of the covariate space as well as the levels of ratings at which pairs of raters tend to agree or disagree. We contrast our approach with the Bayesian methods for studying multirater agreement in Johnson and Albert (1999) and Savitsky and Dalal (2014).
The work in this paper forms the sequel to DeYoreo and Kottas (2015), where we have developed the joint response-covariate modeling approach for binary regression, which requires use of more structured mixture kernels to satisfy necessary identifiability restrictions. The joint response-covariate modeling approach with categorical variables has been explored in a few other papers, including Shahbaba and Neal (2009), Dunson and Bhattacharya (2010), Hannah et al. (2011), and Papageorgiou et al. (2015). Shahbaba and Neal (2009) considered classification of a univariate response using a multinomial logit kernel, and this was extended by Hannah et al. (2011) to accommodate alternative response types with mixtures of generalized linear models. Dunson and Bhattacharya (2010) studied DP mixtures of independent kernels, and Papageorgiou et al. (2015) build a model for spatially indexed data of mixed type (count, categorical, and continuous). However, these models would not be suitable for ordinal data, or, particularly in the first three cases, when inferences are to be made on the association between ordinal response variables.
The rest of the article is organized as follows. In Section 2, we formulate the DP mixture model for ordinal regression, including study of the theoretical properties discussed above (with technical details given in the Appendix). We discuss prior specification and posterior inference, and indicate possible modifications when binary responses are present. The methodology is applied in Section 3 to ozone concentration data, two data examples from econometrics, and a multirater agreement problem. Finally, Section 4 concludes with a discussion.
## 2 Methodology
## 2.1 Model formulation
Suppose that k ordinal categorical variables are recorded for each of n individuals, along with p continuous covariates, so that for individual i we observe a response vector y i = ( y i 1 , . . . , y ik ) and a covariate vector x i = ( x i 1 , . . . , x ip ), with y ij ∈ { 1 , . . . , C j } , and C > j 2. Introduce latent continuous random variables z i = ( z i 1 , . . . , z ik ), i = 1 , . . . , n , such that y ij = l if and only if γ j,l -1 < z ij ≤ γ j,l , for j = 1 , . . . , k , and l = 1 , . . . , C j . For example, in a biomedical application, y i 1 and y i 2 could represent severity of two different symptoms of patient i , recorded on a categorical scale ranging from 'no problem' to 'severe', along with covariate information weight, age, and blood pressure. The assumption that the ordinal responses represent discretized versions of latent continuous responses is natural for many settings, such as the one considered here. Note also that the assumption of random covariates is appropriate in this application, and that the medical measurements are all related and arise from some joint stochastic mechanism. This motivates our focus on building a model for the joint density f ( z x , ), which is a continuous density of dimension k + , which in turn implies a model for the conditional response distribution p f ( y | x ).
To model f ( z x , ) in a flexible way, we use a DP mixture of multivariate normals model, mixing on the mean vector and covariance matrix. That is, we assume ( z i , x i ) | G iid ∼ ∫ N( z i , x i | µ , Σ )d G ( µ , Σ ), and place a DP prior (Ferguson, 1973) on the random mixing distribution G . This DP mixture density regression approach was first proposed by M¨ller u et al. (1996) in the context of continuous variables, and has been more recently studied under different modeling scenarios; see, e.g., M¨ller and Quintana (2010), Park and Dunson u (2010), Taddy and Kottas (2010), Taddy and Kottas (2012), and Wade et al. (2014). Norets and Pelenis (2012) established large support and other theoretical properties for conditional distributions that are induced from finite mixtures of multivariate normals, where the response and covariates may be either discrete or continuous variables.
The hierarchical model is formulated by introducing a latent mixing parameter θ i =
( µ i , Σ i ) for each data vector:
$$( z _ { i }, x _ { i } ) \, | \, \theta _ { i } \stackrel { i n d. } { \sim } N ( \mu _ { i }, \Sigma _ { i } ) ; \ \theta _ { i } \, | \, G \stackrel { i i d } { \sim } G, \ \ i = 1, \dots, n$$
where G | α, ψ ∼ DP( α, G 0 ( · | ψ )), with base (centering) distribution G 0 ( µ , Σ | ψ ) = N( µ | m V , )IW( Σ | ν, S ). The model is completed with hyperpriors on ψ = ( m V S , , ), and a prior on α :
$$m \sim N ( a _ { m }, B _ { m } ), \ \ V \sim I W ( a _ { V }, B _ { V } ), \ \ S \sim W ( a _ { S }, B _ { S } ), \ \ \alpha \sim g a m m a ( a _ { \alpha }, b _ { \alpha } ), \quad ( 2 )$$
where W( a S , B S ) denotes a Wishart distribution with mean a S B S , and IW( a V , B V ) denotes an inverse-Wishart distribution with mean ( a V -( k + ) p -1) -1 B V .
According to its constructive definition (Sethuraman, 1994), the DP generates almost surely discrete distributions, G = ∑ ∞ l =1 p δ l ϑ l . The atoms ϑ l are independent realizations from G 0 , and the weights are determined through stick-breaking from beta distributed random variables with parameters 1 and α . That is, p 1 = v 1 , and p l = v l ∏ l -1 r =1 (1 -v r ), for l = 2 3 , , . . . , with v l iid ∼ beta(1 , α ). The discreteness of the DP allows for ties in the θ i , so that in practice less than n distinct values for the { θ i } are imputed. The data is therefore clustered into a typically small number of groups relative to n , with the number of clusters n ∗ controlled by parameter α , where larger values favor more clusters (Escobar and West, 1995). From the constructive definition for G , the prior model for f ( z x , ) has an almost sure representation as a countable mixture of multivariate normals, and the proposed model can therefore be viewed as a nonparametric extension of the multivariate probit model, albeit with random covariates.
This implies a countable mixture of normal distributions (with covariate-dependent weights) for f ( z | x , G ), from which the latent z may be integrated out to reveal the induced model for the ordinal regression relationships. In general, for a multivariate response Y = ( Y , . . . , Y 1 k ) with an associated covariate vector X , the probability that Y takes on the values l = ( l 1 , . . . , l k ), where l j ∈ { 1 , . . . , C j } , for j = 1 , . . . , k , can be expressed as
$$\Pr ( \boldsymbol Y = l \, | \, x, G ) = \sum _ { r = 1 } ^ { \infty } w _ { r } ( x ) \int _ { \gamma _ { k, l _ { k } - 1 } } ^ { \gamma _ { k, l _ { k } } } \cdots \int _ { \gamma _ { 1, l _ { 1 } - 1 } } ^ { \gamma _ { 1, l _ { 1 } } } N ( z \, | \, m _ { r } ( x ), S _ { r } ) d z$$
with covariate-dependent weights w r ( x ) ∝ p r N( x | µ x r , Σ xx r ), and mean vectors m x r ( ) = µ z r + Σ zx r ( Σ xx r ) -1 ( x -µ x r ), and covariance matrices S r = Σ zz r -Σ zx r ( Σ xx r ) -1 Σ xz r . Here, ( µ r , Σ r ) are the atoms in the DP prior constructive definition, where µ r is partitioned into µ z r and µ x r according to random vectors Z and X , and ( Σ zz r , Σ xx r , Σ zx r , Σ xz r ) are the components of the corresponding partition of covariance matrix Σ r .
To illustrate, consider a bivariate response Y = ( Y , Y 1 2 ), with covariates X . The probability assigned to the event ( Y 1 = l 1 ) ∩ ( Y 2 = l 2 ) is obtained using (3), which involves evaluating bivariate normal distribution functions. However, one may be interested in the marginal relationships between individual components of Y and the covariates. Referring to the example given at the start of this section, we may obtain the probability that both symptoms are severe as a function of X , but also how the first varies as a function of X . The marginal inference, Pr( Y 1 = l 1 | x , G ), is given by the expression
$$\sum _ { r = 1 } ^ { \infty } w _ { r } ( x ) \left \{ \Phi \left ( \frac { \gamma _ { 1, l _ { 1 } } - m _ { r } ( x ) } { s _ { r } ^ { 1 / 2 } } \right ) - \Phi \left ( \frac { \gamma _ { 1, l _ { 1 } - 1 } - m _ { r } ( x ) } { s _ { r } ^ { 1 / 2 } } \right ) \right \}$$
where m r ( x ) and s r are the conditional mean and variance for z 1 conditional on x implied by the joint distribution N( z x , | µ r , Σ r ). Expression (4) provides also the form of the ordinal regression curves in the case of a single response.
Hence, the implied regression relationships have a mixture structure with componentspecific kernels which take the form of parametric probit regressions, and weights which are covariate-dependent. This structure enables inference for non-linear response curves, by favoring a set of parametric models with varying probabilities depending on the location in the covariate space. The limitations of parametric probit models - including relative covariate effects which are constant in terms of the covariate and the ordinal level, monotonicity, and the single-crossing property of the response curves - are thereby overcome.
## 2.2 Model properties
In (1), Σ was left an unrestricted covariance matrix. As an alternative to working with correlation matrices under the probit model, identifiability can be achieved by fixing γ j, 2 (in addition to γ j, 1 ), for j = 1 , . . . , k . As shown here, analogous results can be obtained in
the random covariate setting. In particular, Lemma 1 establishes identifiability for Σ as a general covariance matrix in a parametric ordered probit model with random covariates, assuming all cut-offs, γ j, 1 , . . . , γ j,C j -1 , for j = 1 , . . . , k , are fixed. Here, model identifiability refers to likelihood identifiability of the mixture kernel parameters in the induced model for ( Y , X ) under a single mixture component.
Lemma 1 . Consider a multivariate normal model, N ( Z X , | µ , Σ ) , for the joint distribution of latent responses Z and covariates X , with fixed cut-off points defining the ordinal responses Y through Z . Then, the parameters µ and Σ are identifiable in the implied model for ( Y , X ) , and therefore in the kernel of the implied mixture model for ( Y , X ) , provided C > j 2 , for j = 1 , . . . , k .
Refer to Appendix A for a proof of this result. If C j = 2 for some j , additional restrictions are needed for identifiability, as discussed in Section 2.5.
We emphasize that Lemma 1 does not imply that the parameters of the mixture model are identifiable, rather it establishes identifiability for the kernel of the mixture model. Identifiability is a basic model property for parametric probit models (Hobert and Casella, 1996; McCulloch et al., 2000; Koop, 2003; Talhouk et al., 2012), and is achieved for the special one-component case of our model, a multivariate ordered probit model with random covariates, by fixing the cut-offs. However, this may appear to be a significant restriction, as under the parametric probit setting, fixing all cut-off points would prohibit the model from being able to adequately assign probabilities to the regions determined by the cutoffs. We therefore seek to determine if the nonparametric model with fixed cut-offs is sufficiently flexible to accommodate general distributions for ( Y , X ) and also for Y | X . Kottas et al. (2005) provide an informal argument that the normal DP mixture model for multivariate ordinal responses (without covariates) can approximate arbitrarily well any probability distribution for a contingency table. The basis for this argument is that, in the limit, one mixture component can be placed within each set of cut-offs corresponding to a specific ordinal vector, with the mixture weights assigned accordingly to each cell.
Here, we provide a proof of the full support of our more general model for ordinalcontinuous distributions. In addition to being a desirable property on its own, the ramifications of full support for the prior model are significant, as it is a key condition for the study
of posterior consistency (e.g., Ghosh and Ramamoorthi, 2003). Using the KL divergence to define density neighborhoods, a particular density f 0 is in the KL support of the prior P , if P ( K f glyph[epsilon1] ( 0 )) > 0, for any glyph[epsilon1] > 0, where K f glyph[epsilon1] ( 0 ) = { f : ∫ f 0 ( w ) log( f 0 ( w ) /f ( w ))d w < glyph[epsilon1] } . The KL property is satisfied if any density (typically subject to regularity conditions) in the space of interest is in the KL support of the prior.
It has been established that the DP location normal mixture prior model satisfies the KL property (Wu and Ghosal, 2008). That is, if the mixing distribution G is assigned a DP prior on the space of random distributions for µ , and a normal kernel is chosen such that f ( w | G, Σ ) = ∫ N( w | µ , Σ ) dG ( µ ), with Σ a diagonal matrix, then the induced prior model for densities has full KL support. Letting this induced prior be denoted by P , and modeling the joint distribution of ( X Z , ) with a DP location mixture of normals, the KL property yields:
$$\mathcal { P } \left ( \{ f \colon \int f _ { 0 } ( x, z ) \log ( f _ { 0 } ( x, z ) / f ( x, z ) ) \text{d} x \text{d} z < \epsilon \} \right ) > 0, \text{ \ for any } \epsilon > 0$$
for all densities f 0 ( x z , ) on R p + k that satisfy certain regularity conditions.
To establish the KL property of the prior on mixed ordinal-continuous distributions for ( X Y , ) induced from multivariate continuous distributions for ( X Z , ), consider a generic data-generating distribution p 0 ( x y , ) ∈ D ∗ . Here, D ∗ denotes the space of distributions on R p ×{ 1 , . . . , C 1 } × · · · × { 1 , . . . , C k } . Let f 0 ( x z , ) ∈ D , where D denotes the space of densities on R p + k , such that
$$p _ { 0 } ( x, l _ { 1 }, \dots, l _ { k } ) = \int _ { \gamma _ { k, l _ { k } - 1 } } ^ { \gamma _ { k, l _ { k } } } \cdots \int _ { \gamma _ { 1 }, l _ { 1 } - 1 } ^ { \gamma _ { 1, l _ { 1 } } } f _ { 0 } ( x, z _ { 1 }, \dots, z _ { k } ) \, \mathrm d z _ { 1 } \cdots \mathrm d z _ { k }$$
for any ( l 1 , ..., l k ), where l j ∈ { 1 , . . . , C j } , for j = 1 , . . . , k . That is, f 0 ( x z , ) is the density of a latent continuous distribution which induces the corresponding distribution on the ordinal responses. Note that at least one f 0 ∈ D exists for each p 0 ∈ D ∗ , with one such f 0 described in Appendix B2.
The next lemma (proved in Appendix B1) establishes that, as a consequence of the KL property of the DP mixture of normals (5), the prior assigns positive probability to all KL neighborhoods of p 0 ( x y , ), as well as to all KL neighborhoods of the implied conditional
distributions p 0 ( y | x ). We prove the lemma assuming that p 0 satisfies generic regularity conditions such that the corresponding f 0 is in the KL support of the prior. Hence, the result may be useful for other prior model settings for the latent continuous distribution f ( z x , ). Lemma 2 extends to the regression setting analogous KL support results for discrete and mixed discrete-continuous distributions (Canale and Dunson, 2011, 2015). To ensure the KL property is satisfied for the proposed model, we also include in Appendix B2 a set of conditions for p 0 under which a specific corresponding f 0 satisfies (5), in particular, it satisfies the general conditions for the KL property given in Theorem 2 of Wu and Ghosal (2008).
Lemma 2 . Assume that the distribution of a mixed ordinal-continuous random variable, p 0 ( x y , ) ∈ D ∗ , satisfies appropriate regularity conditions such that for the corresponding density, f 0 ( x z , ) ∈ D , we have P ( K f glyph[epsilon1] ( 0 ( x z , ))) > 0 , for any glyph[epsilon1] > 0 . Then, P ( K p glyph[epsilon1] ( 0 ( x y , ))) > 0 and P ( K p glyph[epsilon1] ( 0 ( y | x ))) > 0 , for any glyph[epsilon1] > 0 .
Lemma 2 establishes full KL support for a model arising from a DP location normal mixture, a simpler version of our model. The KL property reflects a major advantage of the proposed modeling approach in that it offers flexible inference for multivariate ordinal regression while avoiding the need to impute cut-off points, which is a major challenge in fitting multivariate probit models.
The cut-off points can be fixed to arbitrary increasing values, which, for instance, may be equally spaced and centered at zero. As confirmed empirically with simulated data (refer to Section 3 for details), the choice of cut-offs does not affect inferences for the ordinal regression relationships, only the center and scale of the latent variables, which must be interpreted relative to the cut-offs.
## 2.3 Prior specification
To implement the model, we need to specify the parameters of the hyperpriors in (2). A default specification strategy is developed by considering the limiting case of the model as α → 0 + , which results in a single normal distribution for ( Z X , ). This limiting model is essentially the multivariate probit model, with the addition of random covariates. The only covariate information we use here is an approximate center and range for each covariate,
denoted by c x = ( c x 1 , . . . , c x p ) and r x = ( r x 1 , . . . , r x p ). Then c x m and r x m / 4 are used as proxies for the marginal mean and standard deviation of X m . We also seek to center and scale the latent variables appropriately, using the cut-offs. Since Y j is supported on { 1 , . . . , C j } , latent continuous variable Z j must be supported on values slightly below γ j, 1 , up to slightly above γ j,C j -1 . We therefore use r / z j 4, where r z j = ( γ j,C j -1 -γ j, 1 ), as a proxy for the standard deviation of Z j .
Under ( Z X , ) | µ , Σ ∼ N( µ , Σ ), we have E( Z X , ) = a m , and Cov( Z X , ) = a S B S ( ν -d -1) -1 + B V ( a V -d -1) -1 + B m , with d = p + k . Then, assuming each set of cut-offs ( γ j, 0 , . . . , γ j,C j ) are centered at 0, we fix a m = (0 , . . . , 0 , c x ). Letting D = diag { ( r / z 1 4) 2 , . . . , ( r / z k 4) 2 , ( r x 1 / 4) 2 , . . . , ( r x p / 4) 2 } , each of the three terms in Cov( Z X , ) can be assigned an equal proportion of the total covariance, and set to (1 / 3) D , or to (1 / 2) D to inflate the variance slightly. For dispersed but proper priors with finite expectation, ν , a V , and a S can be fixed to d + 2. Fixing these parameters allows for B S and B V to be determined accordingly, completing the default specification strategy for the hyperpriors of m V , , and S .
In the strategy outlined above, the form of Cov( Z X , ) was diagonal, such that a priori we favor independence between Z and X within a particular mixture component. Combined with the other aspects of the prior specification approach, this generally leads to prior means for the ordinal regression curves which do not have any trend across the covariate space. Moreover, the corresponding prior uncertainty bands span a significant portion of the unit interval.
## 2.4 Posterior inference
There exist many MCMC algorithms for DP mixtures, including marginal samplers (e.g., Neal, 2000), slice sampling algorithms (e.g., Kalli et al., 2011), and methods that involve truncation of the mixing distribution G (e.g., Ishwaran and James, 2001). If the inference objectives are limited to posterior predictive estimation, which does not require posterior samples for G , the former two options may be preferable since they avoid truncation of G . However, under the joint modeling approach for the response-covariate distribution, posterior simulation must be extended to G to obtain full inference for regression functionals, and
this requires truncation regardless of the chosen MCMC approach. Since our contribution is on flexible mixture modeling for ordinal regression, we view the choice of the posterior simulation method as one that should balance efficiency with ease of implementation.
The MCMC posterior simulation method we utilize is based on a finite truncation approximation to the countable mixing distribution G , using the DP stick-breaking representation. The blocked Gibbs sampler (Ishwaran and Zarepour, 2000; Ishwaran and James, 2001) replaces the countable sum with a finite sum, G N = ∑ N l =1 p δ l ( µ l , Σ l ) , with ( µ l , Σ l ) i.i.d. from G 0 , for l = 1 , ..., N . Here, the first N -1 elements of p = ( p , ..., p 1 N ) are equivalent to those in the countable representation of G , whereas p N = 1 -∑ N -1 l =1 p l . Under this approach, the posterior samples for model parameters yield posterior samples for G N , and therefore full inference is available for mixture functionals.
The truncation level N can be chosen to any desired level of accuracy, using standard DP properties. In particular, the expectation for the partial sum of the original DP weights, E( ∑ N l =1 p l | α ) = 1 -{ α/ α ( +1) } N , can be averaged over the prior for α to estimate the marginal prior expectation, E( ∑ N l =1 p l ), which is then used to specify N given a tolerance level for the approximation. For instance, for the data example of Section 3.1, we set N = 50, which under a gamma(0 5 0 5) prior for . , . α , yields E( ∑ 50 l =1 p l ) ≈ 0 99994. .
To express the hierarchical model for the data after replacing G with G N , introduce configuration variables ( L , . . . , L 1 n ), such that L i = l if and only if θ i = ( µ l , Σ l ), for i = 1 , ..., n and l = 1 , ..., N . Then, the model for the data becomes
$$y _ { i j } = l \quad \text{iff} \quad \gamma _ { j, l - 1 } < z _ { i j } \leq \gamma _ { j, l }, \quad i = 1, \dots, n, \quad j = 1, \dots, k \\ ( z _ { i }, x _ { i } ) \, | \, \{ ( \mu _ { l }, \Sigma _ { l } ) \, \colon l = 1, \dots, N \}, L _ { i } \stackrel { i n d. } { \sim } N ( \mu _ { L _ { i } }, \Sigma _ { L _ { i } } ), \quad i = 1, \dots, n \\ L _ { i } \, | \, p \stackrel { i i d } { \sim } \sum _ { l = 1 } ^ { N } p _ { l } \delta _ { l } ( L _ { i } ), \quad i = 1, \dots, n \\ ( \mu _ { l }, \Sigma _ { l } ) \, | \, \psi \stackrel { i i d } { \sim } N ( \mu _ { l } \, | \, m, V ) \text{IW} ( \Sigma _ { l } \, | \, \nu, S ), \quad l = 1, \dots, N$$
where the prior density for p is given by α N -1 p α -1 N (1 -p 1 ) -1 (1 -( p 1 + p 2 )) -1 ×··· × (1 -∑ N -2 l =1 p l ) -1 , which is a special case of the generalized Dirichlet distribution. The full model is completed with the conditionally conjugate priors on ψ and α as given in (2).
All full posterior conditional distributions are readily sampled, enabling efficient Gibbs
sampling from the joint posterior distribution of the model above. Conditional on the latent responses z i , we have standard updates for the parameters of a normal DP mixture model. And conditional on the mixture model parameters, each z ij , for i = 1 , . . . , n and j = 1 , . . . , k , has a truncated normal full conditional distribution supported on the interval ( γ j,y ij -1 , γ j,y ij ].
The regression functional Pr( Y = l | x , G ) (estimated by the truncated version of (3) implied by G N ) can be computed over a grid in x at every MCMC iteration. This yields an entire set of samples for ordinal response probabilities at any covariate value x (note that x may include just a portion of the covariate vector or a single covariate). As indicated in (4), in the multivariate setting, we may wish to report inference for individual components of Y over the covariate space.
In some applications, in addition to modeling how Y varies across X , we may also be interested in how the distribution of X changes at different ordinal values of Y . As a feature of the modeling approach, we can obtain inference for f ( x | y , G ), for any configuration of ordinal response levels y . We refer to these inferences as inverse relationships, and illustrate them with the data example of Section 3.2.
Under the multivariate response setting, the association between the ordinal variables may also be a key target of inference. In the social sciences, the correlations between pairs of latent responses, corr( Z , Z r s ), are termed polychoric correlations (Olsson, 1979) when a single multivariate normal distribution is used for the latent responses. Under our mixture modeling framework, we can sample a single corr( Z , Z r s ) at each MCMC iteration according to the corresponding p , providing posterior predictive inference to assess overall agreement between pairs of response variables. As an alternative, and arguably more informative measure of association, we can obtain inference for probability of agreement over each covariate, or probability of agreement at each ordinal level. These inferences can be used to identify parts of the covariate space where there is agreement between response variables, as well as the ordinal values which are associated with higher levels of agreement. In the social sciences it is common to assess agreement among multiple raters or judges who are assigning a grade to the same item. In Section 3.4, we illustrate our methods with such a multirater agreement data example, in which both estimating regression relationships and
modeling agreement are major objectives.
## 2.5 Accommodating binary responses
Our methodology focuses on multivariate ordinal responses with C > j 2, for all j . However, if one or more responses is binary, then the full covariance matrix of the normal mixture kernel for ( Z X , ) is not identifiable. In the univariate probit model, it is standard practice to assume that Z | β ∼ N( x β T , 1), that is, identifiability is achieved by fixing Σ zz . DeYoreo and Kottas (2015) show that a similar restriction can be used to facilitate identifiability in the binary probit model with random covariates, which is built from a multivariate normal distribution for the single latent response and covariates.
Under the general setting involving some binary and some ordinal responses, extending the argument that led to Lemma 1, it can be shown that identifiability is accomplished by fixing the diagonal elements of Σ zz that correspond to the variances of the latent binary responses (the associated covariance elements remain identifiable). To incorporate this restriction, the inverse Wishart distribution for the component of G 0 that corresponds to Σ must be replaced with a more structured specification. To this end, a square-root-free Cholesky decomposition of Σ (Daniels and Pourahmadi, 2002; Webb and Forster, 2008) can be used to fix Σ zz in the setting with a single binary response (DeYoreo and Kottas, 2015), and this is also useful in the multivariate setting. This decomposition expresses Σ in terms of a unit lower triangular matrix B , and a diagonal matrix ∆ , with elements δ i > 0, i = 1 , ..., k + p , such that Σ = B -1 ∆ ( B -1 ) T . The key result is that if ( W ,.. . , W 1 r ) ∼ N( µ B , -1 ∆ ( B -1 ) T ), then this implies Var( W i | W ,.. . , W 1 i -1 ) = δ i , for i = 2 , . . . , r . Therefore, if ( Z X , ) ∼ N( µ B , -1 ∆ ( B -1 ) T ), with ( Z , . . . , Z 1 r ) binary, and ( Z r +1 , . . . , Z k ) ordinal, then fixing δ 1 fixes Var( Z 1 ), fixing δ 2 fixes Var( Z 2 | Z 1 ), and so on. The scale of the latent binary responses may therefore be constrained by fixing δ 1 , the variance of the first latent binary response, Z 1 , and the conditional variances ( δ , . . . , δ 2 r ) of the remaining latent binary responses ( Z , . . . , Z 2 r ). The conditional variances ( δ r +1 , . . . , δ k + p ) are not restricted, since they correspond to the scale of latent ordinal responses or covariates, which are identifiable under our model with fixed cut-offs.
The restricted version of the model for binary responses may be appealing because it
implies that the special case of the mixture model consisting of a single mixture component (the parametric probit model) is identifiable. There is some empirical evidence from the literature that forcing the mixture kernel to be identifiable can improve MCMC performance (e.g., Di Lucca et al., 2013; DeYoreo and Kottas, 2015). To establish full support for the restricted prior model, the result of Lemma 2 would need to be extended, since the prior mixture model for the latent continuous distribution is modified. Alternatively, with sufficient care in MCMC implementation, the mixture model with the unrestricted kernel covariance matrix can be applied with binary responses. Note that Lemma 2 does not require the C j > 2 assumption, that is, it is applicable to settings with a combination of binary and more general ordinal responses.
## 3 Data Examples
The model was extensively tested on simulated data, where the primary goal was to assess how well it can estimate regression functionals which exhibit highly non-linear trends. We also explored effects of sample size, choice of cut-offs, and number of response categories. Some results from simulation studies are presented in the supplementary material.
The effect of sample size was observed in the uncertainty bands for the regression functions, which were reduced in width and became smoother with a larger sample size. The regression estimates captured the truth well in all cases, but were smoother and more accurate with more data, as expected. Even with a five-category ordinal response and only 200 observations, the model was able to capture quite well non-standard regression curves.
As discussed previously, the cut-offs may be fixed to arbitrary increasing values, with the choice expected to have no impact on inference involving the relationship between Y and X . To test this point, the model was fit to a synthetic data set containing an ordinal response with three categories, using cut-off points of ( -∞ -, 20 20 , , ∞ ) and ( -∞ -, 5 5 , , ∞ ), as well as cut-offs ( -∞ , 0 0 1 , . , ∞ ), which correspond to a narrow range of latent Z values producing response value Y = 2. The ordinal regression function estimates were unaffected by the change in cut-offs. The last set of cut-offs forces the model to generate components with small variance (lying in the interval (0 , 0 1)), . but the resulting regression estimates are unchanged from the previous ones.
In the data illustrations that follow, the default prior specification strategy outlined in Section 2.3 was used. The posterior distributions for each component of m were always very peaked compared to the prior. Some sensitivity to the priors was found in terms of posterior learning for hyperparameters V and S , however this was not reflected in the posterior inferences for the regression functions, which displayed little to no change when the priors were altered. Regarding the DP precision parameter α , we noticed a moderate amount of learning for larger data sets, and a small amount for smaller data sets, which is consistent with what has been empirically observed about α in DP mixtures. The prior for α was in all cases chosen to favor reasonably large values, relative to the sample size, for the number of distinct mixture components.
In all data examples, a total of at least 100,000 MCMC samples were taken. The first 5,000 to 10,000 were discarded as burn in, and the remaining draws were thinned to reduce autocorrelation. The models were programmed in R and run on a desktop computer with Intel 3.4 GHz. To provide an idea of how long the algorithm takes to run, for the ozone data example, it took approximately 48 seconds (85 seconds) to produce 1,000 MCMC draws with N = 25 ( N = 75). Additional details related to computation are presented in the supplementary material.
## 3.1 Ozone data
Problems in the environmental sciences provide a broad area of application for which the proposed modeling approach is particularly well-suited. For such problems, it is of interest to estimate relationships between different environmental variables, some of which are typically recorded on an ordinal scale even though they are, in fact, continuous variables. This is also a setting where it is natural to model the joint stochastic mechanism for all variables under study from which different types of conditional relationships can be explored.
To illustrate the utility of our methodology in this context, we work with data set ozone from the 'ElemStatLearn' R package. This example contains four variables: ozone concentration (ppb), wind speed (mph), radiation (langleys), and temperature (degrees Fahrenheit). While these environmental characteristics are all random and interrelated, we
pr(medium)
Figure 1: Ozone data. Posterior mean (solid lines) and 95% interval estimates (dashed lines) for Pr( Y = l | x m , G ) (thick black) compared to Pr( γ l -1 < Z ≤ γ l | x m , G ) (red), for l = 1 2 3 and , , m = 1 2 3, giving the probability that ozone concentration is low, medium, , , and high over covariates radiation, temperature, and wind speed.
focus on estimating ozone concentration as a function of radiation, wind speed, and temperature. To apply our model, rather than using directly the observed ozone concentration, we define an ordinal response containing three ozone concentration categories. Ozone concentration greater than 100 ppb is defined as high (ordinal level 3); this can be considered an extreme level of ozone concentration, as only about 6% of the total of 111 observations are this high. Concentration falling between 50 ppb and 100 ppb (approximately 25% of the observations) is considered medium (ordinal level 2), and values less than 50 ppb are assumed low (ordinal level 1). We use this to illustrate a practical setting in which an ordinal response may arise as a discretized version of a continuous response. Moreover, this example enables comparison of inferences from the ordinal regression model with a model based on the actual continuous responses.
Figure 2: Ozone data. Posterior mean estimates for Pr( Y = l | x , x 1 2 , G ), for l = 1 2 3, , , corresponding to, from left to right, low, medium, and high ozone concentration. The spectrum of colors from white to red indicates probabilities that range from 0 to 1.
The model was applied to the ozone data, with response, Y , given by discretized ozone concentration, and covariates, X = ( X , X , X 1 2 3 ) = (radiation, temperature, wind speed). To validate the inferences obtained from the DP mixture ordinal regression model, we compare results with the ones from a DP mixture of multivariate normals for the continuous vector ( Z, X ), since the observations for ozone concentration, Z , are available on a continuous scale. The latter model corresponds to the density regression approach of M¨ller u et al. (1996), extended with respect to the resulting inferences in Taddy and Kottas (2010). Here, it serves as a benchmark for our ordinal regression model, since it provides the best possible inference that can be obtained under the mixture modeling framework if no loss in information occurs by observing Y rather than Z . We compare the univariate regression curves Pr( Y = l | x m , G ) with Pr( γ l -1 < Z ≤ γ l | x m , G ), for l = 1 2 3, and , , m = 1 2 3, , , the latter based on the mixture model for ( Z, X ). Figure 1 compares posterior mean and 95% interval estimates for the regression curves given each of the three covariates. The key result is that both sets of inferences uncover the same regression relationship trends. The only subtle differences are in the uncertainty bands which are overall slightly wider under the ordinal regression model. Also noteworthy is the fact that the ordinal regression mixture model estimates both relatively standard monotonic regression functions (e.g., for temperature) as well as non-linear effects for radiation.
The ability to capture such a wide range of trends for the regression relationships is a feature of the nonparametric mixture model. Another feature is its capacity to accommo-
date interaction effects among the covariates without the need to incorporate additional terms in the model. Such effects are suggested by the estimates for the response probability surfaces over pairs of covariates; for instance, Figure 2 displays these estimates as a function of radiation and temperature.
## 3.2 Credit ratings of U.S. companies
Here, we consider an example involving Standard and Poor's (S&P) credit ratings for 921 U.S. firms in 2005. The example is taken from Verbeek (2008), in which an ordered logit model was applied to the data, and was also used by Chib and Greenberg (2010) to illustrate an additive cubic spline regression model with a normal DP mixture error distribution. For each firm, a credit rating on a seven-point ordinal scale is available, along with five characteristics. Consistent with the analysis of Chib and Greenberg (2010), we combined the first two categories as well as the last two categories to produce an ordinal response with 5 levels, where higher ratings indicate more creditworthiness. The covariates in this application are book leverage X 1 (ratio of debt to assets), earnings before interest and taxes divided by total assets X 2 , standardized log-sales X 3 (proxy for firm size), retained earnings divided by total assets X 4 (proxy for historical profitability), and working capital divided by total assets X 5 (proxy for short-term liquidity).
The posterior mean estimates for the marginal probability curves, Pr( Y = l | x m , G ), for l = 1 , . . . , 5 and m = 1 , . . . , 5, are shown in Figure 3. These estimates depict some differences from the corresponding ones reported in Chib and Greenberg (2010), which could be due to the additivity assumption of the covariate effects in the regression function under their model. Empirical regression functions - computed by calculating proportions of observations assigned to each class over a grid in each covariate - give convincing graphical evidence that the regression relationships estimated by our model fit the data quite well.
To discuss the estimated regression trends for one of the covariates, consider the standardized log-sales variable, which is a proxy for firm size. The probability of rating level 1 is roughly constant for low log-sales values, and is then decreasing to 0, indicating that small firms have a similar high probability of receiving the lowest rating, whereas the larger the firm, the closer to 0 this probability becomes. The probability curves for levels 2, 3,
Figure 3: Credit rating data. Posterior mean estimates for Pr( Y = l | x m , G ), for each covariate m = 1 , . . . , 5. All five ordinal response curves are displayed in a single panel for each covariate.
and 4 are all quadratic shaped, with peaks occurring at larger log-sales values for higher ratings. Finally, the probability of receiving the highest rating is very small for low to moderate log-sales values, and is increasing for larger values. In summary, log-sales are positively related to credit rating, as expected.
This is another example where it is arguably natural to model the joint distribution for the response and covariates (the specific firm characteristics). This allows our model to accommodate interactions between covariates, as we do not assume additivity or independence in the effects of the covariates on the latent response. In addition to the regression curve estimates, we may obtain inference for the covariate distribution, or for any covariate conditional on a specific ordinal rating. These inverse relationships (discussed in Section 2.4) could be practically relevant in this application. It may be of interest to investors and econometricians to know, for example, approximately how large is a company's leverage, given that it has a rating of 2? Is the distribution of leverage much different from that of a level 3 company? Figure 4 plots the estimated densities for book leverage and standardized log-sales conditional on each of the five ordinal ratings. In general, the distribution of book
Figure 4: Credit rating data. Posterior mean (solid lines) and 95% interval estimates (dashed lines) for covariate densities f ( x | Y = l, G ) conditional on ordinal credit rating l = 1 , ..., 5. The top row corresponds to covariate book leverage, and the bottom row to standardized log-sales.
leverage is centered on smaller values as rating increases, and the densities become more peaked supporting a smaller range of leverage values for higher ratings. The interval bands are slightly wider for the distribution associated with Y = 1 than for Y = 2, 3, or 4, and they are much wider for Y = 5, which is consistent with the small number of firms with a rating of 5. The distribution of log-sales has a mode which occurs at increasing values as rating increases, indicating that if one firm has a higher rating than another, it likely also has higher sales.
## 3.3 Standard and Poor's grades of countries
As a second econometrics example, we consider a data set from Simonoff (2003), comprising S&P ratings of n = 31 countries along with debt service ratio and income, the latter recorded on an ordinal scale with levels of low, medium, and high. Ratings range from 1 to 7, with 1 indicating the best rating of AAA, and 7 the worst of CCC. With two covariates and a very small sample size, this example provides an interesting testbed for our modeling approach.
Since income is available as a discrete variable, W , we model it through the (latent) continuous income variable, Z 2 . Therefore, X represents debt service ratio, and Z =
Figure 5: S&P ratings of countries data. Posterior mean (solid lines) and 95% intervals (gray bands) for the probability response curve associated with each rating, ranging from AAA (level 1) to CCC (level 7), as a function of debt service ratio.
( Z , Z 1 2 ), where W arises from Z 2 just as the ordinal rating response Y arises through latent continuous response Z 1 . This is another application where one may be interested in inverse relationships, such as the distribution of debt service ratio and/or income given a specific S&P rating.
The probability response curves as functions of debt service ratio (Figure 5) contain both monotonic trends (decreasing for response categories 1 and 2, and increasing for category 7), as well as non-linear ones, most notably for categories 4 and 5. The interval bands are wider than in earlier examples, given the smaller sample size. Although not shown here, regression curves can also be obtained over discrete income from Pr( Y = j | W = w, G ) = Pr( Y = j, W = w | G / ) Pr( W = w | G ), for w = 1 2 3 (low, medium, and high income), , , where the numerator contains a double integral of a bivariate normal density function. The probability of receiving a top rating of 1, 2, or 3 is highest for high-income countries, the probability of receiving a moderate rating of 4 or 5 is highest for medium-income countries, and the probability of receiving a poor rating is highest for low-income countries. It is highly unlikely for a country to receive one of the top two ratings unless it is high-income, however there is non-negligible probability of a medium-income country receiving one of the two lowest ratings.
Figure 6: S&P ratings of countries data. Left panel: Posterior densities of latent continuous ratings for Australia (solid line) and Canada (dashed line), the two countries with AA rating. Right panel: Posterior densities of latent continuous ratings for the four countries with A rating: Chile (black), Czech Republic (red), Hungary (green), and Slovenia (blue). In both panels, the gray vertical lines indicate the borders for ordinal ratings.
The latent continuous responses represent latent continuous credit rating in this application. The method for posterior simulation involves sampling z i, 1 , for i = 1 , . . . , 31, which represent the country-specific latent ratings. The two countries with AA (level 2) rating are Canada and Australia. Both of these countries have income classified as high, however Canada has no debt, whereas Australia has a debt service ratio which is around 10. This value is not particularly high, but since higher debt service ratio seems to be associated with poorer ratings, we would expect that Canada would be closer to receiving a better rating of AAA than Australia. Indeed, posterior densities of latent continuous ratings for Canada and Australia (Figure 6, left panel) indicate that Canada's ordinal rating of AA is closer to a AAA rating than Australia's.
Next, consider the four countries with A rating (level 3): Chile, Czech Republic, Hungary, and Slovenia. The first three of these countries are classified as medium income, whereas Slovenia is a high income country. The debt service ratios range from 8.9 (Czech Republic) to 15.8 (Chile). The estimated latent response densities are shown on the right panel of Figure 6. We note that Slovenia's latent rating distribution is centered on values close to the cut-off point for a better rating of AA. The other three distributions are fairly similar. Chile appears closest to a BBB rating, which is consistent with its higher debt service ratio. An interesting observation from these results is that differences in income
appear to have a larger effect on the latent rating distributions than differences in debt service ratio.
## 3.4 Analysis of multirater agreement data
A variety of methods exist for analyzing ordinal data collected from multiple raters when the goal is to measure agreement. Such methods range from the commonly used κ statistic (Cohen, 1960) and its extensions (Fleiss, 1971), which are indices calculated from the observed and expected agreement of raters, to model-based approaches involving log-linear models (Tanner and Young, 1985). We do not attempt a comprehensive review here, rather our focus lies in the use of model-based methods for ordinal responses collected from multiple raters along with covariate information. The proposed multivariate ordinal regression model offers flexibility in this setting in terms of the modeling framework and resulting inferences. We focus on a scenario involving a set of expert graders who evaluate student essays, rating them on an ordinal scale. We contrast our approach to the parametric model of Johnson and Albert (1999), from where the specific data example is taken, and the semiparametric approach of Savitsky and Dalal (2014), both developing Bayesian inference built from modeling for latent responses.
Multirater agreement data arises when k raters assign ordinal scores to n individuals, such that y i = ( y i 1 , . . . , y ik ) collects all scores for the i th individual. The raters typically use the same classification levels, and therefore each y ij ∈ { 1 , . . . , C } . This data structure could be summarized in a contingency table, however, we are concerned with problems in which relevant covariate information is available for each individual. We assume that each judge assigns an ordinal rating to individual i , which represents a discretized version of a continuous rating; that is, y ij is determined by z ij , the continuous latent score assigned by judge j on individual i .
This is in contrast to the formulation of Johnson and Albert (1999), where all judges are assumed to agree on the intrinsic worth of each item, such that z ij = w i + glyph[epsilon1] ij , where w i represents the true latent score, and glyph[epsilon1] ij is the error observed by judge j . Then, w i is linearly related to the covariates, assumed normal with mean containing the term x β T i . The normal latent response distribution is not appropriate when grade distributions are
skewed or favor low/high scores over moderate scores. Since the distribution of the z ij does not have a judge-specific mean, random cut-offs are necessary to allow the ratings of a particular subject to vary among judges.
Savitsky and Dalal (2014) note that the assumption of intrinsic agreement among raters may be inappropriate when raters have different beliefs or perspectives that may influence their scoring behavior. They assume a DP mixture model for the judge-specific latent random vectors built from a mixture kernel defined through independent normals. Dependence is therefore introduced over the latent scores of a single rater (albeit under a restrictive product-kernel for the mixture), but the data vectors arising from each rater are assumed independent. Under this model, it is therefore unclear how to extract inference for inter-rater agreement, which is a key inferential objective for applications of this type in the social sciences.
Similar to the approach in Savitsky and Dalal (2014), there is no notion in our model of an intrinsic true score for an individual. An overall score for an individual could be obtained by averaging in some fashion over the latent scores assigned by each rater. However, our main goal here is to obtain inference about relationships between the ordinal scores and the covariates, as well as for inter-rater agreement over both the covariate space and the scores. Our method offers a potentially useful perspective for modeling multirater agreement data, most notably with respect to the generality of the nonparametric mixture model which can accommodate complex dependence among raters and non-linear relationships with the random covariates.
We apply our method to a problem involving three expert graders who evaluate n = 198 student essays, each assigned a rating on an ordinal scale of 1 through 10 (these represent raters 2, 3, and 4 from the data given in Chapter 5 of Johnson and Albert, 1999). Average word length and total number of essay words are used as the p = 2 covariates, to study if they have an effect on grader ratings. The traditional measure of agreement between raters l and m in the social sciences, the polychoric correlation ρ lm = corr( Z , Z l m ), can be assessed through the covariance mixing parameters ( Σ 1 , . . . , Σ N ). As discussed in Section 2.4, the posterior predictive distribution for ρ lm can be obtained by sampling at each MCMC iteration the corresponding ( ρ lm, 1 , . . . , ρ lm,N ) with probabilities ( p , . . . , p 1 N ).
Figure 7: Multirater data. Posterior mean estimates for probability of high and low rating as a function of average word length (two left plots) and total number of words (two right plots), for raters 1 (solid red lines), 2 (dashed blue lines), and 3 (dotted purple lines).
The polychoric correlation predictive distributions for all three pairs of raters favor more heavily positive correlations (raters 1 and 3 appear to agree most strongly), but place substantial probability on negative correlations. We can study where raters l and m tend to agree or disagree by grouping the latent continuous ratings according to the strength and direction of corr( Z , Z l m ). For instance, a plot of E( z il | data) and E( z im | data) arranged by E(corr( z il , z im ) | data) reveals that raters 1 and 2 strongly agree on very low ratings, but disagree when rater 2 assigns low ratings and rater 1 high ratings. It is also the case for the other pairs of raters that they strongly agree mainly at low scores.
The model provides a variety of inferences in this multivariate ordinal regression example, which can be used to assess how ratings vary across covariates, as well as how raters behave in comparison to one another. Defining a high rating as 8 or higher, and a low rating as 3 or lower, Figure 7 plots estimates for the probability of high and low rating in terms of average word length and total number of words. There appears to be a strong trend in rating as a function of number of words for each rater, with rater 2 in particular assigning higher ratings for essays with more words. The regression curves for high ratings associated with rater 2 are somewhat separated from raters 1 and 3, suggesting that overall rater 2 assigns higher ratings.
To identify regions of the covariate space in which raters tend to agree or disagree, Figure 8 plots estimates for the probability of perfect agreement for the three pairs of
Figure 8: Multirater data. Posterior mean (solid lines) and 95% interval estimates (gray bands) for probability of agreement as a function of total number of words. The left panel corresponds to raters 1 and 2, the middle panel to raters 1 and 3, and the right panel to raters 2 and 3.
raters as a function of total number of words. This inference suggests that raters 1 and 2 agree most strongly on grades for essays with few or many words. The trend in probability of agreement is weaker for the other two pairs of raters.
Finally, to assess the strength of agreement between raters on high and low scores, we study the probability that one rater gives a high/low rating, conditional on the rating given by another rater. Table 1 includes posterior means for Pr( Y l | Y m , G ), for l, m ∈ { 1 2 3 , , } , with Y l and Y m taking values of { 8 9 10 , , } (high) or { 1 2 3 , , } (low). Each row represents the event being conditioned on, while each column represents the event a probability is being assigned to. For example, row 1, column 3 contains the posterior mean for Pr( Y 2 ∈ { 8 9 10 , , } | Y 1 ∈ { 8 9 10 , , } , G ). The cells corresponding to disagreement are highlighted with gray. The first two rows give probabilities conditional on rater 1 scores, and indicate that rater 2 has more disagreement with rater 1 than does rater 3. The last two rows suggest that rater 2 disagrees more with rater 3 than does rater 1. Finally, from the middle two rows, we note slightly more disagreement between raters 1 and 2 than between raters 3 and 2.
## 4 Discussion
Seeking to expand Bayesian nonparametric methodology for ordinal regression, we have presented a fully nonparametric approach to modeling multivariate ordinal responses along
Y 1 = H Y 1 = L Y 2 = H Y 2 = L Y 3 = H Y 3 = L
| Y 1 = H | | | 0 . 54 | 0 . 15 | 0 . 46 | 0 . 06 |
|-----------|--------|--------|----------|----------|----------|----------|
| Y 1 = L | | | 0 . 13 | 0 . 48 | 0 . 04 | 0 . 51 |
| Y 2 = H | 0 . 36 | 0 . 18 | | | 0 . 27 | 0 . 14 |
| Y 2 = L | 0 . 08 | 0 . 56 | | | 0 . 07 | 0 . 41 |
| Y 3 = H | 0 . 63 | 0 . 11 | 0 . 55 | 0 . 18 | | |
| Y 3 = L | 0 . 04 | 0 . 78 | 0 . 15 | 0 . 54 | | |
Table 1: Multirater data. Posterior means for agreement and disagreement conditional probabilities for pairs of raters, with disagreement highlighted in gray (see Section 3.4 for details). H refers to high ratings of { 8 9 10 , , } , and L refers to low ratings of { 1 2 3 , , } .
with covariates. The inferential power of the framework lies in the flexible DP mixture model for the latent responses and covariates. The assumption of random covariates is appropriate for many problems, and modeling the covariates along with the latent responses accounts for dependence or interactions among the covariates. This also allows for inference on functionals of the covariate distribution. By establishing its KL support, we have shown that the prior probability model can accommodate general mixed ordinal-continuous distributions, without imputing cut-off points or restricting the covariance matrix of the normal kernel for the DP mixture model. From a practical point of view, this is a particularly appealing feature of the modeling approach relative to the multivariate probit model and related semiparametric extensions.
The multivariate normal mixture kernel can accommodate any type of continuous covariates (using transformation as needed). Discrete ordinal covariates can also be included by introducing latent continuous variables; this was implemented in the example of Section 3.3, in which the continuous covariate income was recorded on an ordinal scale. In order to handle discrete nominal covariates, the kernel can be modified adding appropriate components to the multivariate normal density, using either a marginal or conditional specification (e.g., Taddy and Kottas, 2010).
The version of the multivariate probit model discussed in the Introduction, and the setting we consider for our model, involves a common vector of covariates X = ( X , ..., X 1 p ) for each response vector Y . That is, the covariates are not specific to particular response variables, but rather ( Y , X ) arises as a multivariate vector. An alternative version of the probit model involves p j covariates ( X j, 1 , . . . , X j,p j ) specific to response variable Y j . This
regression setting was described for multivariate continuous responses by Tiao and Zellner (1964), and this is the version of the multivariate binary probit model considered in Chib and Greenberg (1998).
Scenarios which make use of response specific covariates fall broadly into two categories. The first consists of problems in which only a portion of the covariate vector is thought to affect a particular response, but there may be some overlap in the subset of covariates which generate the responses. Chib and Greenberg (1998) considered a voting behavior problem of this kind in which the first of two responses was assumed to be generated by a subset of the covariates associated with the second response. This data structure can also be accommodated by modeling all covariates X jointly with Y , and conditioning on the relevant subset of X in the regression inferences.
The other type of data structure which is occasionally handled with a multivariate regression model with response specific covariates involves univariate ordinal responses that are related in a hierarchical/dynamic fashion. For instance, Chib and Greenberg (1998) illustrate their model with the commonly used Six Cities data, in which Y = ( Y , . . . , Y 1 4 ) represents wheezing status at ages 7 through 10. Such settings are arguably more naturally approached through hierarchical/dynamic modeling. Indeed, a possible extension of the methodology developed here involves dynamic modeling for ordinal regression relationships, such that at any particular time point a unique regression relationship is estimated in a flexible fashion, while dependence is incorporated across time.
## Acknowledgments
The work of the first author was supported by the National Science Foundation under award SES 1131897. The work of the second author was supported in part by the National Science Foundation under award DMS 1310438. The authors wish to thank a reviewer and the editor for constructive feedback and for comments that improved the presentation of the material in the paper.
## Appendix A: Proof of Lemma 1
Under the normal kernel, N( z x , | µ , Σ ), of the DP mixture model for the latent responses, z = ( z , ..., z 1 k ), and covariates, x , the kernel of the implied mixture model for the ordinal responses, y = ( y , ..., y 1 k ), and covariates is given by k ( y x , | µ , Σ ) = ∫ γ k,y k γ k,y k -1 · · · ∫ γ 1 ,y 1 γ 1 ,y 1 -1 N( z x , | µ , Σ )d z 1 . . . d z k . We assume fixed cut-off points, and y j ∈ { 1 , . . . , C j } with C > j 2, for j = 1 , . . . , k .
We establish likelihood identifiability for parameters µ and Σ in k ( y x , | µ , Σ ). That is, from
$$k ( y, x \, | \, \mu _ { 1 }, \Sigma _ { 1 } ) = k ( y, x \, | \, \mu _ { 2 }, \Sigma _ { 2 } ),$$
for all ( y x , ), with y j ∈ { 1 , . . . , C j } and x ∈ R p , we will obtain µ 1 = µ 2 and Σ 1 = Σ 2 .
Marginalizing over y both sides of (7), we obtain N( x | µ x 1 , Σ xx 1 ) = N( x | µ x 2 , Σ xx 2 ), for all x ∈ R p , and thus µ x 1 = µ x 2 ≡ µ x , and Σ xx 1 = Σ xx 2 ≡ Σ xx . We also have from (7) that for each j = 1 , . . . , k , k y ( j | x µ , 1 , Σ 1 ) = k y ( j | x µ , 2 , Σ 2 ), for all y j ∈ { 1 , . . . , C j } and x ∈ R p . Hence, Pr( Y j ≤ l | x µ , 1 , Σ 1 ) = Pr( Y j ≤ l | x µ , 2 , Σ 2 ), for l = 1 , . . . , C j -1, and for all x ∈ R p . That is,
$$\Phi \left ( \frac { \gamma _ { j, l } - \mu _ { 1 } ^ { z _ { j } } - \Sigma _ { 1 } ^ { z _ { j } x } ( \Sigma ^ { x x } ) ^ { - 1 } ( x - \mu ^ { x } ) } { ( \Sigma _ { 1 } ^ { z _ { j } z _ { j } } - \Sigma _ { 1 } ^ { z _ { j } x } ( \Sigma ^ { x x } ) ^ { - 1 } \Sigma _ { 1 } ^ { x z _ { j } } ) ^ { 1 / 2 } } \right ) = \Phi \left ( \frac { \gamma _ { j, l } - \mu _ { 2 } ^ { z _ { j } } - \Sigma _ { 2 } ^ { z _ { j } x } ( \Sigma ^ { x x } ) ^ { - 1 } ( x - \mu ^ { x } ) } { ( \Sigma _ { 2 } ^ { z _ { j } z _ { j } } - \Sigma _ { 2 } ^ { z _ { j } x } ( \Sigma ^ { x x } ) ^ { - 1 } \Sigma _ { 2 } ^ { x z _ { j } } ) ^ { 1 / 2 } } \right ),$$
for all x ∈ R p , and for l = 1 , . . . , C j -1. Because Φ( ) is · an increasing function, its arguments in equation (8) must be equal. The resulting equation can be expressed in the form a x T + b = 0. For this equation to hold true for any x , we must have a = 0 and b = 0, which yield
$$\frac { \Sigma _ { 1 } ^ { z _ { j } x } } { ( \Sigma _ { 1 } ^ { z _ { j } z _ { j } } - \Sigma _ { 1 } ^ { z _ { j } x } ( \Sigma ^ { x x } ) ^ { - 1 } \Sigma _ { 1 } ^ { x z _ { j } } ) ^ { 1 / 2 } } = \frac { \Sigma _ { 2 } ^ { z _ { j } x } } { ( \Sigma _ { 2 } ^ { z _ { j } z _ { j } } - \Sigma _ { 2 } ^ { z _ { j } x } ( \Sigma ^ { x x } ) ^ { - 1 } \Sigma _ { 2 } ^ { x z _ { j } } ) ^ { 1 / 2 } },$$
and
$$\frac { \gamma _ { j, l } - \mu _ { 1 } ^ { z _ { j } } + \Sigma _ { 1 } ^ { z _ { j } x } ( \Sigma ^ { x x } ) ^ { - 1 } \mu ^ { x } } { ( \Sigma _ { 1 } ^ { z _ { j } z _ { j } } - \Sigma _ { 1 } ^ { z _ { j } x } ( \Sigma ^ { x x } ) ^ { - 1 } \Sigma _ { 1 } ^ { x z _ { j } } ) ^ { 1 / 2 } } = \frac { \gamma _ { j, l } - \mu _ { 2 } ^ { z _ { j } } + \Sigma _ { 2 } ^ { z _ { j } x } ( \Sigma ^ { x x } ) ^ { - 1 } \mu ^ { x } } { ( \Sigma _ { 2 } ^ { z _ { j } z _ { j } } - \Sigma _ { 2 } ^ { z _ { j } x } ( \Sigma ^ { x x } ) ^ { - 1 } \Sigma _ { 2 } ^ { x z _ { j } } ) ^ { 1 / 2 } },$$
for l = 1 , . . . , C j -1. Using (9), (10) can be expressed as ( γ j,l -µ z j 1 ) Σ z x j 2 = ( γ j,l -µ z j 2 ) Σ z x j 1 , for each j = 1 , ..., k . Working with 2 of these C j -1 equations, the system can be shown to have solution Σ z x j 1 = Σ z x j 2 ≡ Σ z x j . Then, from (10) and (9), we obtain µ z j 1 = µ z j 2 ≡ µ z j
and Σ z z j j 1 = Σ z z j j 2 ≡ Σ z z j j , respectively.
Notice that we required 2 of the C j -1 equations of the form in (10) to arrive at this solution. If C j = 2 for some j , we are unable to identify the full covariance matrix Σ . In this case, if we fix Σ z z j j , we can identify µ z j and Σ z x j , as in DeYoreo and Kottas (2015). Although we do not require free cut-offs here due to the flexibility provided by the mixture, if C > j 3, the cut-off points γ j, 3 , . . . , γ j,C j -1 are also identifiable.
glyph[negationslash]
glyph[negationslash]
Finally, we must establish identifiability for Σ z z j j ′ , where j = j ′ . Note that (7) implies k y , y ( j j ′ | µ 1 , Σ 1 ) = k y , y ( j j ′ | µ 2 , Σ 2 ), for any j, j ′ ∈ { 1 , . . . , k } , with j = j ′ . Hence, for any y j ∈ { 1 , . . . , C j } and y j ′ ∈ { 1 , . . . , C j ′ } , ∫ γ j ,y j ′ ′ γ j ,y j ′ ′ -1 ∫ γ j,y j γ j,y j -1 N(( z , z j j ′ ) T | ( µ z j , µ z j ′ ) T , V 1 ) d z j d z j ′ = ∫ γ j ,y j ′ ′ γ j ,y j ′ ′ -1 ∫ γ j,y j γ j,y j -1 N(( z , z j j ′ ) T | ( µ z j , µ z j ′ ) T , V 2 ) d z j d z j ′ . Here, matrices V 1 and V 2 have the same diagonal elements (given by Σ z z j j and Σ z j ′ z j ′ ) and off-diagonal element given by Σ z z j j ′ 1 and Σ z z j j ′ 2 , respectively.
glyph[negationslash]
We will therefore obtain Σ z z j j ′ 1 = Σ z z j j ′ 2 , for j = j ′ , if the bivariate normal distribution function is increasing in the correlation parameter ρ (assuming, without loss of generality, zero means and unit variances). To this end, note that ∂ ∂ρ N(( x, y ) T | 0 , R ) = ∂ 2 ∂x∂y N(( x, y ) T | 0 , R ), where R has unit diagonal elements, and off-diagonal element ρ (e.g., Plackett, 1954). Therefore, ∂ ∂ρ ∫ a -∞ ∫ b -∞ N(( x, y ) T | 0 , R ) d x y d = N(( a, b ) T | 0 , R ) > 0, and thus the result is obtained.
## Appendix B: Kullback-Leibler support for the model
We first prove Lemma 2 assuming (as in the statement of the lemma) that the true ordinalcontinuous distribution p 0 ( x y , ) ∈ D ∗ satisfies appropriate regularity conditions which imply that the corresponding density f 0 ( x z , ) ∈ D is in the KL support of the prior. This result is general and it applies to any prior model for the latent continuous distribution which provides full KL support to f 0 . Next, we provide a specific example for an f 0 that enables the connection with p 0 given in (6). And, using Theorem 2 of Wu and Ghosal (2008), we derive conditions for p 0 under which the corresponding f 0 is in the KL support of the DP location normal mixture prior model, thus validating Lemma 2 for our model.
## B1. Proof of Lemma 2
Let KL( f , f 0 ) = ∫ f 0 ( w ) log( f 0 ( w /f w ) ( ))d w be the KL divergence between densities f 0 and f . Based on the chain rule for relative entropy,
$$K L ( f _ { 0 } ( x, z ), f ( x, z ) ) = K L ( f _ { 0 } ( x ), f ( x ) ) + K L ( f _ { 0 } ( z \, | \, x ), f ( z \, | \, x ) )$$
where the KL divergence between conditional densities is defined as KL( f 0 ( z | x ) , f ( z | x )) = ∫ f 0 ( x ) { ∫ f 0 ( z | x ) log( f 0 ( z | x ) /f ( z | x ))d z } d x . Hence, KL( f 0 ( x ) , f ( x )) ≤ KL( f 0 ( x z , ) , f ( x z , )), and thus, for any glyph[epsilon1] > 0, K f glyph[epsilon1] ( 0 ( x z , )) ⊆ K f glyph[epsilon1] ( 0 ( x )) = { f ( x z , ) : KL( f 0 ( x ) , f ( x )) < glyph[epsilon1] } . Using the KL property of the prior model for f ( x z , ), P ( K f glyph[epsilon1] ( 0 ( x ))) ≥ P ( K f glyph[epsilon1] ( 0 ( x z , ))) > 0, such that the prior assigns positive probability to all KL neighborhoods of the true covariate density f 0 ( x ).
The proof relies on the following inequality for two densities g 1 ( t ) and g 2 ( t ), where t ∈ R s , and for general subsets A , ..., A 1 s of R ,
$$\int _ { A _ { s } } \cdots \int _ { A _ { 1 } } g _ { 1 } ( t ) \log \left ( \frac { g _ { 1 } ( t ) } { g _ { 2 } ( t ) } \right ) \mathrm d t \geq \int _ { A _ { s } } \cdots \int _ { A _ { 1 } } g _ { 1 } ( t ) \mathrm d t \times \log \left ( \frac { \int _ { A _ { s } } \cdots \int _ { A _ { 1 } } g _ { 1 } ( t ) \mathrm d t } { \int _ { A _ { s } } \cdots \int _ { A _ { 1 } } g _ { 2 } ( t ) \mathrm d t } \right ).$$
To prove the inequality, let B r = ∫ A s ··· ∫ A 1 g r ( t )d , for t r = 1 2, such that , h r ( t ) = g r ( t ) /B r , r = 1 2, are densities on , A 1 × ... × A s . Then, the left-hand-side of (11) can be written as B 1 ∫ A s ··· ∫ A 1 h 1 ( t ) log ( B h 1 1 ( t ) B h 2 2 ( t ) ) d t = B 1 log( B 1 B 2 ) + B 1 ∫ A s ··· ∫ A 1 h 1 ( t ) log ( h 1 ( t ) h 2 ( t ) ) d t ≥ B 1 log( B 1 B 2 ), since ∫ A s · · · ∫ A 1 h 1 ( t ) log ( h 1 ( t ) /h 2 ( t )) d t is the KL divergence for densities h 1 and h 2 .
Now, consider density f ( x z , ) ∈ K glyph[epsilon1]/ 2 ( f 0 ( x z , )). By the chain rule, KL( f 0 ( x ) , f ( x )) < glyph[epsilon1]/ 2, and KL( f 0 ( z | x ) , f ( z | x )) < glyph[epsilon1]/ 2. We apply (11) with s = k , g 1 ( t ) = f 0 ( z | x ), g 2 ( t ) = f ( z | x ), and A j = ( γ j,y j -1 , γ j,y j ), for j = 1 , . . . , k . This yields
$$\int _ { \gamma _ { k, \gamma _ { k } - 1 } } ^ { \gamma _ { k, \gamma _ { k } } } \cdots \int _ { \gamma _ { 1, \gamma _ { 1 } - 1 } } ^ { \gamma _ { 1, \gamma _ { 1 } } } f _ { 0 } ( z \, | \, x ) \log \left ( \frac { f _ { 0 } ( z \, | \, x ) } { f ( z \, | \, x ) } \right ) \mathrm d z \geq p _ { 0 } ( y \, | \, x ) \log \left ( \frac { p _ { 0 } ( y \, | \, x ) } { p ^ { * } ( y \, | \, x ) } \right ) \quad \ \ \ ( 1 2 )$$
for any configuration of values y = ( y , ..., y 1 k ) for the ordinal responses. Here, p ∗ ( y | x ) = ∫ γ k,y k γ k,y k -1 · · · ∫ γ 1 ,y 1 γ 1 ,y 1 -1 f ( z | x )d z , such that p ∗ ( x y , ) = f ( x ) p ∗ ( y | x ). Next, we sum both sides of (12) over y , then multiply both sides by f 0 ( x ), and finally integrate both sides over x ,
to obtain
$$\int _ { \mathbb { R } ^ { p } } f _ { 0 } ( x ) \int _ { \mathbb { R } ^ { k } } f _ { 0 } ( z \, | \, x ) \log \left ( \frac { f _ { 0 } ( z \, | \, x ) } { f ( z \, | \, x ) } \right ) \mathrm d z \mathrm d x \geq \int _ { \mathbb { R } ^ { p } } f _ { 0 } ( x ) \sum _ { y _ { k } = 1 } ^ { C _ { k } } \cdots \sum _ { y _ { 1 } = 1 } ^ { C _ { 1 } } p _ { 0 } ( y \, | \, x ) \log \left ( \frac { p _ { 0 } ( y \, | \, x ) } { p ^ { * } ( y \, | \, x ) } \right ) \mathrm d x$$
that is, KL( f 0 ( z | x ) , f ( z | x )) ≥ KL( p 0 ( y | x ) , p ∗ ( y | x )). Since KL( f 0 ( z | x ) , f ( z | x )) < glyph[epsilon1]/ 2, we have KL( p 0 ( y | x ) , p ∗ ( y | x )) < glyph[epsilon1]/ 2, which in conjunction with KL( f 0 ( x ) , f ( x )) < glyph[epsilon1]/ 2, further implies KL( p 0 ( x y , ) , p ∗ ( x y , )) < glyph[epsilon1] , by the chain rule. We have thus obtained
$$K _ { \epsilon / 2 } ( f _ { 0 } ( x, z ) ) \subseteq K _ { \epsilon / 2 } ( p _ { 0 } ( y \, | \, x ) ) \quad \text{and} \quad K _ { \epsilon / 2 } ( f _ { 0 } ( x, z ) ) \subseteq K _ { \epsilon } ( p _ { 0 } ( x, y ) )$$
where K p glyph[epsilon1] ( 0 ( x y , )) = { f ( x z , ) : KL( p 0 ( x y , ) , p ∗ ( x y , )) < glyph[epsilon1] } , from which the result emerges using the KL property of the prior model for f ( x z , ).
## B2. Density f 0 and conditions for the true p 0
We first show that there exists at least one f 0 ( x z , ) ∈ D for any p 0 ( x y , ) ∈ D ∗ , as defined in (6). Consider y j ∈ { 1 , . . . , C j } , for j = 1 , ..., k , and define
$$f _ { 0 } ( x, z ) = \sum _ { y _ { 1 } } \cdots \sum _ { y _ { k } } \frac { p _ { 0 } ( x, y _ { 1 }, \dots, y _ { k } ) \prod _ { j = 1 } ^ { k } 1 _ { ( \gamma _ { j, y _ { j } } ^ { * }, \gamma _ { j, y _ { j } } ^ { * } ) } ( z _ { j } ) } { \prod _ { j = 1 } ^ { k } ( \gamma _ { j, y _ { j } } ^ { * } - \gamma _ { j, y _ { j } - 1 } ^ { * } ) } }$$
where γ ∗ j,l = γ j,l if l ∈ { 1 , . . . , C j - } 1 , γ ∗ j, 0 = b j , and γ ∗ j,C j = d j , with -∞ < b j < γ j, 1 and γ j,C j -1 < d j < ∞ , for j = 1 , . . . , k . Then, f 0 ( x z , ) satisfies the relationship in (6).
The final item needed to show that Lemma 2 is applicable to our model is to develop conditions for the true p 0 under which the corresponding f 0 in (13) satisfies (5). To this end, we use Theorem 2 of Wu and Ghosal (2008), which provides nine conditions (B1 to B9) on the true density f 0 and on the kernel of a nonparametric mixture model such that the KL property is satisfied. Conditions B1, B2, B3 and B9 involve the mixture kernel, and as discussed in Wu and Ghosal (2008), are satisfied by the product normal kernel. Condition B8 involves the prior for the mixing distribution and is valid for the DP prior.
Conditions A, B, C and D stated below for the true p 0 are sufficient for the remaining conditions in Theorem 2 of Wu and Ghosal (2008) to hold. For simpler notation, we
consider the special case of the model with a single ordinal response and a single covariate, such that p 0 ( x, y ) is a distribution on R ×{ 1 , ..., C } , with C ≥ 2. The conditions can be extended to the general setting with multiple ordinal responses and covariates.
- A. There exists 0 < M < ∞ such that 0 < p 0 ( x, y ) ≤ M , for all ( x, y ) ∈ R ×{ 1 , ..., C } .
- B. | ∫ p 0 ( x, y ) log( p 0 ( x, y ) /W ) d x < | ∞ , for all y ∈ { 1 , ..., C } , and for any W > 0.
- C. There exists δ > 0 such that ∫ p 0 ( x, y ) log( p 0 ( x, y ) /ϕ δ ( x, y )) d x < ∞ , for all y ∈ { 1 , ..., C } , where ϕ δ ( x, y ) = inf | x ′ - | x <δ p 0 ( x , y ′ ).
- D. There exist η > 0 and d > 0 such that ∫ p 0 ( x, y )( x 2 + d 2 ) η d x < ∞ and ∫ p 0 ( x, y ) x 2 ( x 2 + d 2 ) η d x < ∞ , for all y ∈ { 1 , ..., C } . In addition, ∫ p 0 ( x, y ) { ( x -a ) 2 / (2 v 2 ) } d x < ∞ , for all y ∈ { 1 , ..., C } , and for any a ∈ R and v > 0.
It is straightforward to verify that B4, B5, B6 and B7 in Theorem 2 of Wu and Ghosal (2008) are satisfied by f 0 in (13) under conditions A, B, C and D, respectively. The only additional assumption needed is that the cut-off points take positive values, which is without loss of generality under our model that uses fixed cut-offs.
## Supplementary Materials
All supplemental files listed below are contained in a single .zip file (supplementary-files.zip) and can be obtained via a single download.
Ozone data: ozone data file. (ozone.csv)
Credit rating data: credit rating data. (credit.dat)
Countries data: S&P ratings of countries. (countries.txt)
Multirater data: data used in the ratings of student essays example. (essay.csv)
Code for ozone data: Code for implementing ordinal regression model on ozone data. (ozone-code-paper, R file)
Code for credit rating data: Code for implementing ordinal regression model on credit rating data. (credit-code-paper, R file)
Code for multirater data: Code for implementing ordinal regression model on multirater data. (multirater-code-paper, R file)
MCMC supplement: Additional material related to MCMC and computation. (supplementarymcmc.pdf)
## References
Albert, J. and Chib, S. (1993), 'Bayesian analysis of binary and polychotomous response data.' Journal of the American Statistical Association , 88, 669-679.
Bao, J. and Hanson, T. (2015), 'Bayesian nonparametric multivariate ordinal regression,' Canadian Journal of Statistics , 43, 337-357.
Basu, S. and Mukhopadhyay, S. (2000), 'Bayesian analysis of binary regression using symmetric and asymmetric links,' The Indian Journal of Statistics Series B , 62, 372-387.
Boes, S. and Winkelmann, R. (2006), 'Ordered response models,' Advances in Statistical Analysis , 90, 165-179.
Canale, A. and Dunson, D. B. (2011), 'Bayesian kernel mixtures for counts,' Journal of the American Statistical Association , 106, 1528-1539.
- -(2015), 'Bayesian multivariate mixed-scale density estimation,' Statistics and Its Interface , 8, 195-201.
- Chen, M. and Dey, D. (2000), 'Bayesian analysis for correlated ordinal data models,' in Generalized Linear Models: A Bayesian Perspective , eds. Dey, D., Ghosh, S., and Mallick, B., New York: Marcel Dekker, pp. 135-162.
- Chib, S. and Greenberg, E. (1998), 'Analysis of multivariate probit models,' Biometrika , 85, 347-361.
- -(2010), 'Additive cubic spline regression with Dirichlet process mixture errors,' Journal of Econometrics , 156, 322-336.
Choudhuri, N., Ghosal, S., and Roy, A. (2007), 'Nonparametric binary regression using a Gaussian process prior,' Statistical Methodology , 4, 227-243.
- Cohen, J. (1960), 'A coefficient of agreement for nominal scales,' Educational and Psychological Measurement , 20, 37-46.
Daniels, M. and Pourahmadi, M. (2002), 'Bayesian analysis of covariance matrices and dynamic models for longitudinal data,' Biometrika , 89, 553-566.
| DeYoreo, M. and Kottas, A. (2015), 'A fully nonparametric modeling approach to binary regression,' Bayesian Analysis , 10, 821-847. |
|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Di Lucca, M., Guglielmi, A., M¨ller, u P., and Quintana, F. (2013), 'A Simple Class of Bayesian Autoregression Models,' Bayesian Analysis , 8, 63-88. |
| Dunson, D. and Bhattacharya, A. (2010), 'Nonparametric Bayes regression and classifica- tion through mixtures of product kernels,' Bayesian Statistics , 9, 145-164. |
| Escobar, M. and West, M. (1995), 'Bayesian density estimation and inference using mix- tures,' Journal of the American Statistical Association , 90, 577-568. |
| Ferguson, T. (1973), 'A Bayesian analysis of some nonparametric problems,' The Annals of Statistics , 1, 209-230. |
| Fleiss, J. (1971), 'Measuring nominal scale agreement among many raters,' Psychological Bulletin , 76, 378-382. |
| Ghosh, J. and Ramamoorthi, R. (2003), Bayesian Nonparametrics , New York: Springer. |
| Gill, J. and Casella, G. (2009), 'Nonparametric priors for ordinal Bayesian social science models: Specification and estimation,' Journal of the American Statistical Association , 104, 453-464. |
| Hannah, L., Blei, D., and Powell, W. (2011), 'Dirichlet process mixtures of generalized linear models,' Journal of Machine Learning Research , 1, 1-33. |
| Hobert, J. and Casella, G. (1996), 'The effect of improper priors on Gibbs sampling in hierarchical linear mixed models,' Journal of the American Statistical Association , 61, 1461-1473. |
| Imai, K. and van Dyk, D. (2005), 'A Bayesian analysis of the multivariate probit model using marginal data augmentation,' Journal of Econometrics , 124, 311-334. |
| Ishwaran, H. and James, L. (2001), 'Gibbs sampling methods for stick-breaking priors,' Journal of the American Statistical Association , 96, 161-173. |
| Ishwaran, H. and Zarepour, M. (2000), 'Markov chain Monte Carlo in approximate Dirich- let and Beta two-parameter process hierarchical models,' Biometrika , 87, 371-390. |
| Johnson, V. and Albert, J. (1999), Ordinal Data Modeling , New York: Springer. |
| Kalli, M., Griffin, J., and Walker, S. (2011), 'Slice Sampling Mixture Models,' Statistics and Computing , 21, 93-105. |
| Koop, G. (2003), Bayesian Econometrics , Chichester: John Wiley and Sons. |
| Kottas, A., M¨ller, u P., and Quintana, F. (2005), 'Nonparametric Bayesian modelling for multivariate ordinal data,' Journal of Computational and Graphical Statistics , 14, 610- 625. |
| Lawrence, E., Bingham, D., Liu, C., and Nair, V. (2008), 'Bayesian inference for multi- variate ordinal data using parameter expansion,' Technometrics , 50, 182-191. |
|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Liu, C. (2001), 'Bayesian analysis of multivariate probit models - discussion on the art of data augmentation by Van Dyk and Meng,' Journal of Computational and Graphical Statistics , 10, 75-81. |
| Liu, J. and Wu, Y. (1999), 'Parameter expansion for data augmentation,' Journal of the American Statistical Assocation , 94, 1264-1274. |
| McCulloch, P., Polson, N., and Rossi, P. (2000), 'A Bayesian analysis of the multinomial probit model with fully identified parameters,' Journal of Econometrics , 99, 173-193. |
| Mukhopadyay, S. and Gelfand, A. (1997), 'Dirichlet process mixed generalized linear mod- els,' Journal of the American Statistical Association , 92, 633-639. |
| M¨ller, u P., Erkanli, A., and West, M. (1996), 'Bayesian curve fitting using multivariate normal mixtures,' Biometrika , 83, 67-79. |
| M¨ller, u P. and Quintana, F. (2010), 'Random partition models with regression on covari- ates,' Journal of Statistical Planning and Inference , 140, 2801-2808. |
| Neal, R. (2000), 'Markov Chain Sampling Methods for Dirichlet Process Mixture Models,' Journal of Computational and Graphical Statistics , 9, 249-265. |
| Newton, M., Czado, C., and Chappell, R. (1996), 'Bayesian inference for semiparametric binary regression,' Journal of the American Statistical Association , 91, 142-153. |
| Norets, A. and Pelenis, J. (2012), 'Bayesian modeling of joint and conditional distribu- tions,' Journal of Econometrics , 168, 332-346. |
| Olsson, U. (1979), 'Maximum likelihood estimation of the polychoric correlation coeffi- cient,' Psychometrika , 44, 443-460. |
| Papageorgiou, G., Richardson, S., and Best, N. (2015), 'Bayesian nonparametric models for spatially indexed data of mixed type,' Journal of the Royal Statistical Society, Series B , 77, 973-999. |
| Park, J. H. and Dunson, D. B. (2010), 'Bayesian generalized product partition model,' Statistica Sinica , 20, 1203-1226. |
| Plackett, R. L. (1954), 'A reduction formula for normal multivariate integrals,' Biometrika , 41, 351-360. |
| Savitsky, T. and Dalal, S. (2014), 'Bayesian non-parametric analysis of multirater ordinal data, with application to prioritizing research goals for prevention of suicide,' Journal of the Royal Statistical Society: Series C , 63, 539-557. |
| Sethuraman, J. (1994), 'A constructive definition of Dirichlet priors,' Statistica Sinica , 4, 639-650. |
|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Shahbaba, B. and Neal, R. (2009), 'Nonlinear modeling using Dirichlet process mixtures,' Journal of Machine Learning Research , 10, 1829-1850. |
| Simonoff, J. (2003), Analyzing Categorical Data , New York: Springer-Verlag. |
| Taddy, M. and Kottas, A. (2010), 'A Bayesian nonparametric approach to inference for quantile regression,' Journal of Business and Economic Statistics , 28, 357-369. |
| - (2012), 'Mixture modeling for marked Poisson processes,' Bayesian Analysis , 7, 335- 362. |
| Talhouk, A., Doucet, A., and Murphy, K. (2012), 'Efficient Bayesian inference for multi- variate probit models with sparse inverse correlation matrices,' Journal of Computational and Graphical Statistics , 21, 739-757. |
| Tanner, M. and Young, M. (1985), 'Modeling ordinal scale disagreement,' Psychological Bulletin , 98, 408-415. |
| Tiao, G. and Zellner, A. (1964), 'On the Bayesian estimation of multivariate regression,' Journal of the Royal Statistical Society, Series B , 26, 277-285. |
| Verbeek, M. (2008), A Guide to Modern Econometrics , John Wiley and Sons, 3rd ed. |
| Wade, S., Dunson, D. B., Petrone, S., and Trippa, L. (2014), 'Improving prediction from Dirichlet process mixtures via enrichment,' Journal of Machine Learning Research , 15, 1041-1071. |
| Walker, S. and Mallick, B. (1997), 'Hierarchical generalized linear models and frailty models with Bayesian nonparametric mixing,' Journal of the Royal Statistical Society B , 59, 845-860. |
| Webb, E. and Forster, J. (2008), 'Bayesian model determination for multivariate ordinal and binary data,' Computational Statistics and Data Analysis , 52, 2632-2649. |
| Wu, Y. and Ghosal, S. (2008), 'Kullback Leibler property of kernel mixture priors in Bayesian density estimation,' Electronic Journal of Statistics , 2, 298-331. |
| null |
[
"Maria DeYoreo",
"Athanasios Kottas"
] |
2014-08-05T16:40:06+00:00
|
2016-09-20T15:02:17+00:00
|
[
"stat.ME"
] |
Bayesian Nonparametric Modeling for Multivariate Ordinal Regression
|
Univariate or multivariate ordinal responses are often assumed to arise from
a latent continuous parametric distribution, with covariate effects which enter
linearly. We introduce a Bayesian nonparametric modeling approach for
univariate and multivariate ordinal regression, which is based on mixture
modeling for the joint distribution of latent responses and covariates. The
modeling framework enables highly flexible inference for ordinal regression
relationships, avoiding assumptions of linearity or additivity in the covariate
effects. In standard parametric ordinal regression models, computational
challenges arise from identifiability constraints and estimation of parameters
requiring nonstandard inferential techniques. A key feature of the
nonparametric model is that it achieves inferential flexibility, while avoiding
these difficulties. In particular, we establish full support of the
nonparametric mixture model under fixed cut-off points that relate through
discretization the latent continuous responses with the ordinal responses. The
practical utility of the modeling approach is illustrated through application
to two data sets from econometrics, an example involving regression
relationships for ozone concentration, and a multirater agreement problem.
|
1408.1028v2
|
## A simple proof of the Gaussian correlation conjecture extended to multivariate gamma distributions
T homas R oyen
University of Applied Sciences - Bingen Berlinstrasse 109, D-55411 Bingen, Germany
E-mail: [email protected]
## Abstract
An extension of the Gaussian correlation conjecture (GCC) is proved for multivariate gamma distributions (in the sense of Krishnamoorthy and Parthasarathy). The classical GCC for Gaussian probability measures is obtained by the special case with an integer degree of freedom ν = 1.
Key words and phrases: Probability inequalities, Gaussian correlation conjecture, multivariate gamma distribution.
2010 Mathematics Subject Classification: Primary 60E15
## 1 Introduction
Let P be a probability measure on R n , n > 1, given by a Gaussian density,
$$( 2 \pi ) ^ { - n / 2 } | \Sigma | ^ { - 1 / 2 } \exp ( - \frac { 1 } { 2 } x ^ { \prime } \Sigma ^ { - 1 } x ), \ \ x \in \mathbb { R } ^ { n },$$
with a non-singular covariance matrix Σ . The Gaussian correlation conjecture (GCC) asserts the inequality
$$P ( C _ { 1 } \cap C _ { 2 } ) \geq P ( C _ { 1 } ) P ( C _ { 2 } )$$
for all convex and centrally symmetric sets C , C 1 2 ⊆ R n , (see [2]). The bivariate case was proved in [8]. Further milestones towards a complete proof were the papers [13] and [4]. In [13] the GCC is verified for all centered ellipsoids and for all sufficiently small C , C 1 2 . In [4] a proof is given if only C 1 is a centered ellipsoid. Other recent claims to (very long) proofs can be found in [7] and [9]. The proof presented here is totally different and comparatively short.
According to [13] the GCC is equivalent to
$$P \Big ( \bigcap _ { i = 1 } ^ { n } A _ { i } \Big ) & \geq P \Big ( \bigcap _ { i = 1 } ^ { k } A _ { i } \Big ) P \Big ( \bigcap _ { i = k + 1 } ^ { n } A _ { i } \Big ) \\ \dots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots.$$
with A i = {| X i | ≤ x i } , x , . . . , x 1 n > 0, 1 ≤ k < n , and any N ( n 0 Σ , )-Gaussian vector ( X , . . . , X 1 n ). The inequality (1.2) was independently proved for k = 1 in [5] and [13]. Here, (1.2) is proved for ( X , . . . , X 1 n ) with an n -variate gamma distribution (Γ ( n α, R )-distribution in the sense of Krishnamoorthy and Parthasarathy [6]), defined by its Laplace transform
$$| I _ { n } + R T | ^ { - \alpha }$$
with the identity matrix I n , a non-singular correlation matrix R = ( ρ ij ) and T = diag( t 1 , . . . , t n ), t 1 , . . . , t n ≥ 0. Admissible values for α are 2 α ∈ N , all values 2 α > n -2, n ≥ 2, (for n -2 < 2 α ≤ n -1 see (1.7)) and all α > 0 if the Laplace transform is infinitely divisible, for which sufficient and necessary conditions are found in [1] and [3]. The Γ ( n α, R )-distribution was originally derived from the joint distribution of the diagonal elements of a W n (2 α, R )-Wishart matrix with the Laplace transform | I n +2 RT | -α . The classical GCC is obtained from the special case α = 1 2. A proof of it / could be easily extended inductively for increasing degrees of freedom ν = 2 α by means of convolution integrals. However, the proof provided here also includes non-integer values of 2 α .
We need the non-central gamma probability density function (p.d.f.)
$$g _ { \alpha } ( x, y ) & = e ^ { - y } \sum _ { k = 0 } ^ { \infty } g _ { \alpha + k } ( x ) \frac { y ^ { k } } { k! }, \quad \alpha, x > 0, y \geq 0, \\. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &.. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &.$$
with the central gamma p.d.f. g α ( x ) = (Γ( α )) -1 x α -1 e -x and the corresponding noncentral gamma cumulative distribution function (c.d.f.)
$$& G _ { \alpha } ( x, y ) = e ^ { - y } \sum _ { k = 0 } ^ { \infty } G _ { \alpha + k } ( x ) \frac { y ^ { k } } { k! }, \quad G _ { \alpha } ( x ) = \int _ { 0 } ^ { x } g _ { \alpha } ( \xi ) \, \text{d} \xi. \\ \rightharpoonup. &. &. &.$$
If R is represented by
$$R = \lambda I _ { n } + A A ^ { T } \iff \lambda ^ { - 1 } R = I _ { n } + B B ^ { T }, \ \ B = \lambda ^ { - 1 / 2 } A,$$
with the minimal eigenvalue λ of R and an n × ( n -1) matrix A of rank m ≤ n -1, then the Γ n ( α, R )-p.d.f. and the Γ n ( α, R )-c.d.f., 2 α ∈ N or 2 α > n -2, can be represented by
$$f ( x _ { 1 }, \dots, x _ { n } ; \alpha, R ) = E \left ( \prod _ { j = 1 } ^ { n } \lambda ^ { - 1 } g _ { \alpha } ( \lambda ^ { - 1 } x _ { j }, \frac { 1 } { 2 } b _ { j } S b _ { j } ^ { T } ) \right )$$
and respectively with the rows b j from B and the expectation referring to S , a W n -1 (2 α, I n -1 )-Wishart matrix. These formulas are derived in a slightly more general form in [10] and [11]; they can be verified by a simple calculation using the Laplace transform of the non-central gamma distribution followed by integration over the Wishart (or pseudoWishart) distribution.
$$F ( x _ { 1 }, \dots, x _ { n } ; \alpha, R ) = E \Big { ( } \prod _ { j = 1 } ^ { n } G _ { \alpha } ( \lambda ^ { - 1 } x _ { j }, \frac { 1 } { 2 } b _ { j } S b _ { j } ^ { T } ) \Big { ) } \quad & ( 1. 8 ) \\ \text{ively with the rows } b _ { i } \, & \text{from } B \text{ and the expectation referring to } S, \text{a } W _ { n - 1 } ( 2 \alpha, I _ { n - 1 } ) \text{-}$$
## 2 Proof of the Gaussian correlation conjecture for Γn ( α , R -distributions )
Theorem 1. Let R = ( R 11 R 12 R 21 R 22 ) be a non-singular partitioned n × n -correlation matrix with n i × n i submatrices R ii and R 12 having positive rank. Then, there holds for the Γ ( n α, R ) -c.d.f.
$$F ( x _ { 1 }, \dots, x _ { n } ; \alpha, R ) > F ( x _ { 1 }, \dots, x _ { n _ { 1 } } ; \alpha, R _ { 1 1 } ) F ( x _ { n _ { 1 } + 1 }, \dots, x _ { n } ; \alpha, R _ { 2 2 } ) \quad \ ( 2. 1 )$$
for all positive numbers x , . . . , x 1 n and 2 α ∈ N or 2 α > n -2 .
## Proof. All the matrices
$$R _ { \tau } = \begin{pmatrix} R _ { 1 1 } & \tau R _ { 1 2 } \\ \tau R _ { 2 1 } & R _ { 2 2 } \end{pmatrix}, \quad 0 \leq \tau \leq 1, \quad \int \nolimits \nolimits \nolimits + \int \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimis \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimits \nolimis \nolimits \nolimits \nolimits \nolimits \nolimits \nolimis \nolimits \nolimits \nolimits \nolimits \nolimis \nolimits \nolimits \nolimits \nolimits \nolimits \nolimis \nolimits \nolimis \nolimits \nolimits \nolimits \nolimis \nolimits \nolimits \nolimits \nolimits \nolimis \nolimits \nolimis \nolimits \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nocolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nulimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nopacity \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimmis \nolimis\nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis\nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimimis \nolimimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nu \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimiim \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimsigma \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimIS \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolim\nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimim\nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolimis \nolcomend{pmatrix}$$
are non-singular correlation matrices. Theorem 1 will be proved if we show
$$\frac { \partial } { \partial \tau } F ( x _ { 1 }, \dots, x _ { n } ; \alpha, R _ { \tau } ) > 0, \quad 0 < \tau < 1.$$
By means of the Laplace transform we shall represent the left-hand side of (2.3) by a sum of not identically vanishing non-negative terms.
/negationslash
Let A J denote the submatrices with row and column indices i ∈ J ⊆ { 1 , . . . , n } from any n × n -matrix A . The determinant | I n + RT | is equal to 1 + ∑ J | R J || T J | , ∅ /negationslash = J ⊆ { 1 , . . . , n } . With J = J 1 ∪ J 2 , J 1 = J ∩ { 1 , . . . , n 1 } /negationslash = ∅ , J 2 = J ∩ { n 1 + 1 , . . . , n } = ∅ , R τ,J = ( R J 1 τR J ,J 1 2 τR J ,J 2 1 R J 2 ) , r J ,J 1 2 = rank( R J ,J 1 2 ), and the squared canonical correlations ρ 2 J ,J 1 2 ,i , i = 1 , . . . , r J ,J 1 2 , which are the positive eigenvalues of R -1 2 / J 1 R J ,J 1 2 R -1 J 2 R J ,J 2 1 R -1 2 / J 1 , we find
$$\mathfrak { x } _ { 1 }, \mathfrak { x } _ { 2 } \ \mathfrak { x } _ { 2 }, \mathfrak { x } _ { 1 } \ \mathfrak { x } _ { 1 } \ \mathfrak { y } _ { 1 } \\ | R _ { \tau, J } | & = | R _ { J _ { 1 } } | \ | R _ { J _ { 2 } } | \ | I _ { J _ { 1 } } - \tau ^ { 2 } R _ { J _ { 1 } } ^ { - 1 / 2 } R _ { J _ { 1 }, J _ { 2 } } R _ { J _ { 2 }, J _ { 1 } } ^ { - 1 } R _ { J _ { 1 } } ^ { - 1 / 2 } | \\ & = | R _ { J _ { 1 } } | \ | R _ { J _ { 2 } } | \ \prod _ { i = 1 } ^ { r _ { J _ { 1 }, J _ { 2 } } } ( 1 - \tau ^ { 2 } \rho _ { J _ { 1 }, J _ { 2 }, i } ^ { 2 } ),$$
and
$$\frac { \partial } { \partial \tau } | R _ { \tau, J } | & = | R _ { J _ { 1 } } | \, | R _ { J _ { 2 } } | \, \frac { \partial } { \partial \tau } ^ { r _ { J _ { 1 }, J _ { 2 } } } \prod _ { i = 1 } ^ { r _ { J _ { 1 }, J _ { 2 } } } ( 1 - \tau ^ { 2 } \rho _ { J _ { 1 }, J _ { 2 }, i } ^ { 2 } ) \\ & = - 2 \tau | R _ { \tau, J } | \, \sum _ { i = 1 } ^ { r _ { J _ { 1 }, J _ { 2 } } } \frac { \rho _ { J _ { 1 }, J _ { 2 }, i } ^ { 2 } } { 1 - \tau ^ { 2 } \rho _ { J _ { 1 }, J _ { 2 }, i } ^ { 2 } }.$$
Therefore,
$$\frac { \partial } { \partial \tau } | I _ { n } + R _ { \tau } T | ^ { - \alpha } & = 2 \alpha \tau | I _ { n } + R _ { \tau } T | ^ { - ( \alpha + 1 ) } \sum _ { J } | R _ { \tau, J } | \left ( \sum _ { i = 1 } ^ { r _ { J _ { 1 }, J _ { 2 } } } \frac { \rho _ { J _ { 1 }, J _ { 2 }, i } ^ { 2 } } { 1 - \tau ^ { 2 } \rho _ { J _ { 1 }, J _ { 2 }, i } ^ { 2 } } \right ) \prod _ { j \in J } t _ { j } \\ & = | I _ { n } + R _ { \tau } T | ^ { - ( \alpha + 1 ) } \sum _ { J } c _ { J } ( \tau ) \prod _ { j \in J } t _ { j }, \\ \text{which is the Laplace transform } h ^ { * } ( t _ { 1 } \dots t _ { n } ; \colon \alpha. \, R _ { \tau } ) \text{ of}$$
which is the Laplace transform h ∗ ( t 1 , . . . , t n ; α, R τ ) of
$$h ( x _ { 1 }, \dots, x _ { n } ; \alpha, R _ { \tau } ) = \sum _ { J } c _ { J } ( \tau ) \left ( \prod _ { j \in J } \frac { \partial } { \partial x _ { j } } \right ) f ( x _ { 1 }, \dots, x _ { n } ; \alpha + 1, R _ { \tau } ), \quad ( 2. 5 ) \$$
where
We have to verify that
$$c _ { J } ( \tau ) = 2 \alpha \tau \left | R _ { \tau, J } \right | \sum _ { i = 1 } ^ { r _ { J _ { 1 }, J _ { 2 } } } \frac { \rho _ { J _ { 1 }, J _ { 2 }, i } ^ { 2 } } { 1 - \tau ^ { 2 } \rho _ { J _ { 1 }, J _ { 2 }, i } ^ { 2 } }.$$
$$h ( x _ { 1 }, \dots, x _ { n } ; \alpha, R _ { \tau } ) = \frac { \partial } { \partial \tau } f ( x _ { 1 }, \dots, x _ { n } ; \alpha, R _ { \tau } ).$$
Integration of the left-hand side in (2.4) over [0 , τ ] provides
$$\text{Integration of the left-hand side in (2.4) over [0, \tau ] provides} \\ | I _ { n } + & R _ { \tau } T | ^ { - \alpha } - | I _ { n } + R _ { 0 } T | ^ { - \alpha } \\ & = \int _ { \mathbb { R } _ { + } ^ { n } } \left ( f ( x _ { 1 }, \dots, x _ { n } ; \alpha, R _ { \tau } ) - f ( x _ { 1 }, \dots, x _ { n } ; \alpha, R _ { 0 } ) \right ) \prod _ { j = 1 } ^ { n } e ^ { - t _ { j } x _ { j } } \, d x _ { j } \\ & = \int _ { 0 } ^ { \tau } h ^ { * } ( t _ { 1 }, \dots, t _ { n } ; \alpha, R _ { \vartheta } ) \, d \vartheta \\ & = \int _ { \mathbb { R } _ { + } ^ { n } } \left ( \int _ { 0 } ^ { \tau } h ( x _ { 1 }, \dots, x _ { n } ; \alpha, R _ { \vartheta } ) \, d \vartheta \right ) \prod _ { j = 1 } ^ { n } e ^ { - t _ { j } x _ { j } } \, d x _ { j }, \\ \text{where the change of the order of integration can be justified by Fubini} ^ { \ } s \text{ criterion in the} \\ \tau \, \int _ { \dots } \, \cdots \, \cdots \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \ \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \sqrt { \, \cdcor { \, \sum } } } } } } } } } } } } }$$
where the change of the order of integration can be justified by Fubini's criterion in the following way:
Applying (1.7), the identity
$$\frac { \partial } { \partial x } g _ { \alpha + 1 } ( x, y ) = g _ { \alpha } ( x, y ) - g _ { \alpha + 1 } ( x, y ),$$
decompositions λ -1 ϑ R ϑ = I n + B B ϑ T ϑ -similar to (1.6) - with rows b jϑ in B ϑ , indicator functions e J of J , and the cardinalities | J | of J , we obtain
$$\text{$\text{$c_{J}$} \, \circ \,, \, \text{$a$} \, \cos \, \frac { 1 } { 2 } \, \circ \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \circ \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \circ \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \circ \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \circ \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \circ \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \circ \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \times \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \times \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \times \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \times \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \times \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \times \,, \, \text{$c$} \, \cos \, \frac { 1 } { 2 } \, \times \,, \, \text{$c$} \, \cos \, \delta } \\ & \sim W _ { 1 } ( 2 ( \alpha + 1 ), I _ { n - 1 } ). \\ & \sup \, W _ { 2 } ( 2 ( \alpha + 1 ), I _ { n - 1 } ).$$
S ∼ W n -1 (2( α +1) , I n -1 ).
Therefore, the Laplace transform of | h | is bounded by a linear combination of integrals of the form
$$& \int _ { \mathbb { R } ^ { n } _ { + } } E \left ( \prod _ { j = 1 } ^ { n } \lambda _ { \vartheta } ^ { - 1 } g _ { \alpha + e _ { K } ( j ) } \left ( \lambda _ { \vartheta } ^ { - 1 } x _ { j }, \frac { 1 } { 2 } b _ { j \vartheta } S b _ { j \vartheta } ^ { T } \right ) \right ) \prod _ { j = 1 } ^ { n } e ^ { - t _ { j } x _ { j } } \, \mathrm d x _ { j } = | I _ { n } + R _ { \vartheta } T | ^ { - ( \alpha + 1 ) } \prod _ { i \in K ^ { c } } ( 1 + \lambda _ { \vartheta } t _ { i } ), \\ & K \subset \{ 1 \quad n \} \text{ where the ri\tilde{b}t-hand side of this emotion is obtained in a way similar}$$
K ⊆ { 1 , . . . , n } , where the right-hand side of this equation is obtained in a way similar to the Laplace transform of f ( x , . . . , x 1 n ; α, R ) in (1.7). The coefficients of this linear combination are non-negative continuous functions of ϑ ∈ [0 , τ ]. Integration over [0 , τ ] yields a finite value, which implies (2.7) and (2.6).
Finally, integration over x , . . . , x 1 n in (2.5) leads to
$$\frac { \partial } { \partial \tau } F ( x _ { 1 }, \dots, x _ { n } ; \alpha, R _ { \tau } ) = \sum _ { J } c _ { J } ( \tau ) \left ( \prod _ { j \in J } \frac { \partial } { \partial x _ { j } } \right ) F ( x _ { 1 }, \dots, x _ { n } ; \alpha + 1, R _ { \tau } ) > 0, \\ \intertext { s u n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n D i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i n d i.$$
since the canonical correlations in (2.4) do not identically vanish because of the positive rank of R 12 .
Remarks. By continuity, Theorem 1 holds also for a singular R , at least with ' ≥ ' instead of ' > '.
A modified proof uses the characteristic function ˆ h corresponding to (2.4). It is the continuous limit of the characteristic functions of the signed measures defined by /epsilon1 -1 ( F x , . . . , x ( 1 n ; α, R τ + /epsilon1 ) -F x , . . . , x ( 1 n ; α, R τ ) ) , /epsilon1 → 0. Thus, ˆ h is the characteristic function of the signed measure which is defined by ∂ ∂τ F x , . . . , x ( 1 n ; α, R τ ) and which is the difference of two finite positive measures on R n .
An approximation by a series of univariate integrals for the difference between the two sides in (2.1) is proposed in [12] with identical values x i = x under the additional conditions
$$\text{conditions} & \\ \rho _ { 1 } & = \frac { 2 } { n _ { 1 } ( n _ { 1 } - 1 ) } \sum _ { 1 \leq i < j \leq n _ { 1 } } \rho _ { i j } > 0, \quad \rho _ { 2 } = \frac { 2 } { n _ { 2 } ( n _ { 2 } - 1 ) } \sum _ { n _ { 1 } + 1 \leq i < j \leq n } \rho _ { i j } > 0, \\ \rho ^ { 2 } & = \left ( \frac { 1 } { n _ { 1 } n _ { 2 } } \sum _ { i = 1 } ^ { n _ { 1 } } \sum _ { j = n _ { 1 } + 1 } ^ { n } \rho _ { i j } \right ) ^ { 2 } \leq \rho _ { 1 } \rho _ { 2 }.$$
The error of this approximation tends to zero with a decreasing variability of the correlations within the submatrices R 11 , R 22 , and R 12 . This approximation is recommended in particular for small values of 1 -F x, . . . , x ( ; α, R ). In a modified approximation, ρ 2 is replaced by 1 n n 1 2 ∑ n 1 i =1 ∑ n j = n 1 +1 ρ 2 ij .
Acknowledgement. The author wishes to thank Donald Richards from the Pennsylvania State University for several valuable recommendations which made this paper easier to read.
## References
- [1] Bapat, R.B. (1989). Infinite divisibility of multivariate gamma distributions and M-matrices, Sankhy ¯ 51, 73-78. a
- [2] Das Gupta, S., Eaton, M. L., Olkin, I., Perlman, M., Savage, L.J., and Sobel, M. (1972). Inequalities on the probability content of convex regions for elliptically contoured distributions. Proc. Sixth Berkeley Symp. Math. Statist. Probab. 2, 241-264, Univ. of California Press, Berkeley.
- [3] Griffiths, R.C. (1984). Characterization of infinitely divisible multivariate gamma distributions, J. Multivariate Anal. 15, 13-20.
- [4] Harg´ e, G. (1999). A particular case of correlation inequality for the Gaussian measure. Ann. Probab. 27, 1939-1951.
- [5] Khatri, C.G. (1967). On certain inequalities for normal distributions and their application to simultaneous confidence bounds, Ann. Math. Stat. 38, 1853-1867.
- [6] Krishnamoorthy, A.S., and Parthasarathy, M. (1951). A multivariate gamma-type distribution. Ann. Math. Stat. 22, 549-557.
- [7] Memarian, Y. (2013). The Gaussian correlation conjecture proof, arXiv:1310. 8099v1 [mathPR].
- [8] Pitt, L.D. (1977). A Gaussian correlation inequality for symmetric convex sets. Ann. Probab. 5, 470-474.
- [9] Qingyang, G. (2013). The Gaussian correlation inequality for symmetric convex sets, arXiv:1012.0676v4 [mathPR].
- [10] Royen, T. (1995). On some central and non-central multivariate chi-square distributions, Statist. Sinica 5, 373-397.
- [11] Royen, T. (2007). Integral representations and approximations for multivariate gamma distributions, Ann. Inst. Statist. Math. 59, 499-513.
- [12] Royen, T. (2013). Some upper tail approximations for the distribution of the maximum of correlated chi-square or gamma random variables, Far East J. Theor. Statist. 43, 27-56.
- [13] Schechtman, G., Schlumprecht, T., and Zinn, J. (1998). On the Gaussian measure of the intersection, Ann. Probab. 26, 346-357.
- [14] ˇ id´k, S a Z. (1967). Rectangular confidence regions for the means of multivariate normal distributions, J. Amer. Statist. Assoc. 62, 626-633.
| null |
[
"T. Royen"
] |
2014-08-05T16:41:47+00:00
|
2014-08-13T17:27:59+00:00
|
[
"math.PR",
"math.ST",
"stat.TH",
"60E15"
] |
A simple proof of the Gaussian correlation conjecture extended to multivariate gamma distributions
|
An extension of the Gaussian correlation conjecture (GCC) is proved for
multivariate gamma distributions (in the sense of Krishnamoorthy and
Parthasarathy). The classical GCC for Gaussian probability measures is obtained
by the special case with one degree of freedom.
|
1408.1031v2
|
## Noname manuscript No.
(will be inserted by the editor)
## Text to Multi-level MindMaps: A Novel Method for Hierarchical Visual Abstraction of Natural Language Text
Mohamed Elhoseiny · Ahmed Elgammal
Received: date / Accepted: date
Abstract MindMapping [46] is a well-known technique used in note taking, which encourages learning and studying. MindMapping has been manually adopted to help present knowledge and concepts in a visual form. Unfortunately, there is no reliable automated approach to generate MindMaps from Natural Language text. This work firstly introduces MindMap Multilevel Visualization concept which is to jointly visualize and summarize textual information. The visualization is achieved pictorially across multiple levels using semantic information (i.e. ontology), while the summarization is achieved by the information in the highest levels as they represent abstract information in the text. This work also presents the first automated approach that takes a text input and generates a MindMap visualization out of it. The approach could visualize text documents in multilevel MindMaps, in which a high-level MindMap node could be expanded into child MindMaps. The proposed method involves understanding of the input text and converting it into intermediate Detailed Meaning Representation (DMR). The DMR is then visualized with two modes; Single level or Multiple levels, which is convenient for larger text. The generated MindMaps from both approaches were evaluated based on Human Subject experiments performed on Amazon Mechanical Turk with various parameter settings.
Keywords Text Visualization · Multi-level MindMap Automation
## 1 Introduction
MindMapping was introduced as a visual note taking technique, developed by Tony Buzan in the 1960s [46]. It is a powerful pictorial method for representing knowledge, concepts and ideas. Figure 1 shows a MindMap example that illustrates William Shakespeare's biography. It is not hard to see that by just looking at this picture, it is
Mohamed Elhoseiny, Ahmed Elgammal
Department of Computer Science, Rutgers University
110 Frelinghuysen Road, Piscataway, NJ 08854-8019, USA
E-mail: [email protected], [email protected]
easy to recall many aspects of Shakespeare's life, which is one of the MindMap benefits (i.e. Review, Revision). Creating MindMaps itself is also interesting compared with linear text. It also improves imagination and meanwhile it is fun.
'Linear note taking has served as one of the greatest impediments to learning', says Tony Buzan. However, he claimed that linear note taking constraints the brain's potentials by forcing it to use limited modes of expression [46] and hence it fails to stimulate creativity. According to Buzan [46], the main motivation behind MindMapping technique is that, unlike linear text, using pictorial representation of ideas stimulates both sides of the brain (left and right). However, there are scientific doubts about Buzan's argument on the brain hemispheres. Recently, neurologists at Utah university performed a two-year study on 1000 people [8], which shows that the brain is more complex than this claim of left and right-sided brain. They divided the brain into 7000 regions and they found no evidence that the human subjects had a stronger right or left-sided brain network 1 . Despite that Buzan's justification of MindMaps by brain hemispheres is mere popular psychology and not scientifically justified, there are a lot of studies that shows that the MindMapping is a powerful and effective technique to access information and to encourage learning, studying and reading (e.g. [14,17]).
Fig. 1: Manually Created Shakespeare's life MindMap [23] (best seen electronically)
With today's vast and rapid increase in the use of electronic document readers, smart phone, and tablets, there is an eminent need to develop methods for visualizing and summarizing textual contents. We argue that MindMapping is not only a note-taking tool, but also can be developed to serve as a powerful tool for joint text
1 More discussion on this study is covered by the http://www.theguardian.com/commentisfree/2013/nov/16/left-right-brain -distinction-myth
summarization and visualization. Converting a text paragraph to a MindMap would provide an easier way to visually access the knowledge and ideas in the text.
One of the reasons why many people may not create MindMaps is that it needs a huge mental effort and concentration, specially for MindMapping a big text. Besides, only few people are creative enough to draw good MindMaps. To the best of our knowledge, there is no existing method to build MindMaps from raw text according to the joint text summarization and visualization concept that we develop in this paper. The goal of this work is to build a step in that direction by developing the first framework that takes a text input and generates a MindMap visualization out of it; a former version was presented in [16]. Our framework is constructed to visualize large text documents in multilevel MindMap, in which a high-level MindMap node could be expanded into a child MindMap. Figure 2 shows the MindMap visualization of the Shakespeare example, generated by our method, in a single level and multiple levels ( his life and work information are expanded as two child MindMaps).
## Shakespeare (1564-1616)
ShakesPeare is a great writer in the history of literature:. He was born in 1564. He had 3 children. He was earning his living from the agricultural products. He lived in Strateford. He wrote 38 plays. He loves acting and writing He performed before the queen in 1594. He died in 1616. good
Fig. 2: Top Left: Input text, Top Right: Shakespeare Single-Level (With web image queries has no restriction on images), and Bottom: Multi-Level MindMaps (with web image queries restricted to the set of Clip-Art Images), generated by our Framework and their corresponding input textual description
There are various potential applications for these versions of automatically created MindMaps that we are developing. For example, audio recording of notes can be converted into text that is passed into our framework to generate a MindMap notes. A lecturer can utilize the framework to prepare presentations by converting text to MindMaps. Furthermore, the generated hierarchical MindMaps would help make MindMap presentations more common due to minor effort and time needed to create them. Finally, MindMaps visualization could help users of electronic book readers to better access information through them.
The contributions of this paper are:
- 1. We introduce MindMapping as an approach to jointly visualize and summarize textual information. Visualization is achieved pictorially across multiple levels using semantic information (i.e. ontology), while summarization is achieved by the information in the highest levels. Our work could also be viewed as a method to visually abstract textual information in multiple levels.
- 2. Novel approach to generate Single and Multilevel MindMaps from pure text and an input Ontology.
- 3. Two approaches were proposed for retrieving relevant pictures from the web.
- 4. First comprehensive evaluation of an automated MindMap framework by human subjects.
The rest of this paper is organized as follows. Section 2 presents the related literature. Section 3 to section 7 describe the proposed framework and techniques. Section 8 describes the evaluation procedures and shows the results using different system parameters. Section 9 presents our conclusion and the future work.
## 2 Related Work
Our formulation addresses the problem as a hybrid between text summarization and visualization. In the context of text summarization, the top level MindMap serves as a summarized version/ abstraction of the document. Meanwhile, the detailed information of textual content could be visually accessed through a hierarchy of MindMaps. This section presents related work in each of text summarization, visualization, a connection to Concept/Topic Maps, and MindMapping existing systems.
## 2.1 Text Summarization
Several approaches have been proposed for text summarization [13]; key approaches follow. Conroy and O'leary [11] used Hidden Markov Models (HMMs) to model the problem by sentence extraction/selection from a document considering local dependencies between sentences. Later, Osborne [41] used log-linear models to obviate the feature independence assumption in previous models as he claimed and showed empirically that the system produced better extracts than a naive-Bayes model, with a prior appended to both models. Then, Kaikhah et al [24] trained a Neural Network model from the labels and the features for each sentence of an article, that could infer
the proper ranking of sentences in a test document. Other approaches use deep Natural Language Processing (e.g. [33], where Marcu et al merged the discourse based heuristics with traditional heuristics).
While, text summarization provides a reasonable resolution to the growing information overload, it does not provide an organized access to the document information. This limits its capability to structurally localize the information of user's interest in the given document, which introduces a key motivation for our multilevel MindMap Visualization concept .
## 2.2 Text Visualization
Some of the state-of-the-art text visualization techniques build on tag clouds. Examples include work conducted by IBM in Wordle, which is a web-based text visualization method that creates a tag-cloud-like displays with careful attention to typography, color and decomposition [47]. The generated clouds give greater prominence to words that appear more frequently in the source text. Van Ham et al. [21] presented PhraseNet technique, which diagrams the relationships between different words in the text. It uses a simple form of pattern matching to provide multiple views of the concepts contained in a book, speech, or poem. Recently, Afzal et al [1] integrated spatial constraints to automatically visualize textual information on typograhical maps. There are three significant differences differentiate our work from the aforementioned works. First, their algorithms do not involve deep understanding of the text. Second, each node is visualized with the word itself. However in MindMaps, visual nodes could be mapped into pictures of a concept and its properties (e.g. car with red color attribute). Third, our approach can generate multilevel representation, which is not supported by Wordle, PhraseNet or Afzal's method.
Fig. 3: Wordle (left) [47] and PhraseNet (right) [21].
Topic models and semantic representations have been adopted to help organize and visualize information. In this direction, Feng et al. [18] developed a model for multimodal meaning representation, based on linguistic and visual context. The approach exploits enormous resource of documents and associated images available on the web. A relevant application is the automatic conversion of text to 3D animation. Ma [31] proposed Lexical Visual Semantic Representation (LVSR) which connects linguistic semantics to the visual semantics and is suitable for action execution. This
application involves deep analysis of the text converting it into user defined concept by which it converts the text into actions to be performed on 3D models. In addition, Coyne et al. [12] presented a method that incorporates lexical and real-world knowledge to depict scenes from language. However, these approaches are constrained to limited number of predesigned 3D-models and relations and are not hierarchical. Recently, Dou et al [15] presented an approach to visualize text collections using topic hierarchies. While, their approach provides a reasonable access for visually exporing text corpus, our approach addresses pictorial visualization and abstraction of a chosen textual document by the MindMap Visualization concept. In addition, the explored hierarchy in [15] are topic based, while the hierarchy provided by our visualization is based on semantic abstraction of the textual input content. As an illustration, our approach could be integrated with the approach [15] to visualize a document selected by exploring the topic hierarchies, organized by [15].
## 2.3 Topic Maps and Concept Maps
Concept maps [39] have been used to express people's understanding about a specific topic. For few decades, this tool has attracted people of all ages and different knowledge domains. They include concepts usually enclosed as boxes or circles and relationships between these concepts indicated by a connecting line. Topic maps [32,42] are very similar to Concept Maps, while Topic Maps are standardized. It has been defined as a form of semantic web technology and a lot of work has been done to address interoperability between the W3C's RDF/OWL/SPARQL. The current standard is taking place under the umbrella of the W3C committee.
In contrast to Concept Maps and Topic Maps, MindMapping often includes but is not limited to radial hierarchies (where central ideas is in the center, while connected to related entities around it) and tree structures. Another difference is that a MindMap reflects people's thoughts about a single topic, while Concept Maps/ Topic Maps can present a real or abstract system or a set of concepts, which is not restricted to specific topic.
The majority of the literature has focused on the representation of topic maps and concept maps to enhance some NLP tasks (e.g. parsing, Word Sense Disambiguation [37], etc). Rather, we would like to view these maps as a powerful tool that could be used to jointly visualize and summarize information. In particular, we present the concept of multilevel/hierarchical MindMaps as a notion of visualizing textual information in multiple levels. Higher levels reflects coarse information in the text document, while lower levels present the fine details. We chose to name it multilevel/hierarchical MindMaps because MindMaps are (1) less restrictive on the representation than Concept and Topic Maps, (2) topic-oriented which is relevant in the setting we utilize (i.e. representing information from textual description), and (3) In our experiments, we focus on topics that could be presented as a single central idea, which is consistent with the MindMap notion. However, we argue that most of the topics could be presented a set of central ideas that could be presented by MindMaps. Hence, we don't restrict our definition of Hierarchical MindMaps to topics of a sin-
gle central idea. We focus in this paper in the steps, we designed to generate the multilevel MindMap from a textual description.
## 2.4 Existing MindMapping Systems
There are two ways for creating a MindMap. One way is manually by using a paper, pens and colors, but this requires, beforehand, reading and understanding the text well to design a good MindMap. Another way is to use existing software that just serves as an editing canvas, where user can insert pictures and create relations between them (i.e., manually). Examples include 'I Mind' [40], 'Nova Mind' [45].
Finally, the most relevant work was conducted by Hamdy et al. [22]. They presented a prototype for the MindMap Automation, however, it was just a demonstration on few examples of few sentences and then it was applied in a mobile application in [26]. There are two critical drawbacks in this approach. (1) It is almost an unevaluated demo. (2) It supports only single-level MindMaps. This limitation make the approach incapable of representing information of a larger text. Hence, it can not support our concept of simultaneously visualizing and summarizing textual information.
## 3 Proposed Framework
Our objective is to propose a method to jointly visualize and summarize visual information (we call multilevel MindMap visualization). We begin by presenting the outline of the approach through the data flow diagram in figure 4; we will present the details in later sections.
Fig. 4: Data Flow of the proposed Multilevel MindMap Generation framework
We first extract, what we call Semantically Enriched Parse Trees (SEPTs), from the input text through a Text Semantic Preprocessing step that we implemented. The SEPTs consist of the parse trees for each sentence augmented with disambiguated senses at each terminal node in the parse tree. Furthermore, for each anaphor node (e.g. pronouns), a pointer, to what it refers to, is stored. These SEPTs are then converted into a Detailed Meaning Representation (DMR) that represent all the information in the text as a graph. Then, the DMR is converted to a Multi-Level Meaning Representation (MLMR), which is a tree of meaning representations. Each node in the MLMR tree stores a Meaning Representation (MR) starting from the root node where its MR holds the most abstract information in the text to the leaves where details are presented. The final step in our framework is the MindMap Generation, which converts an input MR into its pictorial representation and dynamically allocate it on the screen. Hence, the MLMR could be visualized as a Multilevel MindMap by traversing the MLMR tree and visualizing the MR of each node. It is worth mentioning that a Single level MindMap could be generated by running the MindMap Generation step directly on the DMR instead of Generating a MindMap for each node in the MLMR. However, If the input text is of a big size, the Single Level MindMap would be quite unclear and unorganized. Figure 5 (middle part) shows an example, where single level Mindmap visualization hinders comprehension speed, simplicity and clarity. However, Multilevel MindMap visualization is much clearer (see figure 5 bottom part).
An ontology is needed for both DMR Generation to understand attributes of objects/subjects while creating the DMR, and for MLMR generation to semantically group related entities in MLMR. Ontology is also used in the MindMap Generation step to determine whether a concepts needs to be pictorially represented; see figure 4. The details of the outlined approach is presented as follows. Section 4 details the Text Semantic Pre-processing (TSP) step. Section 5 presents the DMR generation approach. Section 6 presents the MLMR Generation. Finally, section 7 presents the MindMap Generation step.
## 4 Text Semantic PreProcessing (TSP)
In this step, we extract sentence-wise information from plain text, which we call Semantically Enriched Parse Tree (SEPT). As aforementioned, each SEPT consists of a parse tree, anaphora resolution (discourse analysis) results, and the intended sense for each word in the sentence. In order to compute the SEPTs, four main tasks are performed: (1) morphological analysis [48], (2) parsing [4,10,34] and syntactic structural analysis [27,30], which produces a single parse tree that respects the structure of given sentence. (3) discourse analysis [28,43] outputs anaphora resolution results (4) Finally, Word Sense Disambiguation (WSD) [2,5,37] determines the intended sense for each word in the sentence based on WordNet [36]. Having augmented the parse trees with the senses produced by WSD and the anaphora resolution results, the generation of a SEPT for each sentence in the text is concluded. In this work, we utilized recent approaches for such a well-studied phase in Natural Language Processing, as detailed in [22]. Other methods could be applied as long as they produce the SEPTs.
Maries Topes were born in 1880 in Edinburgh. Her father was an archaeologist. The mother of Maries topes was suffragist. Marie's topes studied at University of London and University of Munich. She became the first female member of the faculty of science in the university in Manchester\_ She shortly married Reginald Ruggles Gates Her first book was It was condemned by churches and the press. Maries Topes then married Humphrey Verdon Roe. She had a son in 1924. She opened a clinic for Planning of family in Holloway and north London in 1921 It was the first in the country . It offered a free service to married women and also gathered data about contraception: She died in 1958
Fig. 5: Top: Textual description of 'Marie Stopes' historical figure, Middle: its corresponding single Level MindMap (Difficult to read), Bottom: Structured Visualization using Multilevel MindMap Visualization of of the same text (Easier to read)
Figure 6 shows TSP output of Shakespeare's example. The bottom left part shows the generated parse trees as the output of the parsing and syntactic structural analysis. The bottom left part shows the discourse analysis output which is anephora resolution. 'HE(2,1)' means the 'HE' word that occurs as the first word of the second
Fig. 6: Text Semantic Preprocessing (Syntactic Analysis, Anaphora Resolution, Word Sense Disambiguation) for Shakespeare Example. The Syntactic Parse trees are augmented with the senses and the discourse analysis results to form SEPTs
statement. The listing in the figure indicates that 'Shakespeare (1,1)', that occurred in the first word of the first statement, is referenced in the succeeding pronouns (i.e. 'He(2,1', 'His(4,4)', etc). The top right part of the figure shows the meaning/sense of each word that occurs in the sentences, while grouped in the data structure for each statement. Augmenting the parse trees with the the discourse information and senses directly produces the SEPTs. We denote the set of SEPTs extracted from the text by P = { p , i i = 1 → N } . We define n be a node in any of the SEPTs, we define n.r as the node that n points to in case it points to a previously declared noun/noun phrase, n.r = NULL otherwise. n.r could point to a node in a possibility different SEPT. We also define n.s as the sense that n means in case it is a terminal node, n.s = NULL otherwise.
## 5 DMR Generation
Designing a meaning representation for Natural Language Text implies understanding its content and representing it in a semantically accessible way. During DMR
Generation, P is converted to a further meaningful representation, which consists of conceptual frames and relations between these them. There are two types of frames in our framework (Entity frames and Action Frames). Entity frames represent entities mentioned in the text (persons, objects etc), while Action Frames represent activities in the text (i.e. verbs). Possible relations between the frames are Case-Role, Domain and Temporal relations. A Case-Role relation represents semantic role of an entity frame in a particular action. For instance, in 'john plays football' sentence, the subject 'john' has an 'agent' case-role in the 'play' action, and the object 'football' frame has a 'theme' case role in the same action. Temporal relations are interpropositional relations that communicates the simultaneity or ordering in time of events or states. Domain relations encode logical connection between action frames (i.e. verbs). Figure 8 illustrates different types of case-roles, domain relations and temporal relations that are designed in our framework. Having detected the frames and the relations, the final output of the DMR generation became a graph, whose nodes are the frames (Action or Entity frames), while edges are the relations. Mathematically speaking, we assume that DMR is a graph G d = ( V E , ) , where E is the set of entity and action frames (i.e. the nodes) and E is the set of relations including Case-Role, Domain, and Temporal Relations (i.e. the edges).
```
}
}
---------------------------------------------------------------------------------------------------------------------
}
\
```
The DMR generation procedure is illustrated in algorithm 1. The set of frames V and relations E are initialized to empty sets. In order to generate the DMR, each p i ∈ P is traversed and the grammatical structure is followed to define subjects, objects, and verbs. Entity frames are created for objects and subjects. While Action Frames are created for verbs. We denote the set of Entity-Frames in the i th sen-
tence by E i = { e , e i 1 i 2 · · · e i end } , and the set of Action-Frame in the same sentence as A i = { a , a i 1 i 2 · · · a i end } . Each frame is augmented with its intended sense from the node information stored in the SEPTs (i.e. n.s ). This step also includes adding attributes of each frame if any. For example, 'small red ball', a 'ball' entity frame is generated with attributes 'small' and 'red'. The attributes were detected using the parse tree structure. In this case, 'small' and 'red' are adjective of 'ball'. In addition, the type of the attributes is determined by the ontology (e.g. red is a color attribute). Simultaneously, the relations (see figure 8) were created following the syntactic structure of the sentences to connect the generated frames. No new frames are created for anaphors, they are linked to the frames they refer exploiting n.r . The newly generated frames for each statement are added to V and they are connected by case-role, domain, and temporal relations following the parse tree structure. These relations are then added to the edges E . We name the function that fills all the frames and relation as 'FillFramesAndRelation', see algorithm 1.
Practically, we implemented the function 'FillFramesAndRelations' by recursively traversing the given SEPT p i starting from the root node. Each node is checked processing each grammar rule in each node in the produced parse trees; This could be possibly any rule that our parser allows. Each rule could be processed differently depending on whether it produces of frames and relations. For instance, Entity Frames are added to E while processing NP node (see figure 6 Syntactic Analyzer Output). Another example is that Domain Relations are added to R D while processing a rule like S = DS (' because ′′ , ' Sothat ′′ | ' So ′′ ) DS , where there are two statement ( DS ) separated by and preposition that indicate a reason on a result (e.g. 'because','So that', and 'So'). Following the same idea, it is not hard to build a function for the rules of interest that to fill the frames and the relations in each statement 2 . A hash table H help retrieve each of these function based on the rule as a key. GetMatchingRuleFunction( p , H ) simply returns the function that processes p 's rule if one exist in H , and NULL otherwise. Having traversed all p i ∈ P by 'FillFramesAndRelations', a complete graph of all the information in text is created which is the DMR (i.e. G d = ( V E , ) ). Figure 7 shows an example meaning representation for the statement 'Ali ate the sandwich because he was hungry'.
Fig. 7: Example Meaning Representation for 'Ali ate the Sandwich in The restaurant because he was hungry'
Fig. 8: Relations in Meaning Representation and their weights
2 we will put the code of the framework online
## 6 MLMR Generation
Figure 9 shows our MLMR recursive generation step. Initially G d meaning representation is passed to Meaning Representation Summarization Algorithm (MRSA) algorithm, detailed later in this section Specifically, MRSA generates the abstracted version of G d , we name parent MR G p , and it connects parent abstract nodes in the G p to regions in the G d that details such abstract nodes in G p . We name those regions as DMR Regions. Then each of the DMR Regions are processed recursively by the MLMR process if there are still big to fit in the screen. This termination criterion could be defined by the number of nodes for instance. In our case, we simply stop after the MR did not change from the previous iteration. The remaining part of this section details the MRSA algorithm, the role of the ontology in the algorithm, and how to make the MLMR Generation process interactive.
Fig. 9: MLMR Recursive Generation (best seen in color). Orange blocks indicates recursion.
## 6.1 Ontology
We assume that the ontology has the the property that semantically related concepts have short paths in the hierarchy of the ontology. We also assume that the concepts in the ontology could be mapped to the senses of the frames stored in the DMR. We opt to study our framework in one domain to prove that the concept works. Another reason is that natural language research is still progressing towards generalizing ontologies so that it covers an arbitrary text. The ontology, utilized in this work, consists of 1300 concepts that we created according to the taxonomy of the historical figures topics. The ontology consists of three root nodes of 'Work', 'Personal Life',
'Political Life'; the ontology is attached in the supplementary materials. Concepts presented in the textual information are assumed to be found as sub-nodes under one of these four root nodes. The main reason behind designing our custom ontology is to focus on the evaluation of our main contribution, Multilevel MindMap visualization, rather than addressing ambiguity problems that exist in the available ontologies (e.g. WordNet [36], SUMO [38], and CYC [29]), which are out of the scope of this work.
It is important to note that, for each concept in the ontology, we added a field called 'MAP-LEX'. This field basically contains all the senses that matches these concepts. The purpose of this field is to easily map frames in a meaning representation to its corresponding concept. Hence, after the DMR is generated, we augmented each frame in the DMR with the corresponding concept in the ontology by finding the concept, whose 'MAP-LEX' field contains its stored sense 3 . Finally, each concept also contains 'Is-Visual' flag which indicates whether the concept could be presented as a picture.
## 6.2 MRSA
The objective of MRSA is to summarize a meaning representation (MR), based on the ontology and the important frames in the MR. Each frame of the summarized MR is connected to regions (i.e., a set of frames Vc ⊂ V and relations E c ⊂ E ) in the input MR that are related to that frame. Frames in the summarized MR/parent MR that map to more than one frame in the input MR are denoted by group frames . Figure 10 shows the block diagram of MRSA algorithm. The rationale behind MRSA is to measure the significance of MR frames by a weight assignment phase. It then selects the most important frames (we call them main frames) by clustering, based on the assigned weights. Then, the main frames persist at the top level, with surrounding actions, conceptually grouped according to a conceptual metric (i.e., distance between the concepts inside the ontology hierarchy). This is achieved by a conceptual based partitioning step.
## 6.2.1 Weight Assignment
In this step, a weight for each frame f ∈ V of G d graph is assigned to indicate its level of significance; the higher the weight, the more significant the frame is. Each of the case roles, domain relations, and temporal relations is assigned a constant weight depending on its type (see figure 8; for example Agent may have a score different from Location or Reason). We observed that important Entity-frames have many important relationships to the other frames. Based on this observation, we model the weight per Entity-Frame as the sum of the weights of its surrounding relations (see Equation 1).
$$W _ { E F } ^ { k } = \sum _ { i } N _ { C R _ { i } } ^ { k } w _ { C R _ { i } } + \sum _ { i } N _ { D R _ { i } } ^ { k } w _ { D R _ { i } } + \sum _ { i } N _ { T R _ { i } } ^ { k } w _ { T R _ { i } },$$
3 All senses are identified by wordNet sense index
Fig. 10: MR Summarization Algorithm (MRSA)
where W k EF is the weight of the k th Entity Frame ( ∈ V ) and N k CR i is the number of Case Role relations of type i in of the k th Entity Frame. Similarly, N k DR i and N k TR i are the number of Domain Relations, Temporal Relations of type i in f k respectively. w CR i , w DR i and w TR i are the constant weights assigned to CaseRole, Domain and Time Relations of type i (figure 8).
On the other hand, we assume that an Action-Frame is important if its is connected to important frames. Hence, the weight of each Action-Frame updated by its neighbors (see Equation 2).
$$W _ { A F } ^ { k } = \sum _ { i } \left ( R _ { C R _ { i } } \sum _ { j } w ( F _ { C R _ { i j } } ^ { k } ) \right ) + \sum _ { i } \left ( R _ { D R _ { i } } \sum _ { j } w ( F _ { D R _ { i j } } ^ { k } ) \right ) + \sum _ { i } \left ( R _ { T R _ { i } } \sum _ { j } w ( F _ { T R _ { i j } } ^ { k } ) \right ),$$
where W k AF is the weight of Action Frame k , R CR i defines the weight of the case role of type i (figure 8). w F ( k CR ij ) is the weight of the j th frame connected to frame k with caserole of type i . Similarly, R TR i and R DR i are the corresponding ratios for temporal and domain relations while w F ( k DR ij ) and w F ( k TR ij ) are the corresponding weights of frames related to frame k with temporal and domain relations respectively.
## 6.2.2 Weight-based Partitioning
The entity frame weights, obtained from the weight assignment, are partitioned into clusters using K-Means++ [3] which has an advantage of the initial seeding of the cluster centers. K-Means++ provides a way to avoid producing poor clustering which is more likely to happen by the standard K-Means algorithm. The first cluster center is chosen uniformly at random as one of the data points, then each subsequent cluster center is chosen from the remaining data points with probability proportional to its squared distance from the point's closest existing cluster center. We used the method in [44] to select the best K with maximum ratio between the intra-cluster distance measure ( intra ) and is the inter-cluster distance measure ( inter ) , as illustrated in
## equation 3.
glyph[negationslash]
$$int r a = \frac { 1 } { N } \sum _ { i = 1 } ^ { K } \sum _ { x \in C _ { i } } \| x - z _ { i } \| ^ { 2 }, \, \text{inter} = m i n ( \| z _ { i } - z _ { j } \| ^ { 2 } ), \, i \neq j$$
where z i and z j are centers of clusters i and j respectively, and C i is the set of points that belong to cluster i . Having performed the clustering, we select the cluster of frames of the highest center as the main frames.
## 6.2.3 Concept-based Partitioning
The objective of this step is to group semantically related frames under one common concept according to the ontology (i.e. abstraction). The associated actions of each main entity frame (extracted from the top cluster of the weight-based partitioning) are passed through the concept-based partitioning and a list of concepts with their corresponding frames is returned. If an action frame of a main entity frame was not grouped with any of the other action frames, the associated entity frames of that action frame are grouped using concept-based partitioning. So, the new summarized meaning representation contains (1) the main entity frames, (2) the grouped concepts associated with them as new action frames, (3) the ungrouped action frames around the main entity frames, and (4) the grouped concepts of entity frames around them as new entity frames. The following steps illustrates how to conceptually partition concepts around each main entity frame.
- 1. Group frames of the exact concept (i.e. ontological distance =0).
- 2. Merge two frames whose concepts has the shortest path in the ontology.
- 3. Repeat step 2, until the count of the groups < = Gth (We used Gth=6).
The bottom-up manner of our conceptual grouping algorithm behaves as a variant of Agglomerative hierarchical clustering. The algorithm mainly analyzes the selected frame(a) and its neighbors and try to do conceptual partitioning if convenient as illustrated in figure 11 for Einstein and figure 2 for Shakespeare. The advantage of this model is that each high level frame could be expanded into details (the DMR regions). The highest level of groups serves as the abstracted Meaning representation. the DMR region for each abstracted frame is generated by including all the frames that are descendants of the given group frame following the Agglomerative hierarchical clusters.
In Shakespeare's example, 'Shakespeare', as a main entity frame, is detected by the weight assignment and partitioning steps. Afterwards, Shakespeare's actions are grouped conceptually using our conceptual-based partitioning through the ontology. For instance, action frames of 'earn' (living), 'wrote' (35 good plays), and 'is' (great literature) are grouped under 'Work' concept in the ontology. While, action frames of 'lived in' (Strateford), 'born' (in 1564), and 'had' (3 children) are grouped under 'Life/Personal Life' concept. From this example, it is intuitive that conceptual partitioning is done in a bottom-up manner, which is adopted in our algorithm. Hence, It takes only 1-2 recursive calls in these examples. In general, the algorithm recursive calls depend on the ontology and the number of concepts in the given instance.
## Einstein (1879-1955)
Einstein was born in Germany in 1879. received doctorate degree from Institute of Technology in Zurich in 1905. He became associate professor of Physics in 1909. He became German citizen in 1914. He received Nobel Prize in 1921 in Physics. He was involved in the establishment of Hebrew University in 1920. He emigrated to USA in 1933. He died in 1955 in Princeton. He
Fig. 11: Einstein Multi Level MindMaps
## 6.3 Interactive MLMR Generation
MLMR Generation could be also interactive as follows. Firstly, the DMR is summarized by MRSA. The summarized MR is then passed to MindMap Generation, detailed in the next section, to pictorially visualize the most abstract level of the MindMap. The user then can interacts with the generated MindMaps by selecting one of the group frames which triggers running MSRA for the subset of DMR referred by the group frame. The Mindmap generation then visualizes the output MR, which represents the details of the selected group frame. The same process could be repeated interactively until the most detailed level is reached. Previous steps could be stored to allow going back to parent levels and down to the detailed levels. In this interactive mode, the generation is also effective since only the top level need to be computed, and group frames are detailed based on the interactions by the user.
## 7 MindMap Generation
In the MindMap generation step, an arbitrary meaning representation is converted into a MindMap that contains images for visual frames and dynamically allocate them on the screen. Visual frames are determined by checking whether the frame's
concept is visual in the ontology. Then, a query text is generated for each visual frame to retrieve a relevant image from Google Image Search [20].
## 7.1 Query Generation and Image Web Retrieval
We propose two approaches to generate the query. In the first approach, the given frame name is associated with its filled attributes. For instance, if there exist an entity frame (Ball: Color= Red, Size=small), the generated query will be 'small red ball'. On the other hand, The second approach involves frame combination. For example, if we have a sentence like 'Shakespeare performed before the queen in December'. If 'the queen' is web-searched as suggested by the first approach, an irrelevant image will likely be retrieved. However, if the query combines 'Shakespeare' and 'queen'. The query is more likely to return a picture of 'Queen Elizabeth', which is meant by this sentence. That is why we tried to combine frames in the query. We implemented the second approach as follows. Given frame i , we add, to the query generated by the first approach, the name of frame(s) within two edges from frame i (i.e., relevant frame). However, we add the names of the frames, only if it is significantly more important. In other words, relative weight of frame i should be high, with respect to frame j . Mathematically speaking, WF i WF j > th where WF i is the weight of frame i , WF j is weight of frame j , th is the relative importance threshold ( th =6 in our experiments).
Finally, the images displayed in the generated MindMap are retrieved using Google Image Search (GoIS). We used the GoIS API to retrieve the relevant images using the generated query and the first image in the results is chosen.
## 7.2 Automatic Layout allocation
A graph drawing algorithm must take into account aesthetics: criteria for making salient characteristics of the graph easily readable. Some aesthetic criteria include minimizing number of edge crossings, maintaining straight edges as possible,maximize display symmetry, etc. In our work, we achieved an acceptable layout by adopting the spring model by Thomas Kamps et al in [25]. Spring model meets two aesthetic principles 4 . First, all edges are prefered to have the same length. Second, the layout should be symmetric as possible. The main idea is to model the cost function by attaching a spring between the edges of the desired length as in equation 4.
$$E = \sum _ { i = 1 } ^ { n - 1 } \sum _ { j = i + 1 } ^ { n } K _ { i j } ( ( x _ { i } - x _ { j } ) ^ { 2 } + ( y _ { i } - y _ { j } ) ^ { 2 } + l _ { i j } ^ { 2 } - 2 l _ { i j } \sqrt { ( x _ { i } - x _ { j } ) ^ { 2 } + ( y i - y j ) ^ { 2 } }$$
where ( x , y i i ) is the location of the the node i . l ij is the desired length between node i and node j (It is not necessary to have an edge between node i and node j ). There are four parameters that we designed while adopting the spring model (1) l
4 Other third party commercial tools/methods could be used also here (e.g. Microsoft AGL http://research.microsoft.com/en-us/projects/msagl/
which is the desired length of directly connected nodes; we set l to 50 . (2) D , which is the diameter of the drawing (i.e. the maximum distance between a pair of nodes in our system); we set D to 600 . (3) l ij = l , if there is an edge between node i and node j . Otherwise, l ij = D d · ij min l,m { d lm } , where d lm is the length of the shortest path between node l and node m . (4) In our system, all-pairs shortest paths are computed , between every pair of nodes, using Floyd Warshall algorithm [19]. K ij is the weight assigned for the cost function between node i and node j , which indicate the strength of the spring associate between node i and j . K ij = 1 d ∗ ij 2 , where d ∗ ij = d ij if there is no edge between node i and node j , d ∗ ij = l otherwise. This indicates less restrictive springs, when nodes are not directly connected.
While nodes are assumed to be points in [25], each node i in our setting is a rectangular area R i of width W i and height H i , centered around the node position ( x , y i i ) . For instance, R i will be the size of the image to be rendered, if node i is associated with an image. To tackle this problem, we add, to each l ij , M i + M j 2 , where M k is the diameter of Rectangle R k of node k . The rationale behind adding M i + M j 2 is that when the distance between the centers of two rectangles is M i + M j 2 , the rectangles does not intersect regardless their positions. Finally, the gradient of the cost function is computed and used to find a local minimum. We used random re-initialization strategy to reach a good solution minimizing equation 4. We run the spring model 10 times with different initial seeds selected randomly on the screen and the layout with the minimum cost is selected.
## 8 Experimental Results
This section presents the evaluation methodology and experiments of our MindMap Automation Framework. The evaluation was based on 4900 Mechanical Turk (MTurk) [9] rating responses, with different system parameters. As indicated in [7], MTurk users have less than 2% of the workers have no high school degree and average education years of 14 9 . . In order to improve the quality of the responses, the workers were exposed to illustrative examples before getting introduced to the instances used for the evaluation. Furthermore, We also required that the MTurk-workers have at least a 95% approval rate and 1,000 approved assignments 5 to work on our experiments. The following subsections details the evaluation dataset, three experiments to evaluate different aspects of our framework.
## 8.1 Historical Figures Dataset and Evaluation Metrics
Since there is no existing datasets to evaluate MindMap Visualization, we created our own dataset to validate our framework. We chose articles of Historical Figures (e.g., Shakespeare, George Washington, etc) to build our dataset. Thirty-five historical figures were chosen from the BBC historical figures section [6] of size between 150 words to 250 words. We used our framework to generate 455 different MindMaps
5 This is based on their previous history that is maintained in MTurk system
from these Thirty-five chosen articles. The four hundred and fifty-five MindMaps indicates that thirteen MindMaps were generated for each article, one for each parameter setting (detailed in this section). The four hundred and fifty-five cases were evaluated based on Human Subject Rankings by 4900 MTurk workers' responses on the following five question survey.
- 1. To what extent does the generated output represent the text (Regardless the pictures) ? Grade 1-5.
- 2. To what extent are the generated pictures relevant? Grade 1-5.
- 3. How many missing actions are there in the shown diagram (if no missing actions, please put 0)?
- 4. How many missing entities are there in the shown diagram (if no missing entities , please put 0)?
- 5. Howmanyrepeated entities/actions are there in the shown diagram (if no repeated entities/actions, please put 0)?
Our evaluation is partitioned into three experiments to evaluate the effect of changing the system parameters. The framework was evaluated upon three main variations. (a) GoIS Parameters (Size, Image Type) [20] for single level MindMap in Experiment 1 (b) Concept combination in Experiment 2 (c) Multilevel MindMaps in Experiment 3. The responses of the experiments were evaluated based on three metrics: (1) Mean, (2) Standard Deviation to indicate the stability of the response, and (3) For grading questions (i.e., Q1,2), Satisfaction Ratio (SR) that is defined as SR = # of responses ≥ 4 # ofresponses . Appendix A shows the generated single-level and multilevel MindMaps for Akhentan, Van Gokh, George Washington, and Newton, a subset of the evaluated outputs ( figures 14, 15, 16, and 17 ). More automatically generated MindMaps are attached in the supplementary materials. While, the framework can allow the user to change the retrieved images, we only evaluate results with images automatically retrieved by our method.
## 8.2 Experiment 1 (315 cases, 3150 responses )
In Experiment 1, we evaluate the MTurk workers' satisfaction with Single Level MindMaps, and test whether their responses are biased by varying the displayed image parameters. We tested nine variations: three different image types (All [any image], ClipArt, LineArt) and three sizes (All, Auto, Small). The auto size is a size mode, we implemented to determine the image size according to the weight of the frame to be visualized, computed in subsection 6.2.1. Our implementation uses medium size image ( biggest suitable size for MindMaps) if the number of relations to it is greater than six, otherwise, the system uses small image size. The MindMap generation phase in this experiment did not involve concept combination (which is evaluated in Experiment 2). Overall, we have 315 (9 × 35) cases with ten MTurk responses each (i.e. 3150 responses). Table 1 (top part) presents the user response statistics grouped by all the nine variations. The results indicate general satisfaction by the MTurk workers.
Table 1: Human Subject Responses for Experiments 1, 2, 3
| Experiment 1 | Experiment 1 | Experiment 1 | Experiment 1 | Experiment 1 | Experiment 1 |
|----------------|----------------|----------------|----------------|----------------|----------------|
| | Q1 | Q2 | Q3 | Q4 | Q5 |
| Mean | 4.42 | 4.34 | 0.54 | 0.59 | 0.48 |
| StdDev | 0.84 | 0.87 | 1.58 | 1.74 | 1.68 |
| SR | 0.86 | 0.84 | | | |
| Experiment 2 | Experiment 2 | Experiment 2 | Experiment 2 | Experiment 2 | Experiment 2 |
| | Q1 | Q2 | Q3 | Q4 | Q5 |
| Mean | 4.51 | 4.27 | 0.45 | 0.5 | 0.23 |
| StdDev | 0.65 | 0.82 | 1.25 | 1.47 | 0.58 |
| SR | 0.92 | 0.79 | | | |
| Experiment 3 | Experiment 3 | Experiment 3 | Experiment 3 | Experiment 3 | Experiment 3 |
| | Q1 | Q2 | Q3 | Q4 | Q5 |
| Mean | 4.48 | 4.33 | 0.59 | 0.61 | 0.25 |
| StdDev | 0.87 | 0.94 | 1.51 | 1.55 | 0.93 |
| SR | 0.89 | 0.84 | | | |
Q1 and Q2 are the most relevant questions to the appearance of the generated MindMap and the generated images. The results of these questions are grouped for each of the nine variations in this experiment. Figure 12 shows the three evaluation metrics for the nine cases on Q1 and Q2, which indicates that changing the image type or size almost does not affect the responses of the MTurk workers. In other words, these responses were not biased to specific image type or size.
Fig. 12: Experiment1 grouped results (Q1,2): Mean and StdDev is on the left while SR is on the right
## 8.3 Experiment 2 (105 cases,1050 responses )
Experiment 2 focuses on evaluating the user satisfaction of the MindMaps in case of Concept Combination mode (described in subsection 7.1). To evaluate whether the MTurk workers were biased to the pictures, we created Concept Combination cases with three variations of image types resulting in 105 cases (3 × 35 historical figures). Each case was evaluated by ten different MTurk anonymous workers (1050 responses).
Table 1 (middle part) shows summarized results for the experiment. Compared to Experiment 1, there is an improvement of 2% in the mean and 7% in the satisfaction ratio for Q1 with the use of concept combination. However the satisfaction ratio of Q2 (pictures satisfaction) decreased by 10%. This indicates that the retrieved image using direct query is relatively better than retrieved image using Concept Combination query. The results were also grouped by the three variations of Experiment 2, which indicate a similar behavior to Experiment 1(i.e. No bias).
It is important to mention that, the relevance of the MindMap images currently depends only on the quality of the firstly-ranked image retrieved by our generated query from GoIS. After careful look on the MTurk responses and our own evaluation of this experiment, we found that most of the retrieved images are relevant using direct queries. However, in some cases, it does not retrieve a relevant image. For example, 'new capital' node in Akhentan example, is assigned to an unsatisfactory image; see figure 14 in Appendix A. Our intuition is that context information is not considered in this retrieved images in direct query mode, which we tried to model by concept combination (CC) queries. Having applied CC mode to Akhentan example, we managed to retrieve an image of a pharaonic land for this 'new capital' query, which is more relevant. However, the CC mode negatively affected the overall performance of remaining queries since it badly affects most of the remaining frames, as illustrated in the results of this experiment.
Table 2: Experiment 2 (Q1,2) grouped results.
| | Q1 | Q1 | Q1 | Q2 | Q2 | Q2 |
|---------|------|--------|-------|------|--------|------|
| | Mean | StdDev | SR | Mean | StdDev | SR |
| All | 4.56 | 0.63 | 0.937 | 4.31 | 0.825 | 0.8 |
| LineArt | 4.53 | 0.61 | 0.939 | 4.24 | 0.827 | 0.76 |
| ClipArt | 4.46 | 0.71 | 0.91 | 4.26 | 0.825 | 0.81 |
## 8.4 Experiment 3 (35 cases,700 responses)
The purpose of this experiment is mainly to rate the partitioning of the information in the multilevel MindMaps and not the evaluation of the retrieved images (evaluated in Experiment 1 and 2). Hence, we selected one case for each historical figure (with the best images retrieved in Experiment 1). Q2 was changed to evaluate of conceptual grouping instead of the pictures as they was already evaluated in Experiments 1 and
- 2. Question 2 became 'To what extent are you satisfied with the structure of the Multi-level MindMap 6 ? Grade 1-5'. To evaluate it interactively, we have generated 35 interactive flash files (one for each historical figure). Each flash file was evaluated by twenty MTurk Users (i.e., 700 responses).
Results are shown in Table 1 (bottom part). There is 1.4% and 3% improvement in the mean and SR respectively compared to the single-level MindMaps. The results show the satisfaction of the users with the hierarchical representation of the MindMap. While both single-level and multilevel approaches give satisfactory outputs, the multilevel approach is the only way to represent all information in large text, where single level generation is not applicable.
In order to give a careful consideration to the criticism of the generated Multilevel MindMaps, we computed the lowest 15% for all the ratings for Q1 and Q2. We also computed the lowest 15% percentile for the ratings of each of the 35 Historical figures and we reported the average overall the historical figures; see table 3. We further investigated the highest errors reported by the MTurk workers, in response to Q3, Q4 and Q5, which indicate missing actions, entities, or redundancy of any of them. For each of 35 historical figure, we select the responses that have Non-zero values for Q3, Q4, and Q5. Then, we compute the average over all the historical figures; see table 4.
Table 3: Experiment 3 worst 15% percentile Q1 and Q2: The first row shows the 15% percentile over the all responses, while, the second row shows the mean of the average of the 15% for each of historical figure
| | Q1 | Q2 |
|-------------------------------------------------------------------|------|------|
| Mean of 15% percentile worst responses | 4 | 3 |
| Mean of the average 15% percentile for each 35 historical figures | 3.78 | 3.57 |
Table 4: The average of the non-zero responses' mean of Q3, Q4 and Q5 for each historical figure in Experiment 3
| | Q3 | Q4 | Q5 |
|-------------------------------|------|------|------|
| Average of Non-zero responses | 3.57 | 2.83 | 2 |
However, this group of responses does not indicate high disagreement with the generated MindMaps as indicated in tables 3 and 4, we carefully investigated the provided comments by the human subjects on the three experiments. We noticed that the main criticism of the least satisfied subjects was due to some missing/redundant pieces of information and irrelevant images in the MindMap. While, the experiments
6 Clarification information included in the correct class (e.g Work, Personal Life, and Political Life). As an incorrect instance, information like the historical figure's birth or death is classified under Work. This clarification was exposed to the MTurk workers with examples
and our investigation indicate that these cases are few, it is an important feedback to consider for our future work, in which targets generalization on larger scales.
## 9 Discussion and Conclusion
Our framework has multiple limitations. For instance, it depends on the quality of the syntactic analyzer to pass information. If a statement could not be parsed, then it will be ignored, and its information will not be included in the generated MindMap accordingly. Another limitation is that the MLMR algorithm assumes that there is a central idea of a given document, which is valid for the historical figures dataset, but not valid in general. In the future, we aim at exploring different methods to generate hierarchy of abstract information in the text and use word vectors (e.g. [35]) to help semantically related information that might not be captured by the ontology. We aim to extend the system such that it is reliable in handling very large text
(e.g., a book). Furthermore, we will investigate different approaches to retrieve more relevant images from the textual query. In order to generate MindMaps for a large in a reasonable time, the framework computational time needs to be significantly reduced. For instance, Multi-level MindMap Generation of a 250 word document take about 40 seconds on a 3.3GHZ-6GB machine; see figure 13). The recorded time is the average time to convert raw text to MultiLevel MindMap including automatic layout allocation. Finally, rather
Fig. 13: Text-to-MultilevelMindMap performance on 3.3 GHZ Intel Xeon Processor-6GB Memory
than restricting our domain to historical figures in this work, we plan to investigate the generalization of MindMap visualization for any text and tackle ontology-related problems.
In conclusion, we have designed and implemented the first automated framework generates a multilevel Mind Map visualization out of English text. The system was comprehensively tested under different parameter settings by MTurk Human Subjects and high satisfaction rates have been recorded.
## Appendices
Appendix A
## Akhentan (1564-1616)
Akhenaten was the son of Amenhotep III. He reigned from 1352BC to 1336BC. He is considered Amenhotep IV. Akhenaten rejected to worship the Aten, which is the sun He built a new capital in Middle Egypt. It was named Akhetaten or TelElAmarna. It was only occupied for 20 years. Akhenaten was principally married to Nefertiti. Smenkhkare and Tutankhamun succeeded him god.
Fig. 14: Akhentan Single Level and Multi Level MindMaps
## Van Gogh (1853-1890)
Van Gogh is great well-known painter. He was born in 1853. He worked in an international firm of art in 1869 in Hague. He jobs in London and then Paris . He was not interested in them. He was dismissed from them in 1876. He became an artist in 1880 He met many artists in 1886. include Theo, Degas, ToulouseLautrec, Pissarro and Gauguin. He painted a large number of selfthreatened Gauguin with a razor in 1888. He suffered from mental problems. He spent time in psychiatric hospitals. He committed suicide in 1890. got They
Fig. 15: Van Gokh Single Level and Multi Level MindMaps
## George Washington (1732-1799)
George Washington was born in 1732 in Virginia. He became the surveyor of Culpeper County in 1749. He won a reputation for bravery during the French and Indian War. He married MarthaCustis in 1759. He consistently opposed the unfair British taxes\_ He became a leading Virginian figure supporting the colonial cause. He became commander of all colonial forces in 1775. The American victory under George Washington in Saratoga forced the French to ally with the Americans. He became leader of the Constitutional Convention in 1787 . He became the first president of the United States in 1789. He was elected for a second term in 1792.finally retired from public life in 1797. He died in Mount Vernon in 1799.
MOUNTVERNON
Fig. 16: George Washington Single Level and Multi Level MindMaps
## Newton (1643-1727)
Newton was an English physicist and mathematician. Newton was born in 1643 in Lincolnshire. His father was a prosperous farmer. Newton went to University of Cambridge in 1661. He was interested in mathematics, optics, physics and astronomy. Newton published The Optics in 1704. Newton was elected as member of the parliament in 1689 for University of Cambridge. He was elected president of the Royal Society in 1703. He was a difficult man and prone to depression. He died in 1727 in Westminster Abbey.
Fig. 17: Newton Single Level and Multi Level MindMaps
## References
- 1. Afzal, S., Maciejewski, R., Jang, Y., Elmqvist, N., Ebert, D.: Spatial text visualization using automatic typographic maps. TVCG 18 (2012)
- 2. Altintas, E., Karsligil, E., Coskun, V.: The Effect of Windowing in Word Sense Disambiguation (2005). DOI 10.1007/11569596 \ 65
- 3. Arthur, D., Vassilvitskii, S.: k-means++: the advantages of careful seeding. In: Proceedings of the ACM-SIAM symposium on Discrete algorithms. Society for Industrial and Applied Mathematics (2007)
- 4. B., B., E., B.: An efficient recursive transition network parser for arabic language. In: World Congress on Engineering (WCE) (2009)
- 5. Banerjee, S., Banerjee, S.: An adapted lesk algorithm for word sense disambiguation using wordnet. In: CICLing (2002)
- 6. BBC: BBC Historical Figures. http://www.bbc.co.uk/history/historic\_figures/ (2012). [Online; accessed January-2012]
- 7. Berinsky, A.J., Huber, G.A., Lenz, G.S.: Evaluating online labor markets for experimental research: Amazon.com's mechanical turk. Political Analysis 20 (2012)
- 8. Berinsky, A.J., Huber, G.A., Lenz, G.S.: An evaluation of the left-brain vs. right-brain hypothesis with resting state functional connectivity magnetic resonance imaging. PLOS ONE 8 (2013)
- 9. Buhrmester, M.D., Kwang, T., Gosling, S.D.: Amazon's mechanical turk: A new source of inexpensive, yet high-quality data? Perspectives on Psychological Science 6 (2011)
- 10. Collins, M.: Head-driven statistical models for natural language parsing. Computational Linguistics 29 (2003)
- 11. Conroy, J.M., O'leary, D.P.: Text summarization via hidden markov models. In: ACM SIGIR (2001)
- 12. Coyne, B., Rambow, O., Hirschberg, J., Sproat, R.: Frame semantics in text-to-scene generation. In: KES. Springer (2010)
- 13. Das, D., Martins, A.F.T.: A survey on automatic text summarization. Literature Survey for the Language and Statistics II Course at CMU (2007)
- 14. Dhindsa, H.S., Makarimi-Kasim, Roger Anderson, O.: Constructivist-Visual Mind Map Teaching Approach and the Quality of Students' Cognitive Structures. Journal of Science Education and Technology 20 (2) (2011)
- 15. Dou, W., Yu, L., Wang, X., Ma, Z., Ribarsky, W.: Hierarchicaltopics: Visually exploring large text collections using topic hierarchies. TVCG 19 (2013)
- 16. Elhoseiny, M., Elgammal, A.: English2mindmap: An automated system for mindmap generation from english text. In: IEEE International Symposium on Multimedia (ISM) (2012)
- 17. Farrand, P., Hussain, F., Hennessy, E.: The efficacy of the 'mind map' study technique. Journal of Medical Education 36 (2002)
- 18. Feng, Y., Lapata, M.: Visual Information in Semantic Representation. ACL (2010)
- 19. Floyd, R.W.: Algorithm 97: Shortest path. Communications of the ACM 5 (1962)
- 20. Google: Google Image Search. http://www.google.com/advanced\_image\_search/ . [Online; accessed January-2013]
- 21. van Ham, F., Wattenberg, M., Viegas, F.B.: Mapping text with phrase nets. TVCG 15 (2009)
- 22. Hamdy, A., ElHoseiny, M.H., Sahn, R.E., Samier, S., Kamal, E.: Mind maps automation (mma) system. In: International Conference on Semantic Web and Web Services (SWWS) (2009)
- 23. http://img2.mappio.com/: Shakespear Mindmap (2014). URL http://img2.mappio.com/ miwisdom/william-shakespeare-short-biography-mind-map-Large.jpg
- 24. Kaikhah, K.: Automatic text summarization with neural networks. In: International IEEE Conference on Intelligent Systems (2004)
- 25. Kamps, T., Kleinz, J., Read, J.: Constraint-based spring-model algorithm for graph layout. In: Proceedings of the Symposium on Graph Drawing (1996)
- 26. Kudelic, R., Konecki, M., Malekovic, M.: Mind map generator software model with text mining algorithm. In: International Conference on Information Technology Interfaces (ICITI) (2011)
- 27. Lappin, S., McCord, M.: A syntactic filter on pronominal anaphora for slot grammar. In: ACL (1990)
- 28. Leass, H.J.: An algorithm for pronominal anaphora resolution. Computational Linguistics 20 (1994)
- 29. Lenat, D.: Cyc: A large-scale investment in knowledge infrastructure. Communications of the ACM 38 (1995)
- 30. Litkowski, K.C.: Syntactic Clues and Lexical Resources in Question-Answering (2001)
- 31. Ma, M.: Automatic conversion of natural language to 3d animation. Ph.D. thesis, University of Ulster (2006)
- 32. Maicher, L., Park, J. (eds.): Charting the Topic Maps Research and Applications Landscape. Springer (2006)
- 33. Marcu, D.C.: The rhetorical parsing, summarization, and generation of natural language texts. PhD dissertation (1998)
- 34. de Marneffe, M.C., MacCartney, B., Manning, C.D.: Generating typed dependency parses from phrase structure trees. In: LREC (2006)
- 35. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed representations of words and phrases and their compositionality. In: Advances in Neural Information Processing Systems, pp. 3111-3119 (2013)
- 36. Miller, G.A.: Wordnet: A lexical database for english. Communications of the ACM 38 (1995)
- 37. Navigli, R.: Word sense disambiguation: A survey. ACM Computing Surveys 41 (2009)
- 38. Niles, I., Pease, A.: Towards a standard upper ontology. In: FOIS (2001)
- 39. Novak, J.D.: The theory underlying concept maps and how to construct them. Tech. rep. (2006)
- 40. novamind.com: IMind tool (2012 (accessed February, 2012)). URL http://www.novamind. com/
- 41. Osborne, M.: Using maximum entropy for sentence extraction. In: ACL Workshop on Automatic Summarization (2002)
- 42. Park, J., Hunting, S. (eds.): XML Topic Maps: Creating and Using Topic Maps for the Web. AddisonWesley Longman Publishing Co., Inc., Boston, MA, USA (2003)
- 43. Qiu, L., yen Kan, M., seng Chua, T.: A public reference implementation of the rap anaphora resolution algorithm. In: LREC (2004)
- 44. Ray, S., Turi, R.H.: Determination of number of clusters in k-means clustering and application in colour image segmentation. ICAPRDT (1999)
- 45. thinkbuzan.com: IMind tool (2012 (accessed February, 2012)). URL http://www.thinkbuzan. com/us/?utm\_nooverride=1&gclid=CPaV7saVkK4CFUff4AodtTRidg
- 46. Tony Buzan, B.B., Harrison, J.: The Mind Map Book: Unlock Your Creativity, Boost Your Memory, Change Your Life. Pearson Education Ltd (2010)
- 47. Viegas, F., Wattenberg, M., Feinberg, J.: Participatory visualization with wordle. TVCG 15 (2009)
- 48. Youzhi, Z.: Research and implementation of part-of-speech tagging based on hidden markov model. In: Asia-Pacific Conference on Computational Intelligence and Industrial Applications (2009)
Mohamed Elhoseiny is a PhD student at the Department of Computer Science, Rutgers, the State University of New Jersey. He is also a member of the Center for Computational Biomedicine Imaging and Modeling (CBIM). He received his B.Sc. and M.Sc. degrees in computer science from University of AinShams, Egypt in 2006 and 2010, respectively. He is currently pursuing his studies in computer vision and multimedia as a PhD student in Computer Science department at Rutgers University.
Ahmed Elgammal is an associate professor at the Department of Computer Science, Rutgers, the State University of New Jersey Since Fall 2002. Dr. Elgammal is also a member of the Center for Computational Biomedicine Imaging and Modeling (CBIM). His primary research interest is computer vision and machine learning. His research focus includes human activity recognition, human motion analysis, tracking, human iden- tification, and statistical methods for computer vision. Dr. Elgammal received the National Science Foundation CAREER Award in 2006. Dr. Elgammal has been the Principal Investigator and Co-Principal Investigator of several research projects in the
areas of Human Motion Analysis, Gait Analysis, Tracking, Facial Expression Analysis and Scene Modeling; funded by NSF and ONR. Dr. Elgammal is Member of the review committee/board in several of the top conferences and journals in the computer vision field. Dr. Elgammal received his Ph.D. in 2002 from the University of Maryland, College Park. He is a senior IEEE member.
| null |
[
"Mohamed Elhoseiny",
"Ahmed Elgammal"
] |
2014-08-01T03:18:56+00:00
|
2014-12-23T06:27:03+00:00
|
[
"cs.CL",
"cs.HC"
] |
Text to Multi-level MindMaps: A Novel Method for Hierarchical Visual Abstraction of Natural Language Text
|
MindMapping is a well-known technique used in note taking, which encourages
learning and studying. MindMapping has been manually adopted to help present
knowledge and concepts in a visual form. Unfortunately, there is no reliable
automated approach to generate MindMaps from Natural Language text. This work
firstly introduces MindMap Multilevel Visualization concept which is to jointly
visualize and summarize textual information. The visualization is achieved
pictorially across multiple levels using semantic information (i.e. ontology),
while the summarization is achieved by the information in the highest levels as
they represent abstract information in the text. This work also presents the
first automated approach that takes a text input and generates a MindMap
visualization out of it. The approach could visualize text documents in
multilevel MindMaps, in which a high-level MindMap node could be expanded into
child MindMaps. \ignore{ As far as we know, this is the first work that view
MindMapping as a new approach to jointly summarize and visualize textual
information.} The proposed method involves understanding of the input text and
converting it into intermediate Detailed Meaning Representation (DMR). The DMR
is then visualized with two modes; Single level or Multiple levels, which is
convenient for larger text. The generated MindMaps from both approaches were
evaluated based on Human Subject experiments performed on Amazon Mechanical
Turk with various parameter settings.
|
1408.1032v1
|
## A case for intranet-based online portal for undergraduate Computer Science teaching
(An approach for computer-assisted-learning)
## K. V. Iyer
Dept. of Computer Science & Engg. National Institute of Technology Tiruchirapalli 620015, India [email protected]
Abstract -Our proposal for selective subjects especially those involving intensive problem-solving assignments and/or tutorials, such as Introduction to Algorithms and Data structures, Discrete Mathematics, Coding Theory, Number theory, Combinatorics and Graph Theory (CGT), Automata theory, is to supplement lectures with a moderated online forum against an intranet portal. By way of illustration we take the example of a restricted view of OEIS (http://oeis.org). The restriction can be w.r.t. sequences in OEIS that are directly relevant to say CGT. N.J.A.Sloane's OEIS is a collection of over 2,39,147 integer sequences and their properties. In particular OEIS contains definitions of many combinatorial structures, dense range of interpretations, generating functions and conjectured ones, cross references within OEIS and to outside resources, references to texts and technical articles, codes in Maple, Mathematica etc. For organizing courses such as the above mentioned, a first task is to partially create an OEIS-like instructor-moderated portal in a university intranet. During the course of lectures and tutorials students are invited to contribute to the portal and these may be augmented/approved by instructors suitably, to find a place in the portal. By this many concepts can be conveyed to the students in an interesting way with the desired results. In the arguments presented, examples related to Combinatorics and Graph Theory are given.
contents from the web, pertaining to a target theme and summarize and classify the content (it is known that the task of abstracting and classifying page-ranked web-search results can be efficiently automated). A major goal in these activities is to enrich the learning experience of students as is the case with intelligent tutoring systems. It is argued that web-based systems for learning have these advantages [2]: long-distance learning, being a complement for traditional teaching and as a support for life-long learning. Conventional intranet-based education relies on a server onto which lecture notes, handouts, hyperlinks to selected Wiki pages, assignments and solutions etc. are preloaded by the instructor. To some extent these can cope-up with the rapid changes in Computer Science as a result of technological advances. However these approaches lack the feature of interactiveness in computerbased problem-solving. The generic framework ActiveMath described in [2] is web-based, adaptive and interactive.
Keywords-Computer-assisted learning; learning portal; OEIS; intranet; undergraduate teaching; Combinatorics and Graph theory; algorithms.
## I. INTRODUCTION
Consequent to the availability of the Internet and hardware systems, web-based computer education, popular in many educational institutions depends on the availability of high quality learning resources such as electronic books [1] and technical articles, multimedia enabled presentations, refereed technical articles etc. In a related context web-based problemsolving is a well-known learning activity where, among other things, students are required to use a search-engine to extract
Techniques in computer-aided-instruction-oriented systems are known to be teacher-assisted or otherwise. These systems often incorporate online assessments of students such as a quiz. On the negative side, use of such systems are known to create disengagement in the users' minds. As mentioned in [3] the different forms of disengagement from the task at hand include engaging in off-task behavior such as net-surfing, exploiting system's features in a quiz such as repeated guessing, working carelessly and using the system in ways unrelated to the stated work. It has been pointed out that boredom and confusion precede such behaviors although such computer-based learning systems are seen to be quite engaging in general and are useful when the number of students in a class is large.
The suggestion described herein is a student-centric use of a university intranet-web portal associated with a course that dynamically evolves with the progress of the course - this approach of describing and expanding knowledge, developing comprehension, analytic ability and inquisitiveness in a subject can be classified under computer-assisted learning in some sense. The approach outlined herein can be regarded as a positive step in enhancing a learning environment (-learning through examples‖) for undergraduate Computer Science teaching. For selective subjects (those involving intensive problem-solving tutorials and exercises), such as Introduction to Algorithms and Data structures, Discrete Mathematics, Coding theory, Number theory, Combinatorics and Graph Theory (CGT), Automata theory, our proposal is to supplement lectures with a teacher-moderated online forum on a university intranet portal with restricted access to a class of students. In this the role of a teacher and his/her team is augmented by the role of students and it promotes collaborative learning as well, in computer-free classrooms. It also supports the future classroom models where students will listen to lectures at their own convenience in his/her environment and will get engaged in group activities in classrooms.
By way of illustration we take the example of a restricted view of the Online Encyclopaedia of Integer Sequences (OEIS: see [4]). The restriction can be w.r.t. sequences in the OEIS that are directly relevant to say CGT. The rest of the paper briefly mentions about OEIS and then about the ways and means a similar portal can be developed in a university intranet for subjects related to Computer Science and perhaps Mathematics for undergraduate teaching. For the purpose of illustrations basic concepts in CGT is required (see for example, [5,6,7]).
## II. THE OEIS AS A LEARNING RESOURCE
First, a few highlights about the OEIS are given. The proposition of an intranet portal for a specific course may be viewed as a miniature form of OEIS although with many differences in content as well as in the method of construction and usage.
Founded in 1964 by N.J.A.Sloane, OEIS, the now-online learning resource is a collection of over 2,39,147 integer sequences and their properties. In particular, as of now the OEIS contains definitions of many combinatorial structures, dense range of their interpretations, generating functions and conjectured ones, cross references within OEIS and to outside resources, references to texts and technical articles, codes in Maple, Mathematica etc. OEIS incorporates a search feature allowing keyword-based search spanning the entire contents. There are other features as well to use the OEIS. Entries and modifications to the OEIS contributed by integer sequences enthusiasts across the world are reviewed and then approved by a team of subject experts. The overall contents of OEIS is to be viewed at a research level. A motivated student can explore the OEIS, fairly systematically and enhance his/her knowledge in Discrete Mathematics with additional learning support. The popularity of OEIS is evident from the numerous citations it has received in serious mathematical literature. A typical entry identified by an OEIS sequence number (e.g., A136328 in this case) looks like the one given in Illustration 1 (modified to remove hyperlinks). This sequence lists the first sixteen values of the Wiener index of Odd graphs. As a prerequisite, this page requires notions such as graphs, sets and related operations, univariate polynomials, cycles in graphs, induced subgraphs permutations of [n] , oddness, Wiener index, distance in graphs, computer algebra systems. At present OEIS is managed by the OEIS foundation.
## III. OEIS-LIKE INTRANET-PORTAL-BASED COMPUTER SCIENCE TEACHING
For organizing courses such as the above mentioned, the first task is to create partially, an OEIS-like portal in a university intranet on a client-server architecture. The portal can be built using XML technology and can be made accessible via standard browsers. The initial entries are due to an interest group with an identified moderating team. During the course of lectures and tutorials, students should be invited to contribute to the portal and these may be augmented/approved by teachers and graduate students to find a place in the portal. Submissions to the portal can be addressed to a designated mailbox monitored by the moderators. Contributions from the instructors are aimed to be at first year graduate level so as to motivate the undergraduates. A calendar of associated tasks drawn by faculty member will set the timelines. Some of the activities such as supplying figures, formatting, suggesting computer codes, cross referencing pointers, verifications, pointers to other resources can be assigned to a skilled groups of students. By this, many concepts in a specific subject can be conveyed to the students in an interesting way with the desired results. An added advantage is that assignments and reading exercises can be against the portal. This is beneficial when the class strength is large, say 100 which is generally the case in foundational courses. We emphasize in later discussions the dynamic nature - in terms of growth and restructuring - of the portal-idea: For example in the OEIS-context addition suggestions to A136328 in Illustration 1 could be: (a) Under COMMENTS: Properties of cycles in On . (b) Under REFERENCES: Pointer to a proof that On is distance regular. (c) Under LINKS: A table of Hosoya-Wiener polynomials of On for different n values. An exceptional student/teacher may like to add a new page to the OEIS on the Wiener index of line graphs of odd graphs.
## A. An A-CGT-portal and its construction
As a detailed example, a portal on Special Graphs and Combinatorial Objects is appropriate to be associated with a first-level course on CGT (in the sequel this is referred to as ACGT-portal). Graphs such as Kn , Km,n , C , S n n , Ladder graphs , binary hypercubes, Fibonacci trees, Fibonacci weighted trees etc., will serve as starting candidate objects. An A-CGT-portal will contain in each logical page, primary attributes such as definition(s), figures, constructions and algorithms to generate the objects, properties, related objects/graphs, a more-toexplore set of references/reading materials, historical notes/significance, and remarks by students/instructors . The properties and references will strive to capture a body of knowledge appropriate to CGT per se and to other courses requiring CGT as a prerequisite. Typical sources of knowledge include student submissions of assignment problems, copyright unprotected and free web/other resources, sources for which explicit permissions have been obtained, materials developed by instructors. Associated with an A-CGT-portal will be a corpus of prerequisite keywords and phrases - e.g., graphs, combinatorial structures, graphs characterized by one or more parameters, induced subgraphs, isomorphism, trees, power set, permutations and combinations, Pascal's triangle, partitions, recurrence relations, generating functions etc. A backward link will be associated with each logical page of an A-CGTportal - the link will be from a perquisite box(s) built from the corpus to the relevant logical page(s). It is possible to recognize the following two types of prerequisite boxes:
Type P1: This comprises those ( lightweight ) keywords/phrases that are acquired in one or two sittings by a student. These are covered by 1-2 lecture equivalents or textbook chapters.
Type P2: This encompasses ( loaded ) keywords or phrases each of which point to 1-4 short write-ups or pages of references e.g., planarity and embeddings, perfect graphs.
A more simple approach is the logical page attribute allowable perquisite courses . A color encoding of a logical page's background or of the primary attributes will indicate which ACGT-portal pages are part of the CGT course and which ones are considered outside the purview of the course. Alternatively and/or as an added feature a degree of relevance indicator computed as a weighted average based on the prerequisite boxes for each logical page of an A-CGT-portal is considered as a logical page attribute. The relevance will be to lecture notes and textbook chapters and sections.
The demarcation between the two types P1 and P2 is rather fuzzy in the case of many key terms and it depends on the knowledge and comprehension already acquired by the students. In general each keyword or phrase, via a hyperlink is made to point to a short page containing relevant definitions and/or concepts, illustrated if appropriate. In certain instances the page pointed at will be already part of the A-CGT-portal.
A typical logical page in an A-CGT-portal will cover a special graph or it can describe a graph class. We illustrate a logical page with the example of a Wheel graph . By definition Wn = Cn-1 + v , where v is a new vertex connected to all vertices in Cn-1 . The construction is implicit in the definition. A reference will be to a Wiki-source page. Some of the properties include:
- (a) For large n , the ratio number of edges to number of edges approaches 2.
(b) Wn is planar with a unique embedding.
(c) Wn is a self-dual graph.
(d) Wn is a Halin graph .
- (e) Wn is a (triangulated) perfect graph.
Property (d) is color coded to inform the students that topics in perfect graphs are part of a second level course. Accordingly we build a prerequisite box. With queries from students, an instructor will provide the definition of perfect graph in an assignment with examples and ask for a proof that Wn is perfect. An addition under related graphs will be the Gear graph which is a wheel with a vertex added between each adjacent pair of vertices in the outer cycle of the wheel. This is likely to be developed by students and will evolve into a separate logical page of an A-CGT-portal. Although this may enhance the scope of the topics in CGT it will be viewed positively by many students. As opposed to a special graph a logical page may describe k -regular graphs, binary trees, binomial trees, spanning trees of complete graphs, R-L-C electrical networks, Flow networks, Reliability networks etc. For an electrical sciences major R-L-C networks is a reading assignment with type P2 prerequisite box comprising phrases such as active networks and Ohm's law, Kirchhoff's current/voltage laws, network theorems, dual networks, Cutsets and Tie-sets. For Computer Science majors recommended readings will be binomial trees and Flow networks pages. It is easy to see that with the progress of a CGT course, an A-CGTportal will evolve, with learning benefits to the students in general.
## B. A-CGT portal's influence on teaching methodology and on students
It is expected that enthusiastic teachers will always offer challenges via an A-CGT-portal. Teacher's challenges in the form of assignments and tests and open questions and students' responses are analysed and best understood against the class average grade. Assuming that the possible student assessment grades are A, B, C, D, E, F it is possible to understand the level of knowledge and perseverance of the students as given below. Let T denote the term GPA of a student so far in all subjects relevant to CGT, cleared by a student. We group the students based on T -value-mapped grades and list a few descriptive attributes and conclusions for each group.
## Group 1 : T ε {A, B}
- ---Highly motivated in CGT, Able to solve a complete problem set by oneself, Able to pose well-formed problems, Self-initiated to use learning resources. Conclusion: Will be able to contribute entries to A-CGTportal e.g., algorithms for generating/drawing graphs. Group 2 : T ε {C, D} ---Fairly motivated in CGT, Able to understand the subject matter well with some effort, Completes exercises with rewards, Uses learning resources in a goal-oriented manner. Conclusion: Will find A-CGT-portal as a learning resource and with guidance can contribute to page attributes. Group 3 : T ε {E, F} ---Poorly motivated in CGT, Finds difficulty in getting started with the problems, Relies much on group activities, Needs significant effort to cross the subject. Conclusion: Will be able to use the prerequisite boxes and will
learn from A-CGT-portal at a slow pace.
The types of problems posed/arising through a typical ACGT-portal page are as mentioned below.
- (a) Easily solvable by 60% of the students.
- (b) Solvable with a knowledge of higher Mathematics developed through the CGT course and prerequisites.
- (c) Solvable via a programming language such as Java/C++ .
- (d) Conjectures for which solution approaches are sought.
(e) Problems and solutions from international contests such as Math Olympiad, ACM/ICPC.
The exercises in tests or in assignments in the context of an ACGT-portal will test the following:
- (i) Understanding of the definitions and key terms, properties etc. listed in an A-CGT-portal page.
- (ii) Understanding of notions and concepts via problems not immediately apparent from an A-CGT-portal page.
For Group 2/3 students hints and suggestions in a tutorial session include:
(a) Clarifications via explanations (maybe in a native tongue) explaining associated concepts/examples and/or via given assignments/quizzes explicitly establishing the connections. (b) Supplementary reading in CGT and suggestions to explore specific A-CGT-portal pages and prerequisite boxes.
In general these will require additional reading of A-CGTportal pages. For Group 3 students counseling is also appropriate e.g., an advice to learn at a slower pace such as learning in a group under a supervision. Based on the effective usage of an A-CGT-portal and contributions to it, a teacher may expect that a Group 3/2 student will slide into Group 2/1 with the progress of a CGT course.
The number of the different types ((a) through (e) mentioned above) of exercises in assignments and tests will be decided by the percentage of students in the different Groups. We illustrate motivational quiz problems with the following two examples:
Ex. 1. A question in CGT asks to find the number of vertices and edges in the graph block\_n defined thus: A basic block is a graph Gr = (Vr, Er); Vr = (r, A, B, C, D, E, F) and Er = (rA, rB, AC, AD, BE, BF, CD, CE, DF, EF) . A block\_1 is built from two copies of the basic blocks rooted at r\_1 and r\_2 with the added edge r1r2 . A modified block\_1 is a block\_1 wherein the edge r1r2 is replaced by the path r1 -x-r2 where x is a newly added vertex. A block\_2 consists of two copies of the modified block\_1 having the new vertices x1 and x2 and with the new edge x1x2 . A modified block\_2 is a block\_2 where the edge x1x2 is replaced by the path x1-y - ; etc. x2 When block\_n page with its properties is added to the A-CGTportal an observation is the following. Redefine the basic block thus by adding a new level of vertices edges: new Vr = old Vr + {G, . . . , N}; new E = old E - {CD, CE, DF, EF} r r + {CG, CH, DI, DJ, EK, EL, FM, FN, GH, IJ, KL MN, GI, HJ, KM, LN}. We have now defined a new family of 3 -regular graphs.
Ex. 2. Another problem in CGT involves the construction of a family of 4 -regular graphs, Gk, k = 1, 2, 3, . . . . To build G1 we start with a square i.e., C4 with vertices A, B, C, D . We next attach to these four vertices four new (vertex) labeled squares AEFH, BHIJ, CKLM, DNPQ. We add the additional edges EG, HJ, KM, QN in the outer squares. Finally we add the edges EH, FI, JK, IL, MN, NP . This completes the description of G1 . To build G2 we start with a copy of G1 with vertex labels as before. At the vertices F, I, L, P we attach new squares FRST, IUVW, LXYZ, PZ1Z2Z3 . We complete the construction of G2 by adding the edges RU, SV, WX, VY, ZZ\_1, YZ2, TZ3, SZ2 . With Gk defined the questions can be (a) With reasoning state how many C3, C4, C6 are present. (b) Give the recurrence relations to determine the number of vertices and edges.
Note that the prerequisite box for the block\_n page and for the Gk page will be those embracing graph fundamentals including terms and phrases such as k -regular graphs, isomorphism, recurrence relations. A student contribution to block\_n page or Gk page can be a set of figures of the defined graph families. With examples 1 and 2 above a Computer Science student will be invited to explore more about regular graphs e.g., Petersen graph, Harary graphs and will look forward to lectures based on textbook-based theorems. A Group 1 student is likely to discover a general principle such as -duplicate copies of an arbitrary graph on n vertexes, selectively vertexjoined via new edges will increase appropriate vertex degrees by 1 giving a random graph on 2n vertices‖ and will give an algorithm for the question -how to construct a (k+1)regular graph from an k -regular graph‖.
Consider again the family block\_n as an A-CGT-portal page and assume that the concept of distance in graphs is known to the students. An assignment exercise now will be: For the block\_n find the sum of the distances between vertices u and v , the sum taken across all distinct (u,v) pairs. Further successor questions will be of the following types:
- (a) For the graphs block\_n and Gk find the ratio of number of edges to number of vertices.
- (b) Upto isomorphism, enumerate all the spanning trees of the Petersen graph and find their Wiener indexes.
- (c) Get the definition of the Boolean hypercube of order n . (i) Code in Java/C++ to compute the Wiener indexes for n = 1 through 5 . (ii) Find a formula for the Wiener index of the n -hypercube.
It can be argued that the more-to-explore attribute and the prerequisite boxes will assist in the teaching/learning process. In the context of the above discussions, the following examples are illustrative:
- (a) Under the more-to-explore page attribute in an appropriate A-CGT-portal page can be a short write-up culled from contemporary literature, to provide possible use of Wienertype indexes in Biochemistry for identification of lead drug candidates. This will add a further impetus for a Group 1 Computer Science student to discover more about total distance and other concepts such as cliques/independent sets in Graph theory.
- (b) It has been widely suggested (see [8] for example) that algorithm based problem-solving involves abstractions and it
is an essential component in software development, a major concern for Computer Science students. An added activity of algorithm development and/or coding, for Computer Science majors, via an A-CGT-portal is therefore desirable. To begin with, the more-to-explore page attribute in an A-CGT-portal will point to pertinent combinatorial algorithms and graph algorithms - these may be to previously seen lectures and will provide more references for Group 1/2 students. In addition, problems suitably guided by P2 prerequisite boxes are likely to suggest programming exercises e.g., computing the Wiener index, given the weighted adjacency matrix of a graph suggesting say breadth-first search or Floyd-Warshall algorithms. In this process one may expect that Group 1 students will also propose new algorithms and program codes.
## IV. CONCLUDING REMARKS
A remark is that summative assessment with regard to students' contributions is a difficult task in the A-CGT-portal context because collaborations and consultations of other sources are rather difficult to quantify. An immediate honorable mention in an A-CGT-portal is a partial solution but it may not bring out the excellence in Group 1 students. Direct questions in programming/traditional assignments and tests based on students' contributions are likely to yield unfavourable feedbacks from all other students when the competitive spirit is at a level beyond a requisite level which again is difficult to measure. A team of experienced moderators knowing the students' previous performance and extent of work is a solution approach.
An outcome of maintaining students' records in a log, on the activities of contributing to page attributes or contributions as new page on the A-CGT-portal is that they can be suitably guided to credit further offered electives such as Coding Theory, Topics in Graph theory based on their major, and do project works in later semesters subject to their interests and career/program goals. At the end of a CGT course a developed A-CGT-portal can be edited further to serve as a profitable learning resource for various interest groups.
The pedagogical value of an A-CGT-portal should be apparent. Its difference from an annotated set of Wiki-pages as a supplementary resource for unsupervised learning can also be appreciated. In a CGT course students' exploration of an ACGT-portal will depend on attitudes, intellect and inquisitiveness. From the example illustrations it is possible to reason that an A-CGT-portal can judiciously contribute to the various levels of students' cognition as viewed from a psychological perspective, although a conclusion demands an experimental study and inference mechanisms. As pointed out in [8] the processes of abstracting, concretizing, synthesizing and analyzing are the mental faculties that are to be perhaps developed in teaching Computer Science - the extent to which an A-CGT-portal can contribute to these processes is an interesting question. Also, from the viewpoint of freshers and some Group 2 and Group 3 students, didactical questions need further study. We have not addressed many other important aspects such as design, formats, presentation, interfaces pertaining to an A-CGT-portal and its ownership and management, copyright issues, incorporation of students' feedbacks, security issues, in the context of portal development. However these are beyond the central theme in this paper. Knowledge representation in an A-CGT-portal is another pertinent factor that will contribute to the ease of learning concepts. It may be argued that the portal-idea suggested is more towards specializing in one associated subject rather than breadthwise knowledge acquisition across the wide spectrum of subjects in Computer Science. This is to be weighed against further dependent courses in the program and the stated program outcomes/ goals.
## ACKNOWLEDGEMENTS
The author would like to thank Dr.S.Sundarrajan, National Institute of Technology, Tiruchirapalli for his supporting this work.
## REFERENCES
- [1] P.Brusilovsky, J.Eklund, E.Schwarz, -Web-based education for all: a tool for developing adaptive courseware,‖ Comp. Netwk. and ISDN Syst., 30, 291-300, 1998.
- [2] E.Melis, E.Andres, J.Budenbender, A.Frischauf, G.Goguadze, P. Libbrecht, M.Pollet, C.Ullrich, -ActiveMath: A Generic and Adaptive Web-Based Learning Environment, ‖ Int'l. J. Artif. Intell. in Educ., 12, 385-407, 2001.
- [3] M.O.Z.San Pedro, R.S.J.d.Baker, M.M.T.Rodrigo, -Carelessness and Affect in an Intelligent Tutoring System for Mathematics,‖ Int'l. J. Artif. Intell. in Educ., 24, 189-210, 2014.
- [4] N..J.A.Sloane, -The Online Encyclopaedia of Integer Sequences,‖ 2014: http://oeis.org.
- [5] R.Balakrishnan, K.Renganathan, A textbook of Graph theory,Springer, 2000.
- [6] G.Chartrand, P.Zhang, Introduction to Graph theory, McGraw-Hill, 2006.
- [7] T.Andreescu, Z.Feng, A path to Combinatorics for Undergraduates, Birkhauser, 2004.
- [8] J.C.Perrenet, -Levels of Thinking in Computer Science: Development in Bachelor Students' Conceptualization of Algorithms,‖ Educ. Infor. Tech.15,87-107, 2010.
----------------------------------------------------------------------------------------------------------------------------------------------------------
A136328 Wiener Index of the Odd graph
0, 3, 75, 1435, 25515, 436821, 7339332, 121782375, 2005392675, 32835436777, 535550923908, 8707954925033, 141270179732500, 2287544190032700, 36988236910737360, 597341791692978975, 9637351741503033075
OFFSET
1,2
COMMENTS
The odd graph O\_n (n>=2) is a graph whose vertices represent the (n-1)-subsets of {1,2,...,2n-1} and two vertices are connected if and only if they correspond to disjoint subsets. It is a distance regular graph. [Emeric Deutsch, Aug 20 2013]
REFERENCES
Kailasam Viswanathan Iyer, Some computational and graph theoretical aspects of Wiener index, Ph.D., dissertation, Dept. of Comp. Sci. & Engg., National Institute of Technology, Trichy, India, 2007.
LINKS
Table of n, a(n) for n=1..17.
R. Balakrishnan, N. Sridharan and K. Viswanathan Iyer, The Wiener index of Odd graphs, J. Ind. Inst. Sci., vol. 86, no. 5, 2006.
FORMULA
A formula can be found in the Balakrishnan et al. reference.
A formula is "hidden" in the 2nd Maple program. B(n) and C(n) are the intersection arrays of O\_n, H(n) is the Hosoya-Wiener polynomial of O\_n, and Wi(n) is the Wiener index of O\_n. [Emeric Deutsch, Aug 20 2013]
EXAMPLE
a(2)=3 is the Wiener index of O\_2 which is C\_3.
a(3)=75 is the Wiener index of O\_3 which is the Petersen graph.
MAPLE
A136328d := proc(k) add( (2*j+1)*binomial(k-1, j)^2/(1+j), j=0..(k/2-1) ); %+2*add( (k-1-j)*binomial(k-1, j)^2/(1+j), j=floor(k/2)..(k-2) ); k*% ; end proc:
A136328 := proc(n) binomial(2*n-1, n-1)*A136328d(n)/2 ; end proc:
seq(A136328(n), n=1..20) ;
# R.J.Mathar, Sep 15 2010
B := proc (n) options operator, arrow: [seq(n-floor((1/2)*m), m = 1 .. n-1)] end proc: C := proc (n) options operator, arrow: [seq(ceil((1/2)*m), m = 1 .. n- 1)] end proc: H := proc (n) options operator, arrow: (1/2)*binomial(2*n-1, n- 1)*(sum((product(B(n)[r]/C(n)[r], r = 1 .. j))*t^j, j = 1 .. n-1)) end proc: Wi := proc (n) options operator, arrow: subs(t = 1, diff(H(n), t)) end proc: seq(Wi(n), n = 2 .. 20);
# Emeric Deutsch, Aug 20 2013
KEYWORD
nonn
AUTHOR
K.V.Iyer, Mar 27 2008
EXTENSIONS
Extended by R.J.Mathar, Sep 15 2010
STATUS
Approved
----------------------------------------------------------------------------------------------------------------------------- ---------------------------
Illustration 1: A sample page from the OEIS
| null |
[
"K. Viswanathan Iyer"
] |
2014-08-04T04:53:22+00:00
|
2014-08-04T04:53:22+00:00
|
[
"cs.CY"
] |
A case for Intranet-based 0nline portal for undergraduate Computer Science education
|
Our proposal for selective subjects especially those involving intensive
problem-solving assignments and/or tutorials, such as Introduction to
Algorithms and Data structures, Discrete Mathematics, Coding Theory, Number
theory, Combinatorics and Graph Theory (CGT), Automata theory, is to supplement
lectures with a moderated online forum against an intranet portal. By way of
illustration we take the example of a restricted view of OEIS
(http://oeis.org). The restriction can be w.r.t. sequences in OEIS that are
directly relevant to say CGT. N.J.A.Sloane's OEIS is a collection of over
2,39,147 integer sequences and their properties. In particular OEIS contains
definitions of many combinatorial structures, dense range of interpretations,
generating functions and conjectured ones, cross references within OEIS and to
outside resources, references to texts and technical articles, codes in Maple,
Mathematica etc. For organizing courses such as the above mentioned, a first
task is to partially create an OEIS-like instructor-moderated portal in a
university intranet. During the course of lectures and tutorials students are
invited to contribute to the portal and these may be augmented/approved by
instructors suitably, to find a place in the portal. By this many concepts can
be conveyed to the students in an interesting way with the desired results. In
the arguments presented, examples related to CGT are given.
|
1408.1033v1
|
## THE ULTRAVIOLET BRIGHTEST TYPE Ia SUPERNOVA 2011de
Peter J. Brown
George P. and Cynthia Woods Mitchell Institute for Fundamental Physics & Astronomy, Texas A. & M. University, Department of Physics and Astronomy, 4242 TAMU, College Station, TX 77843, USA
Draft version August 6, 2014
## ABSTRACT
We present and discuss the UV/optical photometric light curves and absolute magnitudes of the Type Ia supernova (SN) 2011de from the Swift Ultraviolet/Optical Telescope. We find it to be the UV brightest SN Ia yet observed-more than a factor of ten brighter than normal SNe Ia in the midultraviolet. This object is an extreme example of the differences seen in the ultraviolet for objects which do not appear remarkable in the optical. We find that the UV/optical brightness and broad light curves are broadly consistent with additional flux from the shock of the ejecta hitting a red giant companion. SN 2011de is either the first external interaction of a SN Ia discovered in the UV or an extreme example of the intrinsic UV variations in SNe Ia.
Subject headings: supernovae: general - supernovae: individual (SN2011de, SN2004dt, SN2009dc, SN2011aa, SN2012dn) - ultraviolet: general
## 1. INTRODUCTIONS
Type Ia Supernovae (SNe Ia) are useful standardizable candles because their near-uniform optical luminosities can be further calibrated based on their light curve shape and/or colors (Phillips 1993; Riess et al. 1996; Goldhaber et al. 2001). Assuming that the luminosity relations do not evolve with redshift, they can be used to measure cosmological distances and the expansion of the universe (e.g. Betoule et al. 2014).
lengths. Optically normal SNe Ia show a larger dispersion in the UV, and optically bright SNe Ia have an even larger (relative) increase in UV luminosity.
The near-infrared (NIR) magnitudes of SNe Ia are even more uniform than the optical (Meikle 2000; Krisciunas et al. 2004; Barone-Nugent et al. 2012). On the other side of the spectrum, detailed light curve studies at ultraviolet (UV) wavelengths have only recently reached large enough sample sizes for reasonable comparisons. This is mostly driven by the efficient observations of SNe by the Ultraviolet/Optical Telescope (UVOT; Roming et al. 2005) on board the Swift spacecraft (Gehrels et al. 2004). Early light curve samples from the International Ultraviolet Explorer (IUE), Hubble Space Telescope (HST), and UVOT showed a fair degree of uniformity among normal SNe Ia in the near-UV (Kirshner et al. 1993; Milne et al. 2010) with large variations only appearing in the mid-UV shortward of about 2700 ˚ (Brown et al. 2010). A HST observations of several more SNe Ia (Wang et al. 2012) and larger samples from UVOT show significant differences in the near-UV colors of normal SNe (Milne et al. 2013) that may arise from different populations whose relative fractions could evolve with redshift (Milne et al. 2014). SNe which are nearly identical in the optical and near-UV can still show differences in the mid-UV (Foley & Kirshner 2013). The overluminous SN 2009dc, which likely had more than a Chandrasekhar mass of ejecta, was found to be brighter than average in the UV (Silverman et al. 2011) along with three possibly related objects (Brown et al. 2014b; Scalzo et al. 2014). Thus SNe Ia become more similar at longer wavelengths and more different at shorter wave-
In this letter we discuss the UV brightness of a SN 2011de, an SN Ia which was not noted to be exceptional in the optical but is the UV brightest SN Ia yet observed. In Section 2 we present UV/optical photometry from Swift UVOT. In Section 3 we compare the colors and absolute magnitudes with other SNe Ia and examples from other SN types. In Section 4 we compare SN 2011de to a model that includes emission from a shock interaction with a red giant companion star and summarize in Section 5.
## 2. SWIFT OBSERVATIONS OF SN 2011de
SN 2011de was discovered by Newton & Puckett (2011) on 2011 May 22.305. The only reported upper limit prior to discovery is on 2010 March 30, so there are no strong constraints on the explosion date. Marion & Vaz (2011) classified it as a type Ia SN, noting a double bottomed Si II feature. Balam et al. (2011) reported it to be similar to SN 2003du, a SN with high velocity features (Gerardy et al. 2004).
The host galaxy UGC10018 is located at a redshift of 0.029187 (Falco et al. 1999). We adopt a distance modulus of 35.50 ± 0.17 using H =72.0 (Freedman et al. 0 2001). We use a line of sight extinction in the Milky Way corresponding to an E(B-V)=0.03 from Schlafly & Finkbeiner (2011). Due to the exceptionally blue colors we do not correct for any intrinsic reddeningany intrinsic reddening would imply the intrinsic colors and luminosities are even more extreme than what we show here.
SN 2011de was observed with the Swift spacecraft beginning 2011 May 28.2. Photometry was obtained in six UV and optical bands (see Roming et al. 2005 and Poole et al. 2008 for filter details). Swift UVOT data were analyzed using the reduction method of the Swift Optical/Ultraviolet Supernova Archive (SOUSA, Brown et al. 2014a). This includes subtraction of the un-
Figure 1. UVOT light curves of SN 2011de. The optical light curves are quite broad, while the peak brightness in the UV was reached at or before our first observation.
derlying galaxy flux and uses the revised UV zeropoints and time-dependent sensitivity from Breeveld et al. (2011). The photometry in all six filters is given in Table 1. The corresponding light curves are displayed in Figure 1. For simplicity we will focus on three filters with which to study colors and absolute magnitudes. We use uvm2 for the mid-UV (or MUV), u for the near-UV (or NUV), and the v band for the optical. This is similar to Brown et al. (2014b) with u being substituted for uvw1 because the long optical tails of the uvw1 filter complicate the comparison of objects with different spectral shapes.
## 3. RESULTS
## 3.1. Light Curves
The optical light curves of SN 2011de are very broad, comparable to the broadest SNe Ia observed: SNe 2001ay (Krisciunas et al. 2011) and 2011aa (Brown et al. 2014b). Stretching the MLCS2k2 B and V templates Jha et al. (2007), we estimate the time of b maximum to be MJD 55715.7 ± 1 7 and ∆ . m 15 ( B )= 0.70 ± 0.15. The templates give a peak magnitude of 18.9 ± 0 1 . in both the b and v bands. This corresponds to an absolute magnitude of M V = -19 7 . ± 0 2, . comparable to so-called Super-Chandrasekhar SNe Ia 2009dc (Silverman et al. 2011; Taubenberger et al. 2011) and 2006gz (Hicken et al. 2007). The UV light curves are already fading between the first two observations, so the peak brightness is unknown. Even so SN 2011de is brighter in the mid-UV than the possible SuperChandrasekhar SNe Ia studied in Brown et al. (2014b). This will be examined in more detail later.
## 3.2. Colors
In Figure 2 we compare the colors of SN 2011de to several SNe Ia which are brighter than normal in the UV: the Super-Chandrasekhar mass SN 2009dc (Silverman et al. 2011), and the similarly UV-bright SNe 2011aa, 2012dn (Brown et al. 2014b) and LSQ12gdh (Scalzo et al. 2014). We also use SN 2004dt data substituting F220W from Wang et al. (2012) for uvm2 and F330W for the Swift u band. Normal SNe Ia (see Brown et al. 2014b for the exact SNe) are shown in the grey area. We note that this grey area includes SNe which are described as NUV-red
Figure 2. uvm2-u and u-v colors of SN2011de compared to other UV-bright SNe Ia (as labeled) and to normal SNe Ia (grey shaded region).
and NUV-blue SNe by Milne et al. (2013). In u-v, all of the SNe singled out above are as blue or slightly bluer than the bluest normal SNe before maximum light. After max, however, SN 2004dt reddens faster than average, joining the NUV-red SNe by twenty days after maximum light. SN 2011de reddens at a slower rates and remains a few tenths of a magnitude bluer than the others. Its u-v color behavior is similar to SN 2011aa, while the more concrete Super-Chandrasekhar mass SN 2009dc is a little redder.
Things become a bit more varied in the mid-UV. While the uvm2-u colors of normal SNe Ia already show a larger degree of variation, SNe 2011de and 2011aa are extreme in their blue colors, about 1.5 mag bluer than SNe 2009dc and 2004dt at maximum light which are at the blue edge of the normal NUV-blue SNe. Unlike the normal SNe, which become bluer with time, SNe 2011aa, 2011de, and 2012dn get redder until ten days after peak, at which point they begin to get slightly bluer again. SN 2004dt has been getting bluer faster than the NUV-blue normals to end up about 0.5 mag bluer at twenty days after peak. LSQ12gdj, which is the bluest in u-v is the reddest in uvm2-u. Although SN 2011de is the bluest SN Ia in uvm2-u, the flux still drops to the uvm2. Thus it can only be considered 'blue' down to about 3000 ˚ before A the spectral shape becomes very red.
## 3.3. Absolute Magnitudes
The color differences between SN 2011de and the other SNe Ia are also reflected in the absolute magnitudes, as the normal SNe have a scatter which is relatively small in the optical but grows to shorter wavelengths. The
Figure 3. Left: Absolute magnitudes (correcting for distance modulus and MW extinction) of SN 2011de compared to other SNe Ia. The grey region corresponds to normal SNe Ia (see Brown et al. 2014 for details) while brighter, possible Super-Chandrasekhar mass SNe Ia and the UV bright SN 2004dt (Wang et al. 2012) are also plotted. Right: Absolute magnitudes of SN 2011de compared to other SN types. Explosion dates are approximate. SN 2010jl has been shifted to the left by thirty days to fit on the plot.
left panel of Figure 3 shows the absolute magnitude light curves in the mid-UV, near-UV, and optical compared to the distribution of normal SNe Ia and the same UV bright SNe Ia as above. The UV bright SNe show a larger spread at shorter wavelengths, especially the first two uvm2 observations of SN 2011de which are about three magnitudes brighter than the normal SNe Ia. The brightest point is the first observation, so the peak brightness is not actually known. SN 2011de behaves similarly to the other UV-bright SNe after maximum.
While SN 2011de is much brighter in the UV than normal SNe Ia, normal SNe Ia are notoriously faint in the UV (e.g. Kirshner et al. 1993). It is fair to ask how bright these are in the UV compared to other SN types. The right panel of Figure 3 compares SN 2011de to wellobserved examples of several SN types from Brown et al. (2014a). In the optical, SNe Ia are among the brightest, with only exceptionally bright SNe IIn and SLSNe (with peak magnitudes brighter than -21; Gal-Yam 2012) being brighter. The UV brightness of SNe becomes more varied in the UV, but SN 2011de remains amongst the brightest. In the mid-UV, only the bright SN IIn 2010jl, the SLSNII 2008ex, and the shock breakout of SN 2006aj/GRB060218 are brighter in our comparison.
ness, SNe 2011aa and 2012dn were not significantly overluminous or require more than a Chandrasekhar mass. SN2011de is both optical and UV bright. The integrated luminosity of SN 2011de within the UVOT wavelength range (1600 to 6000 A) is ˚ comparable to that of SN 2009dc. If the high luminosity of SN 2011de is caused by radioactive decay, it would also require it to be Super-Chandrasekhar. The large binding energy would likely result in lower velocities, as seen in most Super-Chandrasekhar candidates. SNe 2011aa and 2011de both have flat NIR light curves, similar to SN 2009dc (Friedman et al. 2014). However, SN 2011de was reported to have a high velocity component of Si II (Marion & Vaz 2011). SN 2004dt showed very high velocities and a very steep velocity gradient.
## 4. DISCUSSION
## 4.1. Commonalities and Differences among UV-bright SNe Ia
SN 2009dc had a high optical luminosity that would require more than a Chandrasekhar mass of ejecta (Silverman et al. 2011; Taubenberger et al. 2011). Brown et al. (2014b) and Scalzo et al. (2014) studied several other UV-bright SNe with possible connections to the Super-Chandrasekhar class. Despite their UV bright-
The high expansion velocities also make SNe 2004dt and 2011de exceptions to the trend that shows high velocity SNe Ia to have redder colors than their lower velocity counterparts in the optical (Foley et al. 2011) and UV (Milne et al. 2013). SNe 2004dt and 2011de also lack conspicuous CII which is found in all other NUV-blue SNe (Thomas et al. 2011; Milne et al. 2013). These relations relate to intrinsic optical colors, and these objects could very well have an additional extrinsic component to their luminosity which dominates the UV luminosity. This is less likely for SN 2004dt, which has HST grism spectra showing a broad excess of UV luminosity between 3000-4000 ˚ A rather than a rise in the UV continuum (Wang et al. 2012). This feature was interpreted by Wang et al. (2012) as arising from less iron absorption. The strong mid-UV flux of SN 2011de is more suggestive of additional flux from a hot source, one possible source for which we discuss shortly.
The similarities with SN 2004dt, the most highly polar-
Figure 4. Top Panel: UVOT light curves of SN 2011de compared to a model interaction with a separation distance of 6 × 10 13 cm. Lower Panels: The observed count rates are compared to the predicted count rates from the model in the uvm2, u, and v bands.
ized SN Ia (Wang et al. 2006), bring to question whether the high luminosity of SN 2011de could come from a favorable viewing angle of an aspheric explosion. However, asymmetric models also predict the more luminous SNe would have narrow light curves (Maeda et al. 2011).
## 4.2. Comparison with a companion shock model
One possible source of excess UV luminosity for SNe Ia is shock emission from the SN ejecta interacting with the donor companion. Kasen (2010) presented analytic and numeric models for the expected emission. Because the intrinsic SN emission in the UV was overpredicted in the numerical models, in Brown et al. (2012) we compared the predicted shock brightness in each UVOT filter from the analytic expressions to the observations to put conservative upper limits on the companion separation. The analytic expressions in Kasen (2010) give the time evolution of the shock temperature and luminosity as a function of the companion separation. The companion is assumed to be filling its Roche lobe, so the companion size is determined by the separation. In Brown et al. (2012) we also estimated the angular dependence of the luminosity via a functional form computed from the numerical models. In our two parameter model for the shock, the shock brightness when looking straight down on the interaction (in a line through the SN and the companion) and rate of fading are determined by the separation distance while an increase in the viewing angle reduces the brightness from its maximal value.
The SN component was left out in Brown et al. (2012) due to the unknown early UV behavior of SNe Ia. With the excellent early-time UV observations of SN 2011fe, which showed no evidence for shock interaction (e.g. Nugent et al. 2011), we now add the SN back in. For the SN emission we use the UV/optical spectra of SN 2011fe based on HST and SNFactory observations from Pereira et al. (2013). We use an explosion date of MJD55796.696 (Nugent et al. 2011) and a distance modulus of 29.04 mag (Shappee & Stanek 2011).
Both the shock and the SN 2011fe luminosities are corrected for the extinction and distance to SN 2011de. The comparison is done in the observed SN frame rather than being corrected to the absolute magnitudes because the extinction correction depends on the spectral shape. The spectral shape is only roughly determined by the photometric observations, while the SN shock model and SN spectrum are defined as spectral sequences and can be appropriately reddened. We find good fits to the SN plus shock for an explosion date about 15 days before the Bband peak. With this explosion date, the observations are matched well with a separation distance of 6 × 10 13 cm viewed straight on. For reference, the nominal one solar mass red giant model considered in Kasen (2010) was at 2 × 10 13 cm. The predicted light curves for this model are compared to the observations in the top panel of Figure 4. The lower panels separates the expected count rate contributions from the SN and the shock compared to the observations. In the optical, both contribute comparable amounts of flux near maximum light. In the mid-UV, the SN barely contributes anything even to later times while the shock component dominates.
The model matches the observations surprisingly well considering the small number of adjustable parameters we used (angle, explosion date, companion separation). This does not, however, mean that the model is correct unique. We are already extrapolating beyond the separation distances used in Kasen (2010) and at later epochs. The analytic expressions were favorably compared to the more sophisticated numerical simulations primarily for
the period shortly after explosion. The shock emission from the smaller separations considered in Kasen (2010) faded below the SN emission within five days after explosion. At these later times the ejecta have swept much further past the companion star, and the applicability of the analytic expressions is uncertain.
There are also uncertainties with the SN component of the model and matching it to the observations. The explosion date of SN 2011aa is not well determined, as there were no recent upper limits prior to discovery and the observations began perhaps five days before maximum light. The optical peak itself is quite broad and not wellsampled, but even with a well determined peak time, it is unknown if the rise time would be expected to be average. The fact that the observations begin so late would usually prevent observing the prompt cooling of the hot shock which has been observed in several core-collapse SNe with Swift (Campana et al. 2006; Roming et al. 2009) and possibly the SN Ia 2012cg (Marion et al. 2014, in preparation). The model shown here with a very large companion star actually has a very slowly cooling/fading shock which broadens the light curve rather than creating the characteristic dip and rebrightening usually associated with the shock breakout of a SN. Additionally, the intrinsic emission from SNe varies greatly in the mid-UV (Brown et al. 2010; Milne et al. 2013). So it would be unexpected to perfectly match the intrinsic UV brightness of the SN without the shock. Nevertheless, in the case of SN 2011de the shock emission would be much larger than the intrinsic UV emission from the SN. Further work is needed to determine how much UV emission could come from the SN itself and what other circumstellar material could produce the shock luminosity evolution possibly observed for SN 2011de. H α emission from circumstellar interaction has been observed for several SNe (e.g. Hamuy et al. 2003; Dilday et al. 2012), though UV data on such SNe is sparse (Silverman et al. 2013). The lack of H α emission seen in SN 2011de (H. Marion, 2014, private communication) may not be consistent with the red giant interaction modeled here. Our model includes the interaction radiation as a blackbody spectrum, neglecting line emission. Further work is required to predict the continuum and line fluxes for a larger collection of companion types, separation distances, and viewing angles.
## 5. SUMMARY
We have presented UV and optical photometry of SN 2011de. The first observations near maximum light show it to be extremely blue in the UV-optical, becoming redder with time. In the mid-UV, SN 2011de is about ten times brighter than the brightest normal SNe Ia. Its total integrated luminosity in the UVOT range (1600-6000 ˚ ) A is comparable to the super-Chandrasekhar SN 2009dc, but comparison of the integrated or bolometric luminosity with other bright SNe Ia is limited by the lack of UV data for those objects. We find the brightness and evolution in the UV/optical to be consistent with the superposition of flux from a normal SN Ia and the shock interaction with a large red giant companion star, though other forms of circumstellar interaction may also yield suitable fits.
Conclusively proving the origin of the excess UV flux is hard from photometry alone. The presence of strong features in the UV spectra of SNe 2009dc and 2012dn caused us to conclude that the high UV luminosity had a photospheric origin, with higher temperatures leading to a lowered UV opacity (Brown et al. 2014b). Higher quality UV spectra and a more detailed analysis are needed to confirm this. SN 2011de was too far and faint for UVOT spectra, so HST would have been required to observe the UV spectra. These objects appear to be fairly rare, so the larger aperture of HST or future UV observatories is likely needed to reach a large enough volume to observe such a SN. SN 2011de may be the first SN Ia with an interaction discovered by the UV light, or it may be the most extreme example of the UV variability intrinsic to these 'standard candles.'
The author thanks L. Wang for suggesting comparisons with the shock model. The author was supported in part by the Mitchell Postdoctoral Fellowship and NSF grant AST-0708873. The Swift Optical/Ultraviolet Supernova Archive (SOUSA) is supported by NASA's Astrophysics Data Analysis Program through grant NNX13AF35G. This work made use of public data in the Swift data archive and the NASA/IPAC Extragalactic Database (NED), which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with NASA.
## REFERENCES
Balam, D. D., Bohlender, D., Monin, D., Chisholm, E., & Green,
- D. W. E. 2011, Central Bureau Electronic Telegrams, 2653, 2 Barone-Nugent, R. L., Lidman, C., Wyithe, J. S. B., et al. 2012,
- MNRAS, 425, 1007 Betoule, M., Kessler, R., Guy, J., et al. 2014, ArXiv e-prints Breeveld, A. A., Landsman, W., Holland, S. T., et al. 2011, in American Institute of Physics Conference Series, Vol. 1358, GAMMA RAY BURSTS 2010. AIP Conference Proceedings, ed. J. E. McEnery, J. L. Racusin, & N. Gehrels, 373-376 Brown, P. J., Breeveld, A. A., Holland, S., Kuin, P., & Pritchard, T. 2014a, ArXiv e-prints Brown, P. J., Dawson, K. S., Harris, D. W., et al. 2012, ApJ, 749, 18 Brown, P. J., Roming, P. W. A., Milne, P., et al. 2010, ApJ, 721, 1608 Brown, P. J., Kuin, P., Scalzo, R., et al. 2014b, ApJ, 787, 29 Campana, S., Mangano, V., Blustin, A. J., et al. 2006, Nature, 442, 1008 Dilday, B., Howell, D. A., Cenko, S. B., et al. 2012, Science, 337, 942 Falco, E. E., Kurtz, M. J., Geller, M. J., et al. 1999, PASP, 111, 438 Foley, R. J., & Kirshner, R. P. 2013, ApJ, 769, L1 Foley, R. J., Sanders, N. E., & Kirshner, R. P. 2011, ApJ, 742, 89 Freedman, W. L., Madore, B. F., Gibson, B. K., et al. 2001, ApJ, 553, 47 Friedman, A. S., Wood-Vasey, W. M., Marion, G. H., et al. 2014, ArXiv e-prints Gal-Yam, A. 2012, Science, 337, 927 Gehrels, N., Chincarini, G., Giommi, P., et al. 2004, ApJ, 611, 1005 Gerardy, C. L., H¨flich, P., Fesen, R. A., et al. 2004, ApJ, 607, o 391 Goldhaber, G., Groom, D. E., Kim, A., et al. 2001, ApJ, 558, 359 Hamuy, M., Phillips, M. M., Suntzeff, N. B., et al. 2003, Nature, 424, 651 Hicken, M., Garnavich, P. M., Prieto, J. L., et al. 2007, ApJL, 669, L17 Jha, S., Riess, A. G., & Kirshner, R. P. 2007, ApJ, 659, 122 Kasen, D. 2010, ApJ, 708, 1025 Kirshner, R. P., Jeffery, D. J., Leibundgut, B., et al. 1993, ApJ, 415, 589
Table 1 UVOT Photometry
| SN | Filter | MJD (days) | Mag (mag) | M err (mag) | 3 σ Limit (mag) | Bright Limit (mag) | Rate (counts s - 1 ) | R err (counts s - 1 ) |
|----------|----------|--------------|-------------|---------------|-------------------|----------------------|------------------------|-------------------------|
| SN2011de | UVW2 | 55709.3 | 16.829 | 0.077 | 21.011 | 11.085 | 1.661 | 0.118 |
| SN2011de | UVM2 | 55709.3 | 16.283 | 0.063 | 21.374 | 10.555 | 1.686 | 0.098 |
| SN2011de | UVW1 | 55709.3 | 15.879 | 0.065 | 20.15 | 11.149 | 4.212 | 0.252 |
| SN2011de | U | 55709.3 | 15.063 | 0.053 | 19.673 | 12.073 | 20.452 | 1.004 |
| SN2011de | B | 55709.3 | 16.061 | 0.057 | 20.104 | 12.874 | 16.577 | 0.866 |
| SN2011de | V | 55709.3 | 16.038 | 0.062 | 19.652 | 11.603 | 5.504 | 0.312 |
Note . - The full table is available in the electronic version. The photometry will also be available from the Swift SN website http://swift.gsfc.nasa.gov/docs/swift/sne/swift sn.html.
Krisciunas, K., Suntzeff, N. B., Phillips, M. M., et al. 2004, AJ, 128, 3034
- Poole, T. S., Breeveld, A. A., Page, M. J., et al. 2008, MNRAS, 383, 627
- Krisciunas, K., Li, W., Matheson, T., et al. 2011, AJ, 142, 74 Maeda, K., Leloudas, G., Taubenberger, S., et al. 2011, MNRAS, 413, 3075
- Marion, G. H., & Vaz, A. 2011, Central Bureau Electronic Telegrams, 2728, 2
- Meikle, W. P. S. 2000, MNRAS, 314, 782
- Milne, P. A., Brown, P. J., Roming, P. W. A., Bufano, F., & Gehrels, N. 2013, ApJ, 779, 23
- Milne, P. A., Brown, P. J., Roming, P. W. A., et al. 2010, ApJ, 721, 1627
- Newton, J., & Puckett, T. 2011, Central Bureau Electronic Telegrams, 2728, 1
- Nugent, P. E., Sullivan, M., Cenko, S. B., et al. 2011, Nature, 480, 344
- Pereira, R., Thomas, R. C., Aldering, G., et al. 2013, A&A, 554, A27
- Phillips, M. M. 1993, ApJL, 413, L105
- Riess, A. G., Press, W. H., & Kirshner, R. P. 1996, ApJ, 473, 88 Roming, P. W. A., Kennedy, T. E., Mason, K. O., et al. 2005, Space Science Reviews, 120, 95
- Roming, P. W. A., Pritchard, T. A., Brown, P. J., et al. 2009, ApJ, 704, L118
- Scalzo, R. A., Childress, M., Tucker, B., et al. 2014, ArXiv e-prints
- Schlafly, E. F., & Finkbeiner, D. P. 2011, ApJ, 737, 103
- Shappee, B. J., & Stanek, K. Z. 2011, ApJ, 733, 124
- Silverman, J. M., Ganeshalingam, M., Li, W., et al. 2011, MNRAS, 410, 585
- Silverman, J. M., Nugent, P. E., Gal-Yam, A., et al. 2013, ApJS, 207, 3
- Taubenberger, S., Benetti, S., Childress, M., et al. 2011, MNRAS, 412, 2735
- Thomas, R. C., Aldering, G., Antilogus, P., et al. 2011, ApJ, 743, 27
- Wang, L., Baade, D., H¨flich, P., et al. 2006, ApJ, 653, 490 o Wang, X., Wang, L., Filippenko, A. V., et al. 2012, ApJ, 749, 126
|
10.1088/2041-8205/796/1/L18
|
[
"Peter J. Brown"
] |
2014-08-05T16:49:40+00:00
|
2014-08-05T16:49:40+00:00
|
[
"astro-ph.HE"
] |
The Ultraviolet Brightest Type Ia Supernova 2011de
|
We present and discuss the UV/optical photometric light curves and absolute
magnitudes of the Type Ia supernova (SN) 2011de from the Swift
Ultraviolet/Optical Telescope. We find it to be the UV brightest SN Ia yet
observed--more than a factor of ten brighter than normal SNe Ia in the
mid-ultraviolet. This object is an extreme example of the differences seen in
the ultraviolet for objects which do not appear remarkable in the optical. We
find that the UV/optical brightness and broad light curves are broadly
consistent with additional flux from the shock of the ejecta hitting a red
giant companion. SN~2011de is either the first external interaction of a SN Ia
discovered in the UV or an extreme example of the intrinsic UV variations in
SNe Ia.
|
1408.1034v2
|
## Pion pT spectra in p + p collisions as a function of √ s and event multiplicity
Priyanka Sett 1 and Prashant Shukla 2 Nuclear Physics Division, Bhabha Atomic Research Center 1 2 , , Trombay, Mumbai - 400 085, India Homi Bhabha National Institute 1 2 , , Anushaktinagar, Mumbai - 400094, India
## Abstract
We study the charged pion transverse momentum ( p T ) spectra in p + p collisions as a function of collision energy √ s and event multiplicity using Tsallis distribution. This study gives an insight of the pion production process in p + p collisions. The study covers pion spectra measured in p + p collisions at SPS energies (6.27-17.27 GeV), RHIC energies (62.4 GeV and 200 GeV) and LHC energies (900 GeV, 2.76 TeV and 7 TeV). The Tsallis parameters have been obtained and parameterized as a function of √ s . The study suggests that as we move to higher energy more and more hard processes contribute to the spectra. We also study the charged pion spectra for different event multiplicities in p + p collisions for LHC energies using Tsallis distribution. The variation of the Tsallis parameters as a function of event multiplicity has been obtained and their behavior is found to be independent of collision energy.
## 1 Introduction
The particle spectra measured in hadronic collisions are of utmost interest because of their fundamental nature and simplicity, which allow to verify pQCD [1, 2] calculations and also help to make comprehensive phenomenological studies. The ratios of the particle yields obtained from the measured spectra allow to get the chemical freeze-out conditions, whereas the spectra themselves reflect the conditions at the kinetic freeze-out. The particle spectra provide useful information about the collision dynamics. The low p T region of the spectrum corresponds to the particles originating from low momentum transfer and multi-scattering processes (non-perturbative QCD), whereas, the high p T region comes from the hard-parton-scattering (pQCD) among the initial partons. The transition of this non-perturbative to perturbative dynamics has no sharp boundary, though one can have an estimate from the ' x T -scaling ' [3]. Extensive [4, 5] and non-extensive [6, 7, 8, 9, 10] statistical approaches have been used to characterize particle spectra in terms of thermodynamic variables. Extensive statistics assume thermal and chemical equilibrium of the system at hadronic phase which lead to an exponential distribution of the particle spectra. In experiments, the particle spectra show a power-law behavior at high p T . This
1 email : [email protected]
2 email : [email protected]
behavior is reproduced by the non-extensive approach with an additional parameter. In recent times, the Tsallis [6] statistical approach is widely used to describe the particle spectra obtained in high-energy collisions with only two parameters; the temperature T and q , known as non-extensivity parameter which is a measure of temperature fluctuations or degree of non-equilibrium in the system.
The Tsallis distribution gives an excellent description of p T spectra of all identified mesons measured in p + p collisions at √ s = 200GeV [11]. In a recent work [12, 13], the Tsallis distribution has been used to describe the p T spectra of identified charged hadrons measured in p + p collisions at RHIC and at LHC energies. Such an approach has also been applied to the inclusive charged hadron p + p data in recent publications [14, 15]. It has been shown in Ref. [12, 16] that the functional form of the Tsallis distribution with thermodynamic origin is of the same form as the QCD-inspired Hagedorn formula [17, 18]. This could be the reason of success of Tsallis distribution in p + p collisions which is a power law typical of QCD hard scatterings. The hardness of the spectra is thus related to q and the parameter T governs the contribution from soft collisions.
Using the Tsallis phenomenological function, we review and study the charged pion spectra in p + p collisions in a large energy regime, spanning from SPS [19] (6.27 GeV -17.27 GeV), RHIC [20] (62.4 and 200 GeV) to LHC [21] (900 GeV, 2.76 TeV and 7 TeV) energies. The object of the present work is to study the behaviour of the Tsallis parameters as a function of collision energy. We also study the charged pion spectra for different event multiplicities in p + collisions for LHC energies. Among all hadrons, pions p are chosen because of their abundance in collisions, simple quark structure and availability of the data at different energies.
## 2 Formalism
The transverse momentum spectra of hadrons, obtained from different fixed and collider experiments have shown that, the high p T region of the spectra can be described successfully by the power law,
$$E \frac { d ^ { 3 } N } { d p ^ { 3 } } = C _ { P } p _ { T } ^ { - n },$$
where C P is the normalization constant and n is the power which determines the shape of the spectra at high p T . However, the low p T region of the particle spectra shows an exponential shape and can be described by the Boltzmann-Gibbs [22, 23] statistical approach,
$$E \frac { d ^ { 3 } N } { d p ^ { 3 } } = C _ { B } e ^ { - E / T },$$
where C B is the normalization constant, E is the particle energy and T is the temperature of the system.
In the early 80's, Hagedorn [17] proposed a phenomenological function which describes the particle spectra for both the higher and lower p T regions:
$$E \frac { d ^ { 3 } N } { d p ^ { 3 } } = A \left ( 1 + \frac { p _ { T } } { p _ { 0 } } \right ) ^ { - n },$$
where A p , 0 and n are the fit parameters. The above equation describes an exponential behavior for low p T and a power-law behavior for high p T .
$$\left ( 1 + \frac { p _ { T } } { p _ { 0 } } \right ) ^ { - n } \ \simeq \ \exp \left ( \frac { - n p _ { T } } { p _ { 0 } } \right ), \ \text{ for } p _ { T } \to 0$$
$$\frac { p _ { T } } { p _ { 0 } } \right ) ^ { - n } \ & \simeq \ \exp \left ( \frac { - n p _ { T } } { p _ { 0 } } \right ), \ \ \text{ for } \ p _ { T } \to 0 \\ & \simeq \ \left ( \frac { p _ { 0 } } { p _ { T } } \right ) ^ { n }, \ \ \text{ for } \ p _ { T } \to \infty.$$
The parameter n in this equation is often related to the 'power' in the 'QCD-inspired' quark interchange model [18].
In the late 80's, Tsallis [6] introduced the idea of the non-extensive statistics in place of thermal Boltzmann-Gibbs statistics. This approach includes a parameter q , known as non-extensive parameter which quantifies the temperature fluctuation [24] in the system as : q -1 = V ar (1 /T / T ) 〈 〉 2 . The non-extensive statistics assume Boltzmann-Gibbs form in the limit q → 1. In Tsallis approach, the Boltzmann-Gibbs distribution takes the form
$$E \frac { d ^ { 3 } N } { d p ^ { 3 } } = C _ { q } \left ( 1 + ( q - 1 ) \frac { E } { T } \right ) ^ { \frac { - 1 } { q - 1 } },$$
where C q is the normalization factor. One can use the relation E = m T at mid-rapidity and n = 1 ( / q -1) in Eq. 6 to obtain :
$$E \frac { d ^ { 3 } N } { d p ^ { 3 } } = C _ { n } \left ( 1 + \frac { m _ { T } } { n T } \right ) ^ { - n },$$
where, C n is the normalization factor. Eq. 7 can be re-written as :
$$\frac { 1 } { 2 \pi p _ { T } } \frac { d ^ { 2 } N } { d p _ { T } d y } = C _ { n } \left ( 1 + \frac { m _ { T } } { n T } \right ) ^ { - n },$$
The value of C n can be obtained by integrating Eq. 8 over momentum space :
$$C _ { n } = \frac { d N / d y } { \int _ { 0 } ^ { \infty } \left ( 1 + \frac { m _ { T } } { n T } \right ) ^ { - n } 2 \pi p _ { T } d p _ { T } }, \\ \text{ly } d N / d y \, & \, is \, the \, p _ { T } \, \text{integrated yield. } \, \text{Eq. } 7 \, \text{with the normalization constant}$$
Here the quantity dN/dy is the p T integrated yield. Eq. 7 with the normalization constant takes the form [11] :
$$E \frac { d ^ { 3 } N } { d p ^ { 3 } } \ = \ \frac { 1 } { 2 \pi } \frac { d N } { d y } \frac { ( n - 1 ) ( n - 2 ) } { ( n T + m ( n - 1 ) ) ( n T + m ) } \left ( \frac { n T + m _ { T } } { n T + m } \right ) ^ { - n },$$
Table 1: The center of mass energy and rapidity of the data used for the study.
| Experiments | Center of mass energy (GeV) | Rapidity | Particles Studied |
|---------------|--------------------------------|--------------|---------------------|
| SPS | 6.27, 7.74, 8.76, 12.32, 17.27 | 0.0-0.2 | π - |
| RHIC | 62.4, 200 | | y | < 0.35 | π + , π - |
| LHC | 900, 2760, 700 | y < 1.0 | π + , π - |
|
|
Larger values of q correspond to smaller values of n which imply dominant hard QCD point-like scattering. Both n and q have been interchangeably used in Tsallis distribution [7, 11, 20, 25, 26]. The Tsallis interpretation of parameters T as temperature and q as non-extensivity parameter is more suited for heavy ion collisions while for p + p collisions Hagedorn interpretation in terms of power n and a parameter T = p /n 0 which controls soft physics processes is more meaningful. Phenomenological studies suggest that, for quark-quark point scattering, n ∼ 4 [27, 28] and when multiple scattering centers are involved n grows larger.
There are many other forms of Eq. 10, which are used often to describe particle spectra in literature, see e.g. Refs. [7, 20, 25, 26, 29, 30, 31].
## 3 Results and Discussions
All the studies are performed with Eq. 10 and the fit parameters n , T and dN/dy are obtained. The different experiments, energies, rapidity ranges and particles used in the analysis are summarized in Table 1. For SPS energies only available data is for π -measured by NA61 Collaboration [19], for RHIC and LHC energies we use ( π + + π -)/2. All the data used are measured in mid-rapidity and are given for unit rapidity. The difference in rapidity range is not expected to affect the behaviour of the spectra. CMS experiment presented [21] transverse momentum spectra for different events classified on the basis of number of true tracks referred here as track multiplicity of event or simply multiplicity. Each multiplicity class is represented by average number of tracks ( 〈 N tracks 〉 ).
## 3.1 Tsallis parameters as a function of √ s in p + p system :
In this analysis all the Tsallis parameters are obtained for charged pion spectra as a function of √ s in p + p system for SPS [19], RHIC [20] and LHC [21] energies. Similar study is available in Ref. [12] using RHIC and LHC data and in Ref. [13] for SPS and LHC data.
The pion p T spectra measured in p + p collisions at different √ s are shown in Fig. 1
T
Figure 1: (Color online) The invariant yield spectra of charged pions as a function of ( m T -m ) for SPS [19] energies 6.27 GeV, 7.74 GeV, 8.76 GeV, 12.32 GeV and 17.27 GeV, RHIC [20] energies 62.4 GeV and 200 GeV and LHC [21] energies 900 GeV, 2.76 TeV and 7 TeV. The solid lines are the Tsallis function (Eq. 10). The negative pion yields are plotted for SPS energies and for all other energies, average yield for positive and negative pion are plotted.
Figure 2: (Color online) The variation of the Tsallis parameter n for charged pions as a function of √ s . The solid curve represents the parameterization ( a +( √ s ) -α b ) .
Table 2: Values of the χ /NDF 2 for Tsallis fits of pion spectra at different √ s . √ s
| s | χ 2 /NDF |
|--------------------------------------------------------------------------------|---------------------------------------------------------------------------|
| 6.27 GeV 7.74 GeV 8.76 GeV 12.32 GeV 17.27 GeV 62 GeV 200 GeV 900 GeV 2.76 TeV | 6.31/12 5.80/12 10.68/12 9.25/12 2.65/12 0.74/11 0.48/11 24.33/19 5.59/19 |
along with with Tsallis fits (Eq. 10) shown by solid lines. The spectra are scaled by arbitrary factors (given in figure) for visual clarity. In case of RHIC data, we restrict the p T range to 1.7 GeV/ c 2 to have similar p T range at all energies. It can be noticed that the spectra become harder with increase in √ s which is depictive of occurrence of harder scatterings at higher collision energy. The χ 2 per degree of freedom χ /NDF 2 values for all the fits are given in Table 2. The χ /NDF 2 values are /lessorsimilar 1, which is indicative of good fit quality.
The parameters n and T obtained from this analysis are shown in Fig. 2 and Fig. 3, respectively, as a function of √ s . The variation of dN/dy as a function of √ s is shown in Fig. 4. The parameter n decreases with increasing √ s and starts saturating at LHC energies. The value of T also reduces slowly from SPS energies to LHC energies. The integrated yield dN/dy increases 10 times when going from SPS to highest LHC energy.
Larger value of n (also larger value of T ) suggests that the spectra has contribution from processes involving small momentum transfer arising due to the re-scattering, recombination of partons, fragmentation from strings etc. Whereas, smaller values of n are indicative of harder processes are involved in particle-production. Thus the spectra at SPS energies have large softer contribution and as the collision energy increases more and more contribution from hard processes are added.
All the three parameters can be parametrized by a function of type
$$f ( \sqrt { s } ) = \left ( a + ( \sqrt { s } ) ^ { - \alpha } \right ) ^ { b } \\ 8 \ \alpha = 0 \, 2 + 0 \, 0 6 \, \text{and} \, h = 4 \, 3 6 + 0 \, 2 4 \, \text{for} \ n ( \sqrt { s } ) \ a = 2 \, 6 3 + 0 \, 6 2$$
Here a = 1 33 . ± 0 08, . α = 0 22 . ± 0 06 and . b = 4 36 . ± 0 24 for . n ( √ s ), a = 2 63 . ± 0 62, . α = 0 04 . ± 0 02 and . b = 3 76 . ± 0 49 for . T ( √ s ) and a = 0 65 . ± 0 01, . α = 0 22 . ± 0 01 and . b = -4 78 . ± 0 03 for . dN ( √ s /dy ) . Using the parameterizations for n by Eq. 11 we get n ∼ 3.46 in the limit √ s →∞ . The extrapolated values for n , T and dN/dy for √ s = 14 TeV are, n ∼ 5.09, T ∼ 90.33 MeV and dN/dy ∼ 3.44.
Figure 3: (Color online) The variation of the Tsallis parameter T for charged pions as a function of √ s . The solid curve represents the parameterization ( a +( √ s ) -α b ) .
Figure 4: (Color online) The variation of the integrated yield dN/dy for charged pions as a function of √ s . The solid curve represents the parameterization ( a +( √ s ) -α ) . b
Figure 5: (Color online) The invariant yield spectra of ( π + + π -)/2 [21], as a function of m T -m for p + p collisions at √ s = 900 GeV. The yields are shown for 〈 N tracks 〉 7, 16, 28, 40, 52, 63 and 75. The spectra are scaled up for clarity by a factor of 6 , where i = 0, i 1, 2, 3, 4, 5 and 6. The solid lines show the Tsallis fits.
## 3.2 Tsallis parameters as a function of multiplicity ( 〈 N tracks 〉 ) for LHC energies :
The Tsallis parameters for charged pion spectra are studied as a function of event multiplicity for different LHC energies 900 GeV, 2.76 and 7 TeV. The event multiplicity data was also studied in a recent work [10] but our analysis and interpretations are different.
The invariant yield spectra corresponding to different multiplicities are fitted with Eq. 10, are shown by the solid black lines in Fig. 5 for 900 GeV, in Fig. 6 for 2.76 TeV and in Fig. 7 for 7 TeV center of mass energy. The spectra are scaled up for distinctness. The Tsallis distribution describes all the spectra well, shown by the χ /NDF 2 values given in Table 3. The χ /NDF 2 values are little higher for some of the lower multiplicities due to the deviation of first data point in p T spectra with the curve.
The parameters n and T obtained from the fits are shown in Fig. 8 and Fig. 9 respectively, as a function of 〈 N tracks 〉 . The circles, squares and triangles correspond to the parameter values obtained from data at 900 GeV, 2.76 TeV and 7 TeV, respectively. It is seen that both the parameters n and T decrease rapidly and then start saturating with the increase of 〈 N tracks 〉 for all three energies. This variation (of n and T ) is very similar to the variation which we find as a function of √ s . It means that events with higher multiplicity have larger contribution from hard processes. The value of n for high multiplicity events at 7 TeV is ∼ 4 which is depictive of production from point quark-quark
Figure 6: (Color online) The invariant yield spectra of ( π + + π -)/2 [21], as a function of m T -m for p + p collisions at √ s = 2.76 TeV. The yields are shown for 〈 N tracks 〉 7, 16, 28, 40, 52, 63, 75, 86 and 98. The spectra are scaled up for clarity by a factor of 3 , where i i = 0, 1, 2, 3, 4, 5, 6, 7 and 8. The solid lines show the Tsallis fits.
Figure 7: (Color online) The invariant yield spectra of ( π + + π -)/2 [21], as a function of m T -m for p + p collisions at √ s = 7.00 TeV. The yields are shown for 〈 N tracks 〉 7, 16, 28, 40, 52, 63, 75, 86, 98, 109, 120 and 131. The spectra are scaled up for clarity by a factor of 3 , where i = 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 and 11. i The solid lines show the Tsallis fits.
Table 3: Values of the χ /NDF 2 for the Tsallis fits in different event multiplicities.
| 〈 N tracks 〉 | χ 2 /NDF values for | χ 2 /NDF values for | χ 2 /NDF values for |
|------------------|-----------------------|-----------------------|-----------------------|
| | 900 GeV | 2.76 TeV | 7.0 TeV |
| 7 | 48.89/18 | 57.66/18 | 48.10/18 |
| 16 | 25.21/18 | 26.38/18 | 27.76/18 |
| 28 | 9.97/18 | 11.47/18 | 22.83/18 |
| 40 | 7.20/18 | 6.91/18 | 19.25/18 |
| 52 | 8.00/18 | 6.08/18 | 22.52/18 |
| 63 | 9.18/18 | 6.44/18 | 26.30/18 |
| 75 | 15.01/18 | 9.05/18 | 21.30/18 |
| 86 | | 8.16/18 | 19.24/18 |
| 98 | | 11.91/18 | 23.59/18 |
| 109 | | | 20.82/18 |
| 120 | | | 16.85/18 |
| 131 | | | 19.77/18 |
Figure 8: (Color online) The variation of the Tsallis parameter n for charged pions as a function of 〈 N tracks 〉 . The variation is shown for 900 GeV by black circles, 2.76 TeV by red squares and 7.00 TeV by green triangles. The dashed curve represents the parameterization ( a +( 〈 N tracks 〉 ) -α b ) .
Figure 9: (Color online) The variation of the Tsallis parameter T for charged pions as a function of 〈 N tracks 〉 . The variation is shown for 900 GeV by black circles, 2.76 TeV by red squares and 7.00 TeV by green triangles. The dashed curve represents the parameterization ( a +( 〈 N tracks 〉 ) -α b ) .
Figure 10: (Color online) The variation of the integrated yield dN/dy for charged pions as a function of 〈 N tracks 〉 . The variation is shown for 900 GeV by black circles, 2.76 TeV by red squares and 7.00 TeV by green triangles.
scattering. The variation of n and T as a function of 〈 N tracks 〉 can be described by the same curve given in the figure for all three energies and are parameterized by,
$$f ( ( < N _ { t r a c k s } > ) = \left ( a + ( < N _ { t r a c k s } > ) ^ { - \alpha } \right ) ^ { b } & & ( 1 2 ) \\ 1 1 3 + \theta \, \theta 1 \ \alpha \, - \, \theta \, 8 1 + \theta \, \theta \, \text{and} \, h \, - \, 1 0 \, 3 0 + \theta \, 7 6 \ \text{for} \ n / N. \quad, \, \bigcup \, \text{and} \, a \, -$$
The p T integrated pion yield distribution in different multiplicity classes is shown in Fig. 10 for the three LHC energies. The total p T integrated pion yield for each energy can be obtained by integrating the above distributions over all multiplicity classes. It is noticed that as the energy increases more and more high mutliplicity events are added in the sample with mean of the distribution shifting towards higher 〈 N tracks 〉 .
Here a = 1 13 . ± 0 01, . α = 0 81 . ± 0 04 and . b = 10 32 . ± 0 76 for . n ( 〈 N tracks 〉 ) and a = 2 20 . ± 0 06, . α = 0 56 . ± 0 08 and . b = 5 33 . ± 0 23 for . T ( 〈 N tracks 〉 ).
## 4 Conclusion
This work presented the study of the transverse momentum spectra of the charged pions for different collisional energies and also for different event-multiplicities (at LHC energies) using Tsallis distribution. The Tsallis parameter n decreases with increasing √ s and starts saturating at LHC energies. The value of T also reduces slowly from SPS energies to LHC energies. It means that the spectra at SPS energies have large softer contribution and as the collision energy increases more and more contribution from hard processes are added. The p T integrated pion yield increases with increasing √ s and becomes 10 times when going from SPS to highest LHC energy. The Tsallis parameters are also obtained as a function of event multiplicity for all three LHC energies which can be described by the same curve. The variation of n and T as a function of multiplicity is very similar to the variation which we find as a function of √ s . It means that events with higher multiplicity have larger contribution from hard processes. The value of n for high multiplicity events at 7 TeV is ∼ 4 which is depictive of production from point quarkquark scattering. The p T integrated pion yield distribution for the three LHC energies shows that as the energy increases, more and more high mutliplicity events are added in the sample with mean of the distribution shifting towards higher multiplicity.
## References
- [1] G. Sterman et al. (CTEQ Collaboration), Rev. Mod. Phys. 67 , 157 (1995).
- [2] J.F. Owens et al. , Phys. Rev. D 18 , 1501 (1978).
- [3] R. Blankenbecler, S.J. Brodsky and J.F. Gunion, Phys. Lett. B 42 , 461 (1972).
- [4] E. Fermi, Progress Theor. Phys., 5 , 570 (1950).
- [5] A. Andronic, P. Braun-Munzinger, J. Stachel, Nucl. Phys. A 772 , 167 (2006).
- [6] C. Tsallis, J. Stat. Phys. 52 , 479 (1988).
- [7] T.S. Biro, G. Purcsel and K. Urmossy, Eur. Phys. J. A 40 , 325 (2009).
- [8] W.M. Alberico and A. Lavagno, Eur. Phys. J. A 40 , 313 (2009).
- [9] T.S. Biro and G. Purcsel, Phys. Rev. Lett. 95 , 162302 (2005).
- [10] K. Urmossy, arXiv:1212.0260[hep-ph].
- [11] A. Adare et al. (PHENIX Collaboration), Phys.Rev. D 83 , 052004 (2011).
- [12] P.K. Khandai, P. Sett, P. Shukla and V. Singh, Int. J. Mod. Phys. 28 , 1350066 (2013).
- [13] M. Rybczynski and Z. Wlodarczyk, Eur. Phys. J. C 74 , 2785 (2014).
- [14] M.D. Azmi and J. Cleymans, J. Phys. G 41 , 065001 (2014).
- [15] J. Cleymans et al. , Phys. Lett. B 723 , 351 (2013).
- [16] C.Y. Wong and G. Wilk, Phys. Rev. D 87 , 114007 (2013).
- [17] R. Hagedorn, Riv. del Nuovo Cim. 6N 10 , 1 (1984).
- [18] R. Blankenbecler and S. J. Brodsky, Phys. Rev. D 10 , 2973 (1974).
- [19] N. Abgrall et al. Eur. Phys. C 74 , 2794 (2014).
- [20] A. Adare et al. (PHENIX Collaboration), Phys. Rev. C 83 , 064903 (2011).
- [21] S. Chatrchyan et al. (CMS Collaboration), Eur. Phys. J. C 72 , 2164 (2012).
- [22] R. Hagedorn, Nuovo Cim. Suppl. 3 , 147 (1965).
- [23] D.B. Walton and J. Rafelski, Phys. Rev. Lett. 84 , 31 (2000).
- [24] G. Wilk and Z. Wlodarczyk, Phys. Rev. Lett. 84 , 2770 (2000).
- [25] B.I. Abelev et al. (STAR Collaboration), Phys. Rev. C 75 , 64901 (2007).
- [26] J. Cleymans and D. Worku, Eur. Phys. J. A 48 , 160 (2012).
- [27] R. Blankenbecler, S.J. Brodsky and J. Gunion, Phys. Rev. D 12 , 3469 (1975) .
- [28] S.J. Brodsky, H.J. Pirner and J. Raufeisen, Phys. Lett. B 637 , 58 (2006).
- [29] T.S. Biro, K. Urmossy and G.G. Barnafoldi, J. Phys. G 35 , 044012 (2008).
- [30] K. Aamodt et al. (ALICE Collaboration), Eur. Phys. J. C 71 , 1655 (2011).
- [31] J. Adams et al. (STAR Collaboration), Phys. Lett. B 637 , 161 (2006).
| null |
[
"P. Sett",
"P. Shukla"
] |
2014-08-05T17:01:15+00:00
|
2014-09-03T04:29:27+00:00
|
[
"hep-ph",
"hep-ex",
"nucl-ex",
"nucl-th"
] |
Pion $p_T$ spectra in $p+p$ collisions as a function of $\sqrt{s}$ and event multiplicity
|
We study the charged pion transverse momentum ($p_T$) spectra in $p$+$p$
collisions as a function of collision energy $\sqrt{s}$ and event multiplicity
using Tsallis distribution. This study gives an insight of the pion production
process in $p+p$ collisions. The study covers pion spectra measured in $p+p$
collisions at SPS energies (6.27-17.27 GeV), RHIC energies (62.4 GeV and 200
GeV) and LHC energies (900 GeV, 2.76 TeV and 7 TeV). The Tsallis parameters
have been obtained and parameterized as a function of $\sqrt{s}$. The study
suggests that as we move to higher energy more and more hard processes
contribute to the spectra. We also study the charged pion spectra for different
event multiplicities in $p$+$p$ collisions for LHC energies using Tsallis
distribution. The variation of the Tsallis parameters as a function of event
multiplicity has been obtained and their behavior is found to be independent of
collision energy.
|
1408.1035v1
|
## Synchronization of Limit Cycle Oscillators by Telegraph Noise
Denis S. Goldobin
Department of Physics, University of Potsdam, Postfach 601553, D-14415 Potsdam, Germany Department of Theoretical Physics, Perm State University, 15 Bukireva str., 614990, Russia
Abstract. We study the influence of telegraph noise on synchrony of limit cycle oscillators. Adopting the phase description for these oscillators, we derive the explicit expression for the Lyapunov exponent. We show that either for weak noise or frequent switching the Lyapunov exponent is negative, and the phase model gives adequate analytical results. In some systems moderate noise can desynchronize oscillations, and we demonstrate this for the Van der Pol-Duffing system.
Keywords: Noise; Synchronization; Lyapunov Exponent
PACS: 05.40.-a; 02.50.Ey; 05.45.Xt
## INTRODUCTION
In autonomous systems exhibiting a stable periodic behavior (in other words, limit cycle oscillators), deviations along the trajectory asymptotically in time neither decay nor grow, i.e. they are neutral. This neutrality is due to the time homogeneity, and may disappear as soon as this homogeneity is broken by means of a time-dependent external forcing. The phenomenon of synchronization of oscillators by periodic signal is well known and quite understood, here the oscillations follow the forcing (e.g., they attain the same frequency). When the role of this time-dependent forcing is played by a stochastic noise, the situation becomes less evident.
The first effect of noise on periodic oscillations is the phase diffusion: the oscillations are no more periodic but posses finite correlations [1, 2]. However, a nonlinear system may somehow follow the noisy force. Although it is not so evident how the synchronization between the system response and noise can be detected for one system, this synchronization can be easily detected by looking on whether the responses of a few identical systems driven by the common noise signal are identical or not. With such an approach, synchrony (asynchrony) of driven systems was early treated in the works [3, 4, 5, 6, 7, 8]. The mathematical criterion for synchronization is the negative leading Lyapunov exponent (LE; it measures the average exponential growth rate of infinitesimally small deviations from the trajectory) in the driven system.
In different fields the effect of synchronization of oscillators by common noise is known under different names. In neurophysiology the property of a single neuron to provide identical outputs for repeated noisy input is treated as "reliability" [9]. In the experiments with noise-driven Nd:YAG lasers [10] this synchronization was called "consistency". When driving signal is related to not stochastic but deterministic chaos, one considers generalized synchronization [11]. In the last case the driven system is often chosen to be chaotic. In fact, in the above mentioned examples there is no limit cycle oscillators at the noiseless limit: for experiments described in [9, 10] the noiseless system is stable, i.e. LE is negative, and for generalized synchronization in chaotic systems, LE is positive. Evidently, in this cases, LE preserves its sign at sufficiently weak noise.
A more intriguing situation takes place when LE in noiseless system is zero (limit cycle oscillators). Analytical and numerical treatments for different types of noise show weak noise to play an ordering role: LE shifts to negative values, and oscillators become synchronized [3, 4, 5, 12, 13]. In the work [14] the nonideal situations are considered: slightly nonidentical oscillators driven by an identical noise signal, and identical oscillators driven by slightly nonidentical noise signals; and additionally, positive LE was reported for large noise in some smooth systems similarly to how it was in the works [3, 4].
Note that in [3, 4] LE was calculated for oscillators driven by a random sequence of pulses, in [5, 12, 13, 14] the white Gaussian noise was considered. A noise of other nature is the telegraph one. By a normalized telegraph noise we mean the signal having values ± 1 and switching instantaneously between these values time to time. The distribution of time intervals between consequent switchings is exponential with the average value τ . The case of telegraph noise may be interesting not only because it completely differs from the previous two, but also because it allows to "touch" the question of relations between periodic and stochastic driving, e.g. to compare results for telegraph noise with the average switching time τ and the stepwise periodic signal of the same amplitude and the period 2 τ . This is why we consider synchronization by telegraph noise.
## PHASE MODEL
A limit cycle oscillator with a small external force is known to be able to be well described within the phase approximation [15], where only dynamics of the system on the limit cycle of the noiseless system is considered . The system states on this limit 1 cycle can be parameterized by a single parameter, phase ϕ . With a stochastic force the equation for the phase reads
$$\dot { \varphi } = \omega + \varepsilon f ( \varphi ) \xi ( t )$$
where 2 π / ω is the period of the limit cycle in the noiseless system, ε is the amplitude of noise, f ( ϕ ) is the normalized sensitivity of the system to noise [ ( 2 π ) -1 ∫ 2 π 0 f 2 ( ϕ ) d ϕ = 1], and ξ is a normalized telegraph noise.
## Master equation
Studying statistical properties of the system under consideration, one can introduce two probability density functions W ± ( ϕ , t ) defining the probability to locate the system
1 Noteworthy, the phase approximation is valid not only for a small external force, but also for a moderate one if only the leading Lyapunov exponent of the limit cycle is negative and large enough.
in vicinity of ϕ with ξ = ± 1, correspondingly, at the moment t . Then the Master equations of the system read
$$\frac { \partial W _ { + } ( \varphi, t ) } { \partial t } + \frac { \partial } { \partial \varphi } \left [ ( \omega + \varepsilon f ( \varphi ) ) W _ { + } ( \varphi, t ) \right ] \ = \ \frac { 1 } { \tau } W _ { - } ( \varphi, t ) - \frac { 1 } { \tau } W _ { + } ( \varphi, t ), \quad ( 2 )$$
$$\frac { \stackrel { \nu } { \partial W _ { - } ( \varphi, t ) } } { \partial t } + \frac { \stackrel { \nu } { \partial } } { \partial \varphi } \left [ ( \omega - \varepsilon f ( \varphi ) ) W _ { - } ( \varphi, t ) \right ] \ = \ \frac { \stackrel { \nu } { 1 } } { \tau } W _ { + } ( \varphi, t ) - \frac { 1 } { \tau } W _ { - } ( \varphi, t ). \quad ( 3 )$$
In the terms of W ≡ W + + W -and V ≡ W + -W -the last system takes the form of
$$\dot { W } = - \omega W _ { \varphi } - \varepsilon \left ( f V \right ) _ { \varphi }, \quad \dot { V } = - \omega V _ { \varphi } - \varepsilon \left ( f W \right ) _ { \varphi } - \frac { 2 } { \tau } V.$$
For steady distributions the probability flux S is constant:
$$S = \omega W ( \varphi ) + \varepsilon f ( \varphi ) V ( \varphi ) ;$$
and system (4) with periodic boundary conditions has the solution
$$V ( \varphi ) = - \frac { \varepsilon \omega C } { \omega ^ { 2 } - \varepsilon ^ { 2 } f ^ { 2 } ( \varphi ) } \, \int _ { \varphi } ^ { \varphi + 2 \pi } d \psi f ^ { \prime } ( \psi ) \exp \left ( \frac { 2 } { \tau } \int _ { \varphi } \frac { d \, \theta } { \omega ^ { 2 } - \varepsilon ^ { 2 } f ^ { 2 } ( \theta ) } \right ), \quad \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \ \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \ \,. \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \ } \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \cdots \, \quad \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \. \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \quad \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdclots \, \cdclots \, \cdclots \, \cdclots \, \cdclots \, \cdclots \, \cdclclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard\, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdClipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \. \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclIPboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \,\cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipboard \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \cdclipread \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \delta \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dg \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \dag \, \$$
where C is defined by the normalization condition:
$$\begin{array} { r c l } \text{where} \colon \text{is even any one non-qual} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{out} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{In} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{ing} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{is} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{ in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{ini} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{\text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in}\text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \text{in} \\ \text{The probability flux reads \end{array}$$
The probability flux reads
$$S = \omega \left ( \exp \left ( \frac { 2 } { \tau } \int _ { 0 } ^ { 2 \pi } \frac { d \theta } { \omega ^ { 2 } - \epsilon ^ { 2 } f ^ { 2 } ( \theta ) } \right ) - 1 \right ) C.$$
## Lyapunov exponent
Studying stability of solutions of the stochastic equation (1), one has to consider behavior of a small perturbation α :
$$\dot { \alpha } = \epsilon f ^ { \prime } ( \varphi ) \alpha \xi ( t ).$$
The Lyapunov exponent (LE) measuring the average exponential growth rate of α can be obtained by averaging the corresponding velocity
$$\lambda \ = \ \langle \frac { d } { d t } \ln \alpha \rangle = \langle \varepsilon f ^ { \prime } ( \varphi ) \xi ( t ) \rangle = \varepsilon \underset { 0 } { \overset { 2 \pi } { \int } } f ^ { \prime } ( \varphi ) V ( \varphi ) d \varphi \\ \underset { \varepsilon - 1 } { \overset { 2 \pi } { \sim } } \, \varepsilon + \varepsilon \,.$$
$$= \ - \varepsilon ^ { 2 } \omega C \underset { 0 } { \overset { 2 \pi } { \int } } d \varphi \underset { \varphi } { \overset { \varphi + 2 \pi } { \int } } d \psi \frac { f ^ { \prime } ( \varphi ) f ^ { \prime } ( \psi ) } { \omega ^ { 2 } - \varepsilon ^ { 2 } f ^ { 2 } ( \varphi ) } \exp \left ( \frac { 2 } { \tau } \underset { \varphi } { \overset { \psi } { \int } } \frac { d \theta } { \omega ^ { 2 } - \varepsilon ^ { 2 } f ^ { 2 } ( \theta ) } \right ). \quad \ \ ( 7 ) \\ \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \colon \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \end{matrix}$$
Let us remind that LE determines the asymptotic behavior of small perturbations, and describes whether close states diverge or converge with time. This process is not necessarily monotonous, i.e., close trajectories can diverge at some time intervals while demonstrating asymptotic convergence, and vice versa.
When τ /lessmuch 1 or ε /lessmuch ω , the eq. (7) can be simplified:
$$\lambda _ { \text{app} } = - \frac { \varepsilon ^ { 2 } } { 2 \pi \omega } \left ( \exp \left ( \frac { 4 \pi } { \tau \omega ^ { 2 } } \right ) - 1 \right ) ^ { - 1 } \underset { 0 } { \overset { 2 \pi } { \int } } d \varphi \underset { 0 } { \overset { 2 \pi } { \int } } d \psi f ^ { \prime } ( \varphi ) f ^ { \prime } ( \psi + \varphi ) \exp \frac { 2 \psi } { \tau \omega ^ { 2 } }. \quad ( 8 )$$
The last expression is strictly negative. Indeed, in the Fourier space it reads
$$\lambda _ { \text{app} } = - \omega \tau \varepsilon ^ { 2 } \sum _ { k = 1 } ^ { \infty } | C _ { k } | ^ { 2 } \frac { k ^ { 2 } } { 1 + ( k \tau \omega ^ { 2 } / 2 ) ^ { 2 } },$$
$$\text{where} \, C _ { k } = ( 2 \pi ) ^ { - 1 } \overset { 2 \pi } { \underset { 0 } { \int } } f ( \varphi ) \, e ^ { - i k \varphi } d \varphi.$$
## COMPARISON TO NUMERICAL SIMULATION
We found that either for weak noise or frequent switching LE is negative regardless to the properties of the smooth function f ( ϕ ) (as for weak white Gaussian noise in similar systems [12, 13]). In the works [14], moderate white Gaussian noise was shown to be able to lead to instability even in smooth systems. In the light of above facts, it is interesting (i) what is the region of validity of our analytical theory , (ii) whether there is some footprints of the synchronization by periodic forcing in the stochastic synchronization , and (iii) whether telegraph noise can desynchronize oscillators .
For the two first purpose we performed simulation of a modified Van der Pol oscillator:
where ξ ( ) t is either a telegraph noise with the average switching time τ or a periodic stepwise signal with the period 2 τ , i.e. the constant switching time τ . The forcing-free modified Van der Pol oscillator has the round stable limit cycle of the unit radius for all µ > 0. Nevertheless, the phase equation (1) with ω = 1 and the simple function
$$\cdot \cdot \cdot \cdot \\ \ddot { x } - \mu ( 1 - x ^ { 2 } - \dot { x } ^ { 2 } ) \dot { x } + x = \varepsilon \sqrt { 2 } \xi ( t ), \\ \intertext { a k e n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n e t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t e t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n d a t i o n D t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t tt t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t ttt t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t T t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t a t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t$$
FIGURE 1. Samples of dependencies λ ε ( , τ ) for the modified Van der Pol oscillator (9) at µ = 0 1. The solid lines present the analytical results of phase description, the triangles plot results of the . approximation (8), the circles correspond to numerical simulation of the noisy modified Van der Pol oscillator, and the dashed line corresponds to numerical simulation of the periodically driven one.
FIGURE 2. Samples of dependencies λ ε ( , τ ) for the Van der Pol-Duffing oscillator (10) at µ = 0 1. . The values of the parameters b and τ are indicated above the plots.
f ( ϕ ) = √ 2cos ϕ may be correctly adopted only if the phase speed is near-constant all over the limit cycle, which is possible at small µ only.
In Fig. 1 one can see that our analytical theory is fortunately in good agreement with the results of analytical simulation not only for weak noise; and the dependence λ ε ( , τ ) for the stochastic driving has no footprints of the one for the periodic driving.
While the dynamical system (9) does not exhibit positive LEs at any noise intensity and any µ , they can be observed for a Van der Pol-Duffing model : 2
$$\cdot & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & && & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &
& & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & \\ & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & - & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & ; & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & } & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &. & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & ( & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & + & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & \ & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & _ & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & = & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &, & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & ^ & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & b & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & | & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &$$
where "Duffing parameter" b describes nonisochronicity of oscillations. In Fig. 2 one can see that at large enough b positive LE appears in a certain range of parameters.
2 A similar situation occurs for white Gaussian noise [14].
## CONCLUSIONS
Having considered the phenomenon of synchronization of limit cycle oscillators by common telegraph noise, we can summarize:
- -Either for weak noise or frequent switching the Lyapunov exponent is negative;
- - For some systems, the phase model gives quite adequate results even for moderate noise levels and values of the average switching time;
- -The dependence λ ε ( , τ ) for stochastic driving does not look to have any footprints of the one for periodic driving;
- -In some systems, moderate telegraph noise can desynchronize oscillations.
Here we do not present results for the nonideal situations (like in [14]): slightly nonidentical oscillators driven by an identical noise signal, and identical oscillators driven by slightly nonidentical noise signals. The reason is that for weak noise these results appear to be the same as in [14] but with λ app given by Eq. (8) instead of λ .
## ACKNOWLEDGMENTS
The work has been supported by DFG (SFB 555).
## REFERENCES
- 1. R. L. Stratonovich, Topics in the Theory of Random Noise , Gordon and Breach, New York, 1963.
- 2. A. N. Malakhov, Fluctuations in Self-Oscillatory Systems , Nauka, Moscow, 1968, (In Russian).
- 3. A. S. Pikovsky, Radiophys. Quantum Electron. 27 (5), 576 (1984).
- 4. A. S. Pikovsky, "Synchronization and stochastization of nonlinear oscillations by external noise", in Nonlinear and Turbulent Processes in Physics , edited by R. Z. Sagdeev, Harwood Acad. Publ., Singapore, 1984, pp. 1601-1604.
- 5. A. S. Pikovsky, M. Rosenblum, and J. Kurths, Synchronization-A Unified Approach to Nonlinear Science , Cambridge University Press, Cambridge, UK, 2001.
- 6. L. Yu, E. Ott, and Q. Chen, Phys. Rev. Lett. 65 , 2935 (1990).
- 7. L. Baroni, R. Livi, and A. Torcini, Phys. Rev. E 63 , 036226 (2001).
- 8. P. Khoury, M. A. Lieberman, and A. J. Lichtenberg, Phys. Rev. E 54 , 3377 (1996).
- 9. Z. F. Mainen and T. J. Sejnowski, Science 268 , 1503 (1995).
- 10. A. Uchida, R. McAllister, and R. Roy, Phys. Rev. Lett. 93 , 244102 (2004).
- 11. H. D. I. Abarbanel, N. F. Rulkov, and M. M. Suschik, Phys. Rev. E 53 , 4528 (1996).
- 12. J. Teramae and D. Tanaka, Phys. Rev. Lett. 93 , 204103 (2004)
- 13. D. S. Goldobin and A. S. Pikovsky, Phisica A 351 , 126 (2005).
- 14. D. S. Goldobin and A. Pikovsky, Phys. Rev. E 71 , 045201(R) (2005).
- 15. Y. Kuramoto, Chemical Oscillations, Waves and Turbulence , Dover, New York, 2003.
|
10.1063/1.2138643
|
[
"Denis S. Goldobin"
] |
2014-08-05T17:07:51+00:00
|
2014-08-05T17:07:51+00:00
|
[
"cond-mat.stat-mech",
"nlin.CD"
] |
Synchronization of Limit Cycle Oscillators by Telegraph Noise
|
We study the influence of telegraph noise on synchrony of limit cycle
oscillators. Adopting the phase description for these oscillators, we derive
the explicit expression for the Lyapunov exponent. We show that either for weak
noise or frequent switching the Lyapunov exponent is negative, and the phase
model gives adequate analytical results. In some systems moderate noise can
desynchronize oscillations, and we demonstrate this for the Van der
Pol--Duffing system.
|
1408.1036v2
|
## Graph-induced operators: Hamiltonian cycle enumeration via fermion-zeon convolution
G. Stacey Staples ∗
## Abstract
Operators are induced on fermion and zeon algebras by the action of adjacency matrices and combinatorial Laplacians on the vector spaces spanned by the graph's vertices. Properties of the algebras automatically give information about the graph's spanning trees and vertex coverings by cycles & matchings. Combining the properties of operators induced on fermions and zeons gives a fermion-zeon convolution that recovers the number of Hamiltonian cycles in an arbitrary graph. The mathematics underlying the graph-theoretic interpretation of these operators is provided by Kirchhoff's theorem and by the seminal works of Goulden and Jackson and Liu, who established formulas for enumeration of Hamiltonian cycles and paths using determinants and permanents of adjacency matrices.
AMS Subj. Classifications: 05C30, 05E10, 15A66, 47C05, 81R05 Keywords: cycles, trees, fermions, zeons, Clifford algebras, graph enumeration
## 1 Introduction & Notational Preliminaries
The interface between operator theory and graph theory has been fertile ground for research in recent years. In the context of quantum probability, the interrelationships have been explored in works by Accardi, Obata, Hashimoto, and others [1, 7].
Fermion algebras are isomorphic to Clifford algebras of appropriate signature. In the Clifford algebra context, properties of spinors have been applied to the study of maximal cliques [2] and graph enumeration problems [6]. Although Clifford algebras provide a comprehensive framework for everything appearing here, introducing the more general Clifford algebra formalism is unnecessary for matters at hand.
Zeon algebras can be thought of as commutative analogues of fermion algebras. Zeon algebras were defined and applied to graph theory in [16], although the name 'zeon' first appeared in work by Feinsilver [3]. Weighting the vertices
∗ Department of Mathematics and Statistics, Southern Illinois University Edwardsville, Edwardsville, IL 62026-1653,USA, email: [email protected]
of a graph with zeon generators allows one to construct a nilpotent adjacency matrix , A , whose entries are generators of the algebra. The matrix is very convenient for performing symbolic computations and allows enumeration of cycles by considering traces of matrix powers [13, 16].
Zeons underlie a significant portion of Liu's work on enumeration of Hamiltonian cycles [9], and their combinatorial properties have been developed in a number of the current author's joint works (e.g., [11, 12, 13]). Combinatorial properties of zeons are also useful for defining partition-dependent stochastic integrals [14, 15]. In other recent work, Neto and dos Anjos [10] establish a number of combinatorial identities using Grassmann-Berezin type integration on zeon algebras.
The main results of the current paper are found in Section 3. After a review of essential graph theory and definitions of graph-induced operators, operators are induced by the combinatorial Laplacian and adjacency matrix on the fermion and zeon algebras, respectively.
In Section 3.2, the combinatorial Laplacian of a graph G on n vertices is used to induce an operator on the n -particle fermion algebra. In Proposition 3.3, the normalized trace of the level-( n -1) induced operator is shown to be the number of spanning trees of G .
In Section 3.3, the adjacency matrix of G is used to induce an operator on the n -particle zeon algebra. Proposition 3.4 reveals the trace of the levelk induced operator to be sums of cycle-matching covers of the vertices of k -vertex subgraphs of G . As a corollary, a sort of cycle cover-perfect matching convolution is found as the leveln trace.
Combining the properties of operators induced on fermions and zeons gives a fermion-zeon convolution that recovers the number of Hamiltonian cycles in an arbitrary graph. As shown in Section 3.4, this convolution takes the form of an operator trace.
The combinatorics underlying the graph-theoretic interpretation of these operators is provided by Kirchhoff's theorem and by the seminal works of Goulden and Jackson [5] and Liu [9], establishing formulas for enumeration of Hamiltonian cycles and paths using determinants and permanents of adjacency matrices. The interpretation of Liu's formula as fermion-zeon convolution is original with the current author.
## 2 Fermions & Zeons
Denote by F n the associative algebra over R generated by the collection { f 1 , . . . , f n , f 1 † , . . . , f n † } satisfying the canonical anticommutation relations (CAR):
$$\{ \mathfrak { f } _ { i }, \mathfrak { f } _ { j } \} _ { + } = \underset { \ast \, \ast \, 2 } { \{ \mathfrak { f } _ { i } ^ { \dagger }, \mathfrak { f } _ { j } ^ { \dagger } \} } _ { + } \ = \ \underset { \ast \, \ast \, 2 } { 0 }, \quad \underset { \ast \, \ast \, \ast } { \infty }$$
$$\mathfrak { f } _ { j } ^ { \, 2 } = \mathfrak { f } _ { j } ^ { \, \dagger \, ^ { 2 } } \ = \ 0,$$
$$\stackrel { \cdot } { \{ f _ { i }, f _ { j } \stackrel { \cdot } { \} } _ { + } } \ = \ \delta _ { i j }.$$
The algebra F n is called the n -particle fermion algebra . The generators f i and f i † are referred to as the i th annihilation operator and i th creation operator , respectively.
Zeon algebras can be regarded as commutative subalgebras of fermion algebras. It is not difficult to see that (2.1) implies commutativity of disjoint pairs of fermions; i.e., letting ζ j = f 2 j -1 2 f j ∈ F 2 n for j = 1 , . . . , n ensures that the zeons satisfy the zeon canonical commutation relations (CCR):
$$[ \zeta _ { i }, \zeta _ { j } ] \ & = \ 0, \\ \zeta _ { j } ^ { 2 } \ & = \ 0.$$
Here, ζ j is defined as a pair of fermion annihilators, although creators could have been used as easily.
Observe that the collection { ζ 1 , . . . , ζ n } with multiplication satisfying ζ ζ i j = 0 ⇔ i = j and satisfying the zeon CCR generates a semigroup of order 2 n -1. Appending the unit scalar ζ ∅ = 1, one obtains a semigroup Z n of order 2 n . Adopting multi-index notation, one writes ζ I = ∏ /lscript ∈ I ζ /lscript for any multi-index I ⊆ [ n ].
a number of works in recent years[3, 4, 14, 16].
The n -particle zeon algebra is defined as the semigroup algebra Z n := R Z n . 1 Employing multi-index notation, arbitrary elements of Z n are written in the form u = ∑ I ⊆ [ n ] u I ζ I . Combinatorial properties of zeons have been developed in
## 3 Graph-Induced Operators
Assume G = ( V, E ) is a graph on n vertices. For i = 1 , . . . , n associate vertex v i ∈ V with the i th fermion creation/annihilation pair, γ i = 1 √ 2 ( f i + f i † ) ∈ F n .
Note that the fermion CAR imply { γ , γ i j } + = 0 and γ i 2 = 1 for i = 1 , . . . , n . Hence, there is no cause for concern in defining multi-index notation by writing the ordered product
$$\gamma _ { I } \coloneqq \prod _ { j \in I } \gamma _ { j } = \frac { 1 } { 2 ^ { | I | / 2 } } \prod _ { j \in I } ( \mathfrak { f } _ { j } + \mathfrak { f } _ { j } ^ { \dagger } ).$$
Remark 3.1 . The collection { γ i : 1 ≤ i ≤ n } generates the 2 n -dimensional Euclidean Clifford algebra, commonly denoted C /lscript n . This algebra is isomorphic to fermion toy Fock space .
For convenience, V is regarded as both the vertex set of G and as the vector space generated by the vertices of G . Let L ( V ) denote the space of linear operators on V . Beginning with an operator X ∈ L ( V ), the corresponding
1 Because the zeon algebra is a subalgebra of a Clifford algebra of appropriate signature, it is often denoted by C /lscript n nil .
operator Ψ ∈ L ( F n ) induced by X is defined naturally by linear extension of the following action on basis blades:
$$\Psi ( \gamma _ { I } ) \coloneqq \prod _ { j \in I } X ( \gamma _ { j } ).$$
Operators are induced in similar manner on Z n . For convenience, when Ψ is an operator on the algebra A induced by the operator X acting on the vector space of generators of A , it will be convenient to write X /dblarrowheadright Ψ.
Because the algebras F n and Z n have natural grade decompositions, an operator X on V similarly induces operators Ψ ( /lscript ) and Ξ ( /lscript ) , respectively, on the grade/lscript subspaces of the algebras. More concisely,
$$\Psi ^ { ( \ell ) } ( \gamma _ { I } ) = \begin{cases} \prod _ { j \in I } X ( \gamma _ { j } ) & | I | = \ell, \\ 0 & \text{otherwise}. \end{cases}$$
Using Dirac notation, one writes Ψ ( /lscript ) ( γ I ) = 〈 γ I | Ψ ( /lscript ) . The use of Dirac notation in matrix representations is made clear by the following: Given n × n matrix A , the i th row and j th column of A are given by 〈 | i A j | 〉 . In this context, the matrix A acts on row vectors by right multiplication. While this may be less common in linear algebra, it makes sense when dealing with adjacency matrices, combinatorial Laplacians, and transition matrices associated with Markov chains.
↦
Once again, let G = ( V, E ) be a graph, where V is regarded both as the vertex set of G and also the vector space spanned by vertices of G . Let A denote an operator on V , suppose A /dblarrowheadright Φ ∈ L ( F n ), and suppose A /dblarrowheadright Ξ ∈ L ( Z n ). Under the vertex-fermion and vertex-zeon associations v i → γ i and v i → ζ i , respectively, there should be no ambiguity in adopting the conventions 〈 v I | Φ | v J 〉 = 〈 γ I | Φ | γ J 〉 and 〈 v I | Ξ | v J 〉 = 〈 ζ I | Ξ | ζ J 〉 .
↦
## 3.1 Graph preliminaries
A graph G = ( V, E ) is a collection of vertices V and a set E of unordered pairs 2 of vertices called edges . Two vertices v , v i j ∈ V are adjacent if there exists an edge e = { v , v i j } ∈ E . A graph is finite if V and E are finite sets. A loop in a graph is an edge of the form { v, v } . A graph is said to be simple if it contains no loops and no unordered pair of vertices appears more than once in E .
A k -walk { v , . . . , v 0 k } in a graph G is a sequence of vertices in G with initial vertex v 0 and terminal vertex v k such that there exists an edge { v , v j j +1 } ∈ E for each 0 ≤ j ≤ k -1. A k -walk contains k edges. A self-avoiding walk is a walk in which no vertex appears more than once. A closed k-walk is a k -walk whose initial vertex is also its terminal vertex. A k -cycle is a self-avoiding closed k -walk with the exception v 0 = v k . A Hamiltonian cycle is an n -cycle in a graph on n vertices; i.e., it contains V.
2 When the edges are ordered pairs, the graph is said to be a directed graph , or digraph .
A tree is a connected graph that contains no cycles. A spanning tree in G is a subgraph that contains all vertices of G and is a tree.
Given a graph G = ( V, E ), a matching of G is a subset E 1 ⊂ E of the edges of G having the property that no pair of edges in E 1 shares a common vertex. The largest possible matching on a graph with n vertices consists of n/ 2 edges, and such a matching is called a perfect matching .
## 3.2 Spanning trees via fermion trace
When G = ( V, E ) is a simple graph on n vertices, the combinatorial Laplacian of G is the n × n matrix L = ( /lscript ij ) defined by
$$\ell _ { i j } = \begin{cases} \deg ( v _ { i } ) & \text{if $i=j$,} \\ - 1 & \text{if $\{v_{i},v_{j}\} \in E$,} \\ 0 & \text{otherwise.} \end{cases} \\ \text{is the diagonal matrix of vertex degrees and} \\ L = D - A.$$
Letting λ 0 ≤ λ 1 ≤ · · · ≤ λ n -1 denote the eigenvalues of L , known properties of L include (but are not limited to) the following:
Equivalently, if D is the diagonal matrix of vertex degrees and A is the adjacency matrix of G , then L = D -A .
- 1. The minimum eigenvalue λ 0 is always zero. 3
- 2. The number of times zero appears as an eigenvalue is equal to the number of connected components of G .
- 3. When G has multiple connected components, L is block-diagonal.
Let Tr( ) · denote normalized operator trace. In particular, if A acts on a finite-dimensional vector space V , then
$$\text{Tr} ( A ) \coloneqq \frac { 1 } { \dim ( V ) } \sum _ { i = 1 } ^ { \dim ( V ) } \langle \mathbf v _ { i } | A | \mathbf v _ { i } \rangle \\ \text{normal basic } \mathbf v _ { i } \cdot 1 < i < n \mathbf \lambda \ \mathbf f \ V }$$
for any orthonormal basis { v i : 1 ≤ i ≤ n } of V .
The following well known result of Kirchhoff [8] gives meaning to the normalized trace of operators on F V induced by the combinatorial Laplacian. It is recalled here without proof.
Theorem 3.2 (Kirchhoff) . For a given connected graph G with n labeled vertices, let λ , . . . , λ 1 n -1 be the non-zero eigenvalues of its Laplacian matrix. Then the number of spanning trees of G is
$$t _ { G } = \frac { 1 } { n } \lambda _ { 1 } \lambda _ { 2 } \cdots \lambda _ { n - 1 }.$$
Equivalently the number of spanning trees is equal to any cofactor of the Laplacian matrix of G .
3 Let v 0 = (1 , . . . , 1) and observe that L ( v 0 ) = 0 .
Viewing the Laplacian as an operator on the vector space V spanned by vertices of G , the combinatorial properties of fermions now allow numbers of spanning trees to be recovered from the trace of the induced operator.
Proposition 3.3. Let L denote the combinatorial Laplacian associated with a finite graph, G = ( V, E ) , and let Φ be the operator on the fermion algebra F n induced by L ; i.e., L /dblarrowheadright Φ . Then, the normalized trace of the induced map at level n -1 satisfies the following:
$$\text{Tr} \left ( \Phi ^ { ( n - 1 ) } \right ) = \ddagger \{ \text{spanning trees of } G \}. \\ \text{for any $k\times n$ matrix } B = ( h _ { \cdot } ) \text{ associate the ith row}$$
denote the coefficient submatrix of B whose columns are indexed by elements of I . The fermion CAR guarantee that the coefficient of γ I in the expansion of
Proof. First, for any k × n matrix B = ( b ij ), associate the i th row of B with the fermion vector b i := n ∑ j =1 b ij √ 2 ( f j + f j † ). For multi-index I of cardinality k , let B I
$$& \text{or } I. \, \text{me in room } \mathcal { A } \text{ guarantees any} \, \text{and} \\ & \text{$i = \prod_{\ell = 1}^{k} b_{\ell}$ is then $\langle_{i},\gamma_{i}\rangle = \det(B_{I})$.} \\ & \text{By definition of } L \, \twoheadrightarrow \, \Phi, \, \text{one immedi}$$
By definition of L /dblarrowheadright Φ, one immediately finds 〈 v I | Φ ( n -1) | v I 〉 = det( L I ), which is a cofactor of the combinatorial Laplacian of G . By Kirchhoff's theorem, the number of spanning trees of a connected graph G on n vertices is equal to any cofactor of the Laplacian of G . Summing over the n diagonal elements of the level n -1 induced operator then gives the stated result.
## 3.3 Cycle-matching covers via zeon trace
Let A be the adjacency matrix of a graph with vertex set V of cardinality n . Regarding A as a linear operator on span( V ), an operator Ξ is induced on Z n by multiplication. Graph-theoretic properties of Ξ are immediately seen.
Proposition 3.4. Let A denote the adjacency matrix of a simple graph with vertex set V , viewed as a linear transformation on the vector space generated by V . Let Ξ ( k ) denote the corresponding operator induced on the gradek subspace of the zeon algebra Z V . For fixed subset I ⊆ V , let X I denote the number of disjoint cycle covers of the subgraph induced by I . Similarly, let M J denote the number of perfect matchings on the subgraph induced by J ⊆ V (nonzero only for J of even cardinality). Then,
$$\text{tr} ( \Xi ^ { ( k ) } ) = \sum _ { \substack { I \subset V \\ | I | = k } } \sum _ { J \subseteq I } X _ { I \vee J } M _ { J }.$$
Proof. Given any k × n matrix B = ( b ij ), associate the i th row of B with the zeon vector b i := n ∑ j =1 b ij ζ j . For multi-index I of cardinality k , let B I denote
$$& \text{the coefficient submatrix of $B$ whose columns are indexed by elements of $I$} \text{. } The \\ & \text{coefficient of $\zeta_{I}$ in the expansion of $j = \prod_{\ell=1}^{k}b_{\ell}$ is then $\langlej,\zeta_{I}\rangle = \text{per}(B_{I})$.} \\ & \text{For $n\times n$ adjacency matrix $A$, one immediately finds $\langlev_{I}|\Xi^{(k)}|v_{I}\rangle = \text{per}(A_{I})$,} \\ & \text{$.........................................................................................................................................................................................................$$
For n × n adjacency matrix A , one immediately finds 〈 v I | Ξ ( k ) | v I 〉 = per( A I ), where A I is the k × k submatrix of A whose rows and columns are indexed by elements of I . Writing A I = (Ξ ij ) and applying the definition of the matrix permanent,
$$\text{per} ( A _ { I } ) = \sum _ { \sigma \in S _ { k } } \prod _ { \ell = 1 } ^ { k } \Xi _ { \ell \sigma ( \ell ) }. & & ( 3. 3 ) \\ \text{elementary result of $\sigma$run theory. every termination $\sigma \in S$.}$$
Recall that by an elementary result of group theory, every permutation σ ∈ S k has a unique (up to order) factorization as a product of disjoint cycles. Further recall that A I is the adjacency matrix of the subgraph induced by vertex set v I . Observing that disjoint 2-cycles (i.e. transpositions) represent edges with no shared vertices (i.e. matchings of subgraphs), it follows that
$$\text{per} ( A _ { I } ) = \sum _ { J \subseteq I } X _ { I \vee J } M _ { J }, \quad \text{ (3.4)} \\ \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \colon$$
where X I \ J and M J are defined as in the statement of the theorem. Summing over all multi-indices of size k then gives the trace of the induced operator.
The following corollary is immediate.
Corollary 3.5 (Cycle-matching convolution) . Let A denote the adjacency matrix of a simple graph with vertex set V , and suppose A /dblarrowheadright Ξ ∈ L ( Z n ) . Let X I and M I ′ be defined as in the statement of Proposition 3.4, where I ′ is the complement of I in [ n ] . Then,
$$\langle \mathbf v _ { [ n ] } | \Xi | \mathbf v _ { [ n ] } \rangle = \text{tr} ( \Xi ^ { ( n ) } ) = \sum _ { I \subseteq V } X _ { I } M _ { I ^ { \prime } }.$$
## 3.4 Hamiltonian cycles via fermion-zeon convolution
By combining properties of induced operators on fermions and zeons, one is able to count the Hamiltonian cycles in an arbitrary graph. The following result of Goulden and Jackson [5] lies at the heart of the main result. The statement of the theorem has been adapted to the notation developed herein.
Theorem 3.6 (Goulden-Jackson) . Let H c be the number of directed Hamiltonian circuits in a digraph on n vertices with adjacency matrix A . Then
$$H _ { c } = \sum _ { I } ( - 1 ) ^ { | I | } \det ( A _ { I } ) \text{per} ( A _ { I ^ { \prime } } ) \\ \dots \cdots \cdots \cdots \cdots \cdots \cdots \cdots$$
where the sum is over all I ⊆ [ n ] \ { c } for any c ∈ [ n ] , and det( A ∅ ) := 1 .
Liu [9] generalized Goulden and Jackson's result to obtain a formula somewhat better suited for consideration as fermion-zeon convolution. In particular, the requirement of summing over subsets omitting one generator can be avoided. In the following theorem, the graph is assumed to be undirected, and all parameters of Liu's original formulation are assumed to be zero. Division by 2 is seen as correction for cycles appearing in two orientations.
Theorem 3.7 (Liu) . Let G be a finite graph with adjacency matrix A , and let H c denote the number of Hamiltonian cycles in G . Then,
$$H _ { c } = \frac { 1 } { 2 n } \sum _ { I \subseteq [ n ] } ( - 1 ) ^ { n - | I | } | I | \text{per} ( A _ { I } ) \text{det} ( A _ { I ^ { \prime } } ). \\. \quad \. \quad \. \quad \. \quad \.$$
To make the concept of fermion-zeon convolution rigorous, let ϕ ∈ L ( F n ), let ξ ∈ L ( Z n ), and define combinatorial integrals on L ( F n ) ⊗L ( Z n ) by
$$\int ( \varphi \ast \xi ) ( \mathbf v _ { I } ) \, d \gamma _ { I } \, d \zeta _ { I ^ { \prime } } \ \coloneqq & = \ \int ( \varphi ( \gamma _ { I } ) \otimes \xi ( \zeta _ { I ^ { \prime } } ) ) \, d \gamma _ { I } \, d \zeta _ { I ^ { \prime } } \\ = \ \langle \varphi ( \gamma _ { I } ) \otimes \xi ( \zeta _ { I ^ { \prime } } ), \gamma _ { I } \otimes \zeta _ { I ^ { \prime } } \rangle, \\ \dots \ \int \nolimits _ { 1 } \, \left \langle \, r \, \colon = \, \alpha \, \right \rangle \, \dots \, \dots \, \alpha \, \left \rangle \, \left \rangle \, \left \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle \, \right \rangle.$$
where I ′ = [ n ] \ I is the complement of I in the n -set. In other words, the value of the integral is the coefficient of γ I ⊗ ζ I ′ in the canonical expansion of ( ϕ γ ( I ) ⊗ ξ ζ ( I ′ )) ∈ F n ⊗ Z n . The fermion-zeon convolution of ϕ and ξ is then defined by
$$\Phi _ { \mathfrak { S } 3 } ^ { \dots } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi ^ { \dots } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \cdot } \Phi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi ^ { \dots } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi ^ { \ dots } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Ps i _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathsf { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \Psi _ { \mathfrak { S } } \\ \begin{matrix} \Phi _ { \mathfrak { S } 3 } ^ { \dots } \Phi _ { \mathfrak { S } 3 } ^ { \dots } \Phi _ { \mathfrak { S } 3 } ^ { \dots } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S S } } \Phi _ { \mathfrak { S } } \Ph i _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathsf { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } & \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \math frak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathr { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \\ \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _{ \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \mathfrak { S } } \Phi _ { \$$
To see how Liu's result naturally appears as fermion-zeon convolution, one first defines the star dual of Ξ ∈ L ( Z n ) by
$$\langle \zeta _ { I } | \Xi ^ { * } | \zeta _ { J } \rangle \coloneqq \langle \zeta _ { I ^ { \prime } } | \Xi | \zeta _ { J ^ { \prime } } \rangle$$
for I, J ⊆ [ n ], where I ′ and J ′ denote the set complements of I and J , respectively, in the n -set. Second, for operators X Y , on spaces of equal finite dimension, denote the componentwise product of operators X Y , by X /circledot Y . In this way, the /circledot product of Ψ ∈ L C ( /lscript Q ( V )) and Ξ /star ∈ L ( Z n ) satisfies
$$\langle \mathbf v _ { I } | \Psi \odot \Xi ^ { * } | \mathbf v _ { J } \rangle = \langle \mathbf v _ { I } | \Psi | \mathbf v _ { J } \rangle \langle \mathbf v _ { I } | \Xi ^ { * } | \mathbf v _ { J } \rangle$$
for all I, J ⊆ [ n ].
Letting σ denote the diagonal operator on the power set 2 V defined by
$$\langle \mathbf v _ { I } | \sigma | \mathbf v _ { I } \rangle \coloneqq ( - 1 ) ^ { | I | } | I ^ { \prime } |,$$
the unifying theorem can now be stated.
Theorem 3.8 (FZ Convolution) . Let G be a simple graph on n vertices having adjacency matrix A . Suppose A /dblarrowheadright Φ ∈ L ( F n ) and A /dblarrowheadright Ξ ∈ L ( Z n ) . Then, the number H c of Hamiltonian cycles in G is given by the following:
$$H _ { c } = \frac { 1 } { 2 n } \text{tr} ( \sigma \Phi \odot \Xi ^ { * } ).$$
Proof. The proof of Theorem 3.8 is now an easy corollary of Theorem 3.7.
$$\begin{smallmatrix} 1 \\ - \end{smallmatrix}$$
$$\colon The proof of Theorem 3.8 is now an easy corollary of Theorem 3.7 \\ H _ { c } \ & = \ \frac { 1 } { 2 n } \sum _ { I \subseteq [ n ] } ( - 1 ) ^ { n - | I | } | I | \text{per} ( A _ { I } ) \text{det} ( A _ { I ^ { \prime } } ) \\ & = \ \frac { 1 } { 2 n } \sum _ { I \in 2 ^ { [ n ] } } ( - 1 ) ^ { n - | I | } | I | \langle \mathbf v _ { I ^ { \prime } } | \Phi | \mathbf v _ { I ^ { \prime } } \rangle \langle \mathbf v _ { I } | \Xi | \mathbf v _ { I } \rangle \\ & = \ \frac { 1 } { 2 n } \sum _ { J \in 2 ^ { [ n ] } } ( - 1 ) ^ { | J | } | J ^ { \prime } | \langle \mathbf v _ { J } | \Phi | \mathbf v _ { J } \rangle \langle \mathbf v _ { J ^ { \prime } } | \Xi ^ { * } | \mathbf v _ { J } \rangle \\ & = \ \frac { 1 } { 2 n } \sum _ { J \in 2 ^ { [ n ] } } ( - 1 ) ^ { | J | } | J ^ { \prime } | \langle \mathbf v _ { J } | \Phi | \mathbf v _ { J } \rangle \langle \mathbf v _ { J } | \Xi ^ { * } | \mathbf v _ { J } \rangle \\ & = \ \frac { 1 } { 2 n } \text{tr} ( \sigma \Phi \odot \Xi ^ { * } ).$$
The fermion-zeon convolution interpretation is made clear by observing the following:
$$\text{following} \colon & \quad \cdot \quad \cdot \quad \cdot \quad \cdot \\ H _ { c } & \ = \ \frac { 1 } { 2 n } \sum _ { J \in 2 ^ { [ n ] } } ( - 1 ) ^ { | J | } | J ^ { \prime } | \langle \mathbf v _ { J } | \Phi | \mathbf v _ { J } \rangle \langle \mathbf v _ { J } | \Xi ^ { * } | \mathbf v _ { J } \rangle \\ & \ = \ \frac { 1 } { 2 n } \sum _ { J \in 2 ^ { [ n ] } } \langle \sigma ( \mathbf v _ { J } ), \mathbf v _ { J } \rangle \int ( \Phi ( \mathbf v _ { J } ) \otimes \Xi ( \mathbf v _ { J ^ { \prime } } ) ) ( d \gamma _ { J } \otimes d \zeta _ { J ^ { \prime } } ) \\ & \ = \ \frac { 1 } { 2 n } \sum _ { J \in 2 ^ { [ n ] } } \int ( \sigma \Phi ( \mathbf v _ { J } ) \otimes \Xi ( \mathbf v _ { J ^ { \prime } } ) ) ( d \gamma _ { J } \otimes d \zeta _ { J ^ { \prime } } ) \\ & \ = \ \frac { 1 } { 2 n } \int _ { \mathfrak { R } \mathfrak { S } \mathfrak { S } } ( \sigma \Phi * \Xi ) \, d \gamma \, d \zeta. \\ & \quad \text{Conclusion}$$
## 4 Conclusion
The results above lend a deeper physical interpretation of the author's earlier results developed using nilpotent adjacency matrices [11, 16]. It was shown previously by the author that labeling a simple graph's vertices with zeon generators { ζ i : i = 1 , . . . , n } and constructing the associated nilpotent adjacency matrix A allows one to count Hamiltonian cycles by considering the trace of the n th power of A :
$$\text{tr} ( \mathfrak { A } ^ { n } ) = 2 n \, H _ { c } \, \zeta _ { [ n ] }.$$
$$\text{By Theorem 3.8, the coefficient of $\zeta_{[n]}$ in tr} ( \mathfrak { A } ^ { n } ) \text{ is now given by fermion-zeon} \\ \text{convolution as} & & \langle \text{tr} ( \mathfrak { A } ^ { n } ), \zeta _ { [ n ] } \rangle = \text{tr} ( \sigma \Phi \odot \Xi ^ { * } ). & & ( 4. 1 )$$
## Acknowledgment
The author thanks Philip Feinsilver for comments and discussions.
## References
- [1] L. Accardi, A. Ben Ghorbal, N. Obata, Monotone independence, comb graphs and Bose-Einstein condensation , Infin. Dimens. Anal. Quantum Probab. Relat. Top. 7 (2004), 419-435.
- [2] M. Budinich, P. Budinich, A spinorial formulation of the maximum clique problem of a graph, J. Math. Phys. , 47 (2006), 043502.
- [3] P. Feinsilver. Zeon algebra, Fock space, and Markov chains, Commun. Stoch. Anal. 2 (2008), 263-275.
- [4] P. Feinsilver, J. McSorley. Zeons, permanents, the Johnson scheme, and generalized derangements, International Journal of Combinatorics , Article ID 539030, 29 pages, (2011).
- dx.doi.org/10.1155/2011/539030
- [5] I.P. Goulden, D.M. Jackson. The enumeration of directed closed Euler trails and directed Hamiltonian circuits by Lagrangian methods, Europ. J. Combinatorics , 2 (1981), 131-135.
- [6] G. Harris, G.S. Staples. Spinorial formulations of graph problems, Advances in Applied Clifford Algebras , 22 (2012), 59-77. dx.doi.org/10.1007/s00006-011-0298-0 .
- [7] Y. Hashimoto, A. Hora, N. Obata. Central limit theorems for large graphs: Method of quantum decomposition , J. Math. Phys. 44 (2003), 71-88.
- [8] G. Kirchhoff. Uber die Aufl¨sung der Gleichungen, auf welche man bei der ¨ o untersuchung der linearen verteilung galvanischer Str¨me gef¨hrt wird, o u Ann. Phys. Chem. , 72 (1847), 497-508.
- [9] C.J. Liu. Enumeration of Hamiltonian cycles and paths in a graph, Proc. Amer. Math. Soc. , 111 (1991), 289-296.
- [10] A. F. Neto. P.H.R. dos Anjos, Zeon algebra and combinatorial identities, SIAM Review , 56 (2014), 353-370.
- [11] R. Schott, G.S. Staples, Nilpotent adjacency matrices, random graphs, and quantum random variables, J. Phys. A: Math. Theor. , 41 155205, (2008).
- [12] R. Schott, G.S. Staples. Partitions and Clifford algebras, Eur. J. Comb. 29 (2008) 1133-1138.
- [13] R. Schott, G.S. Staples. Complexity of counting cycles using zeons, Computers and Mathematics with Applications , 62 1828-1837, (2011). dx.doi.org/10.1016/j.camwa.2011.06.026
- [14] R. Schott, G.S. Staples. Zeons, lattices of partitions, and free probability, Comm. Stoch. Anal. , 4 (2010), 311-334.
- [15] G.S. Staples. Graph-theoretic approach to stochastic integrals with Clifford algebras, J. Theor. Prob. , 20 257-274, (2007).
- [16] G.S. Staples. A new adjacency matrix for finite graphs, Advances in Applied Clifford Algebras , 18 979-991, (2008).
|
10.1007/s10773-017-3381-z
|
[
"G. Stacey Staples"
] |
2014-08-05T17:15:11+00:00
|
2015-01-18T15:28:16+00:00
|
[
"math.CO",
"math.RA",
"math.RT",
"math.SP",
"05C30, 05E10, 15A66, 47C05, 81R05"
] |
Graph-induced operators: Hamiltonian cycle enumeration via fermion-zeon convolution
|
Operators are induced on fermion and zeon algebras by the action of adjacency
matrices and combinatorial Laplacians on the vector spaces spanned by the
graph's vertices. Properties of the algebras automatically give information
about the graph's spanning trees and vertex coverings by cycles \& matchings.
Combining the properties of operators induced on fermions and zeons gives a
fermion-zeon convolution that recovers the number of Hamiltonian cycles in an
arbitrary graph. The mathematics underlying the graph-theoretic interpretation
of these operators is provided by Kirchhoff's theorem and by the seminal works
of Goulden and Jackson and Liu, who established formulas for enumeration of
Hamiltonian cycles and paths using determinants and permanents of adjacency
matrices.
|
1408.1037v1
|
## Assessment of the dependence of ( δS/δV ) on the heat influx for a well-stirred two-phase system with interfacial boiling
Denis S. Goldobin 1, 2, 3 and Anastasiya V. Pimenova 1
1 Institute of Continuous Media Mechanics, UB RAS, Perm 614013, Russia 2 Department of Mathematics, University of Leicester, Leicester LE1 7RH, UK 3 Perm State National Research University, Perm 614990, Russia
For a well-stirred multiphase fluid systems the mean interface area per unit volume, ( δS/δV ), is a significant characteristic of the system state. In particular, it is important for the dynamics of systems of immiscible liquids experiencing interfacial boiling. We estimate the value of parameter ( δS/δV ) as a function of heat influx ˙ Q V to the system.
## I. INTRODUCTION
For a well-stirred multiphase fluid systems the mean interface area per unit volume, ( δS/δV ), is an important characteristic of the state. It becomes even more significant for the systems where this interface is active chemically or in some other way. The systems of immiscible liquids experiencing interfacial boiling [1-14] are an example of the systems where parameter ( δS/δV ) becomes especially important. The problem of calculation of ( δS/δV ) cannot be addressed rigourously and any direct numerical simulation, being extremely challenging and CPU-time consuming, will provide results pertaining to a quite specific system set-ups. Some generale assessments on ( δS/δV ) can be highly beneficial. In this paper we perform these assessments for the process of direct contact boiling in a system of two immiscible liquids.
At the direct contact interface, a vapour layer grows and produces bubbles which breakaway of the interface and rise. The presence of vapour bubbles change the fluid buoyancy and perform a 'stirring' of the system. This stirring enforces increase of the contact area S , while surface tension and gravitational segregation of two liquids counteract the increase of the contact area.
For a two-liquid system experiencing direct contact boiling, the quantity of our interest depends on parameters of liquids and characteristics of the evaporation process, which are controlled by the mean overheating and the bubble production rate [9]. In this work the volumes of both components are assumed to be commensurable, no phase can be considered as a medium hosting dilute inclusions of the other phase. The characteristic width of the neighborhood of the vapour layer, beyond which the neighborhood of another vapour layer lies, is
The relationship between H 1 and H 2 is
$$H _ { 1 } + H _ { 2 } & \sim \left ( \frac { \delta S } { \delta V } \right ) ^ { - 1 }. \\ \text{and } H _ { \lambda } \text{ is}$$
$$\frac { H _ { 1 } } { H _ { 2 } } = \frac { \phi _ { 1 } } { \phi _ { 2 } } = \frac { \phi _ { 1 } } { 1 - \phi _ { 1 } } \,, \\.. \quad c \,. \quad \dots \,. \quad \dots \,.$$
where φ j is the volumetric fraction of the j -th liquid in the system. It will be convenient to use
$$H _ { j } \sim \phi _ { j } \left ( \frac { \delta S } { \delta V } \right ) ^ { - 1 }.$$
The process of boiling of a mixture above the bulk boiling temperature of the more volatile liquid is well-addressed in the literature [3-9]. Hydrodynamic aspects of the process of boiling below the bulk boiling temperature [13, 14] has to be essentially similar at the macroscopic level; rising vapour bubbles drive the stirring of system, working against the gravitational stratification into two layers with a flat horizontal interface, the surface tension forces tending to minimize the interface area, and viscous dissipation of the flow kinetic energy. Specifically, the behaviour of parameter ( δS/δV ) depending on macroscopic characteristics of processes in the system should be the same as for systems with superheating of the more volatile component. In what follows, we perform an analytical assessment of the dependence of ( δS/δV ) on the evaporation rate (or heat influx) for a well-stirred system.
## II. ENERGY FLUX BALANCE IN A WELL-STIRRED SYSTEM
Let us attempt to derive the rough relationships between the macroscopic parameter ( δS/δV ) of the system state and the heat influx rate per unit volume ˙ Q V = δQ/ δV δt ( ) for a statistically stationary process of interfacial boiling.
The flow and consequent stirring in the system are enforced by the buoyancy of the vapour bubbles, while other mechanisms counteract the stirring of the system. These other mechanisms are gravitational stratification of two liquids, surface tension tending to minimise the interface area and viscous dissipation of the flow energy. Since the latent heat of phase transitions and heat of temperature inhomogeneities are enormously large compared to the realistic values of the kinetic energy of microscopic motion and gravitational potential energy [15], the latter can be neglected in consideration of the heat balance. Hence, all the heat inflow into the system can be considered to be spent for the vapour generation; ˙ Q V V -→ (Λ 1 n (0) 1 +Λ 2 n (0) 2 ) ˙ V v , where V is the system volume, ˙ V v is the volume of the vapour produced in the system per unit time, Λ j is the enthalpy of vaporization per one molecule of liquid j , and n (0) j is the saturated vapour pressure of liquid j . Thus,
$$\dot { V } _ { v } = \frac { \dot { Q } _ { V } \, V } { \Lambda _ { 1 } n _ { 1 } ^ { ( 0 ) } + \Lambda _ { 2 } n _ { 2 } ^ { ( 0 ) } } \,.$$
The potential energy of buoyancy of rising vapour bubbles ρ V gh/ l v 2 (where h is the linear size of the system, h ∼ V 1 3 / , ρ l is the average density of liquids, the vapour density is zero compared to the liquid density) is converted into the kinetic energy of liquid flow, the potential energy of a stirred state of the two-liquid system, the surface tension energy and dissipated by viscosity forces. In a statistically stationary state, the mechanical kinetic and potential energies do not change averagely over time and all the energy influx is to be dissipated by viscosity;
$$\rho _ { l } V _ { v } g h / 2 \longrightarrow \dot { W } _ { l, k } \tau \,,$$
where ˙ W l,k is the rate of viscous dissipation of energy, τ is the time of generation of the vapour volume V v , V v = ˙ V τ v . Hence,
$$\rho _ { l } \dot { V } _ { v } g \frac { h } { 2 } \sim \dot { W } _ { l, k } \,.$$
Let us estimate the viscous dissipation of the kinetic energy of flow W l,k ;
$$\stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stackrel { \sim } { \sim } \stack | \text{$\mathcal{ } } \\ \dot { W } _ { l, k } = \int _ { V } \vec { v } \cdot \vec { f } _ { \text{vis} } \mathrm d V \sim \int _ { V } \vec { v } \cdot \left ( - \eta _ { l } \frac { \vec { v } } { H ^ { 2 } } \right ) \mathrm d V \\ \sim - \frac { \eta _ { l } } { \rho _ { l } } \frac { 2 } { \left ( \frac { H _ { 1 } + H _ { 2 } } { 2 } \right ) ^ { 2 } } \int _ { V } \frac { \rho _ { l } v ^ { 2 } } { 2 } \mathrm d V \sim - 8 \nu _ { l } \left ( \frac { \delta S } { \delta V } \right ) ^ { 2 } W _ { l, k } \,. \\ \intertext { i t is the liquid velocity, \vec { f } _ { \text{vis} } \, is the viscous force per unit volume, \, H \, is the spatial scale, in the case half dictionary, \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \text{th} \, \end{cases}$$
Here /vector v is the liquid velocity, /vector f vis is the viscous force per unit volume, H is the spatial scale of flow inhomogeneity, which is the half-distance between the sheets of the folded interface between liquid components, η l and ν l are the characteristic dynamic and kinematic viscosities of liquids, respectively.
Further, we have to establish the relationship between the flow kinetic energy and the mechanical potential energy in the system. Rising vapour bubbles pump the mechanical energy into the system, while its stochastic dynamics is governed by interplay of its flow momentum and the forces of the gravity and the surface tension on the interface. In thermodynamic equilibrium, the total energy is strictly equally distributed between potential and kinetic energies related to quadratic terms in Hamiltonian (this statement is frequently simplified to a less accurate statement, that energy is equally distributed between kinetic and potential energies associated with each degree of freedom). Being not exactly in the case where one can rigorously speak of thermalization of the stochastic Hamiltonian system dynamics, we still may assess the kinetic energy of flow to be of the same order of magnitude as the mechanical potential energy of the system. Thus,
$$W _ { l, k } \sim W _ { l, p g } + W _ { l, p \sigma } \,,$$
where W l,pg and W l,pσ are the gravitational potential energy and the surface tension energy, respectively. We set the zero levels of these potential energies at the stratified state of the system with a flat horizontal interface.
The gravitational potential energy of the well-stirred state with uniform distribution of two phases over hight is
$$W _ { l, p g } \sim \Delta \rho _ { l } V g \frac { h } { 2 } \,, \\ \dots \, \quad \, \dots \, \quad \, \dots$$
where ∆ ρ l is the component density difference. The surface tension energy is
$$W _ { l, p \sigma } & \sim ( \sigma _ { 1 } + \sigma _ { 2 } ) V \left ( \frac { \delta S } { \delta V } \right ) \,, \\ \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \ \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cd.$$
where we neglected the interface area of the stratified state compared to the area V ( δS/δV ) in the well-stirred state. Due to the presence of the vapour layer between liquids the effective surface tension coefficient of the interface is ( σ 1 + σ 2 ) but not σ 12 as it would be in the absence of the vapour layer.
## III. THE VALUE OF ( δS/δV ) YIELDING BALANCE OF ENERGY FLUXES
Collecting Eqs. (1)-(4), one finds
$$\rho _ { l } \frac { \dot { Q } _ { V } \, V } { \Lambda _ { 1 } n _ { 1 * } ^ { ( 0 ) } + \Lambda _ { 2 } n _ { 2 * } ^ { ( 0 ) } } g \frac { h } { 2 } \approx 8 \nu _ { l } \left ( \frac { \delta S } { \delta V } \right ) ^ { 2 } \left [ \Delta \rho _ { l } V g \frac { h } { 2 } + ( \sigma _ { 1 } + \sigma _ { 2 } ) V \left ( \frac { \delta S } { \delta V } \right ) \right ] \,.$$
This equation can be simplified to
$$\dot { Q } _ { V } \approx B \left ( \frac { \delta S } { \delta V } \right ) ^ { 2 } \left [ 1 + \frac { 2 } { k _ { 1 2 } ^ { 2 } h } \left ( \frac { \delta S } { \delta V } \right ) \right ] \,, \\ \dots \quad \dots$$
where B = 8 ν l (Λ 1 n (0) 1 +Λ 2 n (0) 2 )∆ ρ /ρ l l and k 12 = √ ( ρ 2 -ρ 1 ) g/ σ ( 1 + σ 2 ). Noteworthy, the relative importance of the first and second terms in the brackets in Eq. (5) depends on the system size h .
For the n-heptane-water system, B ≈ 1 5 J . / (m · s) and l k 12 ≡ 1 /k 12 ≈ 0 5 cm [14]. . For a well-stirred system the distance between sheets of the folded interface ( δS/δV ) -1 /lessmuch h . The average compound of these two values can be either small or large compared to l k 12 ; (1) h · ( δS/δV ) -1 /lessmuch l 2 k 12 corresponds to the case of the surface tension dominated system, (2) h · ( δS/δV ) -1 /greatermuch l 2 k 12 corresponds to the case of a gravity-driven system.
Cubic equation (5) possesses only one positive solution which is real-valued for any value of ˙ Q /B V ;
$$\left ( \frac { \delta S } { \delta V } \right ) = \left ( \frac { \delta S } { \delta V } \right ) _ { g } \cdot G \left ( \frac { \left ( \dot { Q } _ { V } / B \right ) ^ { 1 / 2 } } { k _ { 1 2 } ^ { 2 } V ^ { 1 / 3 } } \right ),$$
where ( δS/δV ) g = ( ˙ Q /B V ) 1 2 / is the value of parameter ( δS/δV ) for a gravity-driven system and function G s ( ) = (6 s ) -1 ( R + R -1 -1), R = ( √ 27 s + √ 27 s 2 -1) 2 3 / ; G (0) = 1 and G s ( /greatermuch 1) = (2 s ) -1 3 / .
## IV. CONCLUSION
Basing on the energy flux balance in the system and assumption of the system stochastization, we have assessed the value of ( δS/δV ). Expression (6) allows estimating the value of parameter ( δS/δV ) as a function of heat influx ˙ Q V to the system.
The work has been financially supported by the Russian Scientific Foundation (grant no. 14-21-00090).
- [1] E. Krell, Handbook of Laboratory Distillation , 2nd ed. (Elsevier, 1982), Chapter 4.3.
- [2] C. J. Geankoplis, Transport Processes and Separation Process Principles , 4th ed. (Prentice Hall, 2003).
- [3] H. C. Simpson, G. C. Beggs, M. Nazir, Evaporation Of Butane Drops In Brine , Desalination 15 , 11 (1974).
- [4] G. P. Celata, M. Cumo, F. D'Annibale, F. Gugliermetti, G. Ingui', Direct contact evaporation of nearly saturated R 114 in water , Int. J. Heat Mass Transfer 38 , 1495 (1995).
- [5] M. L. Roesle, F. A. Kulacki, An experimental study of boiling in dilute emulsions, part A: heat transfer , Int. J. Heat Mass Transfer 55 , 2160 (2012).
- [6] M. L. Roesle, F. A. Kulacki, An experimental study of boiling in dilute emulsions, part B: visualization , Int. J. Heat Mass Transfer 55 , 2166 (2012).
- [7] S. Sideman, J. Isenberg, Direct Contact Heat Transfer with Change of Phase: Bubble Growth in Three-Phase Systems , Desalination 2 , 207 (1967).
- [8] A. A. Kendoush, Theory of convective drop evaporation in direct contact with an immiscible liquid , Desalination 169 , 33 (2004).
- [9] G. Filipczak, L. Troniewski, S. Witczak, in Evaporation, Condensation and Heat transfer , Ed. by A. Ahsan (InTech, 2011).
- [10] K. F. Gordon, T. Singh, E. Y. Weissman, Boiling heat transfer between immiscible liquids , Int. J. Heat and Mass Transfer 3 , 90 (1961).
- [11] C. B. Prakash, K. L. Pinder, Direct contact heat transfer between two immiscible liquids during vaporisation , Can. J. Chem. Engineering 45 , 210 (1967).
- [12] C. B. Prakash, K. L. Pinder, Direct contact heat transfer between two immiscible liquids during vaporization: Part II: Total evaporation time , Can. J. Chem. Engineering 45 , 215 (1967).
- [13] A. V. Pimenova, D. S. Goldobin, Boiling at the Boundary of Two Immiscible Liquids below the Bulk Boiling Temperature of Each Component , JETP 119 (1), 91 (2014).
- [14] A. V. Pimenova, D. S. Goldobin, Boiling of the Interface between Two Immiscible Liquids below the Bulk Boiling Temperatures of Both Components , unpublished (2014). E-print: arXiv:1407.4725, http://arxiv.org/abs/1407.4725
- [15] Indeed, the energy of thermal motion of atoms corresponds to characteristic atom velocities 10 2 -10 3 m s, while nothing comparable can be imagined for macroscopic flow velocities in / realistic situations. The latent heat of water evaporation is even significantly bigger than the kinetic energy of thermal motion of its atoms at T = 300K.
| null |
[
"Denis S. Goldobin",
"Anastasiya V. Pimenova"
] |
2014-08-05T17:18:39+00:00
|
2014-08-05T17:18:39+00:00
|
[
"cond-mat.stat-mech",
"physics.flu-dyn"
] |
Assessment of the dependence of $(δ{S}/δ{V})$ on the heat influx for a well-stirred two-phase system with interfacial boiling
|
For a well-stirred multiphase fluid systems the mean interface area per unit
volume, $(\delta{S}/\delta{V})$, is a significant characteristic of the system
state. In particular, it is important for the dynamics of systems of immiscible
liquids experiencing interfacial boiling. We estimate the value of parameter
$(\delta{S}/\delta{V})$ as a function of heat influx $\dot{Q}_V$ to the system.
|
1408.1038v1
|
## Induced topological order at the boundary of 3D topological superconductors
Peter Finch, James de Lisle, Giandomenico Palumbo, and Jiannis K. Pachos School of Physics and Astronomy, University of Leeds, Leeds, LS2 9JT, United Kingdom (Dated: August 6, 2014)
We present tight-binding models of 3D topological superconductors in class DIII that support a variety of winding numbers. We show that gapless Majorana surface states emerge at their boundary in agreement with the bulk-boundary correspondence. At the presence of a Zeeman field the surface states become gapped and the boundary behaves as a 2D superconductor in class D. Importantly, the 2D and 3D winding numbers are in agreement signifying that the topological order of the boundary is induced by the order of the 3D bulk. Hence, the boundary of a 3D topological superconductor in class DIII can be used for the robust realisation of localised Majorana zero modes.
PACS numbers: 73.20.At, 74.20.Rp, 03.65.Vf, 71.10.Pm, 74.90.+n
Introduction:1D and 2D tight-binding topological superconductors (TS) are a commonly employed medium to probe localised Majorana zero modes (Majoranas) with non-Abelian statistics. 1D tight-binding TS are employed to theoretically investigate the properties of Majoranas and experimentally verify their existence [1]. 2D tight-binding TS provide exactly solvable models, where Majoranas exhibit rich behaviour [2-4]. Despite the success of these models little work has been done in relation to 3D tight-binding TS [5, 6].
Here we consider a general class of 3D tight-binding models of fermions positioned at the vertices of a cubic lattice. We allow for tunnelling and pairing interactions between nearest and next-to-nearest neighbouring sites of the lattice. When we impose time-reversal (TR) and particle-hole (PH) symmetries we obtain explicit realisations of 3D TS of type DIII [7]. For periodic boundary conditions in all three directions we provide a variety of TS that support winding numbers ν 3D = 0 , ± 1 , ± 2 , ± 3 , ± 4. Similar models that exhibit ν 3D = ± 2 [8] and ν 3D = ± 1 [9] and higher [10] have recently been presented. In our model, the higher values of ν 3D are obtained while keeping fixed the size of the unit cell and the range of interactions. Subsequently, we impose open boundary conditions in one direction and identify edge modes with dispersion relation that crosses the band gap [11], which correspond to gapless Majorana surface states. These states acquire a gap when a Zeeman field is applied at the boundary that breaks TR symmetry [12-14]. Thus, the boundary behaves as a TS of the class D. It is known that both the 3D class DIII and the 2D class D TS have a Z topological invariant [7]. Here we show that under certain conditions the actual values of these topological invariants that describe the bulk and the boundary physics of a 3D class DIII TS are equal. We demonstrate this both numerically, for all the tight-binding models presented here, and theoretically, based on an effective topological field theory. This protection of the boundary topological order from the topological character of the bulk provides the means for the fault-tolerant realisation of localised Majoranas [15].
These Majoranas can be employed for topological quantum computation [16] that is resilient against erroneous perturbations or thermal fluctuations.
DIII lattice:We now introduce the lattice model. We consider two species of fermion, a 1 and a 2 , canonically ordered on a cubic lattice, as shown in Fig. 1 (Left). The unit cell, positioned at j = ( j x , j y , j z ), consists of two sites lying along the x -axis. The Hamiltonian is given by
$$H = \sum _ { j } \left ( \sum _ { k } \mu a _ { k j } ^ { \dagger } a _ { k j } + \sum _ { k, k ^ { \prime }, s } t _ { k k ^ { \prime } s } a _ { k j } ^ { \dagger } a _ { k ^ { \prime } j + s } \\ + \sum _ { k, k ^ { \prime }, s } \Delta _ { k k ^ { \prime } s } a _ { k j } a _ { k ^ { \prime } j + s } \right ) + \text{h.c.}, \quad ( 1 )$$
where t kk ′ s and ∆ kk ′ s are the tunnelling and pairing couplings, respectively, µ is the chemical potential and s is a vector connecting interacting unit cells. The interactions are taken to be at most between next-to-nearest neighbours. For periodic lattice we introduce the Fourier transformation a k, j = ∑ p e i p j · a k, p to obtain H = ∑ p ψ h † p ( p ) ψ p , where ψ p = ( a 1 , p , a † 1 , -p , a 2 , p , a † 2 , -p ) T , p ∈ BZ = [0 2 , π ) × [0 , 2 π ) × [0 , 2 π ) and the kernel h ( p ) is a 4 × 4 hermitian matrix. To impose TR and PH symmetries we introduce the unitary operators C TR and C PH and demand that
$$\underset { \substack { \text{Iajo-} \\ \text{en a} \\ \text{en a} } } { \text{Iajo-} } } \quad C _ { \text{TR} } ^ { \dagger } h ^ { * } ( - \text{p} ) C _ { \text{TR} } = h ( \text{p} ), \ \underset { \text{p} } { C _ { \text{PH} } } ^ { \dagger } h ^ { * } ( - \text{p} ) C _ { \text{PH} } = - h ( \text{p} ) \ ( 2 )$$
with C T TR = -C TR and C T PH = C PH . These requirements guarantee that our model belongs in class DIII. Next we interpret the species index k = 1 2 as spin-, 1 / 2 components. For simplicity we restrict to Hamiltonians such that, in the basis ψ p = ( ia 1 , p -a † 1 , -p , a 2 , p -ia † 2 , -p , ia 1 , p + a † 1 , -p , a 2 , p + ia † 2 , -p ) T / √ 2, the kernel takes the spin-triplet TS form [19, 20]
$$\Phi _ { \quad } h ( p ) = \left ( \begin{array} { c c c c } \epsilon ( p ) \mathbb { I } & \Theta ( p ) \\ \Theta ( p ) ^ { \dagger } & - \epsilon ( p ) \mathbb { I } \end{array} \right ), \quad \quad ( 3 )$$
with glyph[epsilon1] ( p ) denoting the normal state, I is the identity 2 × 2 matrix and Θ( p ) the spin-triplet pairing function
FIG. 1: (Colour online) (Left) The fermionic cubic lattice. The unit cell (dashed square) consists of two fermions a 1 (white site) and a 2 (black site). Tunnelling and pairing interactions are assumed between nearest and next-to-nearest neighbouring sites. (Right) The energy gap min p ( | E ( p ) ) | as function of µ and ∆ when t = 1. The winding number is depicted as a function of µ and ∆ taking values ν 3D = 0 , ± 1. Changes in the winding number are accompanied by quantum phase transitions.
Θ( p ) = i ( d p ( ) · σ ) σ y . Here, σ = ( σ , σ x y , σ z ) are the Pauli matrices and d p ( ) = ( d x ( p ) , d y ( p ) , d z ( p )) are odd functions. The corresponding doubly degenerate spectrum is given by E ( p ) = ± √ glyph[epsilon1] ( p ) 2 + | d p ( ) | 2 . We take the system to be prepared in its lowest energy, where both negative valance bands are completely occupied. The topological nature of the gapped regions is identified by the winding number ν 3D ∈ Z that characterises the mapping between the toroidal Brillouin zone T 3 and the sphere S 3 defined by the normalised 4D vector ( ( glyph[epsilon1] p ) , d p ( )) / E | ( p ) . | This can be evaluated in terms of projectors to the two lowest eigenstates of kernel (3) [7].
We now take a specific coupling configuration that gives rise to a non-trivial winding number. More concretely, for a particular set of interactions { s } we can have glyph[epsilon1] ( p ) = t [cos( p x -p z ) + cos( p x ) + 2cos( p y -p z ) + cos( p y )] -µ , d x ( p ) = ∆[sin( p x -p z ) + sin( p x ) + 2 sin( p y -p z ) -sin( p y )], d y ( p ) = 2∆sin( p y ) and d z ( p ) = 2∆sin( p x + p y ) with t, µ, ∆ ∈ R . The energy gap as function of µ and ∆ for t = 1 as well as the winding number, ν 3D , corresponding to each gapped phase are shown in Fig. 1 (Right). It is possible to evaluate the winding number in terms of the set of momenta p ∗ satisfying d p ( ∗ ) = 0 from ν 3D = 1 2 ∑ d p ( ∗ )=0 sgn[ ( glyph[epsilon1] p ∗ )]sgn { det[ ∂ d j i ( p ∗ )] } [20]. This expression shows explicitly the dependence of its sign, sgn( ν 3D ), in terms of the product of the signs of the couplings ∆ and µ , as shown in Fig. 1 (Right).
Boundary properties:Let us now consider the case where a boundary is introduced. For concreteness we take the lattice to extend between two disconnected planes. The Bottom plane (B) positioned at z = 1 and the Top plane (T) positioned at z = l , where l is a positive integer. The Hamiltonian of the system with
glyph[negationslash]
FIG. 2: (Left) An energy dispersion cross section at p y = 0 for the ν 3D = 1 model ( t = 1, ∆ = 2, µ = 3) with open boundaries at z = 1 and z = 20. A single cone appears with a double degeneracy corresponding to the two Majorana edge modes, one for each surface. (Right) A boundary Zeeman field with B y = 0 generates an energy gap, ∆ E , to the edge modes. (Inset) The energy gap ∆ E as a function of B y .
a boundary is given by
$$H \ = \ & \sum _ { \bar { p }, z, s } \psi _ { z, \bar { p } } ^ { \dagger } h ( \bar { p } ; z, z + s ) \psi _ { z + s, \bar { p } } \\ & + \sum _ { \bar { p } } \left ( \psi _ { 1, \bar { p } } ^ { \dagger } h _ { B } \psi _ { 1, \bar { p } } + \psi _ { l, \bar { p } } ^ { \dagger } h _ { T } \psi _ { l, \bar { p } } \right ) \quad ( 4 )$$
where ψ z p , ¯ = ( a 1 , z p , ¯ , a † 1 , z , -¯ p , a 2 , z p , ¯ , a † 2 , z , -¯ p ) T with z a vector in the z -direction, ¯ p ∈ [0 , 2 π ) × [0 , 2 π ) the momentum on the x y -plane and h B and h T are interaction terms corresponding to the Bottom and Top planes, respectively. These terms are introduced to give an energy gap to the boundary states and do not affect the properties of the bulk. While it is possible to consider independent interaction terms at each plane, for uniformity we choose them to be equal, given by h T = h B = B · σ ⊗ I , where B is a 3D vector. These terms can be viewed as an effective Zeeman field. They correspond to interactions between the fermionic modes a 1 and a 2 within the same unit cell.
glyph[negationslash]
Initially, let us take the case where B = 0. When the winding number ν 3D is non-zero then the bulk-boundary correspondence necessitates that gapless helical modes are present at each edge of the system [7]. These are manifested as 2D gapless Majorana cones in the dispersion relation. This is verified in Fig. 2 (Left), where the two edge modes, one per surface, appear in the spectrum as two degenerate conical states. Next, we introduce a magnetic field, B = 0, on the boundary. Non-zero values of B x,z change the position of the Majorana cones in momentum space. The B y term corresponds to complex tunnelling couplings so it breaks TR symmetry at the boundary. As a result it gives an energy gap to the 2D Majorana surface states that appear as paraboloids in Fig. 2 (Right). We can evaluate the 2D winding number in terms of the projectors onto these gapped surface states [17, 18]. For N such states with the same helic-
ity the winding number is given by ν b = ± N/ 2 where b = T, B , that we call partial winding number. The sign of ν b depends on the helicity of the edge states as well as the sign of B y that generates their gap.
While each plane constituting the boundary can be treated independently, the condition h T = h B allows us to consider the entire boundary as a 2D TS dislocated between the two planes. The effective Hamiltonian that describes the low energy limit of these 2D superconducting states breaks TR symmetry, due to the presence of non-zero B y , so it behaves as a class D system [7]. Subsequently, we can define the sum of the partial winding numbers, ν 2D = ∑ b ν b , which characterises the topological phase of the boundary as a whole.
We would now like to see how the winding number ν 2D of the boundary relates to the winding number ν 3D of the bulk when both bulk and boundary are gapped. To make the comparison legitimate we choose ∆, the order parameter of the bulk, and B y , the order parameter of the boundary, to have the same sign, i.e. sgn( B y ) = sgn(∆). Let us first look at the ν 3D = 0 case. We find that each plane supports two pairs of Majorana edge modes with opposite helicities. As a result the 2D winding number is ν 2D = 0. When ν 3D = 1 each plane supports N = 1 Majorana cone. When gapped, each partial winding number contributes ν b = 1 2 to the 2D winding number, / such that ν 2D = 1. A similar result holds when ν 3D = -1. In analogy to the 2D SC in class D [15] we expect that when ν 3D = ± 1 the boundaries can support Majoranas localised at the endpoints of vortex strings that terminate on the boundaries [21].
Higher winding numbers:We now present models that support higher winding numbers without the need to increase the size of the unit cell or the range of interactions. These models are obtained by searching among a variety of possible configurations of interactions that respect the TR and PH symmetries (2), so they are in class DIII. Using the notation of (3), we present Hamiltonian H 2 with glyph[epsilon1] ( p ) = t [cos( p x )+cos( p x -p z )] -µ d , x ( p ) = ∆[sin( p x )+sin( p x -p z )], d y ( p ) = 2∆sin( p y ) and d z ( p ) = 2∆sin( p x + p y ) that supports topological phases with ν 3D = 0 , ± 2, Hamiltonian H 3 with glyph[epsilon1] ( p ) = - -t [ cos( p x )+ cos( p x -p z )+cos(2 p y ) -cos( p y + p z )+cos( p y )]+ µ d , x ( p ) = ∆[ -sin( p x ) -sin( p x -p z ) + sin(2 p y ) -sin( p y ) -sin( p y -p z )], d y ( p ) = 2∆sin( p y ) and d z ( p ) = 2∆sin( p x -p y ) that supports ν 3D = 0 , ± 1 , ± 3 and Hamiltonian H 4 with glyph[epsilon1] ( p ) = t [cos( p x + p z ) -2 cos( p y )]+ µ d , x ( p ) = -∆sin( p x + p z ), d y ( p ) = -2∆sin( p z ) and d z ( p ) = 2∆sin( p x + p y ) that supports ν 3D = 0 , ± 2 , ± 4.
The phase diagrams of the H 2 , H 3 and H 4 models are shown in Fig. 3 (Top Row). In each case, multiple gapped topological phases are separated by gapless phase transitions. Wethen introduce open boundary conditions in the z -direction and observe gapless Majorana states localised at each surface. When the winding number ν 3D is non-zero we find that for all models N = | ν 3D | many such
FIG. 3: (Colour online) (Top Row) The energy gaps of H 2 , H 3 and H 4 , ordered from left to right, as a function of µ and ∆ for t = 1. The winding number associated with each gapped phase is indicated. (Bottom Row) The centre of the surface Majorana cones are given for the model corresponding to the gap diagram above it, when each model is in the max( ν 3D ) phase. All plots were generated with µ = 0 1, ∆ = . t = 1.
glyph[negationslash]
surface states exist at each edge with the same helicity. Next we consider each model in the phase with max( ν 3D ). The position of the zero energy points of their Majorana edge states is shown in Fig. 3 (Bottom Row). A pair of Majorana modes (one per edge) corresponds to each dot in the Figure. Finally, we introduce a boundary Zeeman field with B y = 0. We numerically find that for all cases the induced 2D winding number satisfies ν 2D = ν 3D . In other words, if a phase transition in the bulk occurs such that the new phase has a different 3D winding number, then the number and/or helicity of Majorana cones on the boundary and the 2D winding number change accordingly.
Bulk-boundary correspondence:The bulkboundary correspondence witnessed here as the agreement between ν 2D and ν 3D is not accidental. We now present an analytic link between these two topological invariants by looking at the thermal Hall conductivity (THC). In particular, we establish a relationship between the THC of a 2D TS in the class D, that is proportional to ν 2D , and the THC induced on the boundary (composite system of two surfaces) of a 3D TS in the class DIII, that is shown to be proportional to ν 3D .
The thermal properties of TS can be encoded in an effective topological field theory (ETFT) obtained by introducing a fictitious gravitational field, as described by Luttinger theory [22]. In the 2D class D model, the ETFT is a gravitational Chern-Simons theory as showed in [15, 23] (see also (8) below). From this description and from independent arguments [24] it was possible to show that the THC is given by κ 2D th = ν C 12 π T , for temperatures T → 0, where ν C is the Chern number of the TS ( glyph[planckover2pi] = k B = 1). In our case, the THC of the composite boundary of the 3D TS is the sum of the THC of the two
surfaces, thus we have
$$\kappa _ { t h } ^ { \text{boundary} } = \frac { \nu _ { 2 D } } { 1 2 } \pi \, T, \quad \quad \ ( 5 ) \quad \text{in} \,.$$
where ν 2D = ν T + ν B .
We now calculate the boundary THC in an alternative way, by starting from the 3D bulk properties. The 3D model (1) of a certain topological phase can be adiabatically connected to a model with a low energy description given by massive Dirac fermions [25] with action
$$S _ { \psi } = \int _ { M } d ^ { 4 } x \, \bar { \psi } ( \gamma ^ { \mu } \partial _ { \mu } + m ) \psi, \quad \quad ( 6 ) \quad \text{$\ a p$} \\ \text{sur}!$$
where µ = 0 1 2 3, , , , ¯ = ψ γ ψ 0 , γ µ are the Dirac matrices and M is the (3 + 1)-dimensional spacetime. In order to calculate the ETFT, we introduce a curved background to the fermionic action S ψ . This is done by coupling the fermions with spin connection, ω µ , and tetrads that naturally encode the geometric properties of curved spaces [26]. The effective action S eff that describes the model purely in terms of the spin connection can be derived by integrating out the fermions in the corresponding partition function. The topological part of S eff that faithfully describes the low-energy behaviour of the model is given by [27, 28]
$$S _ { \text{eff} } ^ { M, \text{top} } = \frac { 1 } { 2 } \frac { \theta } { 7 6 8 \pi ^ { 2 } } \int _ { M } d ^ { 4 } x \, \epsilon ^ { \mu \nu \alpha \beta } \text{tr} ( R _ { \sigma \mu \nu } ^ { \rho } R _ { \rho \alpha \beta } ^ { \sigma } ), \quad ( 7 ) \quad \text{can e} \\ \quad \text{ticular}$$
where R ρ σµν is the Riemann tensor given in terms of the spin connection ω µ , while θ = ν 3D π mod 2 π [29]. This is the ETFT of the 3D TS. The topological behaviour of the gapped boundary can be obtained from (7) by considering that the spatial part of M has a non-empty boundary with spacetime boundary manifold given by ∂M . By applying Stokes' theorem we have
$$S _ { \text{eff} } ^ { \partial M, \text{top} } & = \frac { 1 } { 2 } \frac { \theta } { 1 9 2 \pi ^ { 2 } } \int _ { \partial M } d ^ { 3 } x \, \epsilon ^ { \mu \nu \lambda } \text{tr} \left ( \omega _ { \mu } \partial _ { \nu } \omega _ { \lambda } + \frac { 2 } { 3 } \omega _ { \mu } \omega _ { \nu } \omega _ { \lambda } \right ). \quad \text{$Never$} \\ \longrightarrow \quad. \quad \sim \partial M \text{ton} \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad.$$
The action S ∂M, top eff corresponds to the gravitational Chern-Simons theory. By general arguments [15, 27, 28] one can connect the coefficient of this Chern-Simons theory to the THC on the boundary of a 3D TS. In our system the boundary consists of two disconnected surfaces, the Top and the Bottom. Therefore the total boundary THC, κ boundary th , is given by
$$\kappa _ { \text{th} } ^ { \text{boundary} } = 2 \times \frac { \nu _ { 3 \text{D} } } { 2 4 } \pi T. \quad \ \ ( 9 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
Comparing equations (5) and (9), we deduce that
$$\nu _ { 3 D } = \nu _ { 2 D } \text{ \quad \ \ } ( 1 0 )$$
as it was also numerically verified for all the presented models. Note that a relative sign freedom in (10) is possible due to the orientation of the boundary surface employed during the application of the Stokes' theorem. This is the geometric equivalent of the freedom we had in choosing the sign of B y .
Protected 2D topological order:We have presented 3D TS models in class DIII that exhibit a large variety of winding numbers ν 3D = 0 , ± 1 , ± 2 , ± 3 , ± 4. Numerical and theoretical analysis showed that the boundary of these models behaves as a 2D TS of class D with winding number ν 2D = ν 3D . Nevertheless, there is an intriguing difference between the boundary of a 3D TS and a purely 2D TS system. In the 3D case the 2D topological boundary is delocalised between two spatially separated surfaces. This non-local encoding of topological order, together with relation (10), provide a protection that is not present in the purely 2D system, as we analyse below.
Firstly, external perturbations cannot affect the topological nature of the boundaries. As an adversary mechanism consider placing on a ν b = 1 2 edge another 2D / lattice system with Chern number ν Ch = ± 1. The latter can be effectively described by two massive Majorana fermions [2]. Perturbative interactions between this system and the edge can cause the edge Majorana fermion to pair with one of the Majorana fermions of the appended system creating a non-topological Dirac fermion [30]. This leaves behind an edge with a single Majorana fermion that is still described by ν b = ± 1 / 2, where one can easily compensate for a change in the sign. In particular, this mechanism cannot cause ν b to become zero.
Secondly, phase transitions due to a thermal environment are suppressed at low enough temperatures. Let us consider local thermal fluctuation at the boundary of the TS manifested as a vortex string in the bulk with both endpoints (a vortex-antivortex pair) residing on the same surface. As these endpoints bear localised Majoranas, a finite density of such thermal errors could cause a quantum phase transition due to vortex nucleation [31, 32]. Nevertheless, these vortex strings have a finite energy per string length that generates a string tension [33]. As a consequence, the vortex and the antivortex will be attracted to each other and annihilate, thus self-correcting the thermal error [34]. Only vortex strings with endpoints at opposite surfaces can be stable [35, 36]. Hence, for low enough temperatures, the vortex string tension will cause the density of vortices on the surface to be zero, and its topological phase will remain intact. These errorresilience characteristics make the boundaries of 3D class DIII TS a promising laboratory for performing topological quantum computation [16].
Acknowledgements:We would like to thank Ville Lahtinen, Joel Moore and Steven Simon for inspiring discussions. This work was supported by EPSRC.
- [2] A. Y. Kitaev, Ann. Phys. 321 , 1 (2005).
- [3] V. Lahtinen, G. Kells, A. Carollo, T. Stitt, J. Vala and J. K. Pachos, Ann. Phys. 323 , 9 (2008).
- [4] V. Lahtinen and J. K. Pachos, New J. Phys. 11 , 093027 (2009) .
- [5] S. Ryu, Phys. Rev. B 79 075124 (2009).
- [6] Z. Nussinov, G. Ortiz and E. Cobanera, Phys. Rev. B 86 , 085415 (2012).
- [7] A. P. Schnyder, S. Ryu, A. Furusaki and A. W. W. Ludwig, Phys. Rev. B 78 , 195125 (2008).
- [8] A. P. Schnyder, S. Ryu and A. W. W. Ludwig, Phys. Rev. Lett. 102 , 196804 (2009).
- [9] L. Fu and E. Berg, Phys. Rev. Lett. 105 , 097001 (2010).
- [10] D.-L. Deng, S.-T. Wang and L.-M. Duan, Phys. Rev. B 89 , 075126 (2014).
- [11] M. Z. Hasan and C. L. Kane, Rev. Mod. Phys. 82 , 3045 (2010).
- [12] M. Sato, Y. Takahashi and S. Fujimoto, Phys. Rev. Lett. 103 , 020401 (2009).
- [13] J. Alicea, Phys. Rev. B 81 , 125318 (2010).
- [14] L. Fu and C. L. Kane, Phys. Rev. Lett. 100 , 096407 (2008).
- [15] N. Read and D. Green, Phys. Rev. B 61 , 10267 (2000).
- [16] J. K. Pachos, Topological Quantum Computation , Cambridge University Press (2012).
- [17] D. J. Thouless, M. Kohmoto, M. P. Nightingale, and M. den Nijs, Phys. Rev. Lett. 49 , 405 (1982).
- [18] J. de Lisle, S. De, E. Alba, A. Bullivant, J. J. GarciaRipoll, V. Lahtinen and J. K. Pachos, to appear in NJP, arXiv:1402.3222 (2014).
- [19] M. Sigrist and K. Ueda, Rev. Mod. Phys. 63 , 239 (1991).
- [20] M. Sato, Phys. Rev. B 79 , 214526 (2009).
- [21] J. C.Y. Teo and C.L. Kane, Phys. Rev. Lett. 104 , 046401 (2010).
- [22] J. M. Luttinger, Phys. Rev. 135 , A1505 (1964).
- [23] G. E. Volovik, JETP Lett. 51 , 125 (1990).
- [24] H. Sumiyoshi and S. Fujimoto, J. Phys. Soc. Jpn. 82 , 023602 (2013).
- [25] S. Ryu, A. P. Schnyder, A. Furusaki and A. W. W. Ludwig, New J. Phys. 12 , 065010 (2010).
- [26] We apply the following substitutions in the action (6): ∂ µ → ∂ µ + ω µ , γ µ → γ α µ e α , d x 4 → d x e 4 | | with α, β = 0 1 2 3, , , , where ω µ = i 4 ω αβ µ [ γ α , γ β ] is the spin connection, e µ α are the tetrads and | e | is the corresponding determinant. In particular, we have that g µν = e α β µ e ν η αβ where g µν is the metric tensor and η αβ the Minkowski one, while the spin connection has the Levi-Civita form: ω αβ µ = e α ν glyph[triangleinv] µ e βν , where glyph[triangleinv] µ = e λ µ ( ∂ λ + ω λ ).
- [27] Z. Wang, X.-L. Qi and S.-C. Zhang, Phys. Rev. B 84 , 014527 (2011).
- [28] S. Ryu, J. E. Moore and A. W. W. Ludwig, Phys. Rev. B 85 , 045104 (2012).
- [29] Z. Wang and S.-C. Zhang, Phys. Rev. B 86 , 165116 (2012).
- [30] B. A. Bernevig and T. L. Hughes, Topological Insulators and Topological Superconductors , Princeton University Press (2013).
- [31] V. Lahtinen, A. W. W. Ludwig, J. K. Pachos and S. Trebst, Phys. Rev. B 86 , 075115 (2012).
- [32] B. Bauer, R. M. Lutchyn, M. B. Hastings and M. Troyer, Phys. Rev. B 87 , 014503 (2013).
- [33] M. Eto, Y. Hirono, M. Nitta, and S. Yasui, Prog. Theor. Exp. Phys. 012D01 (2014).
- [34] A. M. J. Schakel, Phys. Rev. E 63 , 026115 (2001).
- [35] P. Hosur, P. Ghaemi, R. S. K. Mong, and Ashvin Vishwanath, Phys. Rev. Lett. 107 , 097001 (2011).
- [36] H.-H. Hung, P. Ghaemi, T. L. Hughes and M. J. Gilbert, Phys. Rev. B 87 , 035401 (2013).
|
10.1103/PhysRevLett.114.016801
|
[
"Peter Finch",
"James de Lisle",
"Giandomenico Palumbo",
"Jiannis K. Pachos"
] |
2014-08-05T17:24:56+00:00
|
2014-08-05T17:24:56+00:00
|
[
"cond-mat.mes-hall",
"cond-mat.supr-con",
"quant-ph"
] |
Induced topological order at the boundary of 3D topological superconductors
|
We present tight-binding models of 3D topological superconductors in class
DIII that support a variety of winding numbers. We show that gapless Majorana
surface states emerge at their boundary in agreement with the bulk-boundary
correspondence. At the presence of a Zeeman field the surface states become
gapped and the boundary behaves as a 2D superconductor in class D. Importantly,
the 2D and 3D winding numbers are in agreement signifying that the topological
order of the boundary is induced by the order of the 3D bulk. Hence, the
boundary of a 3D topological superconductor in class DIII can be used for the
robust realisation of localised Majorana zero modes.
|
1408.1039v1
|
## Silicon-Chip Mid-Infrared Frequency Comb Generation
Austin G. Griffith , Ryan K. W. Lau , Jaime Cardenas , Yoshitomo Okawachi , Aseema 1 2 1 2 Mohanty 1 , Romy Fain , Yoon Ho Daniel Lee , Mengjie Yu , Christopher T. Phare , Carl B. 1 1 2 1 Poitras , Alexander L. Gaeta 1 2,3 , Michal Lipson 1,3,*
1 School of Electrical and Computer Engineering, Cornell University, Ithaca, NY 14850, USA
2
School of Applied and Engineering Physics, Cornell University, Ithaca, NY 14853, USA
3 Kavli Institute at Cornell for Nanoscale Science, Cornell University, Ithaca, NY 14853 USA
*Corresponding author: [email protected]
Optical frequency combs represent a revolutionary technology for high precision spectroscopy due to their narrow linewidths and precise frequency spacing. Generation of such combs in the mid-infrared (IR) spectral region (2-20 μm) is of great interest due to the presence of a large number of gas absorption lines in this wavelength regime . Recently, 1 frequency combs have been demonstrated in the MIR in several platforms, including fiber combs, mode-locked lasers, optical parametric oscillators, and quantum cascade lasers 2-5 . However, these platforms are either relatively bulky or challenging to integrate on-chip. An alternative approach using parametric mixing in microresonators is highly promising since the platform is extremely compact 6-14 and can operate with relatively low powers. However, material and dispersion engineering limitations have prevented the realization of a microresonator comb source past 2.55 μm 13 . Although silicon could in principle provide a
CMOS compatible platform for on-chip comb generation deep into the mid-IR, to date, silicon's linear and nonlinear losses have prevented the realization of a microresonatorbased comb source. Here we overcome these limitations and realize a broadband frequency comb spanning from 2.1 μm to 3.5 μm and demonstrate its viability as a spectroscopic sensing platform. Such a platform is compact and robust and offers the potential to be versatile and durable for use outside the laboratory environment for applications such as real-time monitoring of atmospheric gas conditions.
Recent demonstrations of mid-IR frequency comb platforms have enabled excellent sources for spectroscopy. However, the realization of a robust on-chip integrated mid-IR comb source has proven elusive. On-chip integration and miniaturization of the mid-IR comb is critical, as it will enable high portability for stand-off atmospheric sensing out in the field, and well as monolithic integration with other necessary components, such as resonant cavities for gas sensing and photodetectors for measurements. In particular, a CMOS compatible integrated mid-IR comb source would be inexpensive and straightforward for mass production. Mid-IR combs have been previously demonstrated in platforms such as fiber lasers, mode-locked lasers, and optical parametric oscillators 2-4 , but these platforms are relatively bulky and cannot be integrated. Supercontinuum generation represents another means for generating broad spectrum in the midIR, and has been realized in a number of platforms including silicon waveguides 15 ; however, this requires a high peak power pulsed femtosecond source that can generate a broadband coherent spectrum, and for many applications it is desirable to have comb spacings much larger than the ~100-MHz spacing typically produced by such lasers. Another route to mid-IR comb generation is through the use of quantum cascade lasers (QCLs) , but the active materials used make on5 chip integration difficult.
Microresonator-based combs are promising because they can generate a broad frequency comb in a compact and robust integrated platform. However, material and dispersion engineering limits have prevented the realization of a microresonator comb source past 2.55 μm 13 . Optical frequency combs are a coherent light source consisting of discrete lines that are equally spaced in frequency. Mid-IR comb sources have proved promising for spectroscopy since their broad bandwidth and narrow frequency linewidths make them ideal for probing narrow molecular transitions. With a properly phase-matched geometry, a frequency comb can be generated with a high quality factor microresonator using a single continuous wave (cw) pump laser . Using the 6 parametric χ (3) nonlinear process of four-wave mixing (FWM), energy is transferred from the pump laser into frequency sidebands. Comb lines will be generated at modes supported by the microresonator and lead to an optical frequency comb with a spacing equal to that of the free spectral spacing of the resonant cavity. Microresonator-based frequency comb sources have been demonstrated in a number of platforms, including silica, quartz, fluoride glasses, silicon nitride, Hydex glass, aluminum nitride, and diamond 6-14 . Parametric four-wave mixing and parametric oscillation require a high quality factor resonator and proper dispersion engineering, While such dispersion engineering has allowed for broad-band combs pumped at 1-µm 16 , which has limited the reach of microresonator combs into the mid-IR. In particular, on-chip mid-IR comb generation has not been realized due to the difficulty of creating a highly confined and highquality microresonator in semiconductor thin films.
Although silicon could provide a CMOS-compatible platform for on-chip comb generation deep into the mid-IR, silicon's linear and nonlinear losses have until now prevented the realization of a silicon microresonator-based comb source. Silicon has a wide transparency window from 1.2 µm to past 8 µm 17 and a large third-order optical nonlinearity ( n 2 = 10 -14 cm /W at 2.5 µm 2 wavelength 18 ), which makes it an excellent platform for mid-IR nonlinear optics. In etched silicon microresonators, quality factors have been limited by scattering losses due to roughness in the waveguide sidewalls, which is made worse by the high index contrast between waveguide core and cladding. The dominant nonlinear loss in silicon in the 2.2-3.3 µm region is three photon absorption (3PA) - a process where three photons are simultaneously absorbed to excite an electron-hole pair 19 . The number of photons lost directly to 3PA are small (dominated by linear waveguide losses), but the generated free-carrier population will induce significant optical losses 20 .
In order to overcome silicon's linear losses, we use a novel thermal oxidation based 'etchless' fabrication process to achieve a high quality factor of 590,000 in a silicon microresonator at a wavelength of 2.6 μm. The silicon waveguide is defined in an etchless manner using thermal oxidation, instead of dry etching, to form the waveguide core 21,22 . We deposit low pressure chemical vapor deposition silicon nitride on top of a 500 nm silicon-on-insulator (SOI) wafer. The silicon nitride serves as an oxidation mask. We then pattern and etch only the silicon nitride, then thermally oxidize the wafer. The silicon underneath the nitride mask doesn't oxidize, leaving the silicon waveguide core intact. This method avoids the roughness and absorption sites that can be introduced by reactive ion etching 23 . See methods for full fabrication process. The result is a 590,000 intrinsic quality factor resonator at 2.6 µm, which to our knowledge, is the highest yet achieved for a silicon resonator in the mid-IR. This corresponds to a propagation loss of 0.7 dB/cm. The geometry of the silicon waveguide governs the bandwidth of the frequency comb since the waveguide cross-section determines its dispersion profile 8,24 .
In order to mitigate silicon's nonlinear loss, we embed the resonator in a reverse biased PIN diode junction to sweep out carriers generated from three-photon absorption. Significant free carriers accumulate when pumping a passive silicon waveguide with a cw laser, limited only by the natural free-carrier lifetime of the structure. Here we counteract the carrier generation while using a cw pump is to extract carriers using a PIN junction operated in reverse bias 25 . The PIN junction prevents the electrical injection of carriers into the waveguide while allowing generated free carriers to be swept out - with effective free carrier lifetimes demonstrated as short as 12 ps 26 .
We demonstrate the generation of an optical frequency comb between 2.1 μm and 3.5 μm under reverse bias conditions. In order to generate the frequency comb shown in Fig. 2, we coupled 150 mW of optical power into the bus waveguide composed of a 1.4-µm wide etchless waveguide and tuned the frequency of the pump to be on-resonance with the microresonator. We extract 2.7 mA of current with 10 volts applied to the PIN junction. This relatively high current is an indication that indeed that the three-photon absorption is significant in these structures and therefore the reverse PIN geometry is crucial for achieving oscillations. The generated frequency comb has a line spacing of 130 GHz, corresponding to the FSR of the microresonator. We measure an oscillation threshold with a low power of 3.1 +/- 0.6 mW in the bus waveguide. In order to resolve the full extent of the frequency comb and overcome the limited dynamic range for the optical spectrum analyzer (OSA), we filter the output lines using a series of optical bandpass filters. These filters attenuate the pump and the central part of the comb, allowing us to reach the noise floor of the OSA. The comb spectrum as shown in Fig. 2 is then corrected for the spectral profile of the filters.
We observe a strong influence of the comb spectral shape on the voltage applied to the diode, indicating that indeed the control of the carrier population through the PIN diode is critical for achieving the wide combs. With the voltage source off, combs are generated, however their bandwidth is limited by the natural free-carrier lifetime of the waveguide - which was measured to be 5.23 ns +/- 0.83 ns using a pump-probe experiment. See methods for details. Although a few comb lines can be generated with the PIN turned off, the PIN is required to achieve broadband comb generation above the experiment noise floor - indicated by the dashed line shown in Fig. 3. By increasing the voltage, we find that we can control the comb shape by dynamically changing the carrier lifetime in the ring. In order to simulate the spectral-temporal dynamics of the microresonator combs, we use a recently developed numerical approach 24,27 based on the Lugiato-Lefever equation 28,29 . We show the simulated comb spectra for various free-carrier lifetime values in the right column of Fig. 3 and confirm that a reduced free-carrier lifetime results in broadband comb generation.
In order to demonstrate the feasibility of this comb as a source for spectroscopy, we show proofof-concept gas sensing in the mid-IR. Since our comb source generates lines beyond 3 μm, we have access to a significant number of molecular gases that absorb in that range. Acetylene (C2H2) has strong vibrational absorption in the 3.0 to 3.1 um range, and consists of a set of comb-like absorption lines as shown in Fig. 4A (retrieved from the HITRAN database 30 ). We first filtered the comb to keep only those lines near the acetylene absorption band. We measured the transmission of the comb through an air-filled gas cell, as shown in Fig. 4B using the OSA. We then filled the gas cell with an 8% acetylene/air mixture, and measured the transmission again. Despite the fact that the spacing of the acetylene absorption peaks and the comb lines are different, we were able to measure strong attenuation corresponding to four of the C2H2 absorption lines (Fig. 4C). Note that since we are using an integrated microresonator, the line spacing can in principle be engineered independently of the dispersion profile, to enable detection of numerous spectroscopic lines and therefore high sensitivity detection of a particular gas.
In summary we have demonstrated the first on-chip integrated mid-IR frequency comb source generated in an etchless silicon microresonator platform and the first demonstration of parametric oscillation in a silicon platform. This platform can enable a versatile and straightforward source for mid-IR gas spectroscopy and opens the door for compact and robust mid-IR gas sensors. The ability to achieve comb generation in an integrated platform will enable practical realization of completely integrated mid-IR comb based spectrometers usable in a myriad of environments.
## Methods:
Device fabrication. A thermal oxidation based 'etchless' fabrication process was used to fabricate the devices. Using a commercial silicon-on-insulator wafer with a 500 nm silicon layer and a 3 μm buried oxide, 200 nm of silicon nitride was deposited using low-pressure chemical vapor deposition (LPCVD). MaN-2405 electron-beam resist was used to pattern the nitride layer.
After exposure and development, the silicon nitride was etched in fluorine chemistry. The wafer was then thermally oxidized at 1200 °C to form the etchless waveguide structure. The wafer was oxidized for 20, 50, and 60 minutes of dry, wet, and dry oxidation respectively. The slab in the taper region was then etched away. The wafer was then clad with 2 μm of PECVD silicon dioxide. The oxide was then etched away next to the waveguides, and then the slab was doped with boron and phosphorous on either side to form a PIN diode. Finally, the wafer was metalized, and wires were patterned and etched.
Free-carrier lifetime measurement. For carrier lifetime measurements, we perform pumpprobe experiments with a counter-propagating pulsed pump and CW probe. An Erbium-doped fiber laser (1550nm, 500fs FWHM) is amplified to a peak power of ~200W in an EDFA, then coupled via a circulator into one facet of the chip. A CW probe (1575nm) is coupled into the opposite facet and appears on the third port of the circulator, isolated from the pump. Neither laser overlaps with a ring resonance; we measure only the bus waveguide. We further improve probe isolation by wavelength filtering the third port output with a grating filter, then route this fiber to an optical sampling oscilloscope. The oscilloscope is triggered off the pump laser's fast sampling photodiode, giving us a time-domain measurement of the transmission of the waveguide after strong pump excitation. Absorption increases sharply when the pump-generated free carriers flood the waveguide, then exponentially decreases as they recombine. We take the time constant of this exponential to be the free-carrier lifetime.
Microresonator comb simulation. To model the microresonator comb, we use the LL equation, modified to include contributions from higher-order dispersion, self-steepening, multi-photon absorption, and free-carriers. Free-carrier generation is governed by three-photon absorption, and recombination is modeled with a single exponential with the effective lifetime as the time constant. Full equations and parameter values can be found in Ref. 27.
## Acknowledgements
The devices were fabricated at the Cornell Nano-Scale Science & Technology Facility. The authors gratefully acknowledge support from DARPA for award # W31P4Q 13 1 0016. The ---authors gratefully acknowledge support from AFOSR for award # BAA AFOSR 2012 02 ---supervised by Dr. Enrique Parra. This work made use of the Cornell Center for Materials Research Shared Facilities which are supported through the NSF MRSEC program (DMR1120296). Austin G. Griffith acknowledges the National Defense Science and Engineering Graduate Fellowship (NDSEG) for funding.
## Author Contributions
A.G.G. designed and fabricated the devices, and performed the experiment. A.G.G. analyzed the data. R.K.W.L. performed the comb simulations. J.C. assisted in designing the devices. J.C., A.M., and C.B.P. assisted in fabricating the devices. Y.O., R.K.W.L., M.J., and R.F. assisted in the experiments. Y.H.D.L. and C.T.P. performed the carrier lifetime measurement. J.C., Y.O., R.K.W.L., A.L.G. and M.L. contributed to data analysis. A.L.G. and M.L. supervised the project. All authors discussed the results and commented on the manuscript.
## References
- 1. Schliesser, A., Picqué, N. & Hänsch, T. W. Mid-infrared frequency combs. Nat. Photonics 6, 440-449 (2012).
- 2. Maddaloni, P., Malara, P., Gagliardi, G. & Natale, P. De. Mid-infrared fibre-based optical comb. New J. Phys. 8, 262-262 (2006).
- 3. Sorokin, E., Sorokina, I. T., Mandon, J., Guelachvili, G. & Picque, N. Sensitive multiplex spectroscopy in the molecular fingerprint 2.4 mum region with a Cr(2+):ZnSe femtosecond laser. Opt. Express 15, 16540-5 (2007).
- 4. Adler, F. et al. Phase-stabilized, 15 W frequency comb at 2.8-4.8 μm. Opt. Lett. 34, 1330 (2009).
- 5. Hugi, A., Villares, G., Blaser, S., Liu, H. C. & Faist, J. Mid-infrared frequency comb based on a quantum cascade laser. Nature 492, 229-33 (2012).
- 6. Del'Haye, P. et al. Optical frequency comb generation from a monolithic microresonator. Nature 450, 1214-7 (2007).
- 7. Savchenkov, A. et al. Tunable Optical Frequency Comb with a Crystalline Whispering Gallery Mode Resonator. Phys. Rev. Lett. 101, 093902 (2008).
- 8. Levy, J. S. et al. CMOS-compatible multiple-wavelength oscillator for on-chip optical interconnects. Nat. Photonics 4, 37-40 (2009).
- 9. Razzari, L. et al. CMOS-compatible integrated optical hyper-parametric oscillator. Nat. Photonics 4, 41-45 (2009).
- 10. Papp, S. B. & Diddams, S. a. Spectral and temporal characterization of a fused-quartzmicroresonator optical frequency comb. Phys. Rev. A 84, 053833 (2011).
- 11. Li, J., Lee, H., Chen, T. & Vahala, K. J. Low-Pump-Power, Low-Phase-Noise, and Microwave to Millimeter-Wave Repetition Rate Operation in Microcombs. Phys. Rev. Lett. 109, 233901 (2012).
- 12. Jung, H., Xiong, C., Fong, K. Y., Zhang, X. & Tang, H. X. Optical frequency comb generation from aluminum nitride microring resonator. Opt. Lett. 38, 2810-3 (2013).
- 13. Wang, C. Y. et al. Mid-infrared optical frequency combs at 2.5 μm based on crystalline microresonators. Nat. Commun. 4, 1345 (2013).
- 14. Hausmann, B. J. M., Bulu, I., Venkataraman, V., Deotare, P. & Lončar, M. Diamond nonlinear photonics. Nat. Photonics 8, 369-374 (2014).
- 15. Kuyken, B. et al. An octave spanning mid-infrared frequency comb generated in a silicon nanophotonic wire waveguide. 12 (2014). http://arxiv.org/abs/1404.1893
- 16. Saha, K., Okawachi, Y., Levy, J. & Lau, R. Broadband parametric frequency comb generation with a 1-μm pump source. Opt. Express 20, 369-373 (2012).
- 17. Matavulj, P. S., Yang, P. Y., Bagolini, A., Mashanovich, G. Z. & Miloševic, M. M. Rib waveguides for mid-infrared silicon photonics. 26, 1760-1766 (2009).
- 18. Gholami, F. et al. Third-order nonlinearity in silicon beyond 2350 nm. Appl. Phys. Lett. 99, 081102 (2011).
- 19. Pearl, S., Rotenberg, N. & van Driel, H. M. Three photon absorption in silicon for 23003300 nm. Appl. Phys. Lett. 93, 131102 (2008).
- 20. Lin, Q., Painter, O. J. & Agrawal, G. P. Nonlinear optical phenomena in silicon waveguides: modeling and applications. Opt. Express 15, 16604-44 (2007).
- 21. Desiatov, B., Goykhman, I. & Levy, U. Demonstration of submicron square-like silicon waveguide using optimized LOCOS process. Opt. Express 18, 18592-7 (2010).
- 22. Griffith, A., Cardenas, J., Poitras, C. B. & Lipson, M. High quality factor and high confinement silicon resonators using etchless process. Opt. Express 20, 21341-5 (2012).
- 23. Oehrlein, G. S. Dry etching damage of silicon: A review. Mater. Sci. Eng. B 4, 441-450 (1989).
- 24. Lamont, M. R. E., Okawachi, Y. & Gaeta, A. L. Route to stabilized ultrabroadband microresonator-based frequency combs. Opt. Lett. 38, 3478-81 (2013).
- 25. Rong, H. et al. An all-silicon Raman laser. Nature 433, 292-4 (2005).
- 26. Turner-Foster, A. C. et al. Ultrashort free-carrier lifetime in low-loss silicon nanowaveguides. Opt. Express 18, 3582-91 (2010).
- 27. Lau, R. K., Okawachi, Y., Lamont, M. R. & Gaeta, A. L. Effects of Multiphoton Absorption on Parametric Comb Generation in Silicon Microresonators. Cleo 2014 STh1I.4 (2014). doi:10.1364/CLEO\_SI.2014.STh1I.4
- 28. Lugiato, L. & Lefever, R. Spatial Dissipative Structures in Passive Optical Systems. Phys. Rev. Lett. 58, 2209-2211 (1987).
- 29. Coen, S., Randle, H. G., Sylvestre, T. & Erkintalo, M. Modeling of octave-spanning Kerr frequency combs using a generalized mean-field Lugiato-Lefever model. Opt. Lett. 38, 37-9 (2013).
- 30. Rothman, L. S. et al. The HITRAN2012 molecular spectroscopic database. J. Quant. Spectrosc. Radiat. Transf. 130, 4-50 (2013).
Figure 1| 1(A) Optical microscope image of a ring resonator and metal contacts fabricated using the etchless process. 1(B) Normalized transmission spectrum taken at low input power, showing a resonance with Q=590,000. 1(C) Simulated optical mode at 2.6 um, showing high modal confinement in Si. 1(D) False colored cross-sectional SEM image of silicon waveguide, doped regions, and metal contacts.
Figure 2| Broadband frequency comb generation from 2.1 μm and 3.5 μm in the etchless silicon micro-resonator. This frequency comb is generated with 150 mW of optical power in the bus waveguide, and with a 10 V reverse bias applied on the PIN junction. The frequency spacing of the comb is 130 GHz. Due to the limited dynamic range of the optical spectrum analyzer, the frequency comb was measured using a series of optical filters.
Figure 3 | Rows A-C: The effect of three photon absorption on frequency comb generation depends strongly on the bias voltage of the PIN structure, with a dashed line showing the approximate noise floor. (A) With the pump set at a fixed wavelength and the voltage source off, only a few lines are generated. (B) By applying voltage in reverse bias to the junction, we observe dramatically more comb lines in both experiment (left) and simulation (right). (C) At high voltages, we achieve broad comb generation.
Figure 4| Demonstration of mid-IR gas sensing using frequency comb source. (A) Absorption spectrum of acetylene gas near 3.0 μm from the HITRAN database 30 . (B) Portion of the frequency comb transmitted through empty gas cell. (C) Transmission of the frequency comb when the gas cell contains 8% Acetylene/Air mixture. One can see the strong attenuation of four comb lines, corresponding to four of the acetylene's absorption lines.
|
10.1038/ncomms7299
|
[
"Austin G. Griffith",
"Ryan K. W. Lau",
"Jaime Cardenas",
"Yoshitomo Okawachi",
"Aseema Mohanty",
"Romy Fain",
"Yoon Ho Daniel Lee",
"Mengjie Yu",
"Christopher T. Phare",
"Carl B. Poitras",
"Alexander L. Gaeta",
"Michal Lipson"
] |
2014-08-05T17:28:37+00:00
|
2014-08-05T17:28:37+00:00
|
[
"physics.optics"
] |
Silicon-Chip Mid-Infrared Frequency Comb Generation
|
Optical frequency combs represent a revolutionary technology for high
precision spectroscopy due to their narrow linewidths and precise frequency
spacing. Generation of such combs in the mid-infrared (IR) spectral region
(2-20 um) is of great interest due to the presence of a large number of gas
absorption lines in this wavelength regime. Recently, frequency combs have been
demonstrated in the MIR in several platforms, including fiber combs,
mode-locked lasers, optical parametric oscillators, and quantum cascade lasers.
However, these platforms are either relatively bulky or challenging to
integrate on-chip. An alternative approach using parametric mixing in
microresonators is highly promising since the platform is extremely compact and
can operate with relatively low powers. However, material and dispersion
engineering limitations have prevented the realization of a microresonator comb
source past 2.55 um. Although silicon could in principle provide a CMOS
compatible platform for on-chip comb generation deep into the mid-IR, to date,
silicon's linear and nonlinear losses have prevented the realization of a
microresonator-based comb source. Here we overcome these limitations and
realize a broadband frequency comb spanning from 2.1 um to 3.5 um and
demonstrate its viability as a spectroscopic sensing platform. Such a platform
is compact and robust and offers the potential to be versatile and durable for
use outside the laboratory environment for applications such as real-time
monitoring of atmospheric gas conditions.
|
1408.1040v1
|
## Laser pulse-shape dependence of Compton scattering
Alexander I. Titov a,b , Burkhard K¨ ampfer c,d , Takuya Shibata , Atsushi Hosaka b b,e , and Hideaki Takabe f a Bogoliubov Laboratory of Theoretical Physics, JINR, Dubna 141980, Russia b RCNP, 10-1 Mihogaoka Ibaraki, 567-0047 Osaka, Japan c Helmholtz-Zentrum Dresden-Rossendorf, 01314 Dresden, Germany d Institut f¨r u Theoretische Physik, TU Dresden, 01062 Dresden, Germany e J-PARC Branch, KEK Theory Center, KEK, Tokai, Ibaraki, 319-1106, Japan f
ILE, Yamada-oka, Suita, Osaka 565-0871, Japan
Compton scattering of short and ultra short (sub-cycle) laser pulses off mildly relativistic electrons is considered within a QED framework. The temporal shape of the pulse is essential for the differential cross section as a function of the energy of the scattered photon at fixed observation angle. The partly integrated cross section is sensitive to the non-linear dynamics resulting in a large enhancement of the cross section for short and, in particular, for ultra-short flat-top pulse envelopes which can reach several orders of magnitude, as compared with the case of a long pulse. Such effects can be studied experimentally and must be taken into account in Monte-Carlo/transport simulations of the interaction of electrons and photons in a strong laser field.
PACS numbers: 12.20.Ds, 13.40.-f, 23.20.Nx
## I. INTRODUCTION
The rapidly progressing laser technology [1] offers unprecedented opportunities for investigations of quantum systems with intense laser beams [2]. A laser intensity of ∼ 2 × 10 22 W/cm 2 has been already achieved [3]. Intensities of the order of I L ∼ 10 23 ... 10 25 W/cm 2 are envisaged in near future, e.g. at the CLF [4], ELI [5], and HiPER [6] laser facilities. The high intensities are provided in short pulses on a femtosecond pulse duration level [2, 7], with only a few oscillations of the electromagnetic (e.m.) field or even sub-cycle pulses. (The tight connection of high intensity and short pulse duration is further emphasized in [8]. The attosecond regime will become accessible at shorter wavelengths [9]). These conditions are relevant for the formation of positrons from cascade processes in a photon-electron-positron plasma [10, 11] generated by photon-laser [12], electron-laser [13] or laser-laser interactions [14] (see [15] for surveys). The evaluation of corresponding transport equations needs as an input the probabilities/cross sections for the formation of e + e -pairs (e.g., in the non-linear Breit-Wheeler process) and photons (e.g., in the non-linear Compton scattering).
The solutions, obtained by Reiss, Nikoshov, Ritus, Narozhny and collaborators in a compact form, are valid for an infinite (both in time and space) e.m. background field [16-20]. That means the infinite pulse approximation (IPA) is used often in the above studies. However, the analysis of the non-linear Breit-Wheeler processes in a short plane-wave pulse (we call this the finite pulse approximation (FPA)) performed recently [2124] shows striking differences between results of IPA and FPA which can reach several orders of magnitude depending on the pulse shape, duration, and intensity. Obviously, such significant effects must be properly taken into account in corresponding simulations for the development of seeded cascades [25]. It seems therefore naturally to perform a similar analysis for the Compton scattering, thus further developing recent approaches in [8, 26-31], where FPA effects have been addressed. Most of these papers were focused on fully differential cross sections with relatively simple one-parameter envelope functions, e.g. cos , Gauss, sinh, box or related shapes. It was 2 found that the fully differential cross section has a complicated structure being a rapidly oscillating function of the frequency of the outgoing photons ω ′ at fixed scattering angle θ ′ , especially in the most interesting kinematically forbidden region for one-photon emission enabled by the multi-photon dynamics. The spectral structure of the cross section is rather involved and determined by the properties of the pulse structure, namely its shape and duration, as well as field intensity and kinematics. Since the Compton scattering in the multi-photon (cumulative) region has a basic and applied significance as a source of hard photons it should be interesting and important to answer the following questions (i) what observables and kinematical conditions are preferable for a manifestation of the non-linear dynamics, which may be related to the multi-photon effects, (ii) what is role of the pulse structure (shape and duration), and (iii) under which conditions the predictions of FPA are close to that of IPA which is important for the design of transport approaches. The aim of our paper is to clarify these questions.
For the sake of completeness, we start our analysis from fully differential cross sections which are calculated as a function of the frequency of the outgoing photon at fixed scattering angle. The main difference to the previous studies mentioned above is utilizing a wider class of the pulse envelope functions including flat-top envelopes. Also here, it is shown that the fully differential cross section has a complicated structure being rapidly oscillating function of ω ′ at fixed θ ′ , especially in the kinematically forbidden region. It is clear that experimentally study-
ing the multi-photon dynamics in case of rapidly varying cross sections is a challenging task. Rather integrated observables may overcome this problem.
But here one has to be careful. The totally integrated cross section is not suitable for this aim, since in this case the integration starts from the minimum value of the energy of the outgoing photon, ω ′ 1 , kinetically allowed for the one-photon emission process, and this region dominates in the total cross section, masking the relatively weak effects of multi-photon interactions. To highlight the role of the multi-photon interaction the lower limit of integration ω ′ must be shifted relative to ω ′ 1 : ω > ω ′ ′ 1 . Such partly integrated cross sections are smooth functions of ω ′ and allow to study directly the multi-photon dynamics enabling a clarification of the items (i-iii) formulated above. Going this way we elaborate a method for the calculation of the cross section of Compton scattering in the non-linear (multi-photon) regime accounting for the effect of the finite laser pulse duration with emphasis on different temporal pulse shapes. Our analysis is based on methods developed in [21] for the non-linear Breit-Wheeler processes which is a crossing channel of the non-linear Compton effect. Despite of the similarities between the two processes the physical meaning of the dynamical variables and observables are quite different.
Our paper is organized as follows. In Sect. II we derive the basic expressions for the relevant observables in Compton scattering in FPA. In Sect. III we discuss results of numerical calculations. Our summary is given in Sect. IV.
## II. GENERAL FORMALISM
The Compton process is considered here as the spontaneous emission of one photon off an electron in an external e.m. wave. We employ the four-potential of a circularly polarized laser field in the axial gauge A µ = (0 , /vector A φ ( )) with
$$\mathcal { \cdot } ^ { 1 } \quad \mathcal { \cdot } ^ { 2 } \quad \mathcal { \cdot } ^ { 3 } \quad \mathcal { \cdot } ^ { 4 } \quad \mathcal { \cdot } ^ { 5 } \quad \mathcal { \cdot } ^ { 6 } \quad \mathcal { \cdot } ^ { 7 } \quad \mathcal { \cdot } ^ { 8 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \mathcal { \cdot } ^ { 9 } \quad \\ \vec { A } ( \phi ) & = f ( \phi ) \left ( \vec { a } _ { 1 } \cos ( \phi + \vec { \phi } ) + \vec { a } _ { 2 } \sin ( \phi + \vec { \phi } ) \right ) \, \quad ( 1 ) \quad \text{ for tt} \\ \text{where $\phi$} \ = \ k \cdot x \text{ is invariant phase with four-wave vec-}$$
where φ = k · x is invariant phase with four-wave vector k = ( ω, /vector k ), obeying the null field property k 2 = k · k = 0 (a dot between four-vectors indicates the Lorentz scalar product) implying ω = | /vector k | , /vector a (1 2) , ≡ /vector a ( x,y ) ; | /vector a x | 2 = | /vector a y | 2 = a 2 , /vector a /vector x a y = 0; transversality means /vector k/vector a x,y = 0 in the present gauge. The envelope function f ( φ ) with lim φ →±∞ f ( φ ) = 0 (FPA) accounts for the finite pulse length. (IPA would mean f ( φ ) = 1). To define the pulse duration one can use the number N of cycles in a pulse, N = ∆ /π = 1 2 τω , where the dimensionless quantity ∆ or the duration of the pulse τ are further useful measures. The carrier envelope phase ˜ is particuφ larly important if it is varied in a range comparable with the pulse duration ∆. In IPA it is anyhow irrelevant; in
FPA with ˜ φ /similarequal ∆ the cross section of the photon emission would be determined by an involved interplay of the carrier phase, the pulse duration and pulse shape as well as the intensity of e.m. field as emphasized, e.g., in [32]) (see also [24, 33]). In present work, we drop the carrier phase, thus assuming ˜ φ /lessmuch ∆, and concentrate on the dependence of the cross sections on the parameters responsible essentially for multi-photon effects. A detailed analysis of the impact of ˜ on the photon emission needs a φ separate investigation which is postponed to subsequent work. We also drop a consideration of the pulse focusing (which, however, is more relevant for longer pulses [34] than those covered in the present paper) leaving such an analysis for forthcoming works.
Below, we are going to analyze the dependence of observables on the shape of f ( φ ) for two types of envelopes: the one-parameter hyperbolic secant (hs) shape and the two-parameter symmetrized Fermi (sF) shape
$$\overset { \dots } { \text{lysis} } _ { \text{near} } \quad f _ { h s } ( \phi ) = \frac { 1 } { \cosh \frac { \phi } { \Delta } } \, \quad f _ { s F } ( \phi ) = \frac { \cosh \frac { \Delta } { b } + 1 } { \cosh \frac { \Delta } { b } + \cosh \frac { \phi } { b } } \. \ ( 2 )$$
These two shapes cover a variety of relevant envelopes discussed in literature (for details see [21]). The parameter b in the sF shape describes the ramping time in the neighborhood of φ ∼ ∆. Small ratios b/ ∆cause a flat-top shaping. At b/ ∆ → 0, the sF shape becomes a rectangular pulse [26]. In the following, we choose the ratio b/ ∆ as the second independent parameter for the sF envelope function.
The intensity of the e.m. field is described by the dimensionless parameter ξ 2 = e 2 a 2 m 2 , where m is the electron mass (we use natural units with c = /planckover2pi1 = 1, e / 2 4 π = α ≈ 1 / 137 036). . A second relevant variable is the total energy in an electron-one-photon interaction s = m 2 +2( E + | /vector p | ) ω , where E and ω are the electron energy and the laser background-field photon-frequency in the laboratory system, and we consider head-on collisions.
Using the e.m. potential (1) and the Volkov solution for the electron wave function in that field leads to the following expression for the S matrix element
$$& \stackrel { \mathbb { \sim } } { = } & S = - i e \underset { - \infty } { \overset { \infty } { \int } } d l \, M ( l ) \frac { ( 2 \pi ) ^ { 4 } \delta ^ { 4 } ( p + l k - p ^ { \prime } - k ^ { \prime } ) } { \sqrt { 2 E \, 2 E ^ { \prime } \, 2 \omega ^ { \prime } } } \, \quad ( 3 )$$
where k , k ′ = ( ω , /vector ′ k ′ ), p = ( E, /vector p ) and p ′ = ( E , /vector ′ p ′ ) refer to the four-momenta of the background (laser) field (1), scattered photon, as well as asymptotic incoming (instate) and outgoing (out-state) electrons in the Furry picture. All quantities are considered in the laboratory system. The transition matrix M l ( ) consists of four terms (cf. [26]),
$$M ( l ) = \sum _ { i = 0 } ^ { 3 } M ^ { ( i ) } \, C ^ { ( i ) } ( l )$$
where the transition operators have the form M ( ) i = ¯ u p ′ ˆ M ( ) i u p with
$$\hat { M }$$
$$\hat { M } ^ { ( 0 ) } \ = \ \vec { z } ^ { \$$
Here, u p and ¯ u p ′ are free Dirac spinors depending on the momenta p and p ′ ; and ε ′ denotes the polarization four vector of the scattered photon. Utilizing the prescription of Ref. [21] one can express the coefficients C ( ) i ( ) l through
$$\smile & & \smile _ { \begin{array} { c c c } C ^$$
with ˜ Y l ( z ) = z 2 l ( Y l +1 ( z ) + Y l -1 ( z )) -ξ 2 u u l X z l ( ), where the functions Y l ( z ) and X z l ( ) are defined by
$$\begin{array} {$$
The phase φ 0 is equal to the azimuthal angle of the direction of flight of the outgoing electron, φ 0 = φ e ′ , and is related to the azimuthal angle of the photon momentum as φ γ ′ = φ 0 + . For the dynamical variables in Eqs. (6) and π (7) we use the standard notation: z = 2 lξ √ u u l ( 1 -u u l ) ′ ′ 2
This representation of functions C ( ) i ( ) allows to define l a partial differential cross section with u ≡ ( k · k / k ) ( · p ), u l = l u 0 and u 0 = 2 k · p/m .
$$\frac { d \sigma ( l ) } { d \omega ^ { \prime } \$$
with
$$w ( l ) \ = \ - 2 \widetilde { Y } _ { l } ^ {$$
Equation (9) resembles the corresponding expression for the partial probability of photon emission in the case of
IPA [20] with the substitutions l → n = 1 2 , , · · · and ˜ Y 2 l ( z , Y ) 2 l ( z , Y ) l ( z X ) ∗ l ( z ) → J 2 n ( z ′ ), namely where J n ( z ′ ) denotes Bessel functions with z ′ = 2 nξ √ 1+ ξ 2 √ u u n ( 1 -u u n ) and u n = 2 n k p ( · ) m 2 (1+ ξ 2 ) . Similarly to IPA, the phase φ 0 can be determined through invariants α 1 2 , as cos φ 0 = α /z 1 , sin φ 0 = α /z 2 with α 1 2 , = e ( a 1 2 , · p/k · p -a 1 2 , · p /k ′ · p ′ ).
$$\begin{array} { l l l } I P A \ [ 2 0 ] \text{ with the substitutions } l \to n \ = \ 1, 2, \cdots \text{ and} \\ \widetilde { Y } _ { l } ^ { 2 } ( z ), \, \widetilde { Y } _ { l } ( z ) X _ { l } ^ { * } ( z ) \to J _ { n } ^ { 2 } ( z ^ { \prime } ), \, \text{ namely} \\ \\ w _ { n } \ = \ - 2 J _ { n } ^ { 2 } ( z ^ { \prime } ) + \xi ^ { 2 } ( 1 + \frac { u ^ { 2 } } { 2 ( 1 + u ) } ) \\ \\ \times \ \left ( J _ { n - 1 } ^ { 2 } ( z ^ { \prime } ) + J _ { n + 1 } ^ { 2 } ( z ^ { \prime } ) - 2 J _ { n } ^ { 2 } ( z ^ { \prime } ) \right ) \, \quad ( 1 0 ) \\ \\ \text{where } \, J _ { n } ( z ^ { \prime } ) \ \text{ denotes } \ Bessel \ \text{ functions } \text{ with } \, z ^ { \prime } \ = \\ 2 n \xi \ \sqrt { \ u \ \left ( 1 \quad u \ \right ) } \ \frac { \ } { \ } \ \sim \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \ \lambda \quad \ } \end{array}$$
The dimensionless field intensity ξ 2 can be determined through the average value of the manifestly covariant variable η = T µν p µ p / p ν ( · k ) 2 [35] (cf. also [17]), where T µν is the e.m. stress-energy tensor T µν = g αβ F µα F βν + 1 4 g µν F αβ F αβ and F µν = ∂ µ A ν -∂ A ν µ is e.m. field strength tensor. In the charge's rest frame η = T 00 /ω 2 , where the stress-energy tensor T 00 is equal to the energy density of the e.m. field or to the pulse intensity I L . In IPA the quantity ξ 2 is determined by
$$\xi ^ { 2 } = \frac { e ^ { 2 } } { m ^ { 2 } } \frac { 1 } { \tau _ { I P A } } \underset { - \tau _ { I P A } / 2 } { \overset { \tau _ { I P A } / 2 } { \int } } d t \, \eta = \frac { e ^ { 2 } } { m ^ { 2 } \omega ^ { 2 } } \frac { 1 } { 2 \pi } \underset { - \pi } { \overset { \pi } { \int } } d \phi \ I _ { L } \ ( 1 1 )$$
with the above quoted value e 2 a 2 m 2 , where the averaging interval is set equal to the duration of one cycle, τ IPA = 2 π/ω . The generalization to a finite pulse may be done in a straightforward manner:
$$\xi _ { F P A } ^ { 2 } = \frac { e ^ { 2 } } { m ^ { 2 } } \frac { 1 } { \tau _ { F P A } } \underset { - \infty } { \overset { \infty } { \int } } d t \, \eta = \frac { e ^ { 2 } } { m ^ { 2 } \omega ^ { 2 } } \frac { 1 } { 2 \pi N } \underset { - \infty } { \overset { \infty } { \int } } d \phi \, I _ { L } \. ( 1 2 )$$
Now, the interval τ FPA is determined by the number N of oscillations in a pulse as 2 πN/ω . That is, the quantity ξ 2 , which is included in the expressions for the basic functions (7), cross section (8) and probability (9), can be expressed through the averaged value of the intensity of a finite laser pulse
$$\xi ^ { 2 } = \xi _ { F P A } ^ { 2 } \frac { N } { N _ { 0 } }$$
or
$$( 8 ) \quad & \xi ^ { 2 } = \frac { N } { N _ { 0 } } \frac { e ^ { 2 } } { \omega ^ { 2 } m ^ { 2 } } \langle I _ { L } \rangle \simeq \frac { N } { N _ { 0 } } \frac { 5. 6 \cdot 1 0 ^ { - 1 9 } } { \omega ^ { 2 } [ e V ^ { 2 } ] } \langle I _ { L } \rangle \left [ \frac { W } { \text{cm} ^ { 2 } } \right ] \, \ ( 1 4 ) \\ \quad & \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \end{$$
$$N _ { 0 } = \frac { 1 } { 2 \pi } \underset { - \infty } { \overset { \infty } { \int } } d \phi \left ( f ^ { 2 } ( \phi ) + { f ^ { \prime } } ^ { 2 } ( \phi ) \right ) \text{ \quad \ \ } ( 1 5 )$$
where N I 〈 L 〉 = ( ω/ π 2 ) ∫ ∞ -∞ dt I L . Hence, the normalization factor N 0 defined as has the meaning to renormalize the photon flux in case of the finite pulse and to determine the cross section in
Eq. (8). The factor N 0 is described in some detail in Sect. III.A below. In fact, for the considered envelope functions N 0 /similarequal N and, therefore, ξ 2 /similarequal ξ 2 FPA .
The frequency ω ′ of the emitted photon is related to the auxiliary variable l and the polar angle θ ′ of the direction of the momentum /vector k ′ via
$$\omega ^ { \prime } = \frac { l \, \omega ( E + | \vec { p } | ) } { E + | \vec { p } | \cos \theta ^ { \prime } + l \omega ( 1 - \cos \theta ^ { \prime } ) } \quad \ \ ( 1 6 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\,. \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \, \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \colon \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \ \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \ddot \end{cases}$$
and increases with l at fixed θ ′ since ω ′ is a function of l at fixed θ ′ . For convenience, we also present a similar expression for IPA, where the fermions are dressed and the integer quantity n , together with the field intensity ξ 2 , appear:
$$\omega ^ { \prime } & = \frac { n \, \omega ( E + | \vec { p } | ) } { E + | \vec { p } | \cos \theta ^ { \prime } + \omega ( n + \frac { m ^ { 2 } \xi ^ { 2 } } { 2 ( k \cdot p ) } ) ( 1 - \cos \theta ^ { \prime } ) } \,. \quad ( 1 7 ) \quad \text{$\text{$\text{$two-p$}$}$}, \\ - \quad \dots \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdot \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdclots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \ddots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \dots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \vdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \ \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \ \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots$$
The differential cross section of the one-photon production is eventually
$$\frac { d \sigma } { d \omega ^ { \prime } } = \underset { \zeta } { \int } d l \underset { 0 } { \overset { 2 \pi } { \int } } d \phi _ { e ^ { \prime } } \frac { d \sigma ( l ) } { d \omega ^ { \prime } d \phi _ { e ^ { \prime } } } \delta \left ( l - l ( \omega ^ { \prime } ) \right ) \,. \quad ( 1 8 ) \, \underset { \underset { \underset { \underset { \underset { \text{para} } { \text{on} } } { \text{at} } } } { \text{on} } }$$
The lower integration limit ζ > 0 is defined by kinematics, i.e. by the minimum value of the considered ω ′ , in accordance with Eq. (16). In the IPA case, the variable n = 1 2 , , · · · refers to the contribution of the individual harmonics ( n = 1 with ξ 2 /lessmuch 1 recovers the Klein-Nishina cross section, cf. [19]). The value nω is related to the energy of the background field involved in Compton scattering. Obviously, this value is a multiple of ω . In FPA, the internal quantity l is a continuous variable, implying a continuous distribution of the differential cross section over the ω ′ -θ ′ plane. The quantity lω can be considered as energy of the laser beam involved in the Compton process, which is not a multiple ω . Mindful of this fact, without loss of generality, we denote the processes with l > 1 as a multi-photon generalized Compton scattering, remembering that l is a continuous quantity.
The multi-photon effects become most clearly evident in the partially energy-integrated cross section
$$\tilde { \sigma } ( \omega ^ { \prime } ) = \underset { \omega ^ { \prime } } { \overset { \infty } { \int } } d \bar { \omega } ^ { \prime } \frac { d \sigma ( \bar { \omega } ^ { \prime } ) } { d \bar { \omega } ^ { \prime } } = \underset { l ^ { \prime } } { \overset { \infty } { \int } } d l \frac { d \sigma ( l ) } { d l } \, \quad \ ( 1 9 ) \quad \ \text{array} \text{(} 1 9 ) \quad \text{array} \text{$array} \text{$f$} \\ \text{are $f$} \\ \text{and} \, l ^ { \prime }$$
where dσ l /dl ( ) = ( dσ ω ( ′ ) /dω ′ )( dω ′ ( ) l /dl ), and the minimum value of l ′ is
$$l ^ { \prime } = \frac { \omega ^ { \prime } } { \omega } \, \frac { E + | \vec { p } | \cos \theta ^ { \prime } } { E + | \vec { p } | - \omega ^ { \prime } ( 1 - \cos \theta ^ { \prime } ) } \. \quad \ \ ( 2 0 ) \quad \ f u n c { \underset { \text{velo} } { \text{$e$}$} }.$$
The cross section (19) has the meaning of a cumulative distribution. In this case, the subthreshold, multiphoton events correspond to frequencies ω ′ of the outgoing photon which exceed the corresponding threshold value ω ′ 1 = ω ′ ( l = 1) (cf. Eq. (16)).
## III. NUMERICAL RESULTS
## A. The envelope shapes and the e.m. field structure
Some aspects of the one- and two-parameter envelope functions (2) have been considered in [21]. Here we extend this analysis. The parameter ∆ characterizes the pulse duration 2∆ with ∆ = πN , where N has a meaning of a 'number of oscillations' in the pulse. Certainly, such a definition is rather conditional and is especially meaningful for the flat-top envelope with small values of b/ ∆. In the case of the hs envelope shape, the number of oscillations with small amplitudes may exceed N . Nevertheless, for convenience we call N 'number of oscillations in a pulse' for given f ( φ ), relying on its relation with the shape parameter ∆. The parameter b in the two-parameter sF shape has the meaning of the 'thickness' or ramping time of the pulse shape. It was shown that the properties of the two-parameter sF shape for large values of b/ ∆ /similarequal 0 3 . . . . 0 5 are close to that of the . one-parameter hs shape. Therefore, as mentioned above, in order to stress the difference between one- and twoparameter (flat-top) envelopes we focus our consideration on the choice of b/ ∆ = 0 15 troughout our paper. .
The envelope shape f ( φ ) and the integrand f 2 ( φ ) + f ′ 2 ( φ ) in Eq. (15) as functions of the invariant phase for hs and sF shapes are shown in Fig. 1 in left and right panels, respectively. The numbers in the plot indicate the number N of oscillations in a pulse. The thick solid curves labeled by N are for f ( φ ). The dashed, longdashed, dot-dashed and dot-dot-dashed curves are for f 2 ( φ ) + f ′ 2 ( φ ) with N = 0 5, 2, 5 and 10, respectively. .
FIG. 1: (Color online) The envelope functions f ( φ ) and the integrand f 2 ( φ ) + f ′ 2 ( φ ) in Eq. (15) as the functions of invariant phase φ = kx . The thick solid curves labeled by N are for f ( φ ). The dashed, long-dashed, dot-dashed and dotdot-dashed curves are for f 2 ( φ ) + f ′ 2 ( φ ) with N = 0 5, 2, 5 . and 10, respectively. Left and right panels are for hs and sF envelope shapes, respectively.
For the smooth hs shape the integrand is also a smooth function (cf. Fig. 1, left panel). For the flat-top sF envelope shape and N ≥ 2 both, f ( φ ) and the integrand f 2 ( φ ) + f ′ 2 ( φ ) (the latter one being proportional to the intensity in the course of the pulse), are smooth functions of the invariant phase which more compact as compared with the hs shape with the same value of N . At N = 0 5 . and φ ∼ ∆ the integrand (see dashed red curve in the
right panel) displays some overshoot resulting locally in the height h = 1 4 + ( / ∆ b / 4∆) 2 /similarequal 1 37. . Increasing ∆ (or b/ ∆) leads to a vanishing of this overshoot.
For the hs envelope, the normalization factor in Eq. (15) has the form
$$\Gamma _ { 0 } ^ { h s } = \frac { \Delta } { \pi } \left ( 1 + \frac { 1 } { 3 \Delta ^ { 2 } } \right ) \, \quad \quad \ ( 2 1 ) \\ \Gamma _ { 1 } \Gamma _ { 2 } \Gamma _ { 3 } \Gamma _ { 4 } \Gamma _ { 5 } \Gamma _ { 6 } \Gamma _ { 7 } \Gamma _ { 8 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gammega _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 - 1 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gammu _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9} \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma_ _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9. } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamda _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 0 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamra _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 1 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 1 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Psma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 5 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamria _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 6 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamrema \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ { 9 } \Gamma _ {$$
while for the sF shape one has
$$N _ { 0 } ^ { s F } = \frac { \Delta } { \pi } \left ( F _ { 1 } \left ( t \right ) + F _ { 2 } \left ( t \right ) \frac { b } { \Delta } \right ), \ t = \frac { 1 + \cosh \frac { \Delta } { b } } { \sinh \frac { \Delta } { b } }, \ ( 2 2 ) \\.$$
where
$$\dots \dots \dots & \quad \\ F _ { 1 } ( t ) \ = \ \frac { ( t ^ { 2 } + 1 ) ( - t ^ { 4 } + 1 0 t ^ { 2 } - 1 ) } { 1 6 t } \, & \quad \begin{array} { c } \text{Th} \\ E _ { x } / a \\ \text{by so} \end{array} \\ F _ { 2 } ( t ) \ = \ \frac { 3 t ^ { 1 0 } - 3 5 t ^ { 8 } + 9 0 t ^ { 6 } - 9 0 t ^ { 4 } + 3 5 t ^ { 2 } - 3 } { 2 4 ( t ^ { 2 } - 1 ) ^ { 3 } } \. ( 2 3 ) & \quad \text{per a} \\ \text{The } 1 \\ N = \quad \\ \text{In the limit $\underline{b}$} \to 0. & \quad \dots$$
In the limit b ∆ → 0,
$$\cdot \dots \dots \dots \Delta & \cdot \cdot \cdot \cdot \\ N _ { 0 } ^ { s \text{F} } = \frac { \Delta } { \pi } + \mathcal { O } \left ( \exp [ - \frac { \Delta } { b } ] \right ) \simeq \frac { \Delta } { \pi }. \quad \quad ( 2 4 ) \quad \begin{array} { c } \frac { \frac { \partial } { s } } { \widetilde { \widetilde { \Theta } } ^ { * } } \\ \omega ^ { * } \\ \widetilde { \widetilde { \Theta } } ^ { * } - 0 \end{array} \cdot \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \ dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots\dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \delta }$$
The normalization factor N 0 scaled by N = ∆ /π as a function of N for hs and sF shapes is exhibited in Fig. 2, shown by the dashed blue and solid red curves, respectively.
FIG. 2: (Color online) The scaled normalization factor N /N 0 as a function of the number of oscillations in the pulse, N = ∆ /π , for hs and sF shapes, shown by the dashed blue and solid red curves, respectively.
For the hs shape, N 0 /similarequal N at N ≥ 1 and slightly increases for the sub-cycle envelopes with N < 1 (cf. Eq. (21)). In the case of a flat-top envelope, the ratio N /N 0 is independent of ∆, according to Eq. (23). The contribution of f ′ 2 in (15) is weak and varies from 0.2% to 3.8% for b/ ∆=001and 0.2, respectively. In the limit . b/ ∆ → 0 it vanishes and N 0 → N and, therefore, the overshoot in the integrand does not affect the integral in Eq. (15). But taking into account that very small values of b/ ∆ seems to be not realistic, we restrict our actual calculations to the finite value b/ ∆ = 0 15, where the . overshoot in f 2 ( φ ) + f ′ 2 ( φ ) is minor.
For the sake of completeness, we present also the behavior of e.m. potential /vector A and the electric field strength
/vector E = -∂ /vector A/∂t , where /vector A is given by Eqs. (1) and (2) as functions of the invariant phase φ . The e.m. potential and strength for the oneand two-parameter envelope functions read
$$A _ { x } \, = \, a \, f ( \phi ) \cos \phi \, \quad A _ { y } = a \, f ( \phi ) \sin \phi \, \ \ ( 2 5 )$$
$$E _ { x } \, & = \, \omega \, A _ { x } \left [ - ( \ln f ( \phi ) ) ^ { \prime } + \tan \phi \right ] \,, \quad \quad ( 2 6 ) \\ E _ { \dots } \, & = \, \omega, A _ { \dots } \left [ - \ln f ( \phi ) \vee \omega ^ { \prime } - \cot \theta \right ] \, \quad \quad ( 2 7 )$$
$$E _ { y } \, = \, \omega \, A _ { y } \, [ - ( \ln f ( \phi ) ) ^ { \prime } - \cot \phi ] \, \quad \quad ( 2 7 )$$
with
$$\text{win} \\ - \left ( \ln f ( \phi ) \right ) ^ { \prime } = \begin{cases} \frac { 1 } { \Delta } \tanh \frac { \phi } { \Delta }, & \text{hs}, \\ \frac { 1 } { b } \frac { \sinh \frac { \phi } { b } } { \cosh \frac { \Delta } { b } + \cosh \frac { \phi } { b } }, & \text{SF}. \end{cases}$$
The scaled potentials A /a x and the scaled strengths E /aω x as functions of the invariant phase are exhibited by solid red and dashed blue curves, respectively, in upper and middle panels in Fig. 3 for the hs and sF shapes. The left and right panels correspond to the pulses with N = 2 and 0.5, respectively. The result for the hs shape
FIG. 3: (Color online) The e.m. potentials A/a (solid red curves) and field strengths E/aω (dashed blue curves) as functions of the invariant phase φ . The upper and middle panels correspond to the hyperbolic secant (hs) and symmetrized Fermi (sF) shapes, respectively, for the x components. The lower panels correspond to the y components for the sF envelope shape. The left and right panels are for pulses with N = 2 and N = 0 5, respectively. .
with N = 0 5 is close qualitatively to . that of [8]. One can see that the duration of the pulse increases with increasing number of oscillations. The flat-top sF shape is more compact compared to the hs shape with the same value of the pulse 'scale' parameter ∆.
The result for y components is exhibited in Fig. 3, lower panels, where we restrict ourselves to the example
of the flat-top sF envelope shape. For short pulses with N > 2, the contribution of the first terms in Eqs. (26) and (27) are relatively small and, therefore, the approximate relations A y /similarequal E /ω x and E y /similarequal -ωA x are valid. Both A y and E y are finite. The same is valid for sF shape with N ≥ 2. This is illustrated in Fig. 3, lower panel (left), where the result for the sF envelope with N = 2 is shown. The approximate relations are valid also for sub-cycle pulse with N = 0 5 and the one-parameter hs shape. . In the case of the flat-top envelope for N = 0 5, the above . approximate relations are valid for A y and for the central part of E y (cf. Fig. 3 lower panel (right)). In the border area with φ ≈ ∆ = π 2 , the strength E y has finite narrow peaks with height ˜ = h ( ∆ b ) sin ∆ 4∆ + O (exp( -∆ /b )) /similarequal 1 06. . The height of these peaks decreases with increasing ∆ at fixed b/ ∆ and for N ≥ 2 it becomes negligibly small. This 'pick-like' behavior for the flat-top shape can be compared with the the popular rectangular pulse [26] where the derivative f ′ ( φ ) = θ ′ ( φ -∆) = δ φ ( -∆) is singular at φ = ∆. But such 'pick-like' or even singular behavior of /vector E at the border does not affect the transition matrix M l ( ) in Eq. (4), because it is determined by A and A 2 rather than the e.m. strength.
## B. The differential cross sections
In IPA [19, 20], the cross section of the multi-photon Compton scattering increases with θ ′ towards 180 . o For instance, it peaks at about 170 o for the chosen electron energy of 4 MeV (all quantities are considered in the laboratory frame) and rapidly drops to zero when θ ′ approaches 180 o for the harmonics n > 1 yielding thus the blind spot for back-scattering. Therefore, in our subsequent analysis we choose the near-backward photon production at θ ′ = 170 o and an optical laser with ω = 1 55 eV. Defining one-photon events by . n = 1, this kinematics leads via Eq. (17) to ω ′ 1 ≡ ω ′ ( n = 1 , ξ 2 /lessmuch 1 , θ ′ = 170 ) o /similarequal 0 133 keV which we refer to as a thresh-. old value. Accordingly, ω > ω ′ ′ 1 is enabled by non-linear effects, which in turn may be related loosely to multiphoton dynamics with n > 1 in IPA or l > 1 in FPA where, we remind again, the internal variable l can not be interpreted strictly as number of laser photons involved (cf. [36]). Note that all calculations for IPA are performed in a standard way [19, 20]. The energy of the outgoing photon in IPA is calculated using Eq. (17), where dressing of electrons in the background field is taken into account.
Let us consider first an example of short pulses with moderate intensity, ξ 2 = 10 -3 , similar to a recent experiment of Compton backscattering [37]. Results for the hs and sF shapes are exhibited in Fig. 4. The red curve (marked by boxes) and the blue curve (marked by diamonds) correspond to pulses with N = 2 and 5, respectively. The black stars depict the IPA results, i.e., the harmonics at fixed scattering angle θ ′ . Their positions correspond to integer values of n = 1 , 2 , · · · in accordance with Eq. (17). i.e. the distribution of scattered photon energies is a discrete function of ω ′ . We stress that the cross section at ω ′ > ω ′ 1 is essentially 'sub-threshold', i.e. outside the kinematically allowed region of the KleinNishina process due to multi-photon effects.
FIG. 4: (Color online) Differential cross section dσ/dω ′ | θ ′ =170 o of Compton scattering for ξ 2 = 10 -3 . The red curves (marked by boxes) and blue curves (marked by diamonds) are for N = 2 and 5, respectively. The black stars depict the IPA results for lowest harmonics. Top and bottom panels correspond to hyperbolic secant (hs) and symmetrized Fermi (sF) shapes of the envelopes, respectively.
In the FPA case, the energy distribution becomes a continues function of ω ′ . The actual shape is determined by both the pulse duration and the envelope form. Consider first the case of the hs pulse (cf. Fig. 4, top panel). The cross section displays sharp bumps with peak positions corresponding to integer values of l = n (as in IPA). In the vicinity of the bumps, at l = n ± /epsilon1 , /epsilon1 /lessmuch 1, the cross section is rapidly decreasing. Such a behavior reflects the properties of the functions Y l ( z ) [21] which behave under such conditions as
$$Y _ { n + \epsilon } ( z ) \, \simeq \, \frac { z ^ { n } } { 2 ^ { n } n! } \, \text{e} ^ { - i \epsilon \phi _ { 0 } } F ^ { ( n + 1 ) } ( \epsilon ) \,, \quad ( 2 9 )$$
where F ( n ) ( /epsilon1 ) is the Fourier transform of the function
f n ( φ ). At ξ 2 /lessmuch 1, the contribution of terms ∝ X l is negligible. The behavior of the cross section in the vicinity of the first bump is proportional to F 2 hs ( /epsilon1 ) with
$$F _ { \text{hs} } ( x ) \ = \ \frac { \Delta } { 2 \cosh \frac { 1 } { 2 } \pi \Delta | x | } \, \quad \ \ ( 3 0 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\quad \quad \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bullet \quad \bul. \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
or F hs ( x ) /similarequal ∆exp[ -π ∆ x/ 2]. Thus, the cross section becomes steeper with increasing pulse duration ∆. This result qualitatively agrees with that of Ref. [31].
In the case of the sF shape, the dependence F sF ( /epsilon1 ) is more complicated:
$$\text{ore compiacate} \colon \\ F _ { s F } ( x ) \, = \, \frac { 1 + \exp \left [ - \frac { \Delta } { b } \right ] } { 1 - \exp \left [ - \frac { \Delta } { b } \right ] } \, \frac { b \sin \Delta x } { \sinh \pi b \, x } \,. \quad ( 3 1 ) \\ \text{gether with the overall decrease of the cross section} \quad \text{s}$$
Together with the overall decrease of the cross section proportional to exp[ -2 πb l ( ω ′ )] it also indicates fast oscillations with a frequency ∝ ∆. Such oscillations show up in the cross section as some secondary bumpy structures. These properties are manifest in Fig. 4 (bottom panel): the overall decrease of the cross section decreases with decreasing pulse duration, and the number of the secondary bumps in the region of ω ′ , corresponding to the nearest integer values of l , increases with pulse duration.
In Fig. 5 we present the differential cross sections for different field intensities ξ 2 = 0 01 . , 0 1 and 1, depicted . by red (marked by boxes), blue (marked by diamonds), and green (marked by circles) curves, respectively. The duration of the pulse corresponds to N = 2. The bump positions for FPA in Fig. 11 are shifted relative to the discrete positions of contributions from the individual harmonics in IPA, shown by corresponding symbols. These shifts are a consequence of the electron dressing in IPA which depends on ξ 2 .
For completeness, in Fig. 6 we exhibit the differential cross sections for a sub-cycle pulse with N = 0 5 . for ξ 2 = 0 001 and 1, shown by red curves (marked by . boxes) and green curves (marked by circles), respectively, for the hs (top) and sF (bottom) envelope shapes, respectively. Crosses and pluses depict the IPA results for ξ 2 = 0 001, and 1. . For the hs shape, the cross sections decrease almost monotonically, with a large enhancement of the FPA result compared to IPA for small field intensities ( ξ 2 /lessmuch 1). In case of the flat-top envelope the cross section exhibits some oscillations which point to more complicated spectral properties of the flat-top envelope shape.
To summarize this part we can conclude that predictions for fully differential cross sections for IPA and FPA are quite different. In IPA, the cross section represents the discrete spectrum where the frequencies of the outgoing photons ω ′ are fixed according to Eq. (17). The fully differential cross sections are continuous functions of ω ′ . Some similarities of IPA and FPA can be seen in the case of small field intensities ξ 2 /lessmuch 1 and the smooth oneparameter envelope shape with N = 2 . . . 10. Here, the differential cross sections have a bump structure, where
FIG. 5: (Color online) Differential cross section dσ/dω ′ | θ ′ =170 o of Compton scattering for ξ 2 = 0 01 . , 0 1 . and 1, shown by red (marked by boxes), blue (marked by diamonds), and green (marked by circles)curves, respectively, for N = 2. The symbols 'x', stars and pluses depict the IPA results for the lowest harmonics for ξ 2 = 0 01 . , 0 1 . , and 1, respectively. Top and bottom panels correspond to hyperbolic secant (hs) and symmetrized Fermi (sF) shapes of the envelopes.
the position of bumps and bump heights are close to that predicted by IPA. The situation changes drastically for more complicated (and probably more realistic) flat-top envelope shapes. In this case one can see a lot of additional bumps which reflect the more complicated spectral properties of the flat-top shape; it is difficult to find a relation not only between IPA and FPA, but also within FPA for different pulse durations. Experimentally studying multi-photon effects using rapidly oscillating fully differential cross sections seems to be rather complicated. An analysis of integral observables helps to overcome this problem. In particular, the partly integrated cross sections have a distinct advantage: they are smooth functions of ω ′ and allow to study directly the multi-photon dynamics.
FIG. 6: (Color online) Differential cross section dσ/dω ′ | θ ′ =170 o of Compton scattering for ξ 2 = 10 -3 and 1 shown by red (with boxes) and green (with circles) curves, respectively, for N = 0 5. . Crosses and pluses depict the discrete IPA results for lowest harmonics for ξ 2 = 10 -3 and 1, respectively. Top and bottom panels correspond to hyperbolic secant (hs) and symmetrized Fermi (sF) envelope shapes.
## C. Partly integrated cross sections
Non-linear effects become most transparent in the partially energy-integrated cross section defined in Eq. (19). In this case, the sub-threshold multi-photon events are filtered when the lower limit of integration ω ′ exceeds the threshold value ω ′ 1 = ω ′ ( n = 1 , ξ 2 ) (with ξ 2 /lessmuch 1 for the pure Klein-Nishina process). Thus, events with ω ′ ( ) l /greatermuch ω ′ 1 and l /greatermuch 1 correspond essentially to multiphoton process, where the energy lω /greatermuch ω is absorbed from the pulse. Experimentally, this can be realized by an absorptive medium which is transparent for frequencies above a certain threshold ω ′ . Otherwise, such a partially integrated spectrum can be synthesized from a completely measured spectrum. Admittedly, the considered range of energies with a spectral distribution uncovering many decades is experimentally challenging.
The partially integrated cross sections of Eq. (19) are
FIG. 7: (Color online) The integrated cross section (19) for ξ 2 = 10 -3 . The thin solid curve marked by dots depicts the IPA result. The red (with pluses), blue (with boxes) and green (with crosses) curves correspond to N = 2, 5 and 10, respectively. Top and bottom panels are for hyperbolic secant (hs) and symmetrized Fermi (sF) envelopes.
presented in Fig. 7. The thin solid black curve (marked by dots) depicts IPA results given by
$$\tilde { \sigma } ^ { I P A } ( \omega ^ { \prime } ) = \int _ { l ^ { \prime } ( \omega ^ { \prime } ) } ^ { \infty } \, d l \sum _ { n = 1 } ^ { \infty } \frac { d \sigma _ { n } ^ { I P A } } { d \omega _ { n } ^ { \prime } } \frac { d \omega _ { n } ^ { \prime } } { d n } \theta ( n - l ) \, \quad ( 3 2 )$$
where ω ′ ( n ) is defined by Eq. (17). That is, the partially integrated cross section becomes a step-like function, where each new step corresponds to the contribution of a new (higher) harmonic n , which can be interpreted as n -laser photon process. Results for the finite pulse exhibited by red (marked by pluses), blue (marked by boxes) and green (marked by crosses) curves correspond to N = 2 , 5 and 10, respectively. In the above-threshold region with ω ′ ≤ ω ′ 1 , the cross sections do not depend on the widths and shapes of the envelopes, and the results of IPA and FPA coincide. The situation changes drastically in the deep sub-threshold region, where ω ′ > ω ′ 1 ( l /greatermuch 1) , n /greatermuch 1. For the short pulses with N /similarequal 2, the FPA results exceed that of IPA considerably, and
the excess may reach several orders of magnitude, especially for the flat-top envelope shown by the red curve in Fig. 7 (bottom panel). However, when the number of oscillation in a pulse increases ( N /greaterorsimilar 10) there is a qualitative convergence of FPA and IPA results, independently of the pulse shape. Thus, at N = 10 and ω ′ = 0 6 keV . the difference between predictions for hs and sF shapes is a factor of two, as compared with the difference of the few orders of magnitude at N = 2 for the same value of ω ′ .
To highlight the difference of the hs and sF (flat-top) shapes for short pulse we exhibit in Fig. 8 (top panel) results for N = 2. At l /greaterorsimilar 5, corresponding to ω ′ /greaterorsimilar 0 7 keV, the difference between them is more than two . orders of magnitude.
~
~
FIG. 8: (Color online) The partially integrated cross section (19) for ξ 2 = 10 -3 . Top panel: N = 2, for the hyperbolic secant (hs, dashed blue curve) and symmetrized Fermi (sF, solid red curve) shapes. Bottom panel: The same as in top panel, but for a sub-cycle pulse with N = 0 5. . The crosses and pluses correspond to the asymptotic solutions for hs and sF shapes, respectively, described in the text.
Consider now the case of sub-cycle pulses with N < 1. Our result for N = 0 5 is exhibited . in Fig. 8 (bottom panel). One can see a large enhancement of the cross section with respect to the IPA case for the sub-cycle pulse in the sub-threshold region. The enhancement for the sF
shape is much greater pointing to a sensible dependence on the actual pulse shape. For a qualitative estimate of such a behavior we can drop the φ e ′ dependence by taking φ e ′ = 0. This choice is quite reasonable for the flat-top sF envelope shape and may serve as an upper limit for the cross sections in the case of the smooth hs envelope shape (cf. Sect. III.D). Under the considered conditions the basic function Y l in Eq. (29) can be approximated as
$$\text{the basic function } Y _ { l } \text{ in } E _ { \L }. \text{(29)} \text{ can be approximated as} \\ Y _ { l } \ & \simeq \ \frac { 1 } { 2 \pi } \int d q \, F ( q ) \int d \phi e ^ { i ( l - q ) \phi - i \mathcal { P } ( \phi ) } \\ & \simeq \ \frac { 1 } { 2 \pi } \int d q \, F ( q ) \int d \phi e ^ { i ( l - q - l \beta \xi ) \phi - i \delta } \\ & = \ e ^ { - i \delta } F ( \tilde { l } ) \, \\ \dots \quad \ \quad.$$
where F l ( ) is the Fourier transform of the envelope function, ˜ = l l (1 -βξ ) with β = 2 √ u u l (1 -u u l ) < 1 and δ = z ∫ 0 -∞ dφ cos φf φ ( ). As a result, the cross section is almost completely defined by the square of the Fourier transforms (cf. Eqs. (30) and (31)), i.e. ˜( σ ω ′ ) /similarequal g l ( ( ω ′ )) F 2 ( ˜ l ( ω ′ ) -1), where g ω ( ′ ) is a smooth function of l = ( l ω ′ ) (cf. Eq. (39)). The Fourier transform for the sF shape decreases slower with increasing l . Such a dependence is evident in Fig. 7 (bottom panel). For an illustration, the crosses depict the result of a calculation where the basic functions Y l and X l in the partial probability ω ′ ( ) in Eq. (9) are replaced by their asymptotic l values F (1) ( ˜ l -1) and F (2) ( ˜ l -1). A more detail discussion of the asymptotic result is presented below (cf. Eq. (39)).
The dependence of the partially integrated cross section as a function of ξ 2 at fixed ratio r ≡ ω /ω ′ ′ 1 = 3 for short pulses with N = 0 5 and 2 is exhibited in Fig. 9 . in top and bottom panels, respectively. Note that the minimum value of l ′ ( ω ′ ) is related to r as
$$l ^ { \prime } ( \omega ^ { \prime } ) \ = \ r \frac { E + | \vec { p } | \cos \theta ^ { \prime } } { E + | \vec { p } | \cos \theta ^ { \prime } + \omega ( 1 - r ) ( 1 - \cos \theta ^ { \prime } ) } \, ( 3 4 ) \\. \quad. \quad. \quad. \quad. \quad. \quad.$$
meaning l ′ < r . Similarly, for n min one has n min = x , for I ( x )=x and n min = x +1 for I ( x ) < x with
$$x = \frac { E + | \vec { p } | \cos \theta ^ { \prime } + \frac { \omega m ^ { 2 } \xi ^ { 2 } } { 2 ( k \cdot p ) } ( 1 - \cos \theta ^ { \prime } ) } { E + | \vec { p } | \cos \theta ^ { \prime } + \omega ( 1 - r + \frac { m ^ { 2 } \xi ^ { 2 } } { 2 ( k \cdot p ) } ) ( 1 - \cos \theta ^ { \prime } ) }. \ ( 3 5 )$$
The solid curves and symbols correspond to IPA and FPA, respectively, with different pulse shapes. One can see that the main difference of IPA and FPA, as well as the pulse shape dependence, appears at small field intensities ξ 2 /lessmuch 1, where the dependence of the cross section on the pulse shape and duration is essential.
To explain this result we use the asymptotic solution for ˜ which is obtained by keeping leading terms in σ ξ 2 in Eqs. (9) and (10) and taking into account that the dominant contribution to the integrals of Eqs. (19) and (32) stems from l ∼ l ′ and n ∼ I ( l ′ )+1, respectively. Consider
~
~
FIG. 9: (Color online) The partially integrated cross section as a function of ξ 2 at ω /ω ′ 1 = 3 for short pulses with N = 0 5 . (top) and 2 (bottom). The solid curve and symbols correspond to IPA and FPA (hs and sF envelope functions), respectively.
first the partially integrated cross section in IPA. Using the asymptotical expression for the Bessel functions
$$\dots _ { J } \dots _ { R } \dots \dots \dots \dots \dots \dots \dots \dots \\ J _ { k } ( z ) \simeq \left ( \frac { z } { 2 } \right ) ^ { k } \frac { 1 } { k! } \quad \text{for $z\ll1$} \,, \quad \text{(36)} \\ \text{keeping the leading terms in $Eq.\(10)$ with $J_{n-1}^{2}(z)$} \quad \text{where}$$
and keeping the leading terms in Eq. (10) with J 2 n -1 ( z ) and n = I ( l ′ ) + 1, one obtains
$$\tilde { \sigma } ^ { I P A } & \simeq \frac { 2 \pi \alpha ^ { 2 } } { ( E + | \vec { p } | \cos \theta ^ { \prime } ) | \vec { p } | } \xi ^ { 2 k } \Phi ( k ) \, \quad \ ( 3 7 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\quad \text{k-when } k = \tau / \tau \vee \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee\ \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \ve. \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vet \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vevee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau /\vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau / \vee \ \tau /$$
where k = I ( l ′ ) /similarequal I ( r ) and
$$\Phi ( k ) \ = \ \frac { ( k + 1 ) ^ { 2 ( k + 1 ) } } { ( k + 1 )! ^ { 2 } } \left ( t _ { k } ( 1 - t _ { k } ) \right ) ^ { 2 k } \quad \text{$\text{$solut$} } { \text{thro} } \\ \times \ \left ( 1 + \frac { u } { 2 ( 1 + u ) } - 2 t _ { k } ( 1 - t _ { k } ) \right ) \quad ( 3 8 ) \quad \text{$tran:} } \ \text{mon} \\ \text{with} \, t _ { k } = u / u _ { k }. \ \text{where} \, u = \omega ^ { \prime } ( 1 - \cos \theta ^ { \prime } ) / ( E + | \vec { \beta } | - \omega ^ { \prime } ( 1 - \quad \text{less.} }$$
with t k = u/u k , where u = ω ′ (1 -cos θ ′ ) / E ( + | /vector p |-ω ′ (1 -cos θ ′ )) and u k = 2 kω E ( + | /vector p | ) /m 2 . Within the consid- ered kinematics, t k does not depend on k and can be approximated by t k /similarequal m 2 (1 -cos θ ′ ) / (2( E + | p | cos θ ′ )( E + | p | )) /similarequal 0 35. .
~
FIG. 10: (Color online) The partially integrated cross sections as a function of ξ 2 /lessmuch 1 for sub-cycle pulse with N = 0 5. . The stars are for the full IPA result. The black curves correspond to the asymptotic solution of Eq. (37). The red and blue thin curves marked by pluses and crosses are full calculations for sF and hs shapes, respectively, while the thick red and blue curves are the corresponding asymptotic results of Eq. (39).
The result for asymptotic solution of (37) is shown by the solid black curve in Fig. 10 together with a full calculation depicted by stars. One can see an excellent agreement of these two results.
For FPA in case of sub-cycle pulse with N = 0 5, we . use the asymptotic representation for the basic functions Y l in the form of Eq. (33) which allows to express the partially integrated cross section as
$$\Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ {\Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } \Phi _ { \Phi } ^ { \dagger } ( \Phi _ { \Phi } ^ { \dagger } ) ) \int _ { l ^ { \prime } } ^ { l ^ { \prime } + 1 } d l \, F ^ { 2 } ( \tilde { l } - 1 ) \, \\ \left ( 1 + \frac { u } { 2 ( 1 + u ) } - 2 t _ { l ^ { \prime } } ( 1 - t _ { l ^ { \prime } } ( 1 - t _ { l ^ { \prime } } ) \right ) \, \int _ { l ^ { \prime } } ^ { l ^ { \prime } + 1 } d l \, F ^ { 2 } ( \tilde { l } - 1 ) \, \\ \cdot \quad \Phi _ { \Phi } ^ { \dagger } \, \left ( 1 + \Phi _ { \Phi } ^ { \dagger } \, \left ( 1 + \Phi _ { \Phi } ^ { \dagger } \right ) \,,$$
where F x ( ) is the Fourier transform of the envelope function (cf. Eqs. (30) and (31)). Results for the sub-cycle pulse with N = 0 5 are presented in Fig. 10, where the . red and blue thin curves marked by pluses and crosses are for full calculations for the sF and hs shapes, respectively. The red and blue thick curves are the asymptotic results of Eq. (39) for sF and hs shapes, respectively.
We would like to note that, at ξ 2 /lessmuch 1, our asymptotic solution for sub-cycle pulse weakly depends on ξ only through the weak l (1 -βξ ) dependence in the Fourier transform. The leading ξ 2 dependence of the partial harmonics w l in (9) is compensated by the ξ 2 dependence of the flux factor in the denominator of Eq. (8). Nevertheless, such a weak ξ dependence is in qualitative agreement between full and asymptotic solutions both for sF and hs
~
FIG. 11: (Color online) Results for a short pulse with N = 2. Top panel: The integrand dσ/dl of the partially integrated cross sections for ξ 2 = 2 × 10 -4 . The thin black (with circles) blue (with crosses) and red (with pluses) curves are for the full calculations. The asymptotic predictions (described in the text) are exhibited by the thick black, blue and red curves, respectively. Bottom panel: The black, red and blue thin curves marked by stars, pluses and crosses are full calculation, for IPA and FPA for sF and hs shapes, respectively; the thick black, red and blue curves are the corresponding asymptotic results of Eqs. (37), (39) and (41), respectively.
envelope shapes. Thus, we can conclude that the partially integrated cross section for the sub-cycle pulse at ξ 2 /lessmuch 1 is almost completely determined by the square of the Fourier transform of the envelope function which is a measure of high momentum frequencies generated by the pulse shape.
In the case of a short pulse with N = 2 and ξ 2 = 2 × 10 -4 , the integrand of ˜ for the hs shape for the full calcuσ lation is shown by the blue thin curve marked by crosses in Fig. 11, top panel. The integrand has a bump-like structure with the number of bumps equal to the number of partial harmonics in IPA, similarly to the differential cross section shown in Fig. 4, top panel. For the asymptotic solution, we use in this case the asymptotic expression for the basic functions of Eq. (29). Note that such an expression is valid only for the smooth one-parameter envelope shapes, where the function P ( φ ) defined in Eq. (7) takes a simple form P ( φ ) = z sin( φ -φ 0 ) f ( φ )+ O ( ξ 2 ) [21]. One can see that, if the argument obeys l ′ > I l ( ′ ), then the main contribution to the cross section comes from the two harmonics with
$$Y _ { k, \varepsilon _ { 1 } } ( z ) \text{ and } Y _ { k + 1, \varepsilon _ { 2 } } ( z ) \,,$$
where k = I ( l ′ ), ε 1 = l ′ -I ( l ′ ) ≡ ε > 0, and ε 2 = ε -1 < 0. Then, keeping the leading terms in ξ 2 in (9) one can get the approximate expression for the partly integrated cross section in the form
$$\cosh \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theelta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \thetheta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin \theta \sin
\quad \tilde { \sigma } \sim \frac { 2 \pi \alpha ^ { 2 } } { N _ { 0 } ( E + | \vec { p } | \cos \theta ^ { \prime } ) | \vec { p } | } \xi ^ { 2 ( k - 1 ) } \\ \times \, \left ( \Phi ( k - 1 ) \int d \epsilon ( F ^ { ( k ) } ( \epsilon ) ) ^ { 2 } \\ \frac { \epsilon } { \epsilon } + \xi ^ { 2 } \Phi ( k ) \int d \epsilon ( F ^ { ( k + 1 ) } ( \epsilon ) ) ^ { 2 } \right ), \ ( 4 1 ) \\ \frac { \epsilon } { \epsilon } - 1 \, \end{pmatrix}$$
where F ( m ) is the Fourier transform of m -th power of the envelope function f ( φ ). The integrand of the corresponding cross section is obtained from Eq. (41) by removing the integration and putting ε 1 and ε 2 as the arguments of the squares of the first and second Fourier transforms, respectively. The approximate integrand is shown in Fig. 11, top panel, by the thick blue curve. One can see a reasonable agreement of the approximate and full calculations. The asymptotic expression of the integrand for IPA is equal dσ n /dl = σ n θ n ( -l ) with σ n determined by Eq. (37).
The full and approximate results for ˜ are shown in σ Fig. (11), bottom panel, by crosses and the thick blue curve, respectively. One can see a fairly good agreement of approximate and full results up to ξ 2 = 0 1. .
In the case of the flat-top envelope, the integrand of ˜ has a more complicated structure with a large numσ ber of bumps as shown in Fig. 11, top panel, by the thin red curve marked by pluses. The asymptotic solution for the basic functions of Eq. (29) does not apply here. However, as a first approximation one can use the asymptotic solution of Eq. (33). Then, the cross section ˜ is σ determined by Eq. (39). The asymptotic integrand is defined in this case by Eq. (39) by skipping the integration and putting l = ˆ . l ′ The corresponding integrand is presented in Fig. 11, top panel by thick red line. On can see a satisfactory agreement of full and approximate solutions and, as a consequence, a reasonable agreement of full and approximate calculations of partially integrated cross sections, however, in a very limited range of ξ 2 /lessmuch 1 as shown in Fig. 11, bottom panel, by pluses and the thick red line, respectively.
To summarize this part we note that, in case of short pulses and small field intensities, the partly integrated cross section is determined by the interplay of pulse shape and multi-photon dynamics. For both considered shapes, the cross sections are described by the simple asymptotic expressions which can be used in practical research.
At large values ξ 2 /greatermuch 1, our analysis shows that the dependence on the envelope shape disappears because,
similar to the Breit-Wheeler process [21], only the central part of the envelope becomes important. Formally, under a change of the variable l → l eff = + l m ξ u/ 2 2 2( k · p ), the basic functions Y l ( z ) with l /greatermuch 1 , z /greatermuch 1 become similar to the asymptotic form of the Bessel functions J l ( z ) and, as a consequence, one can get the total production probability (or the total cross section) in the IPA form [19].
## D. Anisotropy
Let us discuss now the φ e ′ (azimuthal angle) dependence of the outgoing electron for sub-cycle pulse. The corresponding cross sections are defined by Eqs. (8) and (18) with fixed azimuthal angle φ e ′ . The anisotropy is manifest most clearly in case of the sub-cycle pulse with finite field intensity, and it is very sensitive to the pulse shape. Thus, for sub-cycle and a smooth one-parameter envelope shape (e.g., for the hs pulse shape) and finite field intensity, the direction of flight of the outgoing electron (photon) is correlated with the coordinate frame of the e.m. field. The production probability would have a maximum if the azimuthal angle of the outgoing electron φ e ′ = φ 0 is equal to zero, or the azimuthal direction of the electron momentum coincides with /vector a x . The explanation of this effect is as follows. The most important dependence of the basic functions Y l ( z ) in (7), which determine the production probability w l ( ) in (9), on azimuthal angle φ e ′ is
$$Y _ { l } ( z ) \in \bigotimes _ { - \infty } ^ { \infty } \, d \phi \, f ( \phi ) \, \text{e} ^ { \quad i l \phi - i z \bigotimes _ { - \infty } ^ { \phi } d \phi ^ { \prime } f ( \phi ^ { \prime } ) \cos ( \phi ^ { \prime } - \phi _ { e ^ { \prime } } ) } \,, \ ( 4 2 )$$
and similarly for X z l ( ) with the substitution f → f 2 in the integrand. The argument in the highly oscillating exponent is
$$\text{exponent is} & & \text{FIG.} \\ R ( \phi ) & = \, i \left [ l \phi - z \cos \phi _ { e ^ { \prime } } \int \, d \phi ^ { \prime } f ( \phi ^ { \prime } ) \cos \phi ^ { \prime } & \text{$N =} \\ - & \text{$z\sin \phi_{e ^ { \prime } } \int \, d \phi ^ { \prime } f ( \phi ^ { \prime } ) \sin \phi ^ { \prime } \right ]. \quad ( 43 ) \\ - & \text{$z\sin \phi_{e ^ { \prime } } \int \, d \phi ^ { \prime } f ( \phi ^ { \prime } ) \sin \phi ^ { \prime } \right ]. \quad ( 43 ) \\ \text{To} \\ \text{The contributions of $Y^{2}_{l}$ and $X^{2}_{l}$ to the probability would} \\ \text{take a maximum when $R$ is minimal.} \, At finite values}$$
The contributions of Y 2 l and X 2 l to the probability would take a maximum when R is minimal. At finite values of l and z , the dominant contribution comes from small φ ′ ∼ 1 /l . Moreover, utilizing the one-parameter subcycle fast-decreasing envelope shape leads to the inequality
$$\underset { - \infty } { \overset { \phi } { \int } } d \phi ^ { \prime } f ( \phi ^ { \prime } ) \cos \phi ^ { \prime } \gg \underset { - \infty } { \overset { \phi } { \int } } d \phi ^ { \prime } f ( \phi ^ { \prime } ) \sin \phi ^ { \prime } \, \quad ( 4 4 ) \quad \text{the s} \\ \text{tativ} \\ \text{where}$$
and therefore the second integral in (43) can be neglected. Then obviously, R would have a minimum at maximum value of cos φ e ′ , i.e. at φ e ′ = 0. This argument does not apply to the sF envelope, where the first and the second integrals in (43) are of the same order of magnitude, and the dependence R on φ e ′ becomes very weak. These properties of the partially integrated cross sections are illustrated in Fig. 12, top panel, where results for the hs and sF envelope shapes are shown by the dashed blue and solid red curves, respectively.
e'
FIG. 12: (Color online) Top panel: Integrated cross sections as a function of azimuthal angle of the outgoing electron for N = 0 5, . ξ 2 = 0 1 and . ω /ω ′ ′ 1 = 3. The dashed blue and solid red curves correspond to the hyperbolic secant (hs) symmetrized Fermi (sF), respectively. Bottom panel: The same as in the top panel, but for the electron anisotropy defined in Eq. (45).
To quantify the anisotropy we define
$$\mathcal { A } ( \phi _ { e ^ { \prime } } ) = \frac { \tilde { \sigma } ( \omega ^ { \prime }, \phi _ { e ^ { \prime } } ) - \tilde { \sigma } ( \omega ^ { \prime }, \pi + \phi _ { e ^ { \prime } } ) } { \tilde { \sigma } ( \omega ^ { \prime }, \phi _ { e ^ { \prime } } ) + \tilde { \sigma } ( \omega ^ { \prime }, \pi + \phi _ { e ^ { \prime } } ) } \. \quad ( 4 5 )$$
In Fig. 12 (bottom panel), the anisotropy is shown as a function of the azimuthal angle φ e ′ in the multi-photon region for ω /ω ′ ′ 1 = 3 and for ξ 2 = 0 1. One can see a dis-. tinct anisotropy for the hs shape and a negligible one for the sF flat-top shape. Our result for the hs shape qualitatively coincides with the result of a recent paper [38], where an analog analysis is done for the one-parameter envelope shape f ( φ ) = cos ( 2 φ/ N 2 ) θ (2 N -| | φ ) which is qualitatively similar to the hyperbolic secant in Eq. (2)
for N ≥ 1. The situation changes for the sF envelope, where the anisotropy is very small for the sub-cycle pulse with N /similarequal 0 5, as shown in Fig. 12, bottom panel. . Therefore, one can conclude that the anisotropy strongly depends on the envelope shape.
## IV. SUMMARY
In summary we have considered the generalized nonlinear (multi-photon) effects in Compton scattering of short and ultra-short (sub-cycle) laser pulses for different pulse shapes. Such pulses might be produced in future facilities. In particular, we have shown that the fully differential cross sections as a function of the frequency of the outgoing photon at fixed production angle are rapidly oscillating functions for short pulses with the duration determined by the number of oscillations N = 2 · · · 10, especially for the flat-top envelope shapes. An experimental study of multi-photon effects in case of rapidly oscillating cross sections seems to be rather challenging. To overcome the problem of such a staggering we suggest to utilize the partly integrated cross section which seems to be a powerful tool for studying the non-linear (multiphoton) dynamics in the sub-threshold region. We find that these cross sections at selected pulse properties (field intensity, pulse duration) are very sensitive to the pulse shape. In the case of small e.m. field intensities, the cross sections may be enhanced by several orders of magnitude as compared to an infinitely long pulse. Such an enhancement is more important for flat-top envelope shapes which generate intensive high-frequency harmonics and play a role of a power amplifier. In the above-threshold region, the partly integrated cross sections manifest some
- [1] G. A. Mourou, T. Tajima, and S. V. Bulanov, Rev. Mod. Phys. 78 , 309 (2006).
- [2] A. Di Piazza, C. M¨ uller, K. Z. Hatsagortsyan, and C. H. Keitel, Rev. Mod. Phys. 84 , 1177 (2012).
- [3] V. Yanovsky et al. , Opt. Express 16 , 2109 (2008).
- [4] www.clf.rl.ac.uk/Facilities/ Astra/Astra+Gemini/12258.aspx
.
'universality', i.e. the independence of the pulse shape structure, where results for FPA and IPA are close to each other. Note that such a 'universality' does not appear in fully differential cross sections, where one can find rapidly oscillating cross sections as a function of ω ′ , especially for the flat-top envelope shape (cf. Figs. 2 and 3). A smooth one-parameter envelope shape leads to a non-trivial anisotropy of the outgoing electrons (photons) for very short pulses. At high field intensity, the central part of envelopes becomes dominant and the integrated cross sections coincide with the results for infinitely long pulses. It provides a rationale for the use of simple analytical expressions of IPA for Monte Carlo transport approaches.
In this work, we consider processes with circularly polarized photon beams. We expect that qualitatively, in the case of a linearly polarized pulse, our main result, i.e. the sensitivity of partially integrated cross sections to the sub-threshold multi-photon interactions and to the pulse structure, would be similar. The main difference is expected for the anisotropy since the momentum of the outgoing electron will be correlated with the direction of pulse polarization. Technically, the case of linear polarization is more complicated because it needs an additional basic function Y 2 l , which is an analog of the A 2 function in IPA (cf. [17]). Therefore we restrict our consideration to the simple and clear example of circular polarization and will analyze the case of linear polarization in follow-up work.
## Acknowledgments
The authors acknowledge fruitful discussions with R. Sauerbrey, T. E. Cowan and D. Seipt.
- [11] N. V. Elkina, A. M. Fedotov, I. Y. Kostyukov, M. V. Legkov, N. B. Narozhny, E. N. Nerush, and H. Ruhl, Phys. Rev. ST Accel. Beams 14 , 054401 (2011).
- [12] A. Ilderton, P. Johansson, and M. Marklund, Phys. Rev. A 84 , 032119 (2011).
- [13] H. Hu, C. M¨ller, and C. H. Keitel, Phys. Rev. Lett. u 105 , 080401 (2010) A. Ilderton, Phys. Rev. Lett. 106 , 020404 (2011).
- [5] \protect\vrule width0pt\protect\href{http://www.eli-beams.eu/}{http://www.eli-beams.eu/} [14] J. G. Kirk, A. R. Bell, and I. Arka, arXiv:0905.0987.
.
- [15] S. V. Bulanov et al. , Nucl. Instrum. Meth. A 660 , 31
- [6] \protect\vrule width0pt\protect\href{http://www.hiper-laser.org}{http://www.hiper-laser.org} . (2011); R. Ruffini, G. Vereshchagin, S.-Sh. Xue, Phys. Rept. 487 , 1 (2010).
- [7] A. L. Cavalieri et al. , New J. Phys. 9 , 242 (2007); Z. Major et al. , AIP Conference Proceedings 1228 , 117 (2010).
- [16] H. R. Reiss, J. Math. Phys. 3 , 59 (1962); Phys. Rev. Lett., 26 , 1072 (1971).
- [8] F. Mackenroth and A. Di Piazza, Phys. Rev. A 83 , 032106 (2011).
- [9] X. Feng, S. Gilbertson, H. Mashiko, He Wang, S. D. Khan, M. Chini, Yi Wu, K. Zhao, and Z. Chang Phys. Rev. Lett. 103 , 183901 (2009).
- [10] A. M. Fedotov, N. B. Narozhny, G. Mourou, and G. Korn, Phys. Rev. Lett. 105 , 080402 (2010).
- [17] A. I. Nikishov and V. I. Ritus, Zh. Eksp. Teor. Fiz. 46 , 776 (1964)[Sov. Phys. JETP 19, 529 (1964)].
- [18] N. B. Narozhnyi, A. I. Nikishov and V. I. Ritus, Zh. Eksp. Teor. Fiz. 47 , 930 (1964)[Sov. Phys. JETP 20, 622 (1965)].
- [19] V. I. Ritus, J. Sov. Laser Res. (United States), 6:5 , 497 (1985).
- [20] V. B. Berestetskii, E. M. Lifshitz, and L. P. Pitaevskii, Quantum Electrodynamics. 2nd ed., (Course of theoretical physics; vol. 4), Pergamon Press Ltd. (1982).
- [21] A. I. Titov, B. Kampfer, H. Takabe and A. Hosaka, Phys. Rev. A 87 , 042106 (2013).
- [22] A. I. Titov, H. Takabe, B. K¨mpfer, and A. Hosaka, Phys. a Rev. Lett. 108 , 240406 (2012).
- [23] T. Nousch, D. Seipt, B. K¨mpfer, and A. I. Titov, Phys. a Lett. B 715 , 246 (2012).
- [24] K. Krajewska and J. Z. Kaminski, Phys. Rev. A 86 , 052104 (2012).
- [25] B. King, N. Elkina and H. Ruhl, Phys. Rev. A 87 , 042117 (2013).
- [26] M. Boca and V. Florescu, Phys. Rev. A 80 , 053403 (2009); M. Boca, V. Dinu and V. Florescu, Phys. Rev. A 86 , 013414 (2012).
- [27] T. Heinzl, D. Seipt, and B. K¨mpfer, Phys. Rev. A a 81 , 022125 (2010).
- [28] D. Seipt and B. K¨ ampfer, Phys. Rev. A 83 , 022101 (2011).
- [29] V. Dinu, T. Heinzl, and A. Ilderton, Phys. Rev. D 86 , 085037 (2012).
- [30] D. Seipt and B. K¨ ampfer, Phys. Rev. D 85 , 101701 (2012).
- [31] K. Krajewska and J. Z. Kaminski, Phys. Rev. A 85 , 062102 (2012).
- [32] F. Mackenroth, A. Di Piazza and C. H. Keitel, Phys. Rev. Lett. 105 , 063903 (2010).
- [33] F. Hebenstreit, R. Alkofer, G. V. Dunne, and H. Gies, arXiv:0910.4457 [hep-ph].
- [34] N. B. Narozhnyi, and M. S. Fofanov, Zh. Eksp. Teor. Fiz. 110 , 26 (1996), [Sov. Phys. JETP 83, 14 (1996)].
- [35] T. Heinzl and A. Ilderton, Opt. Commun. 282 , 1879 (2009).
- [36] D. Seipt and B. K¨ ampfer, Phys. Rev. A, 89 , 023433, (2014).
- [37] A. Jochmann et al., Phys. Rev. Lett. 111 , 114803 (2013).
- [38] D. Seipt and B. K¨ ampfer, Phys. Rev. A 88 , 012127 (2013).
|
10.1140/epjd/e2014-50324-y
|
[
"Alexander I. Titov",
"Burkhard Kampfer",
"Takuya Shibata",
"Atsushi Hosaka",
"Hideaki Takabe"
] |
2014-08-05T17:28:41+00:00
|
2014-08-05T17:28:41+00:00
|
[
"hep-ph"
] |
Laser pulse-shape dependence of Compton scattering
|
Compton scattering of short and ultra short (sub-cycle) laser pulses off
mildly relativistic electrons is considered within a QED framework. The
temporal shape of the pulse is essential for the differential cross section as
a function of the energy of the scattered photon at fixed observation angle.
The partly integrated cross section is sensitive to the non-linear dynamics
resulting in a large enhancement of the cross section for short and, in
particular, for ultra-short flat-top pulse envelopes which can reach several
orders of magnitude, as compared with the case of a long pulse. Such effects
can be studied experimentally and must be taken into account in
Monte-Carlo/transport simulations of %$e^+e^-$ pair production in the
interaction of electrons and photons in a strong laser field.
|
1408.1041v1
|
## Thermalization of strongly interacting bosons after spontaneous emissions in optical lattices
J. Schachenmayer, 1 L. Pollet, 2 M. Troyer, 3 and A. J. Daley 4, 5
1 JILA, NIST, Department of Physics, University of Colorado, 440 UCB, Boulder, CO 80309, USA 2 Department of Physics, Arnold Sommerfeld Center for Theoretical Physics and Center for NanoScience, University of Munich, 80333 Munich, Germany 3 Theoretische Physik, ETH Zurich, 8093 Zurich, Switzerland 4 Department of Physics and SUPA, University of Strathclyde, Glasgow G4 0NG, UK 5 Department of Physics and Astronomy, Department of Physics, University of Pittsburgh, Pittsburgh, PA 15260, USA
We study the out-of-equilibrium dynamics of bosonic atoms in a 1D optical lattice, after the ground-state is excited by a single spontaneous emission event, i.e. after an absorption and reemission of a lattice photon. This is an important fundamental source of decoherence for current experiments, and understanding the resulting dynamics and changes in the many-body state is important for controlling heating in quantum simulators. Previously it was found that in the superfluid regime, simple observables relax to values that can be described by a thermal distribution on experimental time-scales, and that this breaks down for strong interactions (in the Mott insulator regime). Here we expand on this result, investigating the relaxation of the momentum distribution as a function of time, and discussing the relationship to eigenstate thermalization. For the strongly interacting limit, we provide an analytical analysis for the behavior of the system, based on an effective low-energy Hamiltonian in which the dynamics can be understood based on correlated doublon-holon pairs.
## I. INTRODUCTION
An important fundamental question in many-body quantum mechanics is to what extent and under which conditions an isolated system perturbed away from equilibrium will undergo thermalization, in the sense that at long times the system will reach a steady state where simple observables equal the values for a thermal distribution [1-5]. Recently it has been possible to observe integrable dynamics for strongly interacting cold gases confined to move in one dimension [6], and thus reach regimes where systems either do not thermalize in this standard sense [7] or otherwise undergo generalized thermalization [8, 9].
These fundamental questions also have an important impact on the application of cold atoms as quantum simulators [10, 11]. Development in experiments with these systems has reached a stage where it is possible to tailor interesting many-body Hamiltonians, and study the many-body physics of corresponding ground states. However, some of the most interesting physical regimes require the realization of states with small energy gaps, requiring exceptionally low temperatures and entropies, which provide a key challenge for current experiments [12]. This has been particularly true in attempts to observe quantum magnetism in optical lattices within strongly interacting regimes | U | glyph[greatermuch] J , where U is the on-site interaction energy, and J/ glyph[planckover2pi1] indicates the tunneling amplitude for particles moving between neighboring sites. There, magnetic order in multi-species mixtures is driven by terms of the order of J /U 2 , which is typically small for current experiments [12-19].
In this context, it is very important to understand and control the many-body effects of competing heating processes in experiments. Previously, it has often been as- sumed that all of the energy added to the system will be thermalized, causing an effective increase in temperature. However, in regimes in which the system does not thermalize excitations entirely (or for some types of excitations, at all), the behavior is more complex. A simple example where excitations cannot thermalize on typical experimental timescales is given by spontaneous emission events (incoherent light scattering). In such processes, the main contribution to the increase in the average energy of the system as a function of time comes from atoms being excited to higher bands of an optical lattice [2023]. Because the bandgap is usually much larger than the typical energy scales U and J of dynamics in the lowest band, it is not possible for the system to thermalize most of the energy added to the system as a function of time.
At the same time, understanding this process is made more complex by the fact that processes exciting particles to higher bands are rare compared with processes leaving particles in the lowest band in typical experiments [20, 21]. As a result, the dynamics can actually be dominated by heating in the lowest band of the lattice, the basic effects of which have been studied in several recent articles for bosons [20, 21, 24-28] and for fermions [2931]. It is then natural to ask, in particular, whether the system can thermalize after spontaneous emission events that leave atoms in the lowest band.
In this article, we study the interplay between spontaneous emissions and thermalization in detail, asking under what conditions key simple quantities such as the quasimomentum distribution and the kinetic energy relax to values given by a thermal distribution. In each case, we ask this question in a practical context by fixing a thermal distribution such that we take a canonical ensemble with the temperature T chosen so that the ex-
pectation value of the energy in the thermal distribution matches the expectation value of the energy after a spontaneous emission event. In the case of strong interactions, we might expect that the system reaches integrable limits where the system does not thermalize. On the other hand, one might also expect that as in general, the BoseHubbard Hamiltonian is non-integrable, that outside the special cases of strong interactions ( U/J →∞ ) and noninteracting systems ( U = 0), thermalization would essentially always occur.
Instead, we find that this generalization is not universally applicable. It is important to note that spontaneous emissions correspond to local quenches for the many-body state, leading to population of low-lying excited states. As we showed previously [24], after a spontaneous emission, depending on the parameter regime and the corresponding different characteristics of the lowenergy spectrum: either (i) the system relaxes over short times to thermal values of the quasimomentum distribution and kinetic energy, or (ii) on short timescales, the system relaxes to states that are clearly non-thermal, even if all atoms remain in the lowest Bloch band. The latter regime was found to occur if the interactions are strong enough so that the ground-state of the system corresponds to a Mott insulator state. Below we elaborate on these results especially for the quasimomentum distribution, and discuss their interpretation in terms of eigenstate thermalization.
In the limit where the system is close to an ideal Mottinsulator (a product state with a fixed number of particles per site) and in the case of unit filing, one can describe the dynamics of the system in an effective model, where one restricts the local Hilbert space to having a maximum of two particles per site. The low-lying excitations can then be described as correlated pairs of holes ('holons') and doubly occupied sites ('doublons') [26, 32]. Here we use this approximation to show that no thermalization is present for spontaneous emission in this limit. We compare our result to exact numerical calculations either using exact diagonalization in small systems, or by making use of t-DMRG techniques [33-36].
This article is organized as follows. We begin by introducing spontaneous emissions and thermalization in these systems in Section II. Then, in Section III we review the explanation of the basic behavior we find in terms of the eigenstate thermalization. In Section IV we discuss how the behavior can be understood in the strongly interacting limit, making predictions that could be observed for the propagation of excitations when the system does not thermalize in that limit. In Section V we then provide a summary and outlook.
## II. THERMALIZATION OF SPONTANEOUS EMISSIONS
Spontaneous emissions are a fundamental source of heating in optical dipole potentials [22, 23], and one of the key contributing heating sources in current experiments with cold atoms in optical lattices [20, 21]. As shown in Ref. [21], a multi-band many-body master equation can be derived to describe this process in the case of far detuned lattice light. An adiabatic elimination of the electronic excited states leads to an equation for atoms in their electronic ground states. Taking this, and assuming the typical experimental case in which the lattice spacing a is comparable to the wavelength of scattered photons, a = λ/ 2, the dynamics of the many-body density operator ρ follows a master equation of the form ( glyph[planckover2pi1] ≡ 1), ˙ ρ = -i[ H MB , ρ ] + L ρ . Here H MB is a multi-band Bose-Hubbard Hamiltonian and L ρ includes dissipative Lindblad terms with on-site jump operators ∝ ( b [ n ] i ) † b [ m ] i . Here, b [ n ] i is the bosonic annihilation operator at site i , for a particle in band m .
For relatively deep optical lattices, transition rates for inter-band processes coupling neighboring Bloch bands are strongly suppressed by the square of the Lamb-Dicke parameter, η = 2 πa T /λ . For a typical experiment with a lattice depth V 0 ≈ 8 E R [ E R = 4 π 2 glyph[planckover2pi1] 2 / (2 mλ 2 ), where m is the mass of the atom], η 2 ∼ 0 1. . In the usual case of a red-detuned optical lattice the dominant processes are thus intra-band processes, which return the atoms to their initial Bloch band. Provided we initially consider the system to be in the lowest band, processes accessing higher bands are suppressed by a factor of the order η 4 or greater, and the master equation for the 1D model simplifies to:
$$\dot { \rho } = - i \left [ H, \rho \right ] - \frac { \gamma } { 2 } \sum _ { i } \left [ n _ { i }, \left [ n _ { i }, \rho \right ] \right ],$$
where n i = b b † i i is the particle number operator for the lowest band, and γ is the scattering rate. The Hamiltonian is the usual single band Bose-Hubbard Hamiltonian
$$\text{the} \quad H = - J \sum _ { i } \left ( b _ { i } ^ { \dagger } b _ { i + 1 } + b _ { i + 1 } ^ { \dagger } b _ { i } \right ) + \frac { U }$$
In the following we will focus on a system with M sites and N = M particles, so that the filling is ¯ = n N/M = 1. In this case the system undergoes a ground-state quantum phase transition between a superfluid (SF) and a Mott insulating (MI) state at the critical value, which was estimated to be U c ≈ 3 25(5) . J [37, 38]. The underlying physics of the lattice-photon scattering process can be summarized as the environment obtaining information about the position of an atom. Thereby the atoms are localized on the length-scale of a single lattice-site size.
Numerically we can fully simulate the master equation (1) also for large 1D systems by combining t-DMRG techniques with a Monte-Carlo sampling over quantum trajectories [39-42]. There, the density matrix is obtained from a statistical average over many trajectories with randomly applied jump operators [39]. To give an example, in Fig. 1 we plot the dynamics for the evolution under equation (1). We start in the ground-state for a
FIG. 1. Example of heating and relaxation dynamics(a) A system with N = 48 particles on M = 48 sites and U = 2 J is initially in its ground state and subjected to heating with different scattering rates γ for a short time. (b) Depending on γ this leads to an increase in energy. (c) Evolution of the quasi-momentum distribution n q during the heating period ( γ = 0 06 . J ). The tails are lifted due to the localization of particles. (d) Relaxation dynamics of n q after the heating is switched off. The system relaxes to a broadened distribution at tJ = 5 (t-DMRG results, bond-dimension, D = 256, 500 quantum trajectories). (e/f) Relaxation of a large q component of the quasi-momentum distribution after a single spontaneous emission on an arbitrary site (we consider the a weighted ensemble average of jumps on the different sites). The dashed horizontal line indicates the equilibrium value (Monte-Carlo calculation). Thermalization only occurs in the superfluid regime (e), and not in the Mott insulating regime (f) [t-DMRG results, bond-dimension, D = 512 256 , for (e),(f)].
system with U = 2 J (SF regime, M = 48). Then we turn on light scattering for a short time up to tJ = 1 and for different values of γ (illustrated in Fig. 1a). This leads to an increase in energy (Fig. 1b). Once the heating is switched off, the system then relaxes. This relaxation is for example seen when looking at the evolution of the quasi-momentum distribution
$$n _ { q } = \frac { 1 } { M } \sum _ { n, m } e ^ { - i q ( n - m ) } \langle b _ { n } ^ { \dagger } b _ { m } \rangle. \quad \quad ( 3 )$$
During the heating period the localization of particles leads to a 'lifting' of tails of this distribution, as seen in Fig. 1c. Once the heating is switched off, the high quasi-momentum components become redistributed and the system is relaxing to a broadened distribution, depicted in Fig. 1d. The question whether the system thermalizes or not depends now on the shape of this distribution. In the case that the steady state distribution coincides with a Boltzmann distribution for a temperature corresponding to the same mean energy, we call the system thermalized, otherwise we don't. Obviously, the question of thermalization is connected to the observable under consideration. In Ref. [24] it was found that for example the tails of the quasi-momentum and the kinetic energy E kin = 〈 H kin 〉 = -2 J ∑ i 〈 b b † i i +1 〉 + 〈 b † i +1 b i 〉 thermalize on experimentally relevant time-scales in the SF regime and that this thermalization breaks down in the MI.
Here, we demonstrate a specific example of this in Figs. 1e,f. There we show relaxation of a highq component of the quasimomentum distribution, q = 40 π/ 48, in a system with M = 48 lattice sites. We can clearly see that after a short period of time, this component relaxes to its value in a corresponding thermal distribution provided that we initially start in a superfluid state with U/J = 2. In the MI limit, we see that the behavior is similar, in so far as there is a rapid relaxation of this value of quasimomentum on short times. However, this leads to a value that is markedly different to the corresponding thermal value. It is naturally possible that in this regime, there are processes on very long timescales that will eventually lead to thermalization of this quantity. But it is clear that on the MI side of the phase transition, the system exhibits a qualitatively different behavior in terms of how close it is to thermal distributions on the scale of a few tunnelling times. This will significantly impact the ability of the system to thermalize energy added by heating on the timescales of the heating process itself.
It is important to note that in the regime where the system is relaxing, the thermalization and all of the corresponding dynamics is determined by unitary closed system dynamics alone. In the remainder of this paper we will simplify the question of thermalization to the coherent dynamics after a single spontaneous emission event, i.e. we ask whether expectation values of observables relax to thermal values in a unitary time-dynamics after we apply a single quantum jump. Thus, we consider a unitary time-evolution of an initial state
$$| \psi _ { l } ( t = 0 ^ { + } ) \rangle = \frac { n _ { l } \, | \psi _ { g } \rangle } { | | n _ { l } \, | \psi _ { g } \rangle$$
where | ψ g 〉 denotes the ground-state of the system, and then take a stochastic average over the sites l , which is weighted proportional to the square of the density on site l , 〈 ψ g | n 2 l | ψ g 〉 . The mechanism of this thermalization of closed quantum systems can be explained via eigenstate thermalizaton, a mechanism which we will review in the next section.
## III. EIGENSTATE THERMALIZATION
The question whether unitary time-dynamics leads to thermalization or not can be re-phrased in terms of the 'eigenstate thermalization' [1-3], which we will briefly review here. Consider the situation where the initial state is not an eigenstate of the Hamiltonian (such as it is the case after a spontaneous emission). The state can be expanded into energy eigenstates defined via H α | 〉 = E α | α 〉 , | ψ 〉 = ∑ α c α | α 〉 , and the time-evolution is thus given by
$$| \psi ( t ) \rangle = \sum _ { \alpha } e ^ { - \frac { i } { h } E _ { \alpha } t } c _ { \alpha } | \alpha \rangle. \quad \quad ( 5 ) \quad \quad \quad$$
Then the time-dependent expectation value of any observable ˆ is O
$$\langle \hat { O } \rangle _ { t } = \langle \psi ( t ) | \hat { O } | \psi ( t ) \rangle = \sum _ { \alpha, \beta } \mathrm e ^ { - \frac { 1 } { h } ( E _ { \alpha } - E _ { \beta } ) t } c _ { \beta } ^ { * } c _ { \alpha } \langle \beta | \hat { O } | \alpha \rangle. \begin{pmatrix} \text{$\text{$for\text{ in}$} \\ ( U \lesssim \colon \\ \alpha, \text{ and} \\ \text{ of the} \end{pmatrix}.$$
If the system relaxes into a steady state, this steady state must be identical to the long-time average
$$\overline { \langle \hat { O } \rangle _ { t } } = \lim _ { T \to \infty } \frac { 1 } { T } \int _ { 0 } ^ { T } d t \langle \psi ( t ) | \hat { O } | \psi ( t ) = \sum _ { \alpha } | c _ { \alpha } | ^ { 2 } \langle \alpha | \hat { O } | \alpha \rangle. \quad \text{
} .
<?> </?> <? > + - &?>
<?>" | \end{ <! " )$$
glyph[negationslash]
Thus, the value of the steady state expectation (if such a steady state is developed) only depends on two variables: i) The expansion coefficients of the initial state, c α , and ii) the expectation values of the energy-eigenstates with the particular observable, 〈 α O α | ˆ | 〉 . The eigenstate thermalization hypothesis states, that if the the eigenenergy expectation values of an observable vary smoothly in the energy window defined by the c α (and the off-diagonal 〈 α O β | ˆ | 〉 , α = β are small), the system relaxes into a state for which the observable can be described with a micro-canonical ensemble [1, 2].
In our situation, spontaneous emission only adds a small amount of energy to the system. Thus, the question whether this energy thermalizes or not will be determined by the low-energy spectrum. Thus, we can use an exact numerical diagonalization to obtain the lowest energy eigenstates in systems with N = M = 10 sites and particles. As an example, in panels (a)-(c) of Fig. 2 we show the matrix elements 〈 α O β | ˆ | 〉 for the kinetic energy in the system. As observable we use the kinetic energy ˆ = O H kin . The eigenvalues are ordered ascending according to their energy. We find that only in the SF case of U = 2 J , the diagonal elements vary smootly as function, while the off-diagonal elements are zero. This verifies the findings of Ref. [24], and is consistent with the breakdown of the low-energy thermalization when entering the MI regime.
In addition we plot a direct comparison of the two expectation values in Fig. 2d. We can calculate the expectation value on the one hand from the diagonal ensemble
FIG. 2. Eigenstate thermalization of spontaneous emissions . (a)-(c) The matrix-elements of the kinetic energy ( N = M = 10), 〈 α H | kin | β 〉 /M , in the energy eigenbasis (sorted in ascending order according to their energy). Different panels are for increasing on-site interaction U . Only in the SF regime ( U glyph[lessorsimilar] 3 25(5) . J [37]) the diagonal elements vary smoothly with α , and allow for eigenstate thermalization. (d) The difference of the expectations between the diagonal and the canonical ensemble corresponding to the same mean energy after a single spontaneous emission at site i = M/ 2 in a system of M = 8 and M = 10 sites (errorbars are due to the interpolation to find the canonical ensemble corresponding to the correct energy). Although finite size effects are large the change in behavior at the quantum phase transition is evident.
and on the other hand from a canonical ensemble with the same mean-energy. As initial state we choose the state after a single spontaneous emission on site i = M/ 2. Although finite size effects play an important role in this small system, it is already obvious that a discrepancy between the two ensembles arises when entering the MI regime. As was found by t-DMRG calculations in [24], this also holds for larger systems. In the next section we will calculate the low-lying energy eigenstates exactly analytically in the limit of large interactions in the thermodynamic limit.
## IV. DYNAMICS IN THE STRONGLY INTERACTING LIMIT
In this section we will discuss how we can understand the dynamics analytically for the case of strong interactions. In the extreme case of infinite U , at unit filling the particles will be in an idealized Mott insulator state, i.e., | ψ U →∞ 〉 = ∏ n | 1 〉 n . This state is an eigenstate of the jump operator with eigenvalue one, so that light scattering leaving particles in the lowest band does not change the state or introduce any extra energy to the system.
## A. Hard-core bosons
Going beyond this trivial limit, one can use an approximation in which we assume that the bosons are hard-core
bosons (HCBs) [43]. The idea is to restrict the Hilbert space to local states with a maximum of one particle per site. Mathematically, this can be achieved by using the standard Bose-Hubbard model (2) while additionally imposing fermionic anti-commutation relations to the creation/annihilation operators for bosons on the same site, { b , b i † i } = 1. These particles are not real fermions, since they still commute on different sites, [ b , b i † j ] = 1. However this can be fixed by a Jordan Wigner transformation [44]
$$b _ { i } = \prod _ { \alpha = 1 } ^ { i - 1 } e ^ { i \pi c _ { \alpha } ^ { \dagger } c _ { \alpha } } c _ { i }, \quad \quad ( 8 ) \quad \text{$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$\,$$$
where the new quasi-particles now are real fermions, { c , c i † j } = δ i,j . In this picture the Hamiltonian becomes a Hamiltonian for non-interacting fermions:
$$H = - J \sum _ { i } ( c _ { i } c _ { i + 1 } ^ { \dagger } + c _ { i } ^ { \dagger } c _ { i + 1 } ), \quad \ \ ( 9 ) \quad \ \ t e r r _ { \substack { \text{of $t$} \\ \cdot \cdot \cdot } } ^ { \infty \infty }$$
and the ground-state takes the product form | ψ G 〉 = ∏ ∑ N n N i P i,n | i 〉 (where | i 〉 denotes the state of a particle at site i ). The jump-operators -as is true for all site-local operators - are unaffected by the JordanWigner transformation and so in this transformed notation, b b † i i → c c † i i .
If the system has fewer particles than lattice sites, spontaneous emissions will give rise to heating in the lowest band, as the states are not eigenstates of c c † i i , which does not commute with the Hamiltonian. Simple results in this case are discussed in Ref. [39]. However, in the special case of unit filling, which we would like to treat here, the only state described by this simple HCB form is again the trivial state with a single particle on each site, which is an eigenstate of all operators c c † i i and c c † i i +1 . It is then clear that we need to go beyond this treatment in order to understand heating in a Mott Insulator state with finite U .
## B. Doublon/Holon calculation
To study thermalization of spontaneous emission we have to go to the next order in the approximation and allow for states with 0, 1, and 2 particles per site. It turns out that also in this case approximate analytical calculations can be derived [26]. Assuming that the state of the system remains close to the Mott insulator state with ¯ n = 1, to first order excitations will be given by doubly occupied sites on the one hand, and holes on the other. It is now possible to introduce creation/annihilation operators with a vacuum given by the ideal MI state | vac 〉 i = 1 | 〉 i as
$$\tilde { d } _ { i } ^ { \dagger } \left | \text{vac} \right \rangle _ { i } & = \left | 2 \right \rangle _ { i } & ( 1 0 ) \\ \tilde { \iota } ^ { \dagger } \left | \text{d} \right \rangle _ { \lambda } & = \left | \text{d} \right \rangle & \left | \text{d} \right \rangle \iota \iota \iota$$
Since there can only be one hole or doublon on each site these quasi-particles must obey the hard-core constraint, which can again be expressed with on-site anticommutation relations { ˜ d , d i ˜ † i } = 1 = { ˜ h , h i ˜ † i } . Note that in principle one also has to add a constraint in order not to have a doublon and holon on the same site. However, based on the assumption that the total number of doublyoccupied sites and holes remains low we will neglect these terms, as was originally done in Ref. [26]. As in the case of the HCB model one can now turn the quasi-particles into proper fermions via the Jordan-Wigner transformation
$$\tilde { h } _ { i } ^ { \dagger } \left | \text{vac} \right \rangle _ { i } = \left | 0 \right \rangle _ { i } \quad ( 1 1 ) \\ \text{in only be one hole or doublon on each}$$
$$\tilde { d } _ { i } = \prod _ { \alpha = 1 } ^ { i - 1 } e ^ { i \pi d _ { \alpha } ^ { \dagger } d _ { \alpha } } d _ { i }$$
$$\tilde { h } _ { i } = \prod _ { \alpha = 1 } ^ { i - 1 } e ^ { i \pi h _ { \alpha } ^ { \dagger } h _ { \alpha } } h _ { i }.$$
In the picture of these fermionic quasi-particles, the Hamiltonian reads
$$\underset { \substack { \text{suec}, \\ \text{s.w.s.}} } { \text{s.low.} } \quad H _ { d h } = U \sum _ { i } d _ { i } ^ { \dagger } d _ { i } - J \left ( 2 \sum _ { i } d _ { i } d _ { i + 1 } ^ { \dagger } + \sum _ { i } h _ { i } h _ { i + 1 } ^ { \dagger } + \text{h.c.} \right ) \\ \text{s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.} } \quad - \sqrt { 2 } J \left ( \sum _ { i } d _ { i } ^ { \dagger } h _ { i + 1 } ^ { \dagger } - \sum _ { i } h _ { i } d _ { i + 1 } \right ). \quad \text{(14)} \\ \underset { \text{s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s. s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s. s.ss.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.ss.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.ss. s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.ss.ss.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s. s.s s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.sss.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s. ss.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s
}$$
A discrete Fourier transformation into quasi-momentum space ( x j = 1 √ M ∑ q e i jqa x q , for the operator x j , q ∈ { 0 2 , π/M, 2 × 2 π/M ... 2 π } ) gives the quadratic Hamiltonian (up to a constant)
$$\stackrel { \text{w} } { \text{ate} } \quad \tilde { H } _ { d h } = \sum _ { q } \left ( d _ { q } ^ { \dagger } \ h _ { - q } \right ) \begin{pmatrix} E _ { d } ( q ) & i E _ { d h } ( q ) \\ - i E _ { d h } ( q ) & - E _ { h } ( q ) \end{pmatrix} \begin{pmatrix} d _ { q } \\ h _ { - q } ^ { \dagger } \end{pmatrix}.$$
The corresponding dispersion relations are given by E d ( q ) = - √ 4 J cos( q ) + U , E h ( q ) = -2 J cos( q ), and E dh ( q ) = 2 J 2 sin( q ). This Hamiltonian can be diagonalized straightforwardly by a Bogoliubov transformation
$$H = \sum _ { q } \left ( d _ { q } ^ { \dagger } \ h _ { - q } \right ) U _ { B } ( q ) U _ { B } ( q ) ^ { \dagger } \begin{pmatrix} E _ { d } ( q ) & i E _ { d h } ( q ) \\ - i E _ { d h } ( q ) & - E _ { h } ( q ) \end{pmatrix} U _ { B } ( q ) U _ { B } ( q ) ^ { \dagger } \begin{pmatrix} d _ { q } \\ h _ { - q } ^ { \dagger } \end{pmatrix}$$
$$= \sum _ { q } \left ( d _ { q } ^ { \dagger } \ h _ { - q } \right ) U _ { B } ( q ) \begin{pmatrix} \epsilon _ { d } ( q ) & 0 \\ 0 & - \epsilon _ { h } ( q ) \end{pmatrix} U _ { B } ( q ) ^ { \dagger } \begin{pmatrix} d _ { q } \\ h _ { - q } ^ { \dagger } \end{pmatrix}$$
$$\equiv \sum _ { q } \left ( c _ { d, q } ^ { \dagger } \ c _ { h, - q } \right ) \begin{pmatrix} \epsilon _ { d } ( q ) & 0 \\ 0 & - \epsilon _ { h } ( q ) \end{pmatrix} \begin{pmatrix} c _ { d, q } \\ c _ { h. - q } ^ { \dagger } \end{pmatrix}$$
$$= \sum _ { q } \epsilon _ { q } ( q ) c _ { d, q } ^ { \dagger } c _ { d, q } + \epsilon _ { h } ( q ) c _ { h, - q } ^ { \dagger } c _ { h, - q }$$
where the new fermionic quasi-particles c d/h are linear combinations of the d q and h q , i.e. are correlated pairs of doublons and holons. The dispersion relations of the particles are given by (assuming U > 6 J )
$$\epsilon _ { d / h } ( q ) = \pm J \cos ( q ) + \frac { 1 } { 2 } \sqrt { 3 2 J ^ { 2 } \sin ^ { 2 } ( q ) + [ U - 6 J \cos ( q ) ] ^ { 2 } }. \quad - 1 _ { - 1 } ^ { \infty } \\ ( 2 0 ) \quad \ e v e.$$
The ground state in this approximation is the vacuum | vac , which is defined by 〉 c d/q | vac 〉 = 0 for all q .
## C. Spontaneous emission in the doublon/holon picture
Using the inverse transformation, ( c † d,q , c h, -q ) U † B ( q ) = ( d , h † q -q ), we can express the the local particle number operator for site m n , m in the Bogoliubov frame. Applying this operator to the ground-state, we find
$$& \lambda _ { n } | \text{GS} \rangle = | \text{vac} \rangle \\ & + \frac { 1 } { M } \sum _ { q, q ^ { \prime } } ( u _ { q ^ { \prime } } v _ { q } - u _ { q } v _ { q ^ { \prime } } ) c _ { d, q ^ { \prime } } ^ { \dagger } c _ { h, - q } ^ { \dagger } e ^ { i m a ( q - q ^ { \prime } ) } | \text{vac} \rangle, \ ( 2 1 ) \quad \text{We} \\ &. \quad. \quad. \quad. \quad. \quad. \quad. \quad \text{We cc}$$
where we introduced the matrix elements of the transformation as u q = U B ( q )[1 , 1], and v q = U B ( q )[2 , 1]. The time-evolved state after the jump is therefore
$$\text{$one-conveau ocave avec $u$} \, \text{$u$ jump to increcate} \quad \text{as a} \, \colon \\ \sqrt { \mathcal { N } } | \psi ( t ) \rangle & = | \text{vac} \rangle \\ & \quad + \frac { 1 } { M } \sum _ { q, q ^ { \prime } } e ^ { i m a ( q - q ^ { \prime } ) } ( u _ { q ^ { \prime } } v _ { q } - u _ { q } v _ { q ^ { \prime } } ) \\ & \quad \times e ^ { - i t [ \epsilon _ { d } ( q ^ { \prime } ) - \epsilon _ { h } ( q ) ] } c _ { d, q ^ { \dagger } } ^ { \dagger } c _ { h, - q } ^ { \dagger } | \text{vac} \rangle, \quad ( 2 2 ) \quad \text{$D$.} \\ \text{where $\mathcal{N}$} \ \equiv \ \langle G S | n _ { m } ^ { 2 } | G S \rangle. \quad \text{The initial probability}.$$
where N ≡ 〈 GS | n 2 m | GS . 〉 The initial probablitydistribution of the wavepacket w q, q ( ′ ) ≡ u q ′ v q -u v q q ′ is proportional to
$$\text{is proportional to} & \quad \cdots \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \colon \text{$\nu$} & \quad \text{We} \\ | w ( q, q ) ^ { \prime } | ^ { 2 } \in \sin \left [ \arctan \left ( \frac { 4 \sqrt { 2 } \, J \sin \left ( q ^ { \prime } \right ) } { U - 6 \, J \cos \left ( q ^ { \prime } \right ) } \right ) & \quad \text{$\theta$} \quad \text{with} \\ - \arctan \left ( \frac { 4 \sqrt { 2 } \, J \sin \left ( q \right ) } { U - 6 \, J \cos \left ( q \right ) } \right ) \right ] ^ { 2 } & \quad \text{(23)}$$
FIG. 3. The quasimomentum distribution of the initial state that is created by a spontaneous emission in the doublonholon model for U = 7 J (left panel) and U = 10 J (right panel).
and examples for U = 7 J and U = 10 J are depicted in Fig. 3. The wavepackets are peaked at opposite momenta and therefore, the jump creates holes and double occupations which move in the opposite direction. For example, for U/J = 10, the wavepackets are peaked at qa = -q a ′ ≈ 0 30 . π and we can calculate the groupvelocity from (20) as d dqa glyph[epsilon1] d ( q ) | 0 3 . π ≈ 3 75 . a/J .
We can test this analytical result with an exact tDMRG time-evolution calculation after a single jump. We compare the dynamics after a jump in the center of a system with N = M = 48 site in Fig. 4a. There we plot the time-evolution of the number of doubly occupied sites as a function of time. The white dashed line in Fig. 4a indicates the light-cone according to the analytically expected group-velocity and is in excellent agreement with the numerical calculations.
## D. Behaviour of the kinetic energy and similar variables
We now consider a specific type of observable, such as the kinetic energy, which in quasi-momentum space and in the Bogoliubov frame can be written as O = ∑ q O q with
$$O _ { q } & \equiv O _ { 1 1 } ( q ) c _ { d, q } ^ { \dagger } c _ { d, q } + O _ { 1 2 } ( q ) c _ { d, q } ^ { \dagger } c _ { h, - q } ^ { \dagger } \\ & + O _ { 2 1 } ( q ) c _ { h, - q } c _ { d, q } + O _ { 2 2 } ( q ) c _ { h, - q } c _ { h, - q } ^ { \dagger }, \quad ( 2 4 )$$
FIG. 4. Dynamics after a single spontaneous emission (a) Time-evolution of the probability of doubly occupied sites after a single spontaneous emission in the center of a 48 site system. This observable reveals doublon-holon pair wavepackets, which propagate freely through the system in opposite directions until they reach the boundary (dashed white line: analytical result for the light-cone in the thermodynamic limit). (b) Comparison of the corresponding time-evolution of the kinetic energy energy difference to the ground-state. The grey line shows the time-independent analytical result for the thermodynamic limit. The dots show the numerical result right after the jump and the blue line indicates the values that develop before boundary effects become important ( tJ < 5, see inset). (t-DMRG calculation converged with D = 256 512, , n m = 6)
with matrix elements O ij for each q . Calculating the expectation value of this observable with the time-evolved state gives
$$& \text{state gives} \\ & \langle \hat { O } \rangle ( t ) = \frac { 1 } { \mathcal { N } } \sum _ { q } O _ { 2 2 } & \text{We } \text{kinetic} \\ & + \frac { 1 } { \mathcal { N } } \frac { 1 } { M ^ { 2 } } \sum _ { q, q ^ { \prime } } \left ( O _ { 1 1 } ( q ^ { \prime } ) - O _ { 2 2 } ( q ) + \sum _ { p } O _ { 2 2 } ( p ) \right ) | w ( q, q ^ { \prime } ) | ^ { 2 } & \text{with tl} \\ & \text{(25)} & \text{$two$ w} \\ & \text{ical ca} \\ & \text{we cor} \\ & \text{The important noint to note are that all crocess terms} & \text{$\mathcal{ }N$ - $1$}.$$
The important points to note are that all cross-terms corresponding to the O 12 and O 21 part of the operator disappear. This is because terms such as 〈 vac | c h, -p c d,p c † d,q ′ c † h, -q | vac 〉 = δ p,q ′ δ p,q become zero due to the symmetric form of the specific wave-packet that is created in the jump, w q, q ( ′ ) ≡ u q ′ v q -u v q q ′ , w p, p ( ) = 0. Similarly, Kronecker deltas in the terms of the form 〈 vac | c h, -r c d,r ′ c † d,p c d,p c † d,q ′ c † h, -q | vac 〉 = δ p,q ′ δ r ,p ′ δ r,q lead to a cancellation of time-dependent terms. Thus, we find that any observable of this type is time-independent after the spontaneous emission. This is a remarkable result, given the fact that the state (22) is not an eigenstate of the Hamiltonian. The effect of a jump operator here is that it puts the system into a state whose density matrix is diagonal in the space of single doublon-holon pair excitations, for which case these specific observables don't evolve in time. In particular, the expectation value of observables of the form (24) can be effectively expressed by a density matrix -〈 O 〉 = tr( ρO ) - of the form
$$\rho = \lambda _ { \text{vac} } | \text{vac} \rangle \langle \text{vac} | + \sum _ { q, q ^ { \prime } } \lambda _ { q, q ^ { \prime } } | q, q ^ { \prime } \rangle \langle q, q ^ { \prime } |, \quad ( 2 6 ) \quad \text{dyna} \\ \text{span}$$
where | q, q ′ 〉 ≡ c † d,q ′ c † h,q | vac . 〉 Thus, we conclude that after a single spontaneous emission, observables of the form (24) can immediately be described by the diagonal ensemble, i.e. the effective density matrix for those observables takes a diagonal form in the basis low-lying energy states.
As an example, let us now consider the kinetic energy operator, which can be written as
$$\Gamma _ { q } = T _ { q } = \begin{pmatrix} - 4 J \cos [ q ] & - 2 i \sqrt { 2 } J \sin [ q ] \\ 2 i \sqrt { 2 } J \sin [ q ] & 2 J \cos [ q ] \end{pmatrix}. \quad ( 2 7 )$$
In the thermodynamic limit, we can replace the sum over quasi-momenta with the integral, (1 /M ) ∑ q → ∫ d qa / ( ) (2 π ), which yields
$$\text{matrix} \text{matrix} \text{$\mathfrak{$limit}$} \\ \text{$\mathfrak{$limit}$} \\ \text{$\mathfrak{$total}$} \\ \text{result} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\nu$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$} \\ \text{$\mathfrak{$nreal}$}$$
We compare this analytical result to our time-dependent t-DMRG calculations This is in excellent agreement with our numerical results even for relatively small U/J as shown in Fig. 4b.
## E. Eigenstate expectation values
We can check the eigenstate expectation values of the kinetic energy analytically. In Ref. [24], it was found that, once one enters the MI regime, these do not coincide with the expectation values for Boltzmann distributions with the corresponding energies anymore (cf. Section III). Now we can check whether these agree with our analytical calculation. Therefore, using exact diagonalization we compute the 500 lowest eigenstates in a system of N = M = 10 sites. These are shown as dots in Fig. 5a,b for the case of U = 10 J and U = 20 J , respectively. In addition we add the values we obtain analytically for the ground state (grey dot) and the branch of single doublonholon pair excitations (grey lines). We find good agreement for large interactions. Since the first branch converges towards the analytical calculation, we conclude that this branch indeed due to the single doublon-holon pair excitations, which is further verified by the fact that it contains M 2 -M points. The second branch must thus correspond to two pair excitations. Already soon after entering the MI U ∼ 5 the gap and branches start to appear [24]. This indicates that the doublon-holon calculations are indeed a good approximation even for regimes that are not particularly deep in the MI phase.
## V. SUMMARY AND OUTLOOK
In summary, we have investigated the non-equilibrium dynamics of bosons in an optical lattice in the presence of spontaneous emission events. We show clearly parameter
FIG. 5. Eigenstate expectation values of the kinetic energy Violet dots are results from an exact diagonalization calculation in a system with N = M = 10 particles/sites. Panel (a) is for U = 10 J , panel (b) for U = 20 J . The grey dot is the result for the ground-state kinetic energy from the analytical doublon-holon calculation in the thermodynamic limit. Grey lines are expectation values for the first branch of single doublon-holon pair excitations. For increasing U the numerical results approach the analytical result as expected (differences arise from contributions with more than two-particles per site and finite size effects).
regimes in which the system relaxes to thermal distributions for simple quanitites (especially the quasimomentum distribution), and others where it relaxes on short timescales to non-thermal values. We can understand this behavior by applying eigenstate thermalization considerations to small systems for which we can perform exact diagonalization calculations. We also show that we can understand the dynamics in the strongly inter-
- [1] J.M. Deutsch, Phys. Rev. A 43 , 2046 (1991)
- [2] M. Srednicki, Phys. Rev. E 50 , 888 (1994)
- [3] M. Rigol, V. Dunjko, M. Olshanii, Nature 452 , 854 (2008)
- [4] M.A. Cazalilla, R. Citro, T. Giamarchi, E. Orignac, M. Rigol, Rev. Mod. Phys. 83 , 1405 (2011)
- [5] M. Rigol, M. Srednicki, Phys. Rev. Lett. 108 , 110601 (2012)
- [6] T. Kinoshita, T. Wenger, D.S. Weiss, Nature 440 , 900 (2006)
- [7] M. Rigol, Phys. Rev. Lett. 103 , 100403 (2009)
- [8] A.C. Cassidy, C.W. Clark, M. Rigol, Phys. Rev. Lett. 106 , 140405 (2011)
- [9] M. Rigol, M. Fitzpatrick, Phys. Rev. A 84 , 033640 (2011)
- [10] I. Bloch, J. Dalibard, S. Nascimbene, Nat Phys 8 , 267 (2012)
- [11] J.I. Cirac, P. Zoller, Nat Phys 8 , 264 (2012)
- [12] D.C. McKay, B. DeMarco, Reports on Progress in Physics 74 , 054401 (2011)
- [13] R. J¨ ordens, L. Tarruell, D. Greif, T. Uehlinger, N. Strohmaier, H. Moritz, T. Esslinger, L. De Leo, C. Kollath, A. Georges, V. Scarola, L. Pollet, E. Burovski, E. Kozik, M. Troyer, Phys. Rev. Lett. 104 , 180401 (2010)
acting limit well in terms of propagating doublon-hole pairs, which are analytically tractable.
All of these results are expected to be directly observable in ongoing experiments. In addition to the possibility to measure quasimomentum distributions, the propagation of doublon-holon pairs as investigated here is a key signature for the effects of spontaneous emissions that could be measured in quantum gas microscope experiments with single-site resolution in ways that were previously implemented for quenches within the MI regime [32]. Measurements of this type could be used as a diagnostic tool for heating of many-body states in optical lattices, which could in turn be used to improve the robustness of quantum simulators. In the future, these studies can be continued towards thermalization for fermions in optical lattices undergoing spontaneous emissions[29-31], and the incorporation of partial thermalization from excitations to higher Bloch bands.
## ACKNOWLEDGMENTS
We would like to thank I. Bloch, W. Ketterle, S. Langer, H. Pichler, U. Schneider, and P. Zoller for discussions. This work was supported in part by AFOSR grant FA9550-13-1-0093, by a grant from the US Army Research Office with funding from the DARPA OLE program, and by the Aspen Center for Physics support under NSF grant 1066293. Computational resources were provided by the Center for Simulation and Modeling at the University of Pittsburgh.
- [14] W. Hofstetter, J.I. Cirac, P. Zoller, E. Demler, M.D. Lukin, Phys. Rev. Lett. 89 , 220407 (2002)
- [15] C.J.M. Mathy, D.A. Huse, R.G. Hulet, Phys. Rev. A 86 , 023606 (2012)
- [16] R. J¨ ordens, N. Strohmaier, K. G¨ unter, H. Moritz, T. Esslinger, Nature 455 , 204 (2008)
- [17] T. Paiva, R. Scalettar, M. Randeria, N. Trivedi, Phys. Rev. Lett. 104 , 066406 (2010)
- [18] L. De Leo, J.S. Bernier, C. Kollath, A. Georges, V.W. Scarola, Phys. Rev. A 83 , 023606 (2011)
- [19] S. Fuchs, E. Gull, L. Pollet, E. Burovski, E. Kozik, T. Pruschke, M. Troyer, Phys. Rev. Lett. 106 , 030401 (2011)
- [20] F. Gerbier, Y. Castin, Phys. Rev. A 82 , 013615 (2010)
- [21] H. Pichler, A.J. Daley, P. Zoller, Phys. Rev. A 82 , 063605 (2010)
- [22] J.P. Gordon, A. Ashkin, Phys. Rev. A 21 , 1606 (1980)
- [23] J. Dalibard, C. Cohen-Tannoudji, Journal of Physics B: Atomic and Molecular Physics 18 , 1661 (1985)
- [24] J. Schachenmayer, L. Pollet, M. Troyer, A.J. Daley, Phys. Rev. A 89 , 011601 (2014)
- [25] D. Poletti, J.S. Bernier, A. Georges, C. Kollath, Phys. Rev. Lett. 109 , 045302 (2012)
- [26] P. Barmettler, D. Poletti, M. Cheneau, C. Kollath, Phys. Rev. A 85 , 053625 (2012)
- [27] D. Poletti, P. Barmettler, A. Georges, C. Kollath, Phys. Rev. Lett. 111 , 195301 (2013)
- [28] S. Sorg, L. Vidmar, L. Pollet, F. Heidrich-Meisner, arXiv:1405.5404 (2014)
- [29] J.S. Bernier, P. Barmettler, D. Poletti, C. Kollath, Phys. Rev. A 87 , 063608 (2013)
- [30] S. Sarkar, S. Langer, J. Schachenmayer, A.J. Daley, arXiv:1405.7336 (2014)
- [31] J.S. Bernier, D. Poletti, C. Kollath, arXiv:1407.1098 (2014)
- [32] M. Cheneau, P. Barmettler, D. Poletti, M. Endres, P. Schauß, T. Fukuhara, C. Gross, I. Bloch, C. Kollath, S. Kuhr, Nature 481 , 484 (2012)
- [33] G. Vidal, Phys. Rev. Lett. 93 , 040502 (2004)
- [34] A.J. Daley, C. Kollath, U. Schollw¨ck, G. Vidal, Jouro nal of Statistical Mechanics: Theory and Experiment p. P04005 (2004)
- [35] S.R. White, A.E. Feiguin, Phys. Rev. Lett. 93 , 076401 (2004)
- [36] F. Verstraete, V. Murg, J.I. Cirac, Advances in Physics 57 , 143 (2008)
- [37] J. Sch¨ onmeier-Kromer, L. Pollet, Phys. Rev. A 89 , 023605 (2014)
- [38] T.D. K¨hner, S.R. White, H. Monien, Phys. Rev. B u 61 , 12474 (2000)
- [39] A.J. Daley, arXiv:1405.6694 (2014)
- [40] K. Mølmer, Y. Castin, J. Dalibard, J. Opt. Soc. Am. B 10 , 524 (1993)
- [41] C.W. Gardiner, P. Zoller, Quantum Noise (Springer, Berlin, 2005)
- [42] H.J. Carmichael, An Open Systems Approach to Quantum Optics (Springer, Berlin, 1993)
- [43] M. Rigol, A. Muramatsu, Phys. Rev. A 70 , 031603 (2004)
- [44] P. Jordan, E. Wigner, Z. Phys. 47 , 631 (1928)
|
10.1140/epjqt15
|
[
"Johannes Schachenmayer",
"Lode Pollet",
"Matthias Troyer",
"Andrew J. Daley"
] |
2014-08-05T17:29:31+00:00
|
2014-08-05T17:29:31+00:00
|
[
"cond-mat.quant-gas",
"quant-ph"
] |
Thermalization of strongly interacting bosons after spontaneous emissions in optical lattices
|
We study the out-of-equilibrium dynamics of bosonic atoms in a 1D optical
lattice, after the ground-state is excited by a single spontaneous emission
event, i.e. after an absorption and re-emission of a lattice photon. This is an
important fundamental source of decoherence for current experiments, and
understanding the resulting dynamics and changes in the many-body state is
important for controlling heating in quantum simulators. Previously it was
found that in the superfluid regime, simple observables relax to values that
can be described by a thermal distribution on experimental time-scales, and
that this breaks down for strong interactions (in the Mott insulator regime).
Here we expand on this result, investigating the relaxation of the momentum
distribution as a function of time, and discussing the relationship to
eigenstate thermalization. For the strongly interacting limit, we provide an
analytical analysis for the behavior of the system, based on an effective
low-energy Hamiltonian in which the dynamics can be understood based on
correlated doublon-holon pairs.
|
1408.1043v1
|
Nuclear Physics A
Nuclear Physics A 00 (2014) 1-4
## Charge-dependent anisotropic flow studies and the search for the Chiral Magnetic Wave in ALICE
R. Belmont (for the ALICE Collaboration) 1
Wayne State University
## Abstract
Theoretical calculations have shown the possibility of P-violating bubbles in the QCD vacuum, which in combination with the strong magnetic field created in o GLYPH<11> -central heavy-ion collisions lead to novel e GLYPH<11> ects such as the Chiral Magnetic E GLYPH<11> ect (CME) and the Chiral Separation E GLYPH<11> ect (CSE). A coupling between the CME and the CSE produces a wave-like excitation called the Chiral Magnetic Wave (CMW). The CMW produces a quadrupole moment that always has the same sign and is therefore present in an average over events. In this talk we present a series of charge-dependent anisotropic flow measurements in Pb-Pb collisions at p sNN = 2.76 TeV in ALICE, using two- and three-particle correlators with unidentified hadrons. The relation of these measurements to the search for the CMW is discussed.
Keywords: Flow, Parity, Local Charge Conservation, Chiral Magnetic Wave
## 1. Introduction
O -central heavy-ion collisions create an almond-shaped overlap region, which extends above and below the reacGLYPH<11> tion plane. The pressure gradients in-plane are greater than those out-of-plane, which creates an azimuthal anisotropy in momentum space as the system expands. Additionally, the spectator protons can be thought of as small but very dense currents. Since there are two currents close-by pointing in opposite directions, the induced magnetic field from each current adds linearly in the region between them, creating a very large and relatively homogeneous magnetic field in the same location in configuration space as the medium created in the overlap region. The interaction of the magnetic field with the produced particles in a region of space with topologically non-trivial gluon field configurations leads to novel e GLYPH<11> ects like the Chiral Magnetic E GLYPH<11> ect (CME) [1] and the Chiral Separation E GLYPH<11> ect (CSE) [2].
The CME and CSE can be summarized succinctly in a pair of equations [3],
$$\vec { J } _ { V } \ = \ \frac { N _ { c } e } { 2 \pi ^ { 2 } } \mu _ { A } \vec { B },$$
$$\vec { J } _ { A } \ = \ \frac { \stackrel { \cdot } { \frac { N _ { c } e } { 2 \pi ^ { 2 } } } \mu _ { V } \vec { B }.$$
Equation 1 is for the CME, and indicates that a vector current JV (for example an electric current) is coupled to an axial chemical potential GLYPH<22> A , oriented along the the magnetic field B . Equation 2 is for the CSE, and indicates that an axial current JA is coupled to a vector chemical potential GLYPH<22> V (for example the scalar electric potential), again oriented along the the magnetic field.
1 A list of members of the ALICE Collaboration and acknowledgements can be found at the end of this issue.
One can tell by inspection that the two currents are coupled. By changing to the chiral basis, V = R + L and A = R GLYPH<0> L , one can derive two equations indicating two electric currents that always point in opposite directions, leading to an electric quadrupole moment that always has the same sign. A detailed explanation and derivation is given in [3]. This e GLYPH<11> ect is shown schematically in Figure 1.
Figure 1. Cartoon showing the basic picture of the Chiral Magnetic Wave for the two cases GLYPH<22> V > 0 and GLYPH<22> V < 0.
In Figure 1, the upper row shows an example where the plasma has a positive electric charge state, i.e. GLYPH<22> V > 0. This causes a CSE current pointing upwards, leading to an excess of right-handed particles above the reaction plane and an excess of left handed particles below it. This means that above the reaction plane one has GLYPH<22> A > 0 and below it one has GLYPH<22> A < 0. This then leads to two oppositely directed CME currents, each pointed away from the reaction plane. This leads to a positive electric quadrupole, with excess positive charges out of plane at the poles and excess negative charges in plane at the equatorial region. Finally, under the hydrodynamic expansion of the medium, the equatorial region has a larger flow velocity due to the larger pressure gradients in plane, and therefore one observes large v 2 for negative particles than positive particles, i.e. v GLYPH<0> 2 > v + 2 . The lower row of the figure shows the same schematic for the opposite case with GLYPH<22> V < 0, which leads to exactly the same e GLYPH<11> ect with all signs flipped. A detailed explanation and derivation of how the presence of CMW a GLYPH<11> ects the final state observables in this way is given in [4]
## 2. Methodology and observables
From the previous section, the most intuitive observable would be v 2 (or vn to explore possible higher harmonic e GLYPH<11> ects) as a function of the event charge asymmetry A , which is defined as A = N + GLYPH<0> N GLYPH<0> N + + N GLYPH<0> , where N + and N GLYPH<0> represent the number of positive and negative particles, respectively, measured in some well-defined region of phase space, i.e. for some pT selection and, more importantly, some GLYPH<17> acceptance. Indeed, this observable has been proposed theoretically [4] and measured experimentally in Au-Au collisions at p sNN = 200 GeV by the STAR collaboration [5]. However, one of the issues with this observable is that the slope of vn vs A is not independent of experimental e GLYPH<11> ects (for example tracking e GLYPH<14> ciency) and therefore requires a correction factor.
In this analysis, we use a novel three particle correlator [6] that is independent of e GLYPH<14> ciency and therefore requires no correction. The three particle correlator is h cos( n ( GLYPH<30> 1 GLYPH<0> GLYPH<30> 2 )) q 3 i , where GLYPH<30> 1 and GLYPH<30> 2 are the azimuthal angles of particles 1 and 2, and q 3 is the charge ( GLYPH<6> 1) of particle 3. The cos( n ( GLYPH<30> 1 GLYPH<0> GLYPH<30> 2 )) part is estimated using the cumulant method and denoted as cn f 2 . In the absence of charge dependent correlations, the correlator should factorize, i.e. g
$$\langle \cos ( n ( \phi _ { 1 } - \phi _ { 2 } ) ) q _ { 3 } \rangle - \langle q _ { 3 } \rangle \langle \cos ( n ( \phi _ { 1 } - \phi _ { 2 } ) ) \rangle = 0.$$
Note that when the charge of the third particle is averaged over all particles in the event (in the specified kinematic acceptance), the mean is equal to the charge asymmetry, i.e. h q 3 i = h A i .
## 3. Results
Figure 2. Three particle correlator for the second (left panel) and third (right panel) harmonic as a function of centrality.
Figure 2 shows the three-particle correlator for the second (left panel) and third (right panel) harmonic as a function of centrality. For the second harmonic one sees a substantial increase in the correlation strength as the collisions become more peripheral. This could be caused by any possible combination of several factors. The magnetic field strength increases as the impact parameter increases and thus the current gets stronger. This would cause the correlations due to the CMW to get stronger. Additionally, e GLYPH<11> ects due to local charge conservation (LCC, i.e. the creation of balanced charge pairs close by in position space) could play a role [7]. Since central collisions have more combinatoric (uncorrelated) pairs, the correlations due to LCC su GLYPH<11> er combinatorial dilution. Note that neither of these necessarily comes at the expense of the other. Non-flow 3-particle correlations may also contribute, for example from jet-like correlations and 3-body decays. However, a comparison to HIJING Monte Carlo was performed and the threeparticle correlator was found to be statistically insignificant, indicating these e GLYPH<11> ect play at most a very small role. For the third harmonic one sees a small but non-zero correlation. The physical origin is not clear, though background from LCC is likely a contributor. Additional factors could also contribute, such as interference between the second and third harmonics, higher order multipole moments of the P-violating e GLYPH<11> ects, and other possible e GLYPH<11> ects.
ALI-PREL-70961
Figure 3. Three particle correlator for the second harmonic. The left panel shows the results with charge independent subtraction and the right panel shows the results with charge dependent subtraction.
Figure 3 shows the three-particle correlator for the second harmonic as a function of GLYPH<1> GLYPH<17> = GLYPH<17> 1 GLYPH<0> GLYPH<17> 3 . The left panel shows the result where h q 3 i is subtracted, meaning the same average charge is subtracted for each charge. This correlator is likely proportional to v n 2 B ( GLYPH<1> GLYPH<17> ) dN d = GLYPH<17> , where B ( GLYPH<1> GLYPH<17> ) is the charge balance function. See [7] for a treatment of LCC and the charge balance function as related to searches for local P-violating e GLYPH<11> ects and [6] for a discussion of those e GLYPH<11> ects on this particular observable. Contrariwise, the right panel shows the charge-dependent subtraction of h q GLYPH<6> 3 i , meaning the correlation between the the charge of particle 1 ( q 1 ) and q 3 is taken into account. One sees a substantial reduction in the e GLYPH<11> ect in terms of both strength and in range, i.e. the length of the correlation in GLYPH<1> GLYPH<17> . From this we may conclude that this subtraction removes some amount of the LCC contribution, though a detailed theoretical treatment is needed.
Figure 4. Three particle correlator for the third harmonic. The left panel shows the results with charge independent subtraction and the right panel shows the results with charge dependent subtraction.
Figure 4 shows the three-particle correlator for the third harmonic as a function of GLYPH<1> GLYPH<17> . As for the previous figure, the left panel shows the charge independent subtraction and the right shows the charge dependent subtraction. Note that the correlation strength for the unsubtracted is rather strong, whereas the charge dependent subtraction removes the correlation almost entirely.
## 4. Conclusions
Anovel three-particle correlator is employed to search for the CMW. The results for the second and third harmonic were shown as a function of centrality and GLYPH<1> GLYPH<17> . Although LCC is thought to be a major background to these measurements, the charge dependent subtraction can reduce this e GLYPH<11> ect. Even after the reduction some charge dependent signal is still observed, which may indicate that some LCC e GLYPH<11> ect remains, or which may be due to P-violating e GLYPH<11> ects like the CMW, or some combination of factors. Further input from theory is needed to give detailed constraints on the magnitude and range of LCC vs CMW correlations.
## References
- [1] D. Kharzeev, L. McLerran, and H. Warringa, Nucl. Phys. A 803 (2008) 227
- [2] D. Son and P. Surowka, Phys. Rev. Lett. 103 (2009) 191601
- [3] D. Kharzeev and H.-U. Yee, Phys. Rev. D 83 (2011) 085007
- [4] Y. Burnier, D. Kharzeev, J. Liao, and H.-U. Yee, arXiv:1208.2537
- [5] G. Wang (for the STAR Collaboration), Nucl. Phys. A 904-905 (2013) 248c
- [6] S. Voloshin and R. Belmont, these proceedings
- [7] S. Schlichting and S. Pratt, Phys. Rev. C 83 (2011) 014913
|
10.1016/j.nuclphysa.2014.09.070
|
[
"Ron Belmont"
] |
2014-08-05T17:29:58+00:00
|
2014-08-05T17:29:58+00:00
|
[
"nucl-ex",
"nucl-th"
] |
Charge-dependent anisotropic flow studies and the search for the Chiral Magnetic Wave in ALICE
|
Theoretical calculations have shown the possibility of P-violating bubbles in
the QCD vacuum, which in combination with the strong magnetic field created in
off-central heavy-ion collisions lead to novel effects such as the Chiral
Magnetic Effect (CME) and the Chiral Separation Effect (CSE). A coupling
between the CME and the CSE produces a wave-like excitation called the Chiral
Magnetic Wave (CMW). The CMW produces a quadrupole moment that always has the
same sign and is therefore present in an average over events. In this talk we
present a series of charge-dependent anisotropic flow measurements in Pb--Pb
collisions at $\sqrt{s_{NN}}$ = 2.76 TeV in ALICE, using two- and
three-particle correlators with unidentified hadrons. The relation of these
measurements to the search for the CMW is discussed.
|
1408.1044v2
|
## Fast Heuristics for Power Allocation in Zero-Forcing OFDMA-SDMA Systems with Minimum Rate Constraints
## Abstract
We investigate in this paper the optimal power allocation in an OFDM-SDMA system when some users have minimum downlink transmission rate requirements. We first solve the unconstrained power allocation problem for which we propose a fast zero-finding technique that is guaranteed to find an optimal solution, and an approximate solution that has lower complexity but is not guaranteed to converge. For the more complex minimum rate constrained problem, we propose two approximate algorithms. One is an iterative technique that finds an optimal solution on the rate boundaries so that the solution is feasible, but not necessarily optimal. The other is not iterative but cannot guarantee a feasible solution. We present numerical results showing that the computation time for the iterative heuristic is one order of magnitude faster than finding the exact solution with a numerical solver, and the non-iterative technique is an additional order of magnitude faster than the iterative heuristic. We also show that in most cases, the amount of infeasibility with the non-iterative technique is small enough that it could probably be used in practice.
## I. INTRODUCTION
Due to the increasing bandwidth requirements of wireless users, new sophisticated spectrum management techniques are required. An approach that has recently attracted a lot of interest to increase throughput is to exploit the spatial, frequency and multi-user diversity dimensions by combining the orthogonal frequency division multiplexing access (OFDMA) and spatial division multiple access (SDMA) techniques.
OFDMA provides multi-user frequency diversity by dividing the available bandwidth into independent subchannels and then using a channel-aware scheduler to grant access to the users with the best conditions for each subchannel. Meanwhile, the SDMA technique assigns the same frequency subchannel to a group of users with compatible channel vectors, thus increasing the
system's spectral efficiency. Zero forcing (ZF) is a practical SDMA beamforming i.e., precoding technique where the beamforming vectors are simply computed using the pseudo-inverse matrix [1]. This cancels inter-user interference and allows simultaneous parallel transmission to the selected users. The ZF precoding technique is not optimal, but its implementation in practical systems is much simpler than other precoding techniques and is quasi-optimal at higher signal to noise ratio (SNR) or when users have quasi-orthogonal channel vectors [2].
The constrained resource allocation (RA) problem for a system using the ZF OFDMA-SDMA technique is made up of two parts. First, select optimally which users to assign to each subcarrier, and choose their transmit power allocation to maximize the sum rate subject to a total power budget and minimum rate requirements for a subset of users.
The first part of the RA problem - user selection - can be solved using heuristic methods that scan the users' spatial vectors and pick the best users for each subchannel. The second part - power allocation (PA) - can be transformed into a one-dimensional root finding problem when formulated in the dual domain for the case of one total power constraint and yields the well-known water-filling power allocation scheme [3]. Heuristics are proposed in [4], [5], [6], to obtain the optimal water level efficiently for the case of single input single output (SISO) wireless link. For the case of a multiple input single output (MISO) wireless link, such as for multi-antenna SDMA transmission, the combined heuristics for subchannel user selection and power allocation give results close to the optimum [7], [8].
When the base station (BS) supports best effort (BE) and real-time (RT) traffic simultaneously, the RA algorithm should guarantee a certain minimum rate for the RT users, while the BE users with good channel conditions should be assigned resources to increase the sum rate. In this case, the user selection should pick the subchannels to insure that the RT users get their minimum rates. Heuristic methods have been proposed in [9], [10] for user selection. Efficient power allocation is also a central tool to improve the performance of these heuristics [10]. After an initial user selection, PA is performed and the user rates are computed. If these rates are lower than the rate requirements for the RT users, a new user selection is done and PA is performed again until a good enough solution is found.
For a given user selection, the resulting PA problem is convex due to the use of the pseudoinverse technique. It can be solved optimally using a number of standard optimization techniques [11]. This can take some time when the number of subchannels is large, such as in
LTE-Advanced systems [12]. In addition, the PA problem with fixed subchannel assignment is solved many times by the subchannel assignment heuristics [9], [10] when looking for additional subchannels to satisfy the RT constraints. For this reasons, we need efficient heuristics that solves the PA problem for fixed subchannel assignment much faster than the standard algorithms and with an achieved rate not too different from optimal solution. The objective of this paper is the design and evaluation of such a PA heuristic for ZF OFDMA-SDMA systems with RT users with minimum rate requirements.
Constant power distribution among users has been used as a simple heuristic for PA. In a scenario where the minimum rate constraints are lower than the rates achieved by the unconstrained sum rate maximization solution, the performance obtained is close to the optimal for BE traffic [13]. However, when the minimum rates constraints are active, the method cannot always provide feasible points. An adaptable PA scheme should decrease the power of the BE users and give it to the RT users to support their minimum rates. Obviously, this is not possible in a constant power allocation scheme. This is partially solved in [14], by assigning constant power to the BE users and performing water-filling PA on the RT users. In this paper, we go several steps beyond this and use multi-level water-filling PA over all users and an efficient heuristic that operates in the dual domain.
Because we use the pseudo-inverse approximation, the resulting PA problem is convex. If there are no rate constraints, it can be solved exactly and efficiently with the methods proposed in [5], [15], [6]. The fastest is that of [5] by one order of magnitude. Several heuristic methods have been proposed to solve the power minimization problem under rate constraints for SISO systems [16], [17]. But, to the best of our knowledge, fast methods to solve the problem under minimum rate constraints have not been reported for ZF OFDMA-SDMA systems.
One common method to solve convex problems is dual Lagrange decomposition [18]. We define dual variables for the power and rate constraints. For problems like ours, we can get a closed-form expression of the dual function and the users' powers. The dual function is continuous but not differentiable and we could use the subgradient algorithm to get the optimal dual variables. The dual variable for the power constraint indicates the water level and the ones for the rate constraints the superimposed multi-levels [19]. The size of the problem grows with the number of RT users, since each rate constraint is assigned a dual variable, which can lead to large computation times. Instead of solving the dual problem to optimality, our method finds
an approximate solution that generates a feasible point to the problem which is much faster to compute.
The main contribution of this paper is a method to find the dual variables that satisfy the power and rate constraints much faster than solving the dual problem optimally using iterative methods. As our numerical results indicate, the resulting dual variables generate a feasible point that is close to the optimal solution and the computation times are several orders of magnitude lower than for the optimal method.
In section II, we describe the system under consideration and mathematically formulate the PA optimization problem. Then we present in section III some algorithms for the unconstrained maximum throughput power allocation. Next, we show in section IV how to compute fast approximations when rate constraints are present. We then numerically evaluate the proposed algorithm in section V, both in terms of accuracy and CPU time. Finally, we present our conclusions in section VI.
## II. PROBLEM FORMULATION AND OPTIMAL SOLUTION
We consider the resource allocation problem for the downlink transmission in a multi-carrier multi-user MISO system with a single BS. There are K users, some of which have RT traffic with minimum rate requirements, while the others have non real-time (nRT) traffic that can be served on a best-effort basis. The BS is equipped with M transmit antennas and each user has one receive antenna. The system's available bandwidth W is divided into N subchannels whose coherence bandwidth is assumed larger than W/N , thus each subchannel experiences flat fading. In the system under consideration the BS transmits data in the downlink direction to different users on each subchannel by performing linear beamforming precoding. At each OFDM symbol, the BS changes the beamforming vector for each user on each subchannel to maximize a weighted sum rate.
To simplify the RA problem, we assume that the beamforming vectors are chosen according to the ZF criteria, which is known to be nearly optimal when the SNR is high [2]. We assume that we have already chosen a set of users S n out of the total number of users K for each subchannel n , such that g n . = | S n | ≤ M . Let us define the subcarrier channel matrix H n = [ h T n,s n (1) . . . h T n,s n ( g n ) ] T where h n,k ∈ C 1 × M is the channel row M -vector between the BS and
user k on subchannel n , and
$$\wedge \emph { u u n a m i n }, \emph { a n } \\ \beta _ { n, k } \doteq \begin{cases} \left [ ( H _ { n } ^ { \dagger } ) ^ { H } H _ { n } ^ { \dagger } \right ] _ { j, j } & \text{if $k=s_{n}(j)$,} & \forall j \in \{ 1, \dots, g_{n} \} \\ 0, & \text{otherwise.} \end{cases} \\ \text{r objective is to find a fast method to assign the transmit power to users. First, let us define} \\ \text{album normalize}$$
Our objective is to find a fast method to assign the transmit power to users. First, let us define the problem parameters
- ˇ P Total power available;
- c k Scheduling weight of user k ;
- β n,k Effective channel gain of user k on subcarrier n ;
- ˇ d k Minimum rate required by the real-time user k .
The decision variable are defined as
- p n,k Power assigned to user k on subchannel n .
We also define r n,k The transmission rate achieved by user k on subchannel n for one OFDM symbol. We assume that capacity achieving channel coding is employed on each subcarrier such that
$$r _ { n, k } = \log _ { 2 } \left ( 1 + p _ { n, k } \right ).$$
The objective of the PA algorithm is to find the p n,k to maximize the weighted sum rate of the users subject to the power and minimum rate constraints. The users scheduling weights c k and the minimum rate constraints are determined by a higher layer scheduler. The optimization problem is formulated as follows
$$& \max _ { p _ { n, k } } \sum _ { n = 1 } ^ { N } \sum _ { k = 1 } ^ { K } c _ { k } \log _ { 2 } ( 1 + p _ { n, k } ) \\ & \stackrel { N } { \longrightarrow } \stackrel { K } { \longrightarrow } \,..$$
$$\sum _ { n = 1 } ^ { \dots \cdots \cdot } & \quad k \in \mathcal { D } \\ \sum _ { n } \cdots & \quad \mathfrak { n } \cdots \cdots \cdot & \quad \mathfrak { k } \in \mathcal { D } \\ \mathfrak { n } \cdots = \mathfrak { n } \quad \forall n \cdots \cdots & \quad \mathfrak { k } \in \mathfrak { k }$$
$$\sum _ { n = 1 } ^ { N } \sum _ { k = 1 } ^ { K } \beta _ { n, k } \ p _ { n, k } & \leq \check { P } \\ \frac { N } { 2 }$$
$$p _ { n, k } = 0, \quad \forall n, k \notin S _ { n }$$
$$p _ { n, k } \geq 0, \quad \forall n, k \in S _ { n }.$$
Constraint (4) is the total power constraint and constraints (5) are the RT users' minimum rate constraints. Note that this problem is convex since it maximizes a concave function over the convex set defined by (4-7). The solution also has the following property.
## Lemma 1. The solution of problem (3-7) satisfies the power constraint (4) with equality.
Proof: This is a straightforward consequence of the fact that the rate function (2) is an increasing function of p n,k . Suppose p ( a ) is a feasible point that satisfies (4) with strict inequality and satisfied (5). Then there exists a ∆ ≥ 0 such that both the power constraint (4) and the rate constraints (5) are feasible for p ( b ) = p ( a ) + ∆ . But in this case, the values of r n,k will also increase so that p ( a ) cannot be optimal.
In all that follows, we will thus assume that (4) is an equality.
We can solve (3-7) in a number of ways, for instance, with a general-purpose nonlinear programming solver to get solution of the primal, or use a Lagrangian decomposition method and solve the dual by a subgradient technique. However, this is much too long for use in real time applications and we thus need fast approximations. The ones we propose in this paper are based on a fixed-point reformulation of the problem in the dual space.
Because the problem is convex, we can solve it optimally by computing a solution to the first-order optimality equations. Let first define the Lagrange multipliers θ ≥ 0 ∈ R K + for the power constraint (4) and δ k ≥ 0 ∈ R K + for the rate constraints (5). We then write the partial Lagrange function function
$$\text{$x$ unwavant} \\ \mathcal { L } _ { 2 } ( \mathbf p, \theta, \delta ) = - \theta \check { P } + \sum _ { k = 1 } ^ { K } \delta _ { k } \check { d } _ { k } + \sum _ { n = 1 } ^ { N } \sum _ { k = 1 } ^ { K } - ( c _ { k } + \delta _ { k } ) \\ \log _ { 2 } ( 1 + p _ { n, k } ) + \theta \beta _ { n, k } p _ { n, k } - \sum _ { n, k } \mu _ { n, k } p _ { n, k } \\ \text{is the vector of variables $\delta_{k}$. For convenience we have defined $\delta_{k}$ for $k\notin\mathcal{D}$ for which}$$
where δ is the vector of variables δ k . For convenience we have defined δ k for k / ∈ D for which we set ˇ d k = 0 . We get the first-order optimality conditions
$$\frac { \partial \mathcal { L } } { \partial p _ { n, k } } = - \frac { c _ { k } + \delta _ { k } } { 1 + p _ { n, k } } + \theta \beta _ { n, k } = 0$$
which yield the optimal water-filling power allocation as a function of the multipliers θ and δ
/negationslash
$$\bar { p } _ { n, k } = & \left [ \frac { c _ { k } + \delta _ { k } } { \theta \beta _ { n, k } \ln 2 } - 1 \right ] ^ { + }, \quad \forall n, \forall k \, \colon \beta _ { n, k } \neq 0.$$
From this, we get the optimal rate allocation
/negationslash
$$r _ { n, k } = \log _ { 2 } \left ( 1 + \left [ \frac { c _ { k } + \delta _ { k } } { \theta \beta _ { n, k } \ln 2 } - 1 \right ] ^ { + } \right ), \quad \forall n, \forall k \colon \beta _ { n, k } \neq 0. \\ \cdot \cdot \cdot \cdot \cdot \cdot. \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \,,$$
We also define the set of strictly positive allocations for user k
$$\mathcal { B } _ { k } ( \theta, \delta _ { k } ) & = \left \{ n | \overline { p } _ { n, k } > 0 \right \}. \\ \text{the size } \sigma _ { k } ( \theta, \delta _ { k } ) & \text{of this set increases with } \delta _ { k } \text{ and decreases with } \theta.$$
From (10), we see that the size σ k ( θ, δ k ) of this set increases with δ k and decreases with θ .
We have from (10) an expression for the solution in terms of the multipliers. The standard procedure is to replace this into the constraints and solve for the multipliers. This generally yields a set of nonlinear equations and in the case of inequality constraints, it has the added complexity of choosing the set of constraints that are tight at the solution.
We now reformulate the problem in terms of a system of fixed-point equations. This will be used later to get fast approximations and it also yields an algorithm for computing a feasible solution. We can do this by replacing (10) in (4) and (5) and get
$$\tau \, c a n o v a n s & \, \gamma \, \text{p} a n g & \, \colon \, \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \quad \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{p} a n g & \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tau \, \tuple \, \tau \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tuple \, \tul \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \Bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu\, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{\,m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \tuple \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \bigcup \nu \, \text{m} \, \tag* { 1 } \, + \,$$
θ = ∑ K k =1 σ k ( θ, δ k )( c k + δ k ) ( ˇ + P ∑ K k =1 ∑ n ∈B k ( θ,δ k ) β n,k ) ln 2 . (14) This is a set of fixed-point equations because of the presence of B ( θ, δ k ) in the right-hand side. Any set of variables θ , δ that solve these equations will yield an optimal solution via (10). Also note that if B is fixed, each δ is a linear function of θ , and θ is a linear function of δ . This means that both θ and δ can be viewed as piecewise linear functions of the other multipliers.
For now, we want to solve problem (3-7) efficiently. The first approach is to take advantage of the fact that we expect that the number of real-time users will generally be small. This means that if we solve the power allocation without the rate constraints (5), unless the real-time users happen to have particularly bad channels, there is a good chance that the constraints will be met automatically, so that this solution is optimal. We study this problem in section III.
If this solution does not meet the rate constraints, we incorporate them into the problem and solve it approximately. We describe in section IV two heuristic methods for this latter problem. The first one is iterative and yields a feasible solution where the rate constraints are tight. The
second one is based on adjusting the variables to make the solution feasible. It requires no iterations and often finds a feasible point that satisfies the power constraint with equality and the rate constraints with inequality.
## III. POWER ALLOCATION WITHOUT RATE CONSTRAINTS
We want in this section to compute quickly the optimal solution of problem (3-4) and (7) without the rate constraints (5). In this case, we have δ k = 0 but because it will be used in section IV when computing a feasible solution, we present the algorithms with an arbitrary, but fixed, value of δ . We compare two methods that yield an exact solution. The first one finds a root of the power constraint and is guaranteed to converge in a fixed number of iterations. The second method does not have a guaranteed convergence but is faster that the root-finding method.
## A. Root of the Optimality Conditions
This method computes the optimal value of θ by finding a zero of the first-order optimality condition. We replace (10) in (4) and solve for θ the nonlinear equation
$$g _ { 1 } ( \theta ) = \sum _ { k = 1 } ^ { K } \sum _ { n = 1 } ^ { N } \beta _ { n, k } \left [ \frac { ( c _ { k } + \delta _ { k } ) } { \theta \beta _ { n, k } \ln 2 } - 1 \right ] ^ { + } = \check { P }. \\ a _ { 1 } ( \theta ) \text{ is continuous but not differentiable at the corner} \text{ inits } \theta _ { 2 } \text{.} \text{ defined as}$$
The function g 1 ( θ ) is continuous but not differentiable at the corner points θ n,k defined as
$$\theta _ { n, k } = \frac { ( c _ { k } + \delta _ { k } ) } { \beta _ { n, k } \ln 2 }.$$
For a given value of θ , g 1 is the sum of all hyperbolic segments p n,k ( θ ) such that θ ≤ θ n,k . A typical function is shown on Fig. 1. We could solve it using a general-purpose zero-finding algorithm for non-differentiable functions but we can take advantage of the particular structure of the problem to design a more efficient algorithm. This is based on the method presented in [4] that takes advantage of the finite number of corner points. First, we relabel the corner points in increasing values θ , i i = 1 , . . . N K . The sets B k ( θ, δ k ) can be defined in a more compact way
$$\mathcal { B } _ { k } ( \theta, \delta _ { k } ) = \left \{ n | \theta \leq \theta _ { n, k } \right \}.$$
Note that B k ( θ, δ k ) is constant inside any given interval [ θ i -1 , θ i ] .
First we quickly find the interval ∆ where the solution lies by doing a binary search on the set of θ i . At iteration m , denote the two indices that define the interval that contains the solution by
Fig. 1. Solving the first-order equations
( i ( m ) 1 , i ( m ) 3 ) so that ∆ ( m ) = [ θ i ( m ) 1 , θ i ( m ) 3 ] . Obviously, i (0) 1 = 1 and i (0) 3 = NK . Let i ( m ) 2 be the index of the middle point /floorleft ( i ( m ) 3 -i ( m ) 1 ) / 2 /floorright . If g 1 ( θ i ( m ) 2 ) > P ˆ , then i ( m +1) 1 = i ( m ) 2 and i ( m +1) 3 = i ( m ) 3 . If g 1 ( θ i ( m ) 2 ) < P ˆ , then i ( m +1) 3 = i ( m ) 2 and i ( m +1) 1 = i ( m ) 1 . The number of function evaluations is bounded by log ( 2 NK ) since the size of the interval is divided by 2 at each iteration and the algorithm converges in a finite number of iterations. We also need to compute the θ i and sort them.
Once we know that the solution lies in the interval [ θ i -1 , θ i ] , we can compute the value of the sets B k ( θ, δ k ) and find the value of θ directly from (14).
## B. Fixed-Point Algorithm
We can also find the solution of (14), which we write θ = g 2 ( θ ) , by repeated substitution of the left-hand side θ into g 2 . We first compute an initial value of θ by replacing the c k 's and β n,k 's by their average values in (10) and setting the condition p = 0 . Using the current value of θ , we then compute the B k ( θ, δ k ) s. A new value of θ can then be found from the right-hand side of (14) and we iterate until the new value of θ is the same as the previous one. The stopping rule is an equality since various values of θ produce only a finite number of sets B k so that the
Fig. 2. Fixed point iterations
only values produced by the iterations are the θ i . We can see the iteration procedure on Fig. 2 where we have shown the left-hand side in the form y = θ and the right-hand side y = g 2 ( θ ) . The solution is at the intersection of the two curves. We also show the points computed by the substitution algorithm.
The main advantage of this method is that there is no need to compute the θ i and to sort them in advance. The problem with substitution methods is that it is difficult to guarantee convergence. But, in practice we have found that the fixed-point method always converged for all realistic values of the parameters.
We compare the cpu time of the four solution techniques in Table I. We have fixed N = 16 , ˆ = 5 P , c k = 1 and generated the β n,k from a simple but realistic channel model. These values were obtained on a 2.5 GHz Intel Core i5 computer. We have computed the solution 100 times in each case to have a reasonably stable estimate. We can see that the fixed-point method is the fastest of all methods and the number of iterations does not change much with the problem size. The option is then between a technique with guaranteed convergence but somewhat slower than the other technique with no strict convergence guarantee. A compromise could be to use
TABLE I
| M | Minos | Nonlinear | Nonlinear | Binary | Binary | Fixed Point | Fixed Point |
|-----|----------|-------------|---------------|----------|---------------|---------------|---------------|
| | Cpu msec | Cpu msec | No iterations | Cpu msec | No iterations | Cpu msec | No iterations |
| 2 | 23.3 | 19.0 | 20 | 1.5 | 0 | 2.0 | 3 |
| 4 | 71.8 | 43.8 | 21 | 12.6 | 6 | 4.2 | 3 |
| 8 | 292. | 92.5 | 22 | 31.4 | 9 | 7.6 | 3 |
COMPUTATION LOAD WITHOUT RATE CONSTRAINTS
the fixed-point method and switch to the binary search technique if it has not converged within a small number of iterations, say 10.
## IV. POWER ALLOCATION WITH RATE CONSTRAINTS
The problem of allocating power with rate constraints is solved in two steps. First, we try to allocate power without the rate constraints using one of the algorithms of section III since this is very fast. For the case where some of the RT users do not get their minimum rate, we propose in this section two algorithms that try to find a solution that is feasible both for the rate and power constraints, even though it may not be optimal. First we present an iterative algorithm that is guaranteed to produce a feasible solution with the rate constraints at their lower bounds. We then propose a non-iterative procedure that is faster but which is not guaranteed to find a feasible solution.
## A. Feasible Solution at Bounds
To simplify the discussion, consider first a problem with a single rate constraint, say for user 1. The boundaries of the two constraints (4-5) each define an implicit function δ 1 ( θ ) . We can see an example of these boundaries on Fig. 3 for a small network with K = 2 , N = 8 , P = 2 and a single real-time user. The set of feasible solution is the region above the rate constraint and below the power constraint.
A remarkable feature of this plot is that the boundary of the rate constraint seems to be linear. We already know that the boundary is piecewise linear but we now prove the following lemma.
Lemma 2. The boundaries of the rate constraints are linear.
Fig. 3. Domain for a single rate constraint
Proof: If θ ≈ 0 , all the p n,k ≈ ∞ and (5) will be a strict inequality. From the complementarity condition, we will then have δ k = 0 . This is consistent with the interpretation of θ as a penalty for the power constraint: If θ is small, there is no penalty for exceeding the power constraint so that there is a lot of power available. We can then choose large values of the p n,k to increase the total rate so that the rate constraint will be an inequality.
As θ increases, there will be a point θ k where each rate k constraint will become tight. This is given by
$$\overline { \theta } _ { k } = \frac { c _ { k } \sum _ { n \in \mathcal { B } _ { k } ( \overline { \theta } _ { k }, 0 ) } \beta _ { n, k } } { \check { d } _ { k } \log 2 }. \quad \quad \quad$$
At that point, we must increase δ k ≥ 0 to stay on the boundary. For all values of θ ≥ θ k , we choose a value of δ k which is a linear function of θ with slope m k
$$\delta _ { k } = [ m _ { k } \theta - c _ { k } ] ^ { + }$$
$$m _ { k } = \ln 2 \left ( 2 ^ { \delta _ { k } } \prod _ { n \in \mathcal { B } ( \theta, \delta _ { k } ) } \beta _ { n, k } \right ) ^ { \sigma ( \theta, \delta _ { k } ) ^ { - 1 } }.$$
With this choice of δ k , we know that the rate constraint in an equality over the whole range θ ≥ θ k and from this, we see that
$$\varepsilon \, \text{that} \\ \overline { p } _ { n, k } & = \left [ \frac { c _ { k } + \delta _ { k } } { \theta \log 2 \beta _ { n, k } } - 1 \right ] ^ { + } \\ & = \left [ \frac { c _ { k } +$$
which is independent of θ . In other words, once θ ≥ θ k , the set B k ( θ, δ k ) does not change and the rate constraint boundary is linear.
We see from Fig. 3 that the point ( θ (1) , 0) where the power constraint boundary would meet the θ axis, is outside the feasible domain. The problem is to quickly find a point both above the rate boundary and below the power boundary and, if possible, a good point.
We make use of the linearity of the boundary to compute a feasible solution quickly by computing the intersection of the rate and power boundaries. First, we compute the slopes m k of the rate boundaries from (20) at the point ( θ (1) , 0) . Given any value of θ , we can then compute all the δ k 's from (19) and from this, the value of the total power P θ ( ) and total rate R θ ( ) from (10). We can see an example of these curves on Fig. 4. We compute the solution by solving P θ ( ) = ˆ P by some numerical root-finding algorithm which is guaranteed to converge asymptotically. This is the best we can do if we impose the conditions that the rate constraints are met with equality since moving to the right on these boundaries means that the power constraint is not tight, which is a condition for an optimal solution.
## B. Fast Heuristic
As we will see later, the time needed to compute a solution on the boundary may turn out to be too large for use in real time. Instead, we propose the following simple, non-iterative algorithm. The algorithm is based on the observation that a standard way to satisfy some constraint k is to increase the corresponding multiplier δ k . From (10), we see that this will increase the p ,nk which will make the constraints more feasible. But this will also make the power constraint infeasible so that we need to increase θ as well. Let us define
- θ (1) the multiplier from the solution without rate constraints
- θ an upper bound on θ given by θ = Wθ (1) with W > 1 .
Fig. 4. Rate and Power Curves vs θ on the Rate Boundaries
Starting from θ = θ (1) , the heuristic basically moves a fixed amount in the direction of increasing θ . At that point, it computes the solution on the rate boundaries and then tries to improve θ one more time as follows:
- 1) Compute the set of unsatisfied users T = { k | r k < d ˇ k } . 2) Compute B k ( θ, 0)
/negationslash
- 3) For users k ∈ T , set δ (2) k = 0 . For the other users, compute δ (2) k from (13) with θ = θ and B k = B k ( θ, δ (2) k ) .
- 4) Compute a new θ (2) from (14) with δ = δ (2) and B k = B ( θ, δ (2) k ) .
## C. Improved Heuristic
The single iteration heuristic uses an arbitrary value W for the step size which may not produce a very good solution. We can improve the results if we choose the parameter W based on the difference between the required rate and the rate achieved by the maximum throughput PA. This increases the chance that the step will lead to a point to the right of the intersection of the two constraint curves. First we compute a different value of W for each real-time user
$$W _ { k } = 2 ^ { \epsilon ( \vec { d } _ { k } - r _ { k } ) }.$$
Fig. 5. Fast Heuristic, Feasible Solution
and then use the maximum W k among all RT users
$$W = \max _ { k \in \mathcal { T } } W _ { k }$$
$$\overline { \theta } = W \theta ^ { ( 1 ) }.$$
This way, we can increase the emphasis we want to put on meeting the rate constraints by increasing the value of /epsilon1 .
## D. Infeasibility
We can see on Fig. 5 the path of the heuristic in the ( θ, δ ) space with /epsilon1 = 0 2 . . In this case, the final point is feasible since it lies above the rate boundary and lies on the power boundary. We say that the algorithm has succeeded. On the other hand, it is possible for the algorithm not to succeed for a variety of reasons. A practical approach could then be to try the fast heuristic first. If it fails, one could either try the zero-finding solution with a limited number of iterations or simply use the infeasible solution.
The first one is that we don't know where the intersection of the two boundaries lie so that we don't know how far to the right we must go to. Choosing a value of /epsilon1 too small will not go
Fig. 6. Fast Heuristic, Infeasible Solution
far enough and the final point will not be feasible. This can be seen on Fig. 6 where we have chosen /epsilon1 = 0 . The point ( θ, 0) is to the right left of the intersection point of the two boundary curves so that moving to the right until we reach the power boundary still leaves us below the rate constraint. In this case, the algorithm has failed.
There is another, more subtle reason, why the algorithm may fail. Consider the final point of figure 5. We can see that the point is slightly to the left of the power boundary. The reason is that we compute θ (2) using B k ( θ, δ (2) k ) which is not the same as B k ( θ (2) , δ (2) k ) , hence the small error.
Another potential problem is that by increasing θ from θ (1) , the heuristic algorithm reduces the power of the users that are not in T . If some of these are RT users and have a rate close to ˇ d k , they may become unfeasible. A simple solution is to add these users to T by increasing their ˇ d k and run the fast heuristic again.
We can see that the algorithm produces a feasible point as follows. We assume that users close to the unfeasibility boundary are all included in set T by increasing their ˇ d k .
Lemma 3. When successful, the fast algorithm produces a feasible point to problem (3-5).
Proof: By construction, the final point lies on the power boundary. It also lies above the rate boundary for the following reason. The point ( θ, δ (2) k ) lies on the rate boundary by construction. Given that the slope of the rate boundary is positive, all the poinst ( θ, δ (2) k ) with θ < θ will also lie above the boundary.
Note also that this method can be used with an arbitrary method for computing the unconstrained power allocation, such as [5], [15], [6], as long as we get the value of θ (1) from the solution.
## V. NUMERICAL EVALUATION
In this section, we compare the three different methods to solve problem (3-7): An exact solution by a nonlinear solver [20], the boundary solution of IV-A and the fast heuristic of IV-C. We look at their accuracy and the cpu time needed to compute a solution. The nonlinear equations were solved by the netlib function bisect through the Scipy Python interface. All results were obtained on an Intel Core i5 cpu running at 2.5 GHz.
## A. Generating Test Cases
Each test case is defined first by the set ( K,N,M,P ) . We use a Rayleigh fading model to generate the user channels such that each component of the channel vectors h k,n are i.i.d. random variables distributed as CN (0 , 1) . We also assume independent fading between users, antennas and subcarriers. We assign the subchannels with a variant of the semiorthogonal user selection (SUS) heuristic [2] and then compute the β n,k from this. After this, we select the number of real-time users R . The set of values we have used is shown in Table II.
The next step is to generate the rate bounds in such a way that the problems are known to be feasible. Also, we would like to control the 'difficulty' of the problem to see how much this impacts on the accuracy and cpu time. First, we maximize the total rate of real-time users by solving problem (3-4) over the p n,k , k ∈ D only and without the rate constraints. Call this problem P 0 and R 0 the total rate of the real-time users. We then know there is enough power to give the real-time users the rates r (0) k produced by the unconstrained solution.
This means that any problem P 1 with rate constraints ˇ d k = r (0) k has a feasible solution with only the real-time users having a positive rate and with the rate constraints at their bounds. If this were not the case, there would be some spare power that could be used to construct a solution to P 0 with a total rate R > R 1 0 which contradicts the fact that R 0 is the optimal value for P 0 .
TABLE II PARAMETERS FOR TEST CASES
| Number | K | N | M | ˆ P | R |
|----------|-----|-----|-----|-------|-----|
| 1 | 20 | 25 | 2 | 5 | 3 |
| 2 | 20 | 25 | 2 | 5 | 10 |
| 3 | 80 | 25 | 2 | 5 | 0 |
| 4 | 80 | 25 | 2 | 5 | 10 |
| 5 | 80 | 25 | 2 | 5 | 20 |
| 6 | 80 | 25 | 2 | 5 | 30 |
| 7 | 80 | 25 | 2 | 5 | 40 |
| 8 | 80 | 25 | 2 | 5 | 50 |
| 9 | 80 | 25 | 2 | 5 | 60 |
We also know that any problem with rate constraints ˇ d k = sr (0) k is feasible if 0 ≤ s ≤ 1 . Furthermore, the value of s is a rough measure of the 'tightness' of the constrained data set. The problem is unconstrained when s = 0 and is tightly constrained when s = 1 where only the real-time users can get some power. The scale s will be the primary parameter used to display results.
## B. Effect of /epsilon1
First we examine the effect of /epsilon1 on the accuracy of the fast heuristic for problem 1. This is show on Fig. 7 where we plot the objective value computed by the heuristic for different values of /epsilon1 and also the optimal value computed by the nonlinear solver as a function of the scale parameter. For each curve corresponding to the heuristic, we can define a range of values of s where the heuristic value is below the optimal rate and one where it is above. In this latter case, this is an indication that the solution computed by the heuristic is not feasible. The curves show that there is a clear tradeoff between the accuracy of the solution and the range of problems where the solution is feasible. For small values of /epsilon1 , the range of problems where the heuristic can produce a feasible solution is quite small but the solutions are quite accurate. This is opposite to the curves with large /epsilon1 where the algorithm can compute feasible solutions over a larger range of problems but where the objective function solution is not as good.
Fig. 7. Effect of /epsilon1 on Heuristic
Fig. 8. Accuracy of Heuristics, Small Network, Few Real-Time Users
Fig. 9. Accuracy of Heuristics, 10 Real-time Users
## C. Solution Accuracy
Next we show the accuracy of the two heuristics by comparing the total rate they produce with the exact value as a function of the scale parameter. In all cases, we choose /epsilon1 = 0 2 . . The first case is for problem 1 shown on Fig. 8. Here, it turns out that all the rate constraints are at their bound in the optimal solution. This is not unexpected if we have a small number of real-time users: There is a good chance that other users may have a better channel so that once the constraints is satisfied, it is more useful to allocate the power to these non real-time users since they will get a better rate.
A similar result is shown for problem 2 on Fig. 9. This has the same parameters as problem 1 but this time with 10 real-time users. Here too we see that the rate constraints are at their bound in the optimal solution. The approximate heuristic does a rather poor job of finding a feasible solution for the value of /epsilon1 = 0 2 . that we have chosen.
The rate bounds in the example of Fig. 9 are computed from the problem data and in a sense they 'fit' the values of β . In practice, the rate requirement of real-time users would be determined by the application independently of the channel conditions. To see the accuracy of
Fig. 10. Accuracy for Fixed Rate Bounds, Problem 4
the solution techniques on these cases, we have produced the bounds somewhat differently. First, we assume that there is a small number of applications, three in the present case, with different rate bounds ˇ d (0) k in the ratio 1:4:16. For example, this could correspond to voice, fast data transfer and video. We then compute the largest scaling factor γ such that the problem is feasible with bounds ˇ d k = γd ˇ (0) k . We can then control how difficult the problem is by selecting some scale between 0 and 1 to scale the bounds ˇ d k as above.
We show on Fig. 10 the accuracy of the algorithm for problem 4 with the given rate bounds. We can see that the solution on the boundary is still very close to the optimal solution and the fast heuristic produces an infeasible solution. Note however that this case has only 10 real-time users out of 80.
The situation is quite different if we have 60 real-time users as in problem 9. These users are divided into 3 groups of 20 and all users in a group have the same requirement in the order 1:4:16. We can see from Fig. 11 that the solution on the rate bounds is much lower than the optimal value for quite a large range of scales and the fast heuristic again produces infeasible solutions.
Fig. 11. Accuracy for Fixed Rate Bounds, Problem 9
Fig. 12. Difference with Rate Bounds
Fig. 13. CPU Time vs No of RT Users
We can get a better understanding of these results from Fig. 12 where we plot, for each realtime user and scale 1, the difference r k -ˇ d k for the optimal solution and the approximation. As expected, in the optimal solution, all users are above zero, since the solution has to be feasible. More important is the fact that for the first two groups, the ones with the lower bounds, most users actually get more that their minimum rate. This explains the poor behavior of the bounds algorithms which computes an optimal solution on the boundary. Also interesting is the approximate solution for the third group, the one with high rate. We see that the approximate solution is infeasible but that the amount of infeasibility is rather small. Note that this is a relatively hard problem in the sense that the problem is barely feasible in the first place. Given that the solution produced by the bounds algorithm is not very good, a reasonable solution would then be to use the approximation and give some real-time users slightly less than what they require. This may be acceptable considering that the channel conditions will change at the next time frame and that it may then be possible to give these users the rate that they require.
## D. CPU Time
We present in Fig. 13 the cpu time required by the nonlinear solver, labelled exact , the algorithm on the rate boundary, labelled bounds and the heuristic, labelled heur . The values of the computation times are plotted as a function of the number of real-time users R for problems 3-9. For each value of R , we have measured the average, maximum and minimum cpu time required for solving the problem over the whole range of values of s . The maximum and minimum values are presented as error bars. We can see that there is about one order of magnitude between each algorithm. The heuristic takes in the order of 10 ms while the bounds algorithm, about 10 times as much.
## VI. CONCLUSIONS
We have looked at the problem of allocating power to a given set of users in a ZF OFDMASDMA system with some real-time users with minimum transmission rate requirements. The approach is to solve the problem in two steps. First, we solve without taking into account the rate constraints. For this, we provide two fast algorithms. The first one is based on the solution of the first-order optimality condition via a zero-finding algorithm. We use the structure of the problem to perform a search over a finite number of intervals for the one containing the solution. From this, we get the exact value of the solution so that the algorithm is guaranteed to find a solution in O (log 2 NK ) steps.
The second algorithm is based on re-writing the optimality conditions as a fixed-point equation which we can solve by repeated substitution. This is faster than the zero-finding method but there is no guarantee that it will converge. We have found that in practice, convergence always occurs in a small number of iterations for all the cases we have tested.
If the solution is feasible, this is the optimal solution for the constrained problem as well. If some real-time users do not get their minimum rate, we propose two approximate techniques to compute a feasible, or nearly feasible solution by adjusting the dual multipliers. First, we show that the boundary of the rate constraints is linear in the ( θ, δ ) plane. We then use this to propose an algorithm to compute a solution on the rate boundaries. This is based on finding a zero of the first-order equations for the value of θ only, since we can use the linearity of the rate boundary to keep the solution feasible.
The second approximation is not iterative and is a one-step adjustment of the multipliers. It is parameterized in such a way that we can emphasize either feasibility, at the cost of having a somewhat less than optimal value for the rate, or efficiency, where the rate is relatively large but some real-time users get somewhat less than their required rate.
Finally, we present some results for the cpu time needed by three algorithms: the exact solution by a nonlinear solver, the bounds algorithm and the approximation. We find that there is roughly one order of magnitude between each, where the bounds algorithm is an order of magnitude faster than the exact solution, and the fast heuristic is an additional order of magnitude faster.
## REFERENCES
- [1] A. Wiesel, Y. Eldar, and S. Shamai, 'Optimal generalized inverses for zero forcing precoding,' in Proc. of 41st Annual Conf. on Information Sciences and Systems , 2007.
- [2] T. Yoo and A. Goldsmith, 'On the optimality of multiantenna broadcast scheduling using zero-forcing beamforming,' IEEE Journal on Selected Areas in Communications , vol. 24, no. 3, pp. 528-541, Mar. 2006.
- [3] X. Wang and G. Giannakis, 'Resource allocation for wireless multiuser OFDM networks,' IEEE Trans. in Information Theory , vol. 57, no. 7, pp. 4359-4372, 2011.
- [4] D. Perez-Palomar and M. Lagunas, 'Joint transmit-receive space-time equalization in spatially correlated MIMO channels: a beamforming approach,' IEEE Journal on Selected Areas in Communications , vol. 21, no. 5, pp. 730-743, 2003.
- [5] X. Ling, B. Wu, P.-H. Ho, F. Luo, and L. Pan, 'Fast water-filling for agile power allocation in multi-channel wireless communications,' IEEE Communications Letters , vol. 16, no. 8, pp. 1212-1215, 2012.
- [6] J. Jang and a. Y. L. K. B. Lee, 'Transmit power and bit allocations for OFDM systems in a fading channel,' in Proc. IEEE Global Communication Conference, Globecom , vol. 2, 2003, pp. 858-862.
- [7] G. Dimic and N. Sidiropoulos, 'On downlink beamforming with greedy user selection: performance analysis and a simple new algorithm,' IEEE Transactions on Signal Processing , vol. 53, no. 10, pp. 3857-3868, 2005.
- [8] D. Perea-Vega, J. Frigon, and A. Girard, 'Near-optimal and efficient heuristic algorithms for resource allocation in MISOOFDM systems,' in IEEE International Conference on Communications ICC , May 2010, pp. 1-6.
- [9] V. Papoutsis and S. Kotsopoulos, 'Resource Allocation Algorithm for MISO-OFDMA Systems with QoS Provisioning,' in Proc. ICWMC, The Seventh International Conference on Wireless and Mobile Communications , Jun. 2011.
- [10] D. Perea-Vega, J. Frigon, and A. Girard, 'Efficient Heuristic for Resource Allocation in Zero-forcing OFDMA-SDMA Systems with Minimum Rate Constraints,' in - In preparation, available on-line at ... , 2013.
- [11] D. Bertsekas, Convex Analysis and Optimization . Athena Scientific - Belmont, MA, 2003.
- [12] 3GPP, '3GPP: Evolved Universal Terrestrial Radio Access (E-UTRA); Base Station radio transmission and reception; Further advancements for E-UTRA physical layer aspects,' 3GPP TR V10.5, Tech. Spec.n Group Radio Access Network.
- [13] J. Jang and K. Lee, 'Transmit power adaptation for multiuser OFDM systems,' IEEE Journal on selected areas in Communications , vol. 21, no. 2, pp. 171-178, Feb. 2003.
- [14] M. Anas, K. Kim, S. Shin, and K. Kim, 'QoS Aware power allocation for combined guaranteed performance and best effort users in OFDMA systems,' in Proc. of International Symposium on Intelligent Signal Processing and Communication Systems , 2004, pp. 477-481.
- [15] W. Yu, W. Rhee, S. Boyd, and J. Cioffi, 'Iterative water-filling for gaussian vector multiple-access channels,' IEEE Transactions on Information Theory , vol. 50, no. 1, pp. 145-152, 2004.
- [16] S. Zhu, G. Lv, and H. Hui, 'A low complexity heuristic adaptive resource llocation algorithm for multiuser OFDM under rate constraints,' in Communications and Networking in China, 2009. ChinaCOM 2009. Fourth International Conference on , Aug 2009, pp. 1-4.
- [17] D. Kivanc, G. Li, and H. Liu, 'Computationally efficient bandwidth allocation and power control for OFDMA,' Wireless Communications, IEEE Transactions on , vol. 2, no. 6, pp. 1150-1158, Nov 2003.
- [18] L. V. S. Boyd, Convex Optimization . Cambridge University Press, 2004.
- [19] M. Tao, Y.-C. Liang, and F. Zhang, 'Resource allocation for delay differentiated traffic in multiuser OFDM systems,' IEEE Trans. on Wireless Communications , vol. 7, no. 6, pp. 2190-2201, 2008.
- [20] B. A. Murtaugh and M. A. Saunders, 'MINOS 5.4 users guide,' Department of Operations Research, Stanford University, Stanford, CA 94305 USA, Tech. Rep. SOL 83-20R, Dec. 1993.
θ
B
| null |
[
"Diego Perea-Vega",
"Andre Girard",
"Jean-Francois Frigon"
] |
2014-08-05T17:32:25+00:00
|
2014-11-01T20:43:44+00:00
|
[
"cs.IT",
"cs.NI",
"math.IT"
] |
Fast Heuristics for Power Allocation in Zero-Forcing OFDMA-SDMA Systems with Minimum Rate Constraints
|
We investigate in this paper the optimal power allocation in an OFDM-SDMA
system when some users have minimum downlink transmission rate requirements. We
first solve the unconstrained power allocation problem for which we propose a
fast zero-finding technique that is guaranteed to find an optimal solution, and
an approximate solution that has lower complexity but is not guaranteed to
converge. For the more complex minimum rate constrained problem, we propose two
approximate algorithms. One is an iterative technique that finds an optimal
solution on the rate boundaries so that the solution is feasible, but not
necessarily optimal. The other is not iterative but cannot guarantee a feasible
solution. We present numerical results showing that the computation time for
the iterative heuristic is one order of magnitude faster than finding the exact
solution with a numerical solver, and the non-iterative technique is an
additional order of magnitude faster than the iterative heuristic. We also show
that in most cases, the amount of infeasibility with the non-iterative
technique is small enough that it could probably be used in practice.
|
1408.1045v3
|
## LOOP PERCOLATION ON DISCRETE HALF-PLANE
## TITUS LUPU
Abstract. We consider the random walk loop soup on the discrete half-plane Z × N ∗ and study the percolation problem, i.e. the existence of an infinite cluster of loops. We show that the critical value of the intensity is equal to 1 2 . The absence of percolation at intensity 1 2 was shown in a previous work. We also show that in the supercritical regime, one can keep only the loops up to some large enough upper bound on the diameter and still have percolation.
## 1. Introduction
We will consider discrete (rooted) loops on Z 2 , that is to say finite paths to the nearest neighbours on Z 2 that return to the origin and visit at least two vertices. The rooted random walk loop measure µ Z 2 gives to each rooted loop of lengths 2 n the mass (2 n ) -1 4 -2 n . It was introduced in [5]. In [3] are considered loops parametrised by continuous time rather than discrete time. µ Z 2 has a continuous analogue, the measure µ C on the Brownian loops on C . Let P t z,z ′ ( ) · be the standard Brownian bridge probability measure from z to z ′ of length t . µ C is a measure on continuous time-parametrised loops on C defined as where dz ¯ ∧ dz 2 i is the standard volume form on C . The measure µ C was introduced in [6].
$$\text{parametrsed loops on $\mathcal{C}$ defined as} \\ \mu _ { \mathbb { C } } ( \cdot ) \coloneqq \int _ { \mathbb { C } } \int _ { t > 0 } \mathbb { P } _ { z, z } ^ { t } ( \cdot ) \frac { d t } { 2 \pi t ^ { 2 } } \frac { d \bar { z } \wedge d z } { 2 i }, \\ \intertext { a t a n d } \text{volume form on $\mathcal{C}$ } \text{The macro } \text{$\mathcal{C}$}$$
Given α > 0 we will denote by L Z 2 α respectively L C α the Poisson ensemble of intensity αµ Z 2 respectively αµ C , called random walk respectively Brownian loop soup. In [5] it was shown that one can approximate L C α by a rescaled version of L Z 2 α . If A is a subset of Z 2 we will denote by L A α the subset of L Z 2 α made of loops contained in A . If U is an open subset of C we will denote by L U α the subset of L C α made of loops contained in U . For δ > 0 we will denote by L A, ≥ δ α respectively L U, ≥ δ α the subset of random walk loops L A α respectively Brownian loops L U α made of loops of diameter greater or equal to δ . Similarly we will use the notation L A, ≤ δ α for the loops of diameter smaller or equal to δ .
We will consider clusters of loops. Two loops γ and γ ′ in a Poisson ensemble of discrete or continuous loops belong to the same cluster if there is a chain of loops γ , γ 0 1 , . . . , γ n in this Poisson ensemble such that γ 0 = γ , γ n = γ ′ and γ i and γ i -1 visit a common point. For all α > 0 , loops in L Z 2 α as well as in L C α form a single cluster. Thus we will consider loops on discrete half-plane H = Z × N ∗ and on continuous half-plane H = { z ∈ C |/Ifractur ( z ) > 0 } , mainly from the angle of existence of an unbounded cluster.
The percolation problem for Brownian loops was studied in [10]. It was shown that there is a critical intensity α H ∗ ∈ (0 , + ∞ ) such that for α ∈ (0 , α H ∗ ] , L H α has only bounded clusters, and for α > α H ∗ the loops in L H α H form one single cluster.
∗
E-mail address : [email protected], [email protected] .
The critical intensity was identified to be equal to 1 . But actually α H ∗ = 1 2 . In [10] the outer boundaries of outermost clusters in a sub-critical Brownian loop soup were identified to be a Conformal Loop Ensemble CLE κ with the following relation between α and κ .
$$\alpha = \frac { ( 3 \kappa - 8 ) ( 6 - \kappa ) } { 2 \kappa }. \\ \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots$$
The critical value of κ corresponds to CLE 4 . Actually the right relation between α and κ is
So the value of α that corresponds to κ = 4 is 1 2 and not 1 . The missing factor 1 2 appears in the Lawler's work [7] (Proposition 2 1 . ). The error in [10] comes from an error in the article [6] by Lawler and Werner. There the authors consider a Brownian loop soup in the half-plane and a continuous path cutting the half-plane, parametrised by the half-plane capacity. For such a path the half-plane capacity at time t equals 2 t . It discovers progressively new Brownian loops and the authors map these loops conformally to the origin. In the Theorem 1 they identify the processes of these conformally mapped Brownian loops to be a Poisson point process with intensity proportional to the Brownian bubble measure. In the identification of intensity there is a factor 2 missing. Actually in the article [6], the Theorem 1 is inconsistent with the Proposition 11 .
$$\alpha = \frac { 1 } { 2 } \frac { ( 3 \kappa - 8 ) ( 6 - \kappa ) } { 2 \kappa }. \\ \dots \dots \dots \dots \dots$$
The problem of percolation by random walk loops was studied in [4], [2], [9] and [1] in more general setting than dimension 2 . We will focus on the percolation by loops in L H α . The probability of existence of an infinite cluster of loops follows a 0 -1 law and there can be at most one infinite cluster ([9]). Moreover for α = 1 2 loops in L H 1 2 do not percolate ([9]). This result was obtained through a coupling with the massless Gaussian free field. By considering just the loops that go back and forth between two neighbouring vertices we get a lower bound on clusters of loops by clusters of an i.i.d. Bernoulli percolation. In particular this implies that for α large enough loops in L H α percolate. Hence as the parameter α increases there is a phase transition and a critical value α H ∗ ∈ [ 1 2 , + ∞ ) of the parameter. Using the results on the clusters of Brownian loops from [10] and the approximation result from [5] we will show in section 2 the following:
Theorem 1. For all α > 1 2 there is an infinite cluster of loops in L H α . In particular α H ∗ = 1 2 . Moreover, given α > 1 2 , there is n ∈ N ∗ large enough, such that L H , ≤ n α percolates too.
That is to say the critical intensity parameter for the two-dimensional Brownian loop soups and random walk loop soups is the same.
Wewill consider 1 -dependent edge percolations on H , ( ω e ( )) e edge . By 1 -dependent percolation we mean that if two disjoint subsets of edges E 1 and E 2 are at graph distance at least 1 then ( ω e ( )) e ∈ E 1 and ( ω e ( )) e ∈ E 2 are independent. According the results on locally dependent percolation by Liggett, Schonmann and Stacey in [8], for all 1 -dependent edge percolations on H with p the probability of an edge to be open, there is an universal ˜( p p ) ∈ [0 , 1) such that the 1 -dependent edge percolation contains an i.i.d. Bernoulli percolation with probability ˜( p p ) of an edge to be open. Moreover the following constraint holds:
$$\lim _ { p \to 1 ^ { - } } \tilde { p } ( p ) = 1.$$
## 2. Critical intensity parameter
Let α, δ > 0 . Given U an open subset of H , we will denote by L U, ≥ δ α respectively L U ∩ H , ≥ δ α the subset of L H α respectively L H α made of loops contained in U and with
diameter greater or equal to δ . We will use the notations L U α and L U ∩ H α when there is a condition on the range but not on the diameter.
Let Q ext and Q int be the following rectangles:
$$Q _ { e x t } \coloneqq ( 0, 6 ) \times ( 0, 3 ), \ \ Q _ { i n t } \coloneqq ( 1, 5 ) \times ( 1, 2 ).$$
We consider the subset of Brownian loops L Q ext , ≥ δ α , which is a.s. finite. We introduce the events C 1 ( L Q ext , ≥ δ α ) , C 2 ( L Q ext , ≥ δ α ) and C 3 ( L Q ext , ≥ δ α ) depending on the loops in L Q ext , ≥ δ α . The event C 1 ( L Q ext , ≥ δ α ) will be satisfied if there is a cluster K 1 of loops in L Q int , ≥ δ α such that in L (0 6) , × (1 2) , , ≥ δ α there is a loop that intersects K 1 and { 1 }× (1 , 2) and a loop that intersects K 1 and { 5 }× (1 , 2) . The two loops may be the same. C 2 ( L Q ext , ≥ δ α ) will be satisfied if there is a cluster K 2 in L (1 2) , 2 , ≥ δ α such that in L (1 2) , × (0 3) , , ≥ δ α there is a loop that intersects K 2 and (1 , 2) ×{ } 1 and a loop that intersects K 2 and (1 , 2) ×{ } 2 . The event C 3 ( L Q ext , ≥ δ α ) is similar to the event C 2 ( L Q ext , ≥ δ α ) where the square (1 , 2) 2 is replaced by the square (4 , 5) × (1 , 2) and the rectangle (1 , 2) × (0 , 3) by the rectangle (4 , 5) × (0 , 3) . Next figure illustrates the event ⋂ 3 i =1 C i ( L Q ext , ≥ δ α ) .
Fig.1: Illustration of the event ⋂ 3 i =1 C i ( L Qext , ≥ δ α ) .
One should imagine that the smooth loops are actually Brownian. Only a set of loops that is sufficient for the event is represented. Full line loops stay inside Q int . Dashed loops cross the boundary of Q int .
We will call the event ⋂ 3 i =1 C i ( L Q ext , ≥ δ α ) special crossing event with exterior rectangle Q ext and interior rectangle Q int . We will also consider translations, rotations and rescaling of Q ext and Q int and deal with special crossing events corresponding to the new rectangles. We are interested in the event ⋂ 3 i =1 C i ( L Q ext , ≥ δ α ) because then the loops in L Q ext , ≥ δ α achieve the three crossings drawn on the figure 2 :
Fig.2: The three crossings we are interested in.
Next we show that if α > 1 2 and δ is small enough then the probability of ⋂ 3 i =1 C i ( L Q ext , ≥ δ α ) is close to 1 .
Lemma 2.1. Let Q be a rectangle of form Q = ( -a, a ) × (0 , b ) . Let α > 0 . Let ( B t ) t ≥ 0 be the standard Brownian motion on C started from 0 and let L Q α be a Poisson ensemble of loops independent from B . Then for all ε > 0 there is t ∈ (0 , ε ) such that B at time t intersects a loop in L Q α .
Proof. First we consider a loops soup in H , L H α , independent of B . Let
$$T \coloneqq \inf \{ t > 0 | B _ { t } \text{ is in the range of a loop in } \mathcal { L } _ { \alpha } ^ { \mathbb { H } } \}.$$
T is a.s. finite. Indeed a loop in L H α delimits a domain with non-empty interior. Since the Brownian motion on C is recurrent, B will visit this domain and thus intersect the loop. Let λ > 0 . The Poisson ensemble of loops L H α is invariant in law under the Brownian scaling
↦
$$( \gamma ( t ) ) _ { 0 \leq t \leq t _ { \gamma } } \longmapsto \lambda ^ { - \frac { 1 } { 2 } } ( \gamma ( \lambda t ) ) _ { 0 \leq t \leq \lambda ^ { - 1 } t _ { \gamma } }.$$
So does the Brownian motion B . Thus λT has the same law as T . It follows that T = 0 a.s.
The set of loops L H α \ L Q α is at positive distance from 0 thus B cannot intersect it immediately. It follows that B intersects immediately L Q α . /square
Lemma 2.2. Let a, α > 0 . There is a.s. a loop in L ( -a,a ) 2 α that intersects the real line R .
Proof. Let L ( n ) α be the subset of L ( -a,a ) 2 α made of loops γ of duration t γ comprised between 2 - -n 1 and 2 -n . The family ( L ( n ) α ) n ≥ 0 is independent. By Brownian scaling, the probability that a loop in L ( n ) α intersects R is the same as a loop in L ( -a 2 n/ 2 ,a 2 n/ 2 ) 2 α of duration comprised between 1 2 and 1 intersects R . This is at least as big as the similar probability for L (0) α . Since the latter probability is non-zero, the intersection events occurs a.s. for infinitely many of L ( n ) α . /square
Lemma 2.3. Let a, α > 0 . There is a.s. a loop in L ( -a,a ) 2 α that intersects the real line R and a loop in L ( -a,a ) × (0 ,a ) α .
Proof. Consider the subset of L ( -a,a ) 2 α made of loops intersecting R . It is non empty according the lemma 2.2. Moreover it is independent of L ( -a,a ) × (0 ,a ) α . The law of a Brownian loop that intersects R is locally, near the point of intersection, absolutely continuous with respect to the law of a Brownian motion started from there. Applying lemma 2.1, we get that it intersects a.s. a loop in L ( -a,a ) × (0 ,a ) α . /square
Lemma 2.4. Let α > 1 2 . Then
$$\hat { \lim _ { \delta \to 0 ^ { + } } \mathbb { P } \left ( \bigcap _ { i = 1 } ^ { 3 } C _ { i } ( \mathcal { L } _ { \alpha } ^ { Q _ { e x t }, \geq \delta } ) \right ) } = 1. \\ \intertext { t o show that the r o b a b i l i t v o f a c h o f t }$$
Proof. It is enough to show that the probability of each of the C i ( L Q ext , ≥ δ α ) converges to 1 as δ tends to 0 . Since the three cases are very similar, we will do the proof only for C 1 ( L Q ext , ≥ δ α ) . According to lemma 2.3 there is a loop γ in L (0 6) , × (1 2) , α that intersects { 1 } × (1 , 2) and a loop γ ′ in L Q int α . Similarly there is a loop ˜ γ in L (0 6) , × (1 2) , α that intersects { 5 } × (1 , 2) and a loop ˜ γ ′ in L Q int α . Since α > 1 2 , γ ′ and ˜ γ ′ belong to the same cluster in L Q int α ([10]). Thus there is a chain of loops ( γ , . . . , γ 0 n ) in L Q int α , with γ 0 = γ ′ and γ n = ˜ γ ′ , joining γ ′ and ˜ γ ′ . If δ is the minimum of diameters of ( γ , . . . , γ 0 n ) and γ and ˜ γ then C 1 ( L Q ext , ≥ δ α ) is satisfied.
Let ¯ δ be maximal value of δ such that C 1 ( L Q ext , ≥ δ α ) is satisfied. ¯ δ is a well defined random variable with values in (0 , + ∞ ) . Then
$$\lim _ { \delta \to 0 ^ { + } } \mathbb { P } ( C _ { 1 } ( \mathcal { L } _ { \alpha } ^ { Q _ { e x t }, \geq \delta } ) ) = \lim _ { \delta \to 0 ^ { + } } \mathbb { P } ( \delta \leq \bar { \delta } ) = 1.$$
/square
Next we recall the result on approximation of Brownian loops by random walk loops from [5]. Let N ∈ N ∗ . We consider the discrete loops γ on Z × N ∗ . We define on these loops a map Φ N to continuous loops on H . Given γ a discrete loop and ( z , . . . , z 0 n -1 , z 0 ) the sequence of the vertices it visits, the continuous loop Φ N γ satisfies:
- · the duration of Φ N γ is n 2 N 2 ;
- · Φ N γ ( n 2 N 2 ) = Φ N γ (0) = z 0 N ;
- · for j ∈ { 0 , . . . , n - } 1 , Φ N γ ( j 2 N 2 ) = z j N ;
- · between the times j 2 N 2 , j ∈ { 0 , . . . , n } , Φ N γ interpolates linearly.
The number of jumps n of a discrete loop γ will be denoted s γ . The life-time of a continuous loop ˜ γ will be denoted by t ˜ γ . Let θ ∈ ( 2 3 , 2) and r ≥ 1 . There is a coupling between L H α and L H α such that except on an event of probability at most cste · ( α +1) r 2 N 2 -3 θ there is a one to one correspondence between the two sets
- · { γ ∈ L H α | s γ > N , γ 2 θ | (0) | < Nr } ,
- · { ˜ γ ∈ L H α | t ˜ γ > N θ -2 , | ˜(0) γ | < r } ,
such that given a discrete loop γ and the continuous loop ˜ γ corresponding to it,
$$& \text{such that given a discrete loop } \gamma \text{ and the continuous loop } \dot { \gamma } \text{ corresponding to it,} \\ & \left | \frac { s _ { \gamma } } { 2 N ^ { 2 } } - t _ { \tilde { \gamma } } \right | \leq \frac { 5 } { 8 } N ^ { - 2 }, \quad \sup _ { 0 \leq u \leq 1 } \left | \Phi _ { N } \gamma \left ( u \frac { s _ { \gamma } } { 2 N ^ { 2 } } \right ) - \tilde { \gamma } ( u t _ { \tilde { \gamma } } ) \right | \leq c s t e \cdot N ^ { - 1 } \log ( N ). \\ & \text{Next we state without proof a lemma that follows immediately from this approxi-} \\ & \text{mat}.$$
∣ Next we state without proof a lemma that follows immediately from this approximation.
Lemma 2.5. Let α > 0 and δ > 0 . As N tends to + ∞ the random set of interpolating continuous loops converges in law to the set of Brownian loops L Q ext , ≥ δ α .
$$\left \{ \Phi _ { N } \gamma | \gamma \in \mathcal { L } _ { \alpha } ^ { N Q _ { e x t } \cap \mathbf H, \geq N \delta } \right \} \\ \text{set of } B r o w n i a n \ l o o p s \ \mathcal { L } _ { \alpha } ^ { Q _ { e x t }, \geq \delta }. \\. \quad. \quad. \quad.$$
We need to show that the above convergence for the uniform norm also implies a convergence of the intersection relations, that is to say that
$$\{ ( \gamma, \gamma ^ { \prime } ) | \gamma, \gamma ^ { \prime } \in \mathcal { L } _ { \alpha } ^ { N Q _ { e x t } \cap \mathfrak H, \geq N \delta }, \gamma \text{ intersects } \gamma ^ { \prime } \}$$
converges in law to
$$\{ ( \tilde { \gamma }, \tilde { \gamma } ^ { \prime } ) | \tilde { \gamma }, \tilde { \gamma } ^ { \prime } \in \mathcal { L } _ { \alpha } ^ { Q _ { e x t }, \geq \delta }, \tilde { \gamma } \text{ intersects } \tilde { \gamma } ^ { \prime } \}.$$
Let j ∈ N . Let γ be a continuous path on C (not necessarily a loop) of lifetime t γ . For r > 0 let
$$T _ { r } ( \gamma ) \coloneqq \inf \{ s > 0 | | \gamma ( s ) | \geq r \} \in ( 0, + \infty ].$$
If T r ( γ ) < + ∞ let
$$e ^ { i \omega _ { r } } \coloneqq \frac { \gamma ( T _ { r } ( \gamma ) ) } { r }.$$
Let I j be the real interval
$$I _ { j } & \coloneqq \left ( \frac { 7 } { 1 2 } 2 ^ { - j }, \frac { 9 } { 1 2 } 2 ^ { - j } \right ). \\ r _ { 2 } \right ) & \text{ be the annulus}$$
For 0 < r 1 < r 2 let A ( r 1 , r 2 ) be the annulus
$$\mathcal { A } ( r _ { 1 }, r _ { 2 } ) \coloneqq \{ z \in \mathbb { C } | r _ { 1 } < | z | < r _ { 2 } \}.$$
For r > 0 let HD r ( ) be the half-disc
$$H D ( r ) \coloneqq B ( 0, r ) \cap \{ z \in \mathbb { C } | \Re ( z ) > 0 \}.$$
We will say that the path γ satisfies the condition C j if
- · T 11 12 2 -j ( γ ) < + ∞ ,
- · after time T 11 12 2 -j ( γ ) < + ∞ , γ hits e i ( ω 2 - -j 1 + π 2 ) I j at a time ˜ t j before hitting the circle S (0 , 2 -j ) ,
- · on the time interval ( T 2 - -j 1 ( γ , t ) ˜ ) j γ stays in the half-disc e iω 2 - -j 1 HD (2 -j ) ,
- · from time ˜ t j the path γ stays in the annulus A ( 7 12 2 -j , 9 12 2 -j ) until surrounding the disc B (0 , 7 12 2 -j ) once clockwise and hitting e i ( ω 2 - -j 1 + ) π I j .
Figure 3 illustrates a path satisfying the condition C j . If this condition is satisfied than γ disconnects the disc B (0 , 7 12 2 -j ) from infinity. Moreover if one perturbs γ by any continuous function f : [0 , t γ ] → C such that ‖ f ‖ ∞ ≤ 1 12 2 -j then the path ( γ s ( ) + f ( s )) 0 ≤ ≤ s t γ disconnects the disc B (0 , 2 - -j 1 ) from infinity. The disconnection is made inside the annulus A (2 - -j 1 , 2 -j ) .
/Bullet
.
Fig.3: Representation of a path γ satisfying the condition C j
Lemma 2.6. Let ( B t ) 0 ≤ ≤ t T be a standard Brownian path on C starting from 0 . Then almost surely it satisfies the condition C j for infinitely many values of j ∈ N .
Proof. Let ˜ B be the Brownian path B continued for t ∈ (0 , + ∞ ) . The events ' ˜ B satisfies the condition C j ' are i.i.d. Indeed such an event is rotation invariant and depends only on ˜ B on the time interval ( T 2 - -j 1 ( ˜ B , T ) 2 -j ( ˜ B )) . Moreover the probability of such an event is non-zero. Thus ˜ B satisfies the condition C j for infinitely many values of j ∈ N . Since so does B .
$$\lim _ { j \to + \infty } T _ { 2 ^ { - j } } ( \widetilde { B } ) = 0,$$
/square
Lemma 2.7. Let z , z 1 2 ∈ C and t 1 , t 2 > 0 . Let ( b (1) s ) 0 ≤ ≤ s t 1 and ( b (2) s ) 0 ≤ ≤ s t 2 be two independent standard Brownian bridges from z 1 to z 1 and z 2 to z 2 respectively. On the event that b (1) intersects b (2) there is a.s. ε > 0 such that for all continuous functions f 1 : [0 , t 1 ] → C and f 2 : [0 , t 2 ] → C of infinity norm ‖ f i ‖ ∞ ≤ ε , ( b (1) s + f 1 ( s )) 0 ≤ ≤ s t 1 intersects ( b (2) s + f 2 ( s )) 0 ≤ ≤ s t 2 .
Proof. Let T (1) 2 be the first time b (1) hits the range of b (2) . If the two path do not intersect each other T (1) 2 = + ∞ . On the event T (1) 2 < + ∞ the conditional
law of ( b (1) T (1) 2 + s -b (1) T (1) 2 ) 0 ≤ ≤ s t 1 -T (1) 2 -ε ( ε > 0 a small constant) given the value T (1) 2 is absolutely continuous with respect the law of a Brownian path starting from 0 . From lemma 2.6 follows that the path ( b (1) T (1) 2 + s -b (1) T (1) 2 ) 0 ≤ ≤ s t 1 -T (1) 2 satisfies the condition C j for infinitely many values of j ∈ N . Let
$$\ell _ { \ } \lambda \left \{ j \in \mathbb { N } | ( b ^ { ( 1 ) } _ { T ^ { ( 1 ) } _ { 2 } + s } - b ^ { ( 1 ) } _ { T ^ { ( 1 ) } _ { 2 } } ) _ { 0 \leq s \leq t _ { 1 } - T ^ { ( 1 ) } _ { 2 } } \text{ satisfies the condition } \mathcal { C } _ { j } \\ \text{ and } \exists s \in [ 0. t \cap l. \, | b ^ { ( 2 ) } _ { \ } - b ^ { ( 2 ) } _ { \ } | > \frac { 1 3 } { } 2 ^ { - j } \right \}.$$
˜ is a r.v. defined on the event where b (1) and b (2) intersect. If f 1 and f 2 are such that ‖ f i ‖ ≤ 1 12 2 -˜ then the path b (1) + f 1 disconnects the disc B b ( (1) T (1) 2 , 2 - -˜ 1 ) from infinity inside the annulus b (1) T (1) 2 + A (2 - -˜ 1 , 2 -˜ ) and the path b (2) + f 2 crosses from the circle S b ( (1) T (1) 2 , 2 - -˜ 1 ) to the circle S b ( (1) T (1) 2 , 2 -˜ ) , so the two must intersect. /square
$$\L ^ { \prime \prime } _ { 0 \leq s \leq t _ { 1 } - \L T _ { 2 } ^ { \prime \prime } } & \quad \L ^ { \prime } \\ \text{ and } \exists s \in [ 0, t _ { 2 } ], | b _ { s } ^ { ( 2 ) } - b _ { T _ { 2 } ^ { ( 1 ) } } ^ { ( 2 ) } | & \geq \frac { 1 3 } { 1 2 } 2 ^ { - j } \Big \}. \\ \text{ here } b ^ { ( 1 ) } \text{ and } b ^ { ( 2 ) } \text{ intersect. If } f _ { 1 } \text{ and } f _ { 0 } \text{ are such}$$
Observe that two discrete loops γ and γ ′ intersect each other if and only if the continuous loops Φ N γ and Φ N γ ′ do. From lemmas 2.5 and 2.7 follows:
Corollary 2.8. Let α > 0 and δ > 0 . As N tends to + ∞ the random set of interpolating continuous loops
$$\sigma \, \upsilon \upsilon _ { P \sigma } \\ \left \{ \Phi _ { N } \gamma | \gamma \in \mathcal { L } _ { \alpha } ^ { N Q _ { e x t } \cap \mathbf H, \geq N \delta } \right \} \\ \ t i o n \ r e l a t i o n s$$
jointly with the intersection relations
$$\{ ( \gamma, \gamma ^ { \prime } ) | \gamma, \gamma ^ { \prime } \in \mathcal { L } _ { \alpha } ^ { N Q _ { e x t } \cap \mathfrak H, \geq N \delta }, \gamma \ i n t e r s e c t s \, \gamma ^ { \prime } \}$$
converges in law to the set of Brownian loops L Q ext , ≥ δ α jointly with the intersection relations
$$\{ ( \tilde { \gamma }, \tilde { \gamma } ^ { \prime } ) | \tilde { \gamma }, \tilde { \gamma } ^ { \prime } \in \mathcal { L } _ { \alpha } ^ { Q _ { e x t }, \geq \delta }, \tilde { \gamma } \ i n t e r s e c t s \ \tilde { \gamma } ^ { \prime } \}.$$
We consider the scaled up rectangle NQ ext and NQ int . The next lemma deals with the probability that the discrete loops L NQ ext ∩ H α realise the special crossing event with exterior rectangle NQ ext and interior rectangle NQ int . See figures 1 and 2 and consider that Q ext is replaced by NQ ext , Q int by NQ int and L Q ext , ≥ δ α by L NQ ext ∩ H α .
Lemma 2.9. Let α > 1 2 . As N tends to + ∞ , the probability that the loops L NQ ext ∩ H α realise a special crossing event with exterior rectangle NQ ext and interior rectangle NQ int converges to 1 .
Proof. Let δ > 0 . The probability that the loops L NQ ext ∩ H α realise the special crossing event with exterior rectangle NQ ext and interior rectangle NQ int is at least as large as the probability that the loops L NQ ext ∩ H , ≥ Nδ α realise the special crossing event with the same interior and exterior rectangle. From the corollary 2.8 follows that the latter probability converges as N → + ∞ to
$$\text{proadly converges as $1v\to$} \\ \mathbb { P } \left ( \bigcap _ { i = 1 } ^ { 3 } C _ { i } ( \mathcal { L } _ { \alpha } ^ { Q _ { e x t }, \geq \delta } ) \right ). \\ \text{the lemma 2.4.}$$
We conclude by applying the lemma 2.4.
/square
To conclude that for α > 1 2 , L H α has an infinite cluster we will use a block percolation construction that will combine special crossing events .
Proof of the Theorem 1. From [9] we know already that α H ∗ ≤ 1 2 . We need to show that for α > 1 2 , L H α has an infinite cluster.
Let α > 1 2 and N ≥ 1 . We consider a dependent edge percolation, denoted ( ω N ( e )) e edge of H , on the discrete half plane H . If e is an edge of form { ( j, k ) , ( j + 1 , k ) } , k ≥ 1 , then ω N ( e ) = 1 (open edge) if L ( NQ int +3 Nj + 3 i Nk ) ∩ H α achieves a special crossing event with exterior rectangle NQ ext +3 Nj + 3 i Nk and interior rectangle NQ int +3 Nj + 3 i Nk . If e is an edge of form { ( j, k ) , ( j, k +1) } , k ≥ 1 , then ω N ( e ) = 1 if L ( iNQ int +3 Nj + 3 i Nk ) ∩ H α achieves a special crossing event with exterior rectangle iNQ ext + 3 Nj + 3 i Nk and interior rectangle iNQ int + 3 Nj + 3 i Nk , where the multiplication by i means rotation by + π 2 . ω N is a 1 -dependent edge percolation: if two disjoint subsets of edges E 1 and E 2 are such that no edge is adjacent to both E 1 and E 2 , then ( ω N ( e )) e ∈ E 1 and ( ω N ( e )) e ∈ E 2 are independent. This is due to the fact that the subsets of loops involved in the definition of special crossing events for edges in E 1 and and edges in E 2 are disjoint. To an open path in ω N corresponds a cluster of L H α whose loops form crossings of related interior rectangles. Thus if ω N has an unbounded cluster, then so does L H α . See next picture.
Fig.4: Crossings achieved by subsets of loops in L H α , corresponding to five open edges in ω N .
The probability P ( ω N ( e ) = 1) is uniform and we will denote it p N . According to the lemma 2.9
## lim N → ∞ + p N = 1 .
Thus for N large enough ˜( p p N ) > 1 2 . 1 2 is the critical probability for the i.i.d. Bernoulli edge percolation on H . So for N large enough ω N contains a supercritical i.i.d. Bernoulli edge percolation and percolates itself. Thus L H α percolates too. Actually, since our construction only uses loops of diameter less or equal to 6 N , we have also percolation for L H , ≤ 6 N α . /square
## Acknowledgements
The author would like to thank Artem Sapozhnikov and Yinshan Chang for their comments on previous versions of this paper.
## References
- [1] Y. Chang. Supercritical loop percolation on Z d for d ≥ 3 . arXiv:1504.07906, Apr. 2015.
- [2] Y. Chang and A. Sapozhnikov. Phase transition in loop percolation. Probab. Theory Related Fields , pages 1-47, Apr. 2015.
- [3] Y. Le Jan. Markov paths, loops and fields. In 2008 St-Flour summer school, L.N. Math. , volume 2026. Springer, 2011.
- [4] Y. Le Jan and S. Lemaire. Markovian loop clusters on graphs. Illinois J. Math. , 57(2):525558, 2013.
- [5] G. F. Lawler and J. A. Trujillo Ferreras. Random walk loop soup. Trans. Amer. Math. Soc. , 359(2):767-787, 2007.
- [6] G. F. Lawler and W. Werner. The brownian loop-soup. Probab. Theory Related Fields , 128:565-588, 2004.
- [7] G.F. Lawler. Partition functions, loop measure, and versions of SLE. J. Stat. Phys. , 134:813837, 2009.
- [8] T. M. Liggett, R. H. Schonmann, and A. M. Stacey. Domination by product measures. Ann. Probab. , 25(1):71-95, 1997.
- [9] T. Lupu. From loop clusters and random interlacements to the free field. To appear in Ann. Probab. arXiv:1402.0298, 2014.
- [10] S. Sheffield and W. Werner. Conformal loop ensembles: the markovian characterization and the loop-soup construction. Ann. of Math. , 176(3):1827-1917, 2012.
|
10.1214/16-ECP4571
|
[
"Titus Lupu"
] |
2014-08-05T17:33:02+00:00
|
2016-01-11T06:57:15+00:00
|
[
"math.PR",
"60K35"
] |
Loop percolation on discrete half-plane
|
We consider the random walk loop soup on the discrete half-plane and study
the percolation problem, i.e. the existence of an infinite cluster of loops. We
show that the critical value of the intensity is equal to 1/2. The absence of
percolation at intensity 1/2 was shown in a previous work. We also show that in
the supercritical regime, one can keep only the loops up to some large enough
upper bound on the diameter and still have percolation.
|
1408.1046v1
|
## Emergent surface superconductivity of nanosized Dirac puddles in a topological insulator
Lukas Zhao, 1 Haiming Deng, 1 Inna Korzhovska, 1 Jeff Secor, 1 Milan
Begliarbekov, 1 Zhiyi Chen, 1 Erick Andrade, 2 Ethan Rosenthal, 2
Abhay Pasupathy, 2 Vadim Oganesyan 3 4 , & Lia Krusin-Elbaum 1 4 ,
1 Department of Physics, The City College of New York, CUNY, New York, NY 10031, USA
2 Department of Physics, Columbia University, New York, New York 10027, USA
3 Department of Engineering Science and Physics,
College of Staten Island, CUNY, Staten Island, NY 10314, USA and
4 The Graduate Center, CUNY, New York, NY 10016, USA
Surfaces of three-dimensional topological insulators have emerged as one of the most remarkable states of condensed quantum matter 1-5 where exotic electronic phases of Dirac particles should arise 1,6-8 . Here we report a discovery of surface superconductivity in a topological material (Sb Te ) with resistive transition 2 3 at a temperature of ∼ 9 K induced through a minor tuning of growth chemistry that depletes bulk conduction channels. The depletion shifts Fermi energy towards the Dirac point as witnessed by about two orders of magnitude reduction of carrier density and by very large ( ∼ 25 000 , cm 2 / V s) carrier mobility. · Direct evidence from scanning tunneling spectroscopy and from magnetic response show that the superconducting condensate forms in surface puddles at unprecedentedly higher temperatures, near 60 K and above. The new superconducting state we observe to emerge in puddles can be tuned by the topological material's parameters such as Fermi velocity and mean free path through band engineering; it could potentially become a hunting ground for Majorana modes 6 and lead to a disruptive paradigm change 9 in how quantum information is processed.
When nontrivial topological order of the electronic structure is concurrent with spontaneous symmetry breaking associated with strong correlations between particles the outcome is a putative state in which superconducting Cooper pairing does not conform to a conventional view. Such novel order has been predicted to arise on the surfaces of three dimensional topological insulators, where metallic conduction channels host helical Dirac fermions 1 that cannot be destroyed by non-magnetic scattering processes and can support unusual electronic phases 1,6-8 when electron correlations are at play. Up to now, the reported superconducting phases obtained at relatively low temperatures by doping 10-12 or under very high pressures 13,14 were found to have increased bulk carrier densities and hence appear to be of bulk origin.
In this report we demonstrate that surface superconductivity in the p -type topological material Sb Te 2 3 can be induced by a small variation in Te vapor pressure during the crystal growth. Tellurium overpressure in a very narrow range, while making no detectable structural changes, acts to compensate bulk carriers so that in the superconducting state the hole carrier density relative to that in the non-superconducting state is reduced, upshifting the Fermi energy from the bulk valence bands towards the vicinity of the Dirac point. From frequency dependent magnetic response and local superconducting gap scanning tunneling spectroscopy (STS) we show that superconductivity originates in surface Dirac puddles at remarkably high temperatures, glyph[greaterorsimilar] 60 K. The superconducting gaps observed in STS can be locally as large as ∼ 25 meV, which in Bardeen-Cooper-Schrieffer (BCS) theory would correspond to transition temperatures above liquid nitrogen. Global superconducting coherence is reached when interpuddle diffusion of depaired quasiparticles 15 establishes a percolative path connecting the puddles, at which point a superconducting resistive transition is observed.
Figure 1a shows resistivity of Sb Te 2 3 synthesized under ∼ 1 4 MPa Te vapor pressure . during the high-temperature step of the crystal growth cycle (see Methods) and measured at ambient pressure. The system undergoes a transition to zero resistance at the onset temperature T CR ∼ = 8 6 . K . We note that this is the highest T C observed in a topological material at ambient pressure after synthesis (see Methods, Figs. S1, S2, and Table S1 for chemical and structural analysis). In the narrow Te pressure range 1 2 . < P < 1 5 MPa, . where superconductivity is found, the hole density is reduced by nearly two orders of magnitude (Figs. 1b and S4) to below = 10 ∼ 18 cm -3 - a finding which should be contrasted with the increased electron density recorded in, e.g. , superconducting Bi Se 2 3 doped with Cu 10 .
In the same narrow pressure range the diamagnetism is strongly enhanced (Fig. 1b), but the expected superconducting transition into Meissner field expulsion state is found to take place at a significantly higher temperature T CD ∼ = 55 K , see Fig. 1c.
The diamagnetic transition detected in dc magnetization is sharp; the measured signal is about 10 -3 of the full Meissner value of 1 / 4 π expected in a superconductor and is nearly flat below ∼ 10 K. And while resistivity does not display any comparably strong features at T CD , scanning tunneling spectroscopy (STS) (Figs. 1d,e and Methods) shows a clear presence of superconducting energy gaps (Fig. 1f and Fig. S3) that can vary locally from 0 to glyph[greaterorsimilar] 20 meV. This patchy gap landscape indicates a laterally inhomogeneous superconducting state which within the BCS theory (using the gap equation 2∆(0) = 3 5 . k B T C , where k B is the Boltzmann constant) has local transition temperatures that can be greater than 60 K, inline with the temperature where the onset of diamagnetic Meissner signal is observed.
The resistive transition downshifts and broadens with increasing external magnetic field (Fig. 2a). By tracking onset temperature T CR we map the field-temperature H -T phase diagram for the two field orientations: applied along the c -axis of the crystal, H ‖ c , and parallel to the ab -plane, H ‖ ab (upper and lower panels of Fig. 2a, respectively). The low-temperature critical field anisotropy ∼ 1 4 at first glance is surprisingly small, given the . large structural anisotropy of Sb Te 2 3 (Figs. S1, S2). This suggests that the limiting field may be of spin rather than orbital origin. We estimate this field 16 by comparing magnetic energy 1 2 N (0) gµ 2 B H 2 p (0) with the superconducting condensation energy at the transition 1 2 N (0)∆ (0). 2 Here N (0) is the density of states at the Fermi energy, g is the Land´ e g -factor, and ∆(0) and H p (0) are zero-temperature superconducting gap and upper critical field, respectively. Using again BCS gap equation and taking the large measured value 17 of g -factor, g ∼ = 50, gives H p (0) = 2 6 . T CR √ g ≈ 3 2 . T , in good agreement with the experiment. We surmise then that depairing is of paramagnetic (Zeeman) origin, with a small anisotropy of the g -factor.
Another prominent feature in the H -T phase diagram is the large high-field lowtemperature region in the superconducting state below the critical field H p where we observe de Haas-van Aphen (dHvA) quantum oscillations (inset in Fig. 2b and Figs. 4a,b) - one of the most direct probes of quasiparticle excitations in metals 18 . This region, bound by the field H glyph[star] ( T ) at which oscillations first appear is obtained from the temperature dependence of the low-field onset of dHvA. Occurrence of dHvA oscillations is known in extreme type
II superconductors in the vortex state 19 , although in conventional superconductors there is a considerable amplitude damping effect below upper critical field H c 2 . The superconducting Sb Te 2 3 shows no additional damping at the critical field (Fig. S5). This suggests two possible reasons for robust quantum oscillations in the superconducting state: one is inhomogeneous superconductivity with some residual unpaired fermion quasiparticles 15 , another is the superconducting gap developing nodes in momentum space 19 .
The diamagnetism of the superconducting samples is also highly unusual. Differential diamagnetic magnetic susceptibility χ (Figs. 3a, b) measured using a relatively low (10 kHz) frequency is strongly field and temperature dependent. Large diamagnetism persists to high ( > 100 K) temperatures as it smoothly and monotonically reduces to nonsuperconducting values. This should be contrasted with dc magnetization (such as shown in Fig. 1b) which has a sharp Meissner transition at ∼ 55 K. This unusual differential response is found to vary strongly with frequency of ac excitation even for the low ( ∼ kHz) frequencies used (Figs. 3d, e). General analyticity considerations at low external field dictate that low frequency corrections enter quadratically, χ ω, T ( ) = χ 0 ( T )+ ( b T ) ω 2 , and such behavior is found in resistively shunted Josephson networks and other cases of two fluid (normal + superfluid) dynamics 20 . With this in mind we analyzed the frequency dependence of χ by performing frequency scans at different temperatures, focussing on the low field regime. The data (inset in Fig. 3g) clearly follows a parabolic frequency dependence. The fit to this simple form uncovers a spectacular dichotomy between temperature variation of the zero frequency value χ 0 ( T ) (Fig. 3g and Supplementary Information) and the dispersion coefficient b T ( ) (Fig. 3f): χ 0 ( T ) shows a sharp diamagnetic onset in the vicinity of ∼ 55 K in close correspondence with the temperature dependence recorded in dc magnetization, while the prefactor b T ( ) smoothly marches toward the near null value at very high temperatures (also Figs. S7, S8 and Supplementary Information Section G). The zero-frequency response χ 0 ( T ) is large, much larger than any non-superconducting diamagnetism, and we identify its variation and magnitude with a sharp mesoscopic Meissner transition observed using SQUID (Fig. 1d), while persistence of the finite frequency response coefficient, b , suggests existence of considerable superconducting fluctuations at smaller scales and significantly higher temperatures.
The large separation between T CR and T CD signifies two separate physical processes at work. The patchy network of large local superconducting gaps detected by STS and strong diamagnetic response above T CR unaccompanied by detectable transport signatures is most
naturally ascribed to isolated and well separated 'puddles' of local superconductivity, which we take to be congregating near the surface of the sample and very thin 21 , to account for the observed large orbital anisotropy of the response (Figs. 3g, f), and also the overall depletion of bulk conducting channels (Figs. 1d, 4c). The relatively sharply defined T CD implies that sufficiently many of these puddles are larger than the superconducting coherence length ξ (roughly estimated 22 at below ∼ 10 nm) so that we may ignore size effects on the local order parameter. Using Ginzburg-Landau phenomenology 16 with T CD being the critical temperature of the puddles, we obtain χ ∼ 1 /λ T ( ) 2 ∼ T CD -T in the strongly type-II limit, valid at high temperatures, in agreement with the approximately linear onset of χ 0 shown in Fig. 3g. Here λ is the magnetic penetration depth.
The resistive transition at T CR requires a mechanism for generation of Josephson (phase) coupling and establishment of global coherence of a sufficient fraction of these puddles. We propose that long-lived quasiparticles that give rise to the observed magneto-oscillations also mediate Josephson coupling among puddles 15,23 . When the separation between puddles is larger than the quasiparticle mean-free path, and there is good screening of charge fluctuations in and out of the puddles 15 , global coherence is attained once the typical puddle separation, a , is comparable to metallic quasiparticles' diffusion length a ≈ L T = √ glyph[planckover2pi1] D/k T B ≈ √ glyph[planckover2pi1] v F glyph[lscript]/k B T , where we have used the semiclassical expression for the diffusion constant D in terms of Fermi velocity v F and mean free path glyph[lscript] , D ≈ v F glyph[lscript] . In this limit T CR is largely controlled by the properties of the metallic matrix and the spatial arrangement of puddles, with the sizes of the puddles, the strength of local order parameter and other details only modifying the criterion above through logarithmic factors 15 . The interpuddle spacing can be then estimated from the observed T CR provided the diffusion constant, D , of the intervening metal is known.
Next we turn to the analysis of dHvA oscillations, from which we can obtain D . Quantum dHvA oscillations (Figs. 2b, 4a,b, and Supplementary Information) are quite remarkable in their own right as they clearly display a beat structure found in two-dimensional quantum well states (2DEG) in semiconductors 24 . Such states have been predicted in topological insulators (TIs) in ab initio calculations 21 and detected in angle-resolved photoemission spectroscopy (ARPES) 25 but, to the best of our knowledge, have never been observed in magneto-oscillations. Lifshitz-Kosevich analysis 18 of these oscillations (see Supplementary Information, Table S2) yields v F glyph[similarequal] 5 3 . · 10 5 m/s, glyph[lscript] ≈ 95 nm, m = 0 065 . m e , k F + =
3 7 . · 10 8 m -1 , k F -= 3 3 . · 10 8 m -1 for Fermi velocity, mean-free path, effective mass, and the two close Fermi wavevectors that induce the beats, respectively. We note that the overall beat structure closely scales with the magnetic field component transverse to the surface as shown in Fig. 4b for the two field orientations and therefore implies a twodimensional origin of the signal. Based on these numbers we obtain the diffusion constant D = v F glyph[lscript]/ 2 = 0 025 m /s, and the typical interpuddle separation . 2 a ≈ 140 nm. Using the actual sample area A to estimate the total number of puddles, and relating the absolute value of the diamagnetic response to the single puddle's response by assuming simply additive contributions of individual monodispersed puddles, we obtain the typical puddle size of about R ≈ 37 nm (see Supplementary Information). It is inline with the size scales of surface Dirac puddles reported in STS studies 26 .
Based on this correspondence and on the observed patchy distribution of superconducting gaps, we associate the high temperature superconductivity in Sb Te 2 3 with the assembly of Dirac puddles and identify the nature of the metallic matrix that mediates global superconductivity at T CR as 2DEG, the two-dimensional electron gas 25 . One reason for this assignment comes from the spin-orbit splitting δ glyph[similarequal] 1 34 meV obtained from the beats in . dHvA quantum oscillations (Fig. S6); it is comparable to spin splitting observed in e.g., a 2DEG InGaAs/InAlAs heterostructures with strong spin orbit coupling 24 . Figure 4c emphasizes another important reason: in the narrow region of Te overpressure the carrier density is hugely reduced, bringing the Fermi level up to just below the Dirac point. Owing to a peculiar dispersion of the bulk valence bands 3 of Sb Te , the finite 2 3 k superconducting pairing spans the δk sliver that includes the combined system of Dirac puddles connected via coherent diffusion in the metallic 2DEG as visualized in Fig. 1d. Yet another clue pointing to the surface 2DEG resides in huge carrier mobility ( ∼ 25 000 cm , 2 / V s) in the superconducting · Sb Te , a factor of 2 3 ∼ 165 larger than in our nonsuperconducting Sb Te 2 3 (Table S2) where the Fermi level is pinned deeply within the valence band, or in the related families of TIs where carrier mobilities at best are on the order of a ∼ 1000 -3000 cm 2 / V s. ·
We remark that two-dimensional superconductivity that has been found in 2D electron gas at the interfaces between two band insulators, LaAlO 3 and SrTiO 3 occurs below 200 millikelvin 27 , a much lower temperature than we observe; the high T C in Sb Te 2 3 is a likely spillover from the superconducting puddles 28 supporting pairing of helical Dirac holes. Puddles have been known to form at low carrier densities in other two-dimensional Dirac
systems 29 owing to nonlinear screening effects 30 . The striking new superconducting surface state that emerges in puddles in a TI is potentially tunable through material's control, e.g. , of the quasiparticle mean free path, as well as the system's Fermi velocity.
## Methods
Single crystals of Sb 2 Te 3 were grown in evacuated quartz tubes in a horizontal gradient furnace heated to 1000 o C and cooled to room temperature in 7-10 days. The starting materials used were cm-sized chunks of Sb (purity 99.9999%) and Te (purity 99.9995%) from Alfa-Aesar used in stoichiometric ratios. The critical parameter was Te pressure during the high temperature segment of the growth process; it was determined from the ideal gas equation P = nRT max /V , where T max is the highest temperature used, V is the enclosed volume, n is the umber of moles of the material, and R is the ideal gas constant. The structure and composition of crystals grown in the 0.4 MPa to 3 MPa pressure range was determined from X-ray diffraction and glow discharge analysis (Evans Analytical Group) that determined impurity content ( < 0 6 ppm wt) of the final crystals . (Supplementary Information). X-ray diffraction was performed in Panalytical diffractometer using Cu K α ( λ = 1 5405 ˚ ) line from Philips high intensity ceramic sealed tube (3 kW) X-ray source . A with a Soller slit (0.04 rad) incident and diffracted beam optics. Structural identity was confirmed by micro-Raman spectroscopy of characteristic phonon modes. Carrier densities were determined from Hall resistivity and Shubnikov-de Haas oscillations (Supplementary Information). Scanning tunneling microscopy (STM) and spectroscopy (STS) measurements were carried out in a homebuilt cryogenic UHV STM at a temperature of 5.8 K. Spectroscopy measurements were performed using a lock-in amplifier running at a frequency of 1.831 kHz and an excitation voltage of 0.2 mV. Spectroscopic imaging was carried out over a grid of points (either 128 × 128 or 256 × 256 pixels) at various energies using the same lock-in amplifier parameters. Fourier transforms (FFTs) are affine transformed to correct for drift and then hexagonally averaged to enhance signal to noise. Transport and susceptibility measurements were performed in a 14 Tesla Quantum Design PPMS system in 1 mT of He gas. For transport, lithographically patterned Au/Ti contacts were fabricated on 100 nm thin crystals exfoliated onto SiO /Si substrates. 2 Differential susceptibility was measured in a compensated pickup-coil detection configuration with the excitation/detection coils designed to align with the applied static field. The ac excitation amplitude was set at 10 -5 T in a frequency range up to 10 kHz. The system was calibrated using paramagnetic Pd standard and superconducting Nb. dc magnetization measurements were performed using Quantum Design SQUID Magnetometer in up to 5.5 Tesla fields.
Acknowledgements: We are grateful to David Huse, Steven Kivelson, Boris Spivak,
Kyungwha Park, Kamran Behnia, Steven Girvin, and Shivaji Sondhi for enlightening discussions. This work was supported in part by NSF under DMR-1122594 (LK) and DMR0955714 (VO), and by DOD-W911NF-13-1-0159 (LK).
## FIGURE LEGENDS
Figure 1 | Superconductivity in a topological insulator Sb Te . 2 3 a, Resistivity of an exfoliated 100 nm thin Sb Te 2 3 crystal synthesized under ∼ 1 4 MPa Te vapor pressure . (see Methods) shows onset of transition to zero resistance at T CR ∼ = 8 6 . K . Inset shows Hall contact configuration used. b, Diamagnetic susceptibility (left) at 1.9 K measured in a 0.2 T field shows huge diamagnetism (Meissner effect) only in the narrow Te vapor pressure range. It is anticorrelated with the measured carrier density (right) which in the superconducting state is decreased by a factor of nearly 100. Outside this range samples are nonsuperconducting. c, dc magnetization measured at 0.2 Tesla in SQUID shows a sharp diamagnetic onset at T CD ≈ 55 K > T CR . The data taken under zero-field-cooled (ZFC) and field-cooled (FC) conditions are essentially identical, consistent with negligible vortex pinning in 2D (ref. [16]). d, Illustration of superconductive puddles (blue, size 2 R ) of Dirac bands, where pairing occurs at high temperature ( T CD ), connected to the 2DEG metallic matrix (grey) which establishes a percolative path (dashed red line) at T CR . The electronic Dirac dispersion and spin-orbit split dispersion of 2DEG are also sketched, see text. d,e,f, STM and STS scans were performed in areas ∼ 25 nm each, spaced ∼ 200 nm apart. e, A typical topograph of a scanned area shows well ordered hexagonal lattice, with differential conductance d I/ d V (from the average of 500 scans) shown in f . Depending on the scan area, the gaps 2∆ vary from zero to glyph[greaterorsimilar] 20 meV, see Fig. S3.
Figure 2 | Magnetic field dependence of the superconducting transition in Sb Te . 2 3 a, Resistive transition temperature downshifts with increasing magnetic field applied (top) along the c -axis of the crystal, H ‖ c , and (bottom) parallel to the ab -plane, H ‖ ab . The onset of superconductivity is indicated by the arrows. b, H -T phase diagram of the superconducting state has relatively small zero-temperature anisotropy of ∼ 1 4. . The critical field values at T → 0 agree with the paramagnetic (Zeeman) depairing field H p , see text. Inset: de Haas van Alfen quantum oscillations (dHvA) shown for H ‖ c at 1.9 K
persist below H p down to the temperature dependent onset field H glyph[star] .
Figure 3 | Unusual high temperature diamagnetism in the superconducting Sb Te . 2 3 a, Differential magnetic susceptibility χ vs . temperature for several values of magnetic field µ H 0 ‖ c -axis. Meissner-like signal is 'flat' below T CR and monotonically vanishes at much higher ( > 100 K) temperatures. χ was measured by applying a small ac excitation field h ac = 10 -5 T at f = 10 kHz. b, Full temperature and field dependence of χ . c, χ vs . magnetic field µ H 0 ‖ c -axis for a series of temperatures. The pronounced dHvA oscillations are apparent at 1.9 K. d, χ strongly depends on frequency. e, Full frequency and temperature dependence of χ . Inset in g : Analysis of the frequency dependence of χ shows it to be quadratic in ω = 2 πf (see text). f, The prefactor b T ( ) in the ω 2 term dominates the total variation of χ at finite frequencies. This variation is consistent with kinetic inductance of the patchy distributed 2DEG network discussed in the text. g, Main panel: The zero frequency χ 0 ( T ) obtained from the fits to χ T ( ) = χ 0 ( T ) + b T ( ) ω 2 shows a sharp onset at the same T CD .
Figure 4 | Signatures of 2D superconductivity in Sb Te . a, 2 3 dHvA oscillations show beats arising from two very close oscillation periods. Inset: Fast Fourier transform (FFT) of the signal. The beats are a signature of spin-splitting by a strong spin-orbit interaction in 2DEG surface regions 25 . b, dHvA oscillations vs. inverse transverse component of magnetic field H ⊥ = H cos θ scale with H ⊥ . c, Hole carrier density n vs. Te pressure in the superconducting Sb Te 2 3 (in the narrow ∆ P vicinity of ∼ 1 4 MPa of Te pressure) . and non-superconducting Sb Te 2 3 (outside ∆ P ) states. In the superconducting region n is reduced, bringing the chemical potential from the deep inside the bulk valence band to just below the Dirac point, as illustrated in sketches of the band structures in all three regions. The conductivity of the superconducting Sb Te 2 3 remains hole-like ( p -type), with Fermi level crossing both, the Dirac bands and a 2DEG sliver of the valence band (upper middle sketch).
a
Fig. 1 LZ
## Supplementary Information: Emergent surface superconductivity of nanosized Dirac puddles
## in a topological insulator
The Supplementary Information is organized into seven sections:
- (A) Structural analysis (X-ray, Raman, TEM)
- (B) Elemental characterization
- (C) Superconducting gap mapping using scanning tunneling spectroscopy
- (D) Variation of carrier density measured via Hall and de Haas-van Alphen effects
- (E) Lifshitz-Kosevich analysis and determination of spin-orbit splitting
- (F) Estimating interpuddle separation
- (G) Frequency and temperature dependence in the inductive linear response
## A. Structural analysis
FIG. S1: (a) X-ray diffraction spectrum for Sb Te 2 3 grown in a sealed quartz tube under Te vapor pressure P = 1 42 MPa collected in Panalytical diffractometer using Cu K . α λ ( = 1 5405 ˚ ) line from . A Philips high intensity ceramic sealed tube (3 kW) X-ray source with a Soller slit (0.04 rad) incident and diffracted beam optics. Rietveld refinement lines shown in red are in full correspondence with the measured spectra. The obtained lattice parameters a = b = 4 26 ˚ and . A c = 30 46 ˚ remain . A unchanged in the 0.5 - 2 MPa pressure range, indicating that in this pressure range there is no structural change. (b) The robustness of the structure is also apparent in the identical micro-Raman spectra for superconducting (red) a non-superconducting (black) Sb Te . 2 3 The E phonon modes are doubly degenerate modes in the ab -plane and the A modes are nondegenerate vibrations with atomic motion along the c -axis. The spectra were taken in ambient conditions in a backscattering geometry with linearly polarized excitation in the ab plan and normalized to the out of plane vibration at 70 cm -1 .
Transmission Electron Microscopy (TEM), X-ray diffraction and micro-Raman spectra of superconducting Sb Te 2 3 indicate absence of any structural changes induced by Te overpressure. We confirmed that the relatively low Te vapor pressure during the synthesis did
not alter the layered rhombohedral van der Waals structure or lattice parameters of Sb Te 2 3 as determined from the X-ray diffraction (XRD) spectra (Figs. S1 and S2).
b
FIG. S2: (a) Rhombohedral layered structure of the crystal constructed from the X-ray diffraction (XRD) spectra (see Fig. S1). The structure remains van der Walls type, with three quantuple layers per unit cell and with the XRD-determined lattice parameters up to ∼ 2 -3 MPa. (b) High resolution transmission electron microscopy images of the crystal are consistent with the XRD. A hexagonal lattice in the ab -plane is shown on the left and a layered van der Walls structure along the c -axis (normal to the (00 ¯ ) cleavage plane is shown on the right. 1
## B. GDMS Elemental analysis
Glow discharge mass spectrometry (GDMS) analysis of superconducting Sb Te 2 3 lists the impurity content in these crystals, see Table S1. The impurity content is the same in the nonsuperconducting Sb Te . 2 3 GDMS was performed by Evans Analytical Group (EAG). It has detection limits on the sub-ppm range for most elements that are nearly matrix-independent. In this technique, collisions between the gas-phase sample atoms and the plasma gas pass energy to the sample atoms, exciting the atoms. The atoms then loose their energy by emitting light with the atom specific wavelength. From the intensity of emitted light the atomic concentration can be determined. Sample atoms are also ionized through collisions
that then are detected by mass spectrometry. The impurity content of all elements detected in Sb Te 2 3 matrix is less that a small fraction of ppm.
TABLE S1
| Element | Concentration (ppm weight) | Concentration (weight ratio) |
|-----------|------------------------------|--------------------------------|
| Li | < 0.005 | < 5×10 -9 |
| Al | < 0.01 | < 1×10 -8 |
| Ca | 0.08 | 8×10 -8 |
| Ti | < 0.005 | < 5×10 -9 |
| V | < 0.005 | < 5×10 -9 |
| Cu | 0.4 | 4×10 -7 |
| Zn | 0.12 | 1.2×10 -7 |
| Se | 2.6 | 2.6×10 -6 |
| Nb | < 0.01 | < 1×10 -8 |
| Mo | < 0.05 | < 5×10 -8 |
| Ru | < 0.05 | < 5×10 -8 |
| Pd | < 0.01 | < 1×10 -8 |
| In | Binder | Binder |
| Sn | < 0.5 | < 5×10 -7 |
| Sb | Matrix | Matrix |
| Te | Matrix | Matrix |
| I | < 0.5 | < 5×10 -7 |
| Ta | Source | Source |
| Hg | < 0.1 | < 1×10 -7 |
| Pb | 0.63 | < 6.3×10 -7 |
## C. Superconducting gap mapping using scanning tunneling spectroscopy
FIG. S3: Scanning tunneling spectroscopy from superconducting surfaces of Sb Te . The scans in 2 3 (a)-(c) were performed in three different surface areas about 25 nm each, spaced ∼ 200 nm apart, as sketched in the outset. (a) Differential conductance d I/ d V ( V ) was obtained from the average of 500 scans. In area 1 it shows well articulated coherence peaks and the gap 2∆ ≈ 10 meV corresponding to T C ∼ 30 K estimated from BCS gap equation. (b) d I/ d V in area 2 has a more complex behavior with two gaps evident: a smaller one 2∆ 1 ≈ 8 meV and a larger one 2∆ 2 glyph[greaterorsimilar] 20 meV corresponding to T C glyph[greaterorsimilar] 70 K. (c) d I/ d V in area 3 shows essentially no gap. The slope in measured d I/ d V in area 3 is also seen in good metals such as copper 31 or gold 32 where it is due to band structure effects. Other samples grown in the Te pressure range where superconductivity is found show similar distribution of gap energies, consistent with superconducting puddles with average T CD ∼ 55 K embedded in the gapless matrix gleaned in area 3.
The variation in the articulation of coherence peaks as well as finite density of states (DOS) at low energies is commonly seen in strongly correlated systems such as highT C cuprates 33 and heavy fermion superconductors 34 where it is a consequence of electronic inhomogeneity and strong Coulomb repulsion. It can also be detected when the Fermi energy is small as in the case of superconductivity at the LAO/STO interfaces 35 .
D. Hall resistivity and de Haas van Alfen (dHvA) quantum oscillations in superconducting Sb Te 2 3
b
18
0.8
0.6
0.4
0.2
0.0
-0.2
-0.4
-0.6
-0.8
-5
n
= 1.4x10
/cc
(a.u.)
-0.5
-1.0
0.10
-4
-3
-2
0.15
0.20
1/(
0
H
0.25
) (1/T)
n
= 1.4 x 10
300nm
-1
0
1
2
18
/cc
3
4
5
0
H
(T)
FIG. S4: (a) Main panel: Hall resistivity of non-superconducting Sb Te 2 3 shows it to be p -type. Carrier density determined from Hall is n = 3 9 . × 10 20 / cc. However, carrier concentration determined from dHvA is n = 2 56 . × 10 19 / cc, over an order of magnitude lower. Similar differences are found in other TIs when chemical potential is located deeply inside either valence or conduction bands. This is reconciled when taking into account the incoherent addition of contributions to Hall conductivity from hexagonal 'pockets' (factor of 6), while the Fermi cross-sections in dHvA are sampled coherently. The additional factor of 2 comes from spin-splitting in the bulk bands, fully accounting for the differences in n . (b) Main panel: Hall resistivity of superconducting Sb Te 2 3 shows it also to be p -type. Here, however, carrier densities determined from Hall and dHvA (top outset) are identical n = 1 4 . × 10 18 / cc, and over an order of magnitude lower than in the superconducting samples. This is consistent with the location of the Fermi level just below the Dirac point, and nearly on top of the valence band (Fig. 4c).
Variation of carrier density with Te pressure (Fig. 1 in the main text) is obtained by Hall effect and dHvA oscillations as shown here for two samples. We note that all Sb Te 2 3 crystals are p -type (the charge carriers are holes). The non-superconducting crystals synthesized at Te vapor pressures P < 1 2 . MPa and P > 1 55 . MPa have metallic-like temperature cm)
(m xy
dependence (d R/ T > d 0) of resistance and carrier densities n ∼ 2 × 10 19 -10 20 cm -3 . In the superconducting Sb Te 2 3 crystals ( P ∼ 1 4 MPa) the carrier density is over an order of . magnitude lower.
## E. de Haas van Alfen oscillations: Lifshitz-Kosevich analysis; spin-orbit splitting
Electronic parameters of Sb Te , such as cyclotron mass, carrier densities, or Fermi ve2 3 locities are obtained from de Haas van Alfen (dHvA ) oscillations using Lifshitz-Kosevich theory 37 , see Figs S2-S4, and Table S2. The most striking aspect of the measured dHvA signal is the absence of any significant amplitude damping upon entering the superconducting state shown in the Dingle plot 38,39 in Fig. S3.
The beating effect in dHvA oscillations implies the existence of two closely spaced frequency components with similar amplitudes (see Fig. 4a). This has been seen in In x Ga 1 -x As /In 0 52 . Al 0 48 . As hetero-structures 24 where a single subband is spin-split by strong spin-orbit coupling. A spin-split Landau level gives rise to two closely spaces frequencies with similar amplitudes leading to a modulation of the dHvA amplitude given by A ∼ cos πν , where ν = δ glyph[planckover2pi1] ω c , and δ is the energy separation between the spin-split Landau levels. Nodes in the beat pattern in dHvA will occur at half-integer values of ν ( ± 0 5 . , ± 1 5 . , ....) where A is zero. The total spin splitting δ can be expressed as δ = δ 0 + δ 1 glyph[planckover2pi1] ω c + δ 2 ( glyph[planckover2pi1] ω c ) 2 + ..., where δ 0 is the zero field spitting, δ 1 glyph[planckover2pi1] ω c is the linear in field splitting, and ω c = eB/ m glyph[planckover2pi1] c . The higher order terms become significant at high fields. Fig. S4a show a plot of δ vs. glyph[planckover2pi1] ω c for the fields corresponding to the nodes in dHvA oscillations in 4b (also Fig. 4a), using cyclotron mass m c = 0 065 . m e obtained from the fit of oscillations to Lifshitz-Kosevich theory 37 ∆ σ xx ( T ) ∆ σ xx (0) = λ T ( ) sinhλ T ( ) shown in (c) using Monte Carlo technique. Here σ xx is the in-plane conductivity for magnetic field applied normal to the cleavage plane and λ T ( ) = 2 πk B T glyph[planckover2pi1] eB m c . The extrapolation to zero field yields zero-field spin splitting δ 0 ≈ 1 34 meV, comparable to spin . splitting found in other 2DEG hetero-structures.
c2
FIG. S5: Comparison of de Haas van Alfen (dHvA) oscillation amplitude damping for V Si and 3 Sb Te . 2 3 Plot shows field dependence of D = ln[ α sinh( X B ) 1 2 / T -1 ] (known as Dingle plot) that shows the change in the dHvA oscillation amplitude upon crossing the superconducting limiting field. Here α T, B ( ) is the amplitude of quantum oscillations, X = 2 π k 2 B T/ ω glyph[planckover2pi1] c and ω c = eB/m c is the cyclotron frequency. The data shown for V Si are from Ref. 3 36 where the field scale was normalized to upper critical field B c 2 . In conventional superconductor V Si there is an additional 3 attenuation of the oscillation amplitude at the transition into the superconducting state observed in many extreme type II superconductors 19 . In a nontrivial superconducting Sb Te 2 3 there is no additional damping at the superconducting limiting field H p . Here we shifted the Dingle scale (in the same units) to overlay the data for both systems in their normal states (red line). The observed modulation of the Dingle factor D in Sb Te 2 3 is a result of beats in dHvA. The origins of the absence of attenuation can be of two sources: (i) inhomogeneous (puddle) superconductivity, and (ii) the superconducting order parameter is nodal 19 .
| | n (cm -3 ) | µ (cm 2 V -1 s -1 ) | l (nm) | k F (Å -1 ) | v F (m/s) | τ 0 (s) |
|----------|--------------|-----------------------|----------|---------------|-------------|-------------|
| S.C. | 1.41x10 18 | 24966 | 95 | 0.0347 | 5.3x10 5 | 1.56x10 -13 |
| Non S.C. | 2.56x10 19 | 152 | 22.8 | 0.2279 | -- | -- |
TABLE S2
c
c
FIG. S6: (a) Determination of spin-splitting due to spin-orbit coupling from (b) the beats in de Haas van Alfen quantum oscillations in superconducting Sb Te . 2 3 (c) Determination of cyclotron mass from the temperature dependence of dHvA oscillations in (b).
## F. Estimating interpuddle separation
Based on the material parameters in Table S2, we obtain the diffusion constant D = v F glyph[lscript]/ 2 = 0 025 m /s, and the typical interpuddle separation . 2 a ≈ 140 nm. From the area A of the sample A = 12 6 mm . 2 we estimate the total number of puddles to be on the order n ≈ A/a 2 ≈ 6 · 10 . 8 Next we compare the absolute value of the diamagnetic response and relate it to the single puddle's response by assuming simply additive contributions of individual monodispersed puddles, V χ 0 ≈ nχ 1 , where V is the sample volume and χ 1 is the average extensive susceptibility of one typical puddle. Using V ≈ 2 5 mm . 3 we obtain χ 1 ≈ 2 10 · -22 at low temperatures. Relating this value to puddles' dimensions, e.g. , radius R and thickness t 21 , is complicated at high temperatures without an independent measurement of the local penetration depth, λ . However, at low temperatures and for sufficiently large puddles we assume field exclusion which significantly simplifies the analysis, yielding χ 1 = -4 R 3 . Independence of this expression of the thickness t is due to large demagnetization correction 20 . The estimated puddle size of about R ≈ 37 nm is comparable to the size scales of surface Dirac puddles observed by scanning tunneling microscopy 26 .
## G. Frequency and temperature variation of the ac response
Complex susceptibility, χ ω, T ( ) = χ ′ ( ω, T ) + iχ ′′ ( ω, T ), is characterized by reactive inphase response, χ ′ (denoted χ in main text), but also has a dissipative out-of-phase component χ ′′ 20 . We can readily identify the dissipative component with eddy current heating χ ′′ ( ω, T ) ∼ ωσ T ( ), where σ T ( ) is the temperature dependent dc conductivity (Fig. S7). Classical current fluctuations can only screen magnetic fields at high frequencies, with χ ′ ( ω → ∞ ) → -1 4 π and χ ′ ( ω → 0) ≈ -( ωτ ) 2 , where τ is the characteristic relaxation time. For example, modeling a conductor as a simple LR-circuit we find τ = L/R . Proper interpretation of finite value of χ ′ (0) < 0 requires quantum mechanics. One may, however, use purely classical phenomenology to capture diamagnetism by positing existence of macroscopic 'perfect inductor' paths, with τ →∞ . For superconductors we may simply think of
perfect inductors as linearized Josephson elements and there exists abundant literature on phases and phase transitions of resistively shunted Josephson networks 40,41 .
Roughly speaking, the normal phase can be thought of in a coarse grained fashion as a macroscopic LR circuit, while the superconductor has R → 0. While this '1-loop' phenomenology correctly captures asymptotic ω → 0 limit of the two phases, it misses the
FIG. S7: (a) The in-phase component of χ is quadratic in frequency χ T ( ) = χ 0 ( T ) + b T ( ) ω 2 , also see Fig. S7 and main text. (b) The dissipative (out-of-phase) component of χ ω ( ) is strictly frequency linear, as expected. (c) The standard eddy current mechanism is responsible for dissipation dominated by the bulk, the out-of-phase component χ ' is proportional to conductivity: the observed value is consistent, up to geometric factors and closely follows the temperature dependence in-plane resistivity ρ xx of the bulk , according to the standard formula for power P = π 2 ( h 2 ac d f 2 2 / ρ 2 xx ( T )) dissipated during the ac excitation cycles. Here f = ω/ π 2 and d is the sample thickness.
low frequency correction in the superconducting phase, both ∼ ω 2 in χ ′ , and, importantly χ ′′ ∼ ω , which may be thought of as the response from a finite normal fluid fraction. Various simple '2-loop' improvements are possible to rectify this situation, e.g., two inductors in parallel (and only one perfect) have a simply additive response, i.e. χ ′ ( ω → 0) = χ 0 + b ω 2 + . . . , with χ 0 coming from perfect inductors and b representative of resistively shunted elements. We have used this phenomenology to extract χ 0 in two different samples, see Fig. 3 in the main text and Fig. S7. The two measurements appear to show different amount of anisotropy for H ‖ ab vs. H ‖ c , which we attribute to the instrumental misalignment of the latter.
1
Qi, X.-L. & Zhang, S.-C.
1057-1110 (2011).
Fu, L., Kane, C. & Mele, E.
2
Topological insulators in three dimensions.
Rev. Mod. Phys.
Phys. Rev. Lett.
83
98
,
,
106803 (2007).
FIG. S8: (a) A sharp diamagnetic transition at ∼ 50 K in the zero frequency response χ 0 ( T ) obtained from the fits to χ T ( ) = χ 0 ( T ) + b T ( ) ω 2 of another Sb Te 2 3 crystal. (b) The prefactor b T ( ) in the ω 2 term is varying smoothly, dominating the total variation of χ at finite frequencies. This variation is consistent with kinetic inductance of the patchy distributed 2DEG network discussed in the main text.
- 3 Zhang, H. et al. Topological insulators in Bi 2 Se , Bi Te 3 2 3 and Sb Te 2 3 with a single Dirac cone on the surface. Nat. Phys. 5 , 438-442 (2009).
- 4 Hsieh, D. et al. A tunable topological insulator in the spin helical Dirac transport regime. Nature 460 , 1101-5 (2009).
- 5 Roushan, P. et al. Topological surface states protected from backscattering by chiral spin texture. Nature 460 , 1106-9 (2009).
- 6 Fu, L. & Kane, C. Superconducting proximity effect and Majorana fermions at the surface of a topological insulator. Phys. Rev. Lett. 100 , 096407 (2008).
- 7 Qi, X.-L., Hughes, T., Raghu, S. & Zhang, S.-C. Time-reversal-invariant topological superconductors and superfluids in two and three dimensions. Phys. Rev. Lett. 102 , 187001 (2009).
- 8 Fu, L. & Berg, E. Odd-parity topological superconductors: Theory and application to Cu Bi Se . x 2 3 Phys. Rev. Lett. 105 , 097001 (2010).
- 9 Nayak, C., Stern, A., Freedman, M. & Das Sarma, S. Non-Abelian anyons and topological quantum computation. Rev. Mod. Phys. 80 , 1083-1159 (2008).
- 10 Hor, Y. S. et al. Superconductivity in Cu x Bi Se 2 3 and its implications for pairing in the undoped topological insulator. Phys. Rev. Lett. 104 , 057001 (2010).
- 11 Sasaki, S. et al. Topological superconductivity in Cu x Bi Se . 2 3 Phys. Rev. Lett. 107 , 217001 (2011).
- 12 Wray, L. A. et al. Observation of topological order in a superconducting doped topological insulator. Nat. Phys. 6 , 855-859 (2010).
- 13 Zhang, J. & Zhang, S. Pressure-induced superconductivity in topological parent compound Bi Te . 2 3 Proc. Natl. Acad. Sci. 108 , 24-28 (2011).
- 14 Zhu, J. et al. Superconductivity in topological insulator Sb Te 2 3 induced by pressure. Sci. Rep. 3 , 2016 (2013).
- 15 Spivak, B., Oreto, P. & Kivelson, S. Theory of quantum metal to superconductor transitions in highly conducting systems. Phys. Rev. B 77 , 214523 (2008).
- 16 Tinkham, M. Introduction to superconductivity (Courier Dover Publications, 2012).
- 17 Analytis, J. et al. Two-dimensional surface state in the quantum limit of a topological insulator. Nat. Phys. 6 (2010).
- 18 Shoenberg, D. Magnetic oscillations in metals (Cambridge University Press, 1984).
- 19 Maniv, T., Zhuravlev, V., Vagner, I. & Wyder, P. Vortex states and quantum magnetic oscil-
lations in conventional type-II superconductors. Rev. Mod. Phys. 73 , 867-911 (2001).
- 20 Alternating-field susceptometry and magnetic susceptibility of superconductors. In Hein, R., Francavilla, T. & Liebenberg, D. (eds.) Magnetic Susceptibility of Superconductors and Other Spin Systems (Springer US, 1991).
- 21 Park, K., Beule, C. D. & Partoens, B. The ageing effect in topological insulators: evolution of the surface electronic structure of Bi 2 Se 3 upon K adsorption. New J. Phys. 15 , 113031 (2013).
- 22 Werthamer, N., Hellfand, E. & Hohenberg, P. Temperature and purity dependence of the superconducting critical field, H c 2 . III. Electron spin and spin-orbit effects. Phys. Rev. 147 , 295 (1966).
- 23 Eley, S., Gopalakrishnan, S., Goldbart, P. & Mason, N. Approaching zero-temperature metallic states in mesoscopic superconductor-normal-superconductor arrays. Nat. Phys. 8 , 59-62 (2011).
- 24 Das, B., Miller, D., Datta, S. & Reifenberger, R. Evidence for spin splitting in In x Ga 1 -x As/In 0 52 . Al 0 48 . As heterostructures as B → 0. Phys. Rev. B 39 , 1411-1414 (1989).
- 25 Bahramy, M. S. et al. Emergent quantum confinement at topological insulator surfaces. Nat. Commun. 3 , 1159 (2012).
- 26 Beidenkopf, H. et al. Spatial fluctuations of helical Dirac fermions on the surface of topological insulators. Nat. Phys. 7 , 939-943 (2011).
- 27 Reyren, N. et al. Superconducting interfaces between insulating oxides. Science 317 , 1196-9 (2007).
- 28 Nandkishore, R., Maciejko, J., Huse, D. A. & Sondhi, S. L. Superconductivity of disordered Dirac fermions. Phys. Rev. B 87 , 174511 (2013).
- 29 Martin, J. et al. Observation of electron-hole puddles in graphene using a scanning singleelectron transistor. Nat. Phys. 4 , 144 (2007).
- 30 Hwang, E., Adam, S. & Das Sarma, S. Carrier transport in two-dimensional graphene layers. Phys. Rev. Lett. 98 , 186806 (2007).
- 31 Crommie, M., Lutz, C. & Eigler, D. Imaging standing waves in a two-dimensional electron gas. Nature 363 , 524-527 (1993).
- 32 Chen, W., Madhavan, V., Jamneala, T. & Crommie, M. Scanning tunneling microscopy observation of an electronic superlattice at the surface of clean gold. Phys. Rev. Lett. 80 , 1469 (1998).
- 33 McElroy, K. et al. Atomic-scale sources and mechanism of nanoscale electronic disorder in
Bi Sr CaCu O 2 2 2 8+ δ . Science 309 , 1048-1052 (2005).
- 34 Jourdan, M., Huth, M. & Adrian, H. Superconductivity mediated by spin fluctuations in the heavy-fermion compound UPd Al . 2 3 Nature 398 , 47-49 (1999).
- 35 Richter, C. et al. Interface superconductor with gap behaviour like a high-temperature superconductor. Nature 502 , 528-531 (2013).
- 36 Corcoran, R., Harrison, N. & Hayden, S. Quasiparticles in the vortex state of V Si. 3 Phys. Rev. Lett. 72 , 701-704 (1994).
- 37 Lifshits, I. & Kosevich, A. On the theory of magnetic susceptibility of metals at low temperatures. Sov. Phys. JETP 2 , 636-645 (1956).
- 38 Dingle, R. B. Some magnetic properties of metals. II. The influence of collisions on the magnetic behaviour of large systems. Proc. R. Soc. A Math. Phys. Eng. Sci. 211 , 517-525 (1952).
- 39 Janssen, T. et al. Quantitative investigation of the de Haas-van Alphen effect in the superconducting state. Phys. Rev. B 57 , 11698-11712 (1998).
- 40 Refael, G., Demler, E., Oreg, Y. & Fisher, D. Superconductor-to-normal transitions in dissipative chains of mesoscopic grains and nanowires. Phys. Rev. B 75 , 014522 (2007).
- 41 Giovannella, C. & Lambert, C. Lectures on superconductivity in networks and mesoscopic systems. AIP Conf. Proc. 427 , 273 (1998).
| null |
[
"Lukas Zhao",
"Haiming Deng",
"Inna Korzhovska",
"Jeff Secor",
"Milan Begliarbekov",
"Zhiyi Chen",
"Erick Andrade",
"Ethan Rosenthal",
"Abhay Pasupathy",
"Vadim Oganesyan",
"Lia Krusin-Elbaum"
] |
2014-08-05T17:33:36+00:00
|
2014-08-05T17:33:36+00:00
|
[
"cond-mat.supr-con",
"cond-mat.mes-hall",
"cond-mat.mtrl-sci"
] |
Emergent surface superconductivity of nanosized Dirac puddles in a topological insulator
|
Surfaces of three-dimensional topological insulators have emerged as one of
the most remarkable states of condensed quantum matter1-5 where exotic
electronic phases of Dirac particles should arise1,6-8. Here we report a
discovery of surface superconductivity in a topological material (Sb2Te3) with
resistive transition at a temperature of ~9 K induced through a minor tuning of
growth chemistry that depletes bulk conduction channels. The depletion shifts
Fermi energy towards the Dirac point as witnessed by about two orders of
magnitude reduction of carrier density and by very large (~25,000 cm^2/Vs)
carrier mobility. Direct evidence from scanning tunneling spectroscopy and from
magnetic response show that the superconducting condensate forms in surface
puddles at unprecedentedly higher temperatures, near 60 K and above. The new
superconducting state we observe to emerge in puddles can be tuned by the
topological material's parameters such as Fermi velocity and mean free path
through band engineering; it could potentially become a hunting ground for
Majorana modes6 and lead to a disruptive paradigm change9 in how quantum
information is processed.
|
1408.1047v2
|
## Remapping simulated halo catalogues in redshift space
## A. J. Mead 1 /star and J. A. Peacock 1
1 Institute for Astronomy, University of Edinburgh, Royal Observatory, Blackford Hill, Edinburgh EH9 3HJ
Accepted 2014 September 18. Received 2014 September 18; in original form 2014 August 5
## ABSTRACT
We discuss the extension to redshift space of a rescaling algorithm, designed to alter the effective cosmology of a pre-existing simulated particle distribution or catalogue of dark matter haloes. The rescaling approach was initially developed by Angulo & White and was adapted and applied to halo catalogues in real space in our previous work. This algorithm requires no information other than the initial and target cosmological parameters, and it contains no tuned parameters. It is shown here that the rescaling method also works well in redshift space, and that the rescaled simulations can reproduce the growth rate of cosmological density fluctuations appropriate for the target cosmology. Even when rescaling a grossly non-standard model with Λ = 0 and zero baryons, the redshift-space power spectrum of standard ΛCDM can be reproduced to about 5% error for k < 0 2 . h Mpc -1 . The ratio of quadrupole-to-monopole power spectra remains correct to the same tolerance up to k = 1 h Mpc -1 , provided that the input halo catalogue contains measured internal velocity dispersions.
Key words: cosmology: theory - large-scale structure of universe
## 1 INTRODUCTION
Numerical simulations are now an essential standard tool in the analysis and interpretation of cosmological surveys. Modern examples of which deliver high statistical power and are seeking to measure cosmological parameters with ever-increasing precision, thus requiring a corresponding rise in the care with which observational selection effects are treated. Practical survey complications are inevitably encountered to some extent, such as difficulty in obtaining spectra for adjacent objects in multiplexed spectroscopic surveys, and these are most robustly treated via Monte Carlo: analysis of a simulated mock dataset where the underlying cosmological parameters are known. This approach is increasingly deployed in order to verify the robustness of parameter estimates, or to identify and eliminate small residual biases. It can be witnessed in action in measurements of Baryon Acoustic Oscillations (BAO) in the galaxy distribution (e.g. Anderson et al. 2014) or in redshift-space distortions of galaxy clustering (e.g. Samushia et al. 2013; de la Torre et al. 2013).
the data. In order for the final parameter constraints to be meaningful (small 'errors on the errors'), the data covariance matrix itself must be precisely measured. This can require an ensemble of ∼ 1000 simulations - or significantly more if the size of the dataset under study is sufficiently large (Taylor et al. 2013).
Such studies also require the generation of large libraries of simulations. An ensemble of mock datasets is always required for a given cosmology, simply in order to generate an estimate of the data covariance matrix; this is used on the assumption of a Gaussian error distribution to calculate the likelihood of a given hypothetical model with respect to
/star E-mail: [email protected]
But the size of the numerical challenge is greatly increased if we accept that, in principle, a new ensemble of simulations is required for each new cosmological model under study. Moreover, mock galaxy data are required, so in principle a galaxy formation code must be run as part of each simulation. The latter issue can be dealt with rapidly using halo occupation distribution (HOD) models (e.g. Zheng et al. 2005), reducing the problem to one of obtaining catalogues of dark-matter haloes for the model under study. A way to speed up this basic generation of dark-matter simulations was presented by Angulo & White (2010; hereafter AW10). They showed that it was possible to rescale an N -body particle distribution so that the results closely approximated the output of a simulation with a different set of cosmological parameters.
In Mead & Peacock (2014; hereafter MP14) we showed that the AW10 method could be applied directly to halo catalogues, yielding a mass function and power spectrum of haloes that were well matched to the target cosmology following rescaling. In fact, this approach is capable of yielding more accurate results, since it retains the mass-dependent halo bias relation and can incorporate a cosmology dependence of halo concentrations. Moreover, this method has
a major practical advantage in terms of data compression: halo catalogues take up several orders of magnitude less storage than the raw particle data, and thus it is often the case that only the halo catalogue is made public (or even stored) by major simulation projects such as that of Rasera et al. (2010).
In MP14 we showed that these methods worked well on haloes in real space; in this paper we extend our approach to redshift space. The algorithm presented here consists of the following steps: (a) Initially the length and time units in the original halo catalogue are rescaled, in order to match the halo mass function, exactly as in the original AW10 algorithm. (b) In MP14 we showed that the particle or halo distribution itself may be used to compute the linear displacement field, from which we modify particle or halo positions and velocities so that they reproduce the correct large-scale clustering in the target cosmology. Eisenstein et al. (2007) showed how to recreate the displacement fields via the overdensity field, and Padmanabhan et al. (2012) used a variant of this approach to improve the sharpness of the BAO feature in BOSS data. (c) Finally, we modify the halo internal physical and velocity structure directly - by 'reconstituting' the density profiles around haloes so that they have the correct sizes and internal structure for the target cosmology. In this way, the rescaled velocity field contains both the linear velocities that contribute coherent redshift-space distortions (Kaiser 1987) and the post-linear effects that cause the 'fingers-of-God' (FOG) distortion.
Our paper is set out as follows: In Section 2 we briefly review the AW10 and MP14 algorithms and explain the extensions required to get good results in redshift space. We discuss the cosmology dependence of the internal physical and velocity structure of haloes. In Section 3 we describe our simulations and our methods for generating halo catalogues. In Section 4 we discuss our conventions and method for generating power spectra in redshift space. In Section 5 we show results for the full redshift-space power spectrum as well as the monopole and quadrupole moments of this. Results are presented for matter, haloes, and particles within haloes, and we show that an accurate recovery of the growth rate may be made from the rescaled distributions. Finally we sum up in Section 6.
## 2 METHOD
## 2.1 Overview of rescaling
We begin with a summary of the main features of the AW10 method and the MP14 extension. Quantities in the target cosmology are denoted with primes while the original quantities are unprimed. The original simulation output at redshift z , in a box of size L , is rescaled to a target simulation at redshift z ′ , in a box of size L ′ = sL . For a given z ′ , s and z are chosen so as to minimize the difference in σ 2 ( R ) (the linear-theory variance in density averaged over spheres of comoving radius R ) between the two cosmologies. The following cost function is convenient:
$$\delta _ { r m s } ^ { 2 } ( s, z \, | \, z ^ { \prime } ) = \frac { 1 } { \ln ( R _ { 2 } ^ { \prime } / R _ { 1 } ^ { \prime } ) } \int _ { R _ { 1 } ^ { \prime } } ^ { R _ { 2 } ^ { \prime } } \frac { d R } { R } \left [ 1 - \frac { \sigma ( R / s, z ) } { \sigma ^ { \prime } ( R, z ^ { \prime } ) } \right ] ^ { 2 } \,, \, ( 1 ) \quad \begin{array} { c } \delta = \cdot \end{array}.$$
with R ′ = sR . R 1 and R 2 are chosen so as to relate to the physical scale of the least and most massive haloes in the original simulation via
$$\text{red} \right ) _ { \text{t all.} } \quad M = \frac { 4 } { 3 } \i R ^ { 3 } \bar { \rho } \,,$$
where ¯ is ρ the mean comoving density. This approach is taken because the halo mass function is approximately universal when expressed in terms of the variable ν = δ /σ R c ( ) where δ c /similarequal 1 686 . (e.g. Press & Schechter 1974; Sheth & Tormen 1999; Sheth et al. 2001), and because the statistics of the nonlinear density field are driven by the mass function of haloes. In detail, the mass function displays deviations from universality at up to around the 10% level, so more accurate results might be obtained by directly matching theoretical predictions for a non-universal mass function (e.g. Luki´ et al. 2007; Reed et al. 2007; Tinker et al. 2008). c We have chosen not to do this because these mass functions are tuned to specific cosmological parameters and we are interested in quite broad shifts in cosmology in this work. Similarly, one might minimize the difference in ν M ( ) = δ /σ M c ( ) where δ c could be taken to vary with cosmology. For standard models, these variations in δ c are negligible (e.g. Eke et al. 1996; Lacey & Cole 1993; Percival 2005), but large variations in δ c are a feature of some modified gravity models (e.g. Schmidt et al. 2009; Li & Efstathiou 2012), owing to screening mechanisms.
In order to conserve mass the scaling in length units simultaneously implies a scaling in mass:
$$\stackrel { \ s u y } { e } \, \text{$\ sigma$} _ { \ } \\. \, \text{We} _ { \. } \, \text{$\ sigma$} _ { \ } \, \text{$\Gamma$} _ { m } M \equiv s _ { m } M \.$$
Note that we use units of h -1 Mpc for length and h -1 M /circledot for mass and the necessary factors of h are included in the scalings.
Additionally the dimensionless velocity units of the simulation must be conserved before and after scaling (see AW10; MP14) which implies a scaling in bulk velocities of particles or haloes such that
$$\inf _ { \text{un} } \inf _ { \text{nally} } \quad \mathbf v ^ { \prime } = s \frac { H ^ { \prime } ( a ^ { \prime } ) f _ { g } ^ { \prime } a ^ { \prime } } { H ( a ) f _ { g } a } \mathbf v$$
where H is the Hubble parameter at the epoch in question; f g ≡ dln g/ dln a is the logarithmic growth rate; g a ( ) is the linear theory growth function; and a is the scale factor.
Following scaling in s and z , the two cosmologies should have similar nonlinear power spectra; but the linear power on large scales will in general not be matched. This difference can be corrected for by using the approximation of Zel'dovich (1970; hereafter ZA) to perturb the particle or halo positions using the displacement field: the phase of each mode of the large-scale displacement field is preserved, but the amplitude is altered to match the target power spectrum. The displacement field f is defined so as to move particles from their initial Lagrangian positions q to their comoving Eulerian positions x :
$$\stackrel { \nu \nu ^ { \nu } } { \text{heres} } \quad x = q + f \.$$
At linear order the displacement field is related to the matter over-density δ via
$$\partial = - \nabla \cdot \mathfrak { f }$$
which in Fourier space is
$$f _ { k } & = - i \frac { \delta _ { k } } { k ^ { 2 } } k \,. & ( 7 ) & \quad \delta \nu _ { k ^ { \prime } } ^ { \prime }$$
If the displacement field in the original simulation is known, then an additional displacement can be specified in Fourier space to reflect the differences in the linear matter power spectra between the two cosmologies:
$$\delta f ^ { \prime } _ { k ^ { \prime } } & = \left [ \sqrt { \frac { \Delta ^ { \prime } _ { \text{lin} } ^ { 2 } ( k ^ { \prime }, z ^ { \prime } ) } { \Delta ^ { 2 } _ { \text{lin} } ( s k ^ { \prime }, z ) } } - 1 \right ] f ^ { \prime } _ { k ^ { \prime } } \. & ( 8 ) & \quad \text{ment} \\ & \quad \text{v} ^ { \prime \prime } =$$
where ∆ 2 is the power spectrum in the form of fractional variance in density per ln k , and f ′ is the linear displacement field after the input simulation has been scaled.
The particle or halo positions in the original simulation can be used to estimate the overdensity field via equation (7). To derive the overdensity of matter from the halo overdensity field, δ H , must be debiased, respecting the relation
$$\delta _ { H } & = b ( M ) \delta \ ; & ( 9 ) & = \underset { \xi \xi } { \text{base} }$$
in practice, we use the bias relations of Sheth & Tormen (1999). We then define a number-weighted effective bias for all haloes:
$$& \text{and allows.} \\ & b _ { \text{eff} } = \frac { \int _ { \nu _ { \min } } ^ { \nu _ { \max } } d \nu \, b ( \nu ) f ( \nu ) / m ( \nu ) } { \int _ { \nu _ { \min } } ^ { \nu _ { \max } } d \nu \, f ( \nu ) / m ( \nu ) } \, \quad & ( 1 0 ) \quad & \text{for } t \\ & \text{as} \quad & \text{ing I} \\ & \text{where } f ( \nu ) \, d \nu \, \text{is the fraction of the total density contributed} \\ & \text{bv halves in the range } d \nu. \text{For very low-mass halves } b ( M ) <$$
where f ( ν ) dν is the fraction of the total density contributed by haloes in the range dν . For very low-mass haloes b M ( ) < 1 and the method could in principle yield an unphysical negative mass density. In practice, this method is of interest for large-volume simulations where haloes of extremely low mass are not resolved, and we always find b eff > 1. Negative densities can always be avoided either by using something more sophisticated than a linear biasing relation, or by using haloes only above a certain mass to recreate the overdensity field.
Equation (7) is only valid for the linear components of both fields, so in practice the reconstructed δ must be smoothed with a window of width the nonlinear scale R nl to remove the nonlinear components. To do this we use a Gaussian filter
$$F ( k ) & = e ^ { - k ^ { 2 } R _ { \mathbf n 1 } ^ { 2 } / 2 } \, & ( 1 1 ) & \quad Z A \ c$$
where σ R ( nl , z ) = 1. Then equation (7) can be used to estimate the linear displacement field in Fourier space and particles can be displaced differentially in order to account for the differing linear power spectra:
$$x ^ { \prime \prime } & = x ^ { \prime } + \delta \dot { f } ^ { \prime } \, & ( 1 2 ) & \quad \text{Byo}$$
where the double dash indicates positions after this displacement has been applied. As noted in MP14, haloes are biased tracers of the density field and their effective displacement fields must therefore also be biased. Thus f H = b M ( ) f for each halo. In MP14, good results for the rescaled halo power spectrum were in general not obtained unless a biased displacement field was used.
The ZA also allows residual differences in linear velocities to be corrected on a mode-by-mode basis. In the ZA the peculiar velocity field ( v ≡ a ˙ ) x is related to the displacement field by v = aHf g f and additional differential changes to the peculiar velocities of particles or haloes are then given by
$$( 7 ) \quad \delta \nu ^ { \prime } _ { k ^ { \prime } } = a ^ { \prime } H ^ { \prime } f ^ { \prime } _ { g } \left [ \sqrt { \frac { \Delta ^ { ^ { \prime } 2 } _ { \ln } ( k ^ { \prime }, z ^ { \prime } ) } { \Delta ^ { 2 } _ { \text{lin} } ( s k ^ { \prime }, z ) } } - 1 \right ] \mathfrak { f } ^ { \prime } _ { k ^ { \prime } } \. \quad \quad ( 1 3 )$$
Note that, in contrast to the displacement field case, the halo velocities are unbiased: the equivalence principle requires haloes of all masses in a given region to share a common large-scale velocity. The final velocities after the displacement field step are then
$$\mathbf v ^ { \prime \prime } = \mathbf v ^ { \prime } + \delta \mathbf v ^ { \prime } \.$$
At this stage the linear power and mass function should be very close to the desired target. According to the halo model (e.g. Peacock & Smith 2000; Seljak 2000; Cooray & Sheth 2002), matching the mass function should also yield the correct quasi-linear clustering. Note that the rescaling method naturally includes effects such as halo exclusion (Smith, Desjacques, & Marian 2011) because it is based on rescaling an exact N -body calculation. However, the highly nonlinear part of the power spectrum is influenced by the internal structure of haloes, and this will also change if the cosmology is altered (e.g. Navarro et al. 1997; Bullock et al. 2001; Dooley et al. 2014). Such effects will alter the mass distribution defined by the haloes, and will also be important in producing mock galaxy catalogues using HOD methods. Here, a given halo is typically assigned a central galaxy and a number of satellite galaxies that are taken to trace the density profile. It is therefore necessary to consider how the internal structure of haloes should be rescaled. Note that such an adjustment is not part of the original AW10 method, which is a further advantage of the current approach.
AW10 was first applied directly to galaxy catalogues by Ruiz et al. (2011). Differences in galaxy formation between cosmological models have been studied using this method by Guo et al. (2013) and the cosmological constraints one can derive from differences in galaxy clustering are discussed by Simha & Cole (2013) (who used sub-halo abundance matching to populate rescaled catalogues with galaxies). In all cases only small scales were investigated and the large-scale ZA correction was not applied to the halo distribution. MP14 was the first work to successfully apply the ZA correction when working directly with a halo catalogue.
## 2.2 Rescaling halo properties
## 2.2.1 Halo catalogues
Beyond positions, velocities and masses of haloes, there are a number of other properties that may plausibly be stored in a halo catalogue (see Table 1). Halo radial density profiles have been shown to be accurately approximated on average by the 'NFW' profile of Navarro, Frenk, & White (1997):
$$\mathfrak { f } \text{ tor} \\ \text{over} \\ \mathfrak { l } \text{ dis-}$$
where r s is a scale radius and ρ N is a normalization to obtain the correct halo mass. Although subsequent work (Moore et al. 1999; Merritt et al. 2005; Merritt et al. 2006) showed this form to be imperfect at small r , it will suffice for our present purpose: mock galaxies are either central at r = 0 exactly, or satellites that tend to be found around r s , where the NFW approximation is good. The density profile
is truncated at the virial radius r v , which will depend on the exact definition of a halo; this can be measured directly or inferred from the halo mass:
$$r _ { \nu } & = \left ( \frac { 3 M } { 4 \pi \Delta _ { \nu } \bar { \rho } } \right ) ^ { 1 / 3 } \,, & \text{scalif} \\ \text{where $\Lambda$... is the average density of a halo with respect to the}$$
where ∆ v is the average density of a halo with respect to the background matter density. We adopt the common value of 178, justified by the near-universality of the halo mass function when defining haloes via the friends-of-friends algorithm with a cosmology-independent linking length.
The density profile is therefore fully specified via a value for r s or alternatively for the halo concentration c = r v /r s . There are many suggested relations for determining the concentration as a function of cosmological parameters (e.g. Navarro et al. 1997; Eke et al. 2001; Bullock et al. 2001; Neto et al. 2007). We adopt the relations of Bullock et al. (2001) because they are couched in term of physical quantities such as halo formation time, rather than being empirical fitting functions that apply only to a limited range of models.
Apart from virial radii and concentration, additional quantities that may be stored with a halo catalogue are the halo velocity dispersion, σ v , and the eigenvalues, λ 2 , and eigenvectors, w , of the moment of inertia tensor, which gives a measure of halo asphericity:
$$I _ { i j } & = \sum _ { k = 1 } ^ { N } ( x _ { k, i } - \bar { x } _ { i } ) ( x _ { k, j } - \bar { x } _ { j } ) \. & ( 1 7 ) \quad \text{$\text{are}$} \\ & \lambda _ { 3 } t \natural \\ & \text{halo} \\ & \text{Here $x_{i}$ is position of the $k^{th}$ halo particle with halo centre} \text{-}$$
Here x k is position of the k th halo particle with halo centreof-mass (CM) position ¯, x k ∈ { 1 , ..., N } and there are N particles in each halo and i, j ∈ { 1 2 3 , , } and label coordinates. Diagonalizing this tensor provides the axial ratios of the halo (via the eigenvalues) and the orientation of the halo (via the eigenvectors). In practice, we use a minimum size of N = 100 particles for this estimation.
We now discuss how each of these may be scaled: Measured concentrations may be scaled by a ratio of theoretical predictions taken from a relation such as Bullock et al. (2001); this corrects the mean relation, and implicitly assumes that the fractional scatter in concentrations of individual haloes has no strong cosmology dependence. Virial radii will be scaled by a factor of s because the virial overdensity criterion is independent of cosmology. For the NFW profile, and assuming isotropic orbits, the halo velocity dispersion is
$$\sigma _ { v } ^ { 2 } = \frac { G M } { 3 r _ { v } } \, \frac { c [ 1 - 1 / ( 1 + c ) ^ { 2 } - 2 \ln ( 1 + c ) / ( 1 + c ) ] } { 2 [ \ln ( 1 + c ) - c / ( 1 + c ) ] ^ { 2 } } \, \quad ( 1 8 ) \quad \text{
< \text{calc} </ \text{ent } | - + \text{nonint . > ( ) " : ; , ' !$$
which would give the scaling a mild concentration dependence. We chose not to use this more complicated formula for this initial investigation, adopting the simple approximation of
$$\sigma _ { v } ^ { 2 } & \simeq \frac { G M } { 3 r _ { v } } \, & ( 1 9 ) \quad \text{one} \, \mathfrak { c } \\ & \text{an} \, N$$
from which a scaling factor for σ v of s √ Ω ′ m / Ω m follows. The difference in σ v between equations (18) and (19) is only /similarequal 10% for concentrations of interest in this work ( c < 10). We do not attempt to solve the Jeans equation for the velocity dispersion as a function of radius (which would be possible given assumptions about orbital anisotropy). The limitations of halo scaling in redshift space that we uncover below would not be removed by such complications. The scalings we use in practice are summarized in Table 1.
## 2.2.2 Reconstitution of haloes
Given a hypothesis for the the density profile of a given halo, one may undertake reconstitution of the halo, either to recreate the dark-matter particles of which the halo is composed, or to populate it with mock galaxies. In the former case, one can use the NFW profile to place particles at random in order to sample the density (ignoring subhaloes). The boundary of a halo can either be taken to be given by r v according to equation (16) with a concentration given by a relation such as Bullock et al. (2001) or these can be taken from rescaled stored values. Haloes can then be assigned a velocity dispersion via either equation (18) or (19). The simplest approach is then to take the velocity components to be Gaussian distributed (see Kazantzidis et al. 2004; Wojtak et al. 2005 for limitations); we examined the approximation of Gaussian internal halo velocities in testing and found it to be accurate for our particular definition of haloes in our simulations (see section 3).
If desired, aspherical reconstitution can be achieved by distorting the halo particle distribution once it has been generated by the spherical halo reconstitution process described above. If the square roots of the eigenvalues are λ 1 , λ 2 and λ 3 then each coordinate of the particles in the reconstituted halo in the CM frame, y , is modified according to
$$\underset { \omega } { \ m e } - \underset { N } { \ m e } - y _ { i } \rightarrow 3 \lambda _ { i } y _ { i } / ( \lambda _ { 1 } + \lambda _ { 2 } + \lambda _ { 3 } )$$
where i ∈ { 1 2 3 , , } . We also considered the prescription y i → λ y / λ λ λ i i ( 1 2 3 ) 1 / 3 etc. but found this to work less well in recovering the shapes of aspherical haloes. The relative to CM position vector of each halo particle is then rotated by the inverse matrix of eigenvectors in order to orient the halo correctly.
## 2.2.3 Restructuring halo particles
Given access to the full particle distribution, two options are available to match the haloes in the target cosmology; one could either remove haloes from the rescaled particle distribution entirely, replacing them with reconstituted haloes in the manner described above - a method we call 'regurgitation'. Alternatively the haloes can be identified in the scaled distribution and reshaped to account for the different cosmology - a method we call 'restructuring'. Again we point out that it would be difficult to implement the biased displacement field for haloes if working from a particle distribution, and so the b M ( ) relation will not be correctly reproduced in this case.
To restructure haloes in a rescaled particle distribution one can proceed as follows: The amount of mass enclosed by an NFW profile at a radius r is given by
$$\text{ows.} _ { \text{only} } \quad M _ { \text{enc} } ( r ) = M \frac { \mu ( r / r _ { s } ) } { \mu ( c ) } \equiv f ( r ) M$$
where
$$\underset { \text{$e$ve} } { e v e } _ { \text{d be} } \quad \mu ( x ) = \ln \left ( 1 + x \right ) - \frac { x } { 1 + x } \.$$
Table 1. Scalings for various quantities that may be contained in a halo catalogue.
| quantity | symbol | scaling | comments |
|------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------|------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Positions Velocities Particle or halo mass Virial radius Halo concentration Halo velocity dispersion Inertia tensor eigenvalues Normalized inertia tensor eigenvectors | x v M r v c σ v λ 2 w | s s H ′ f ′ g a ′ /Hf g s 3 Ω ′ m / Ω m s c ′ th /c th s √ Ω ′ m / Ω m s 2 - | Additionally use equation (8) for linear displacements Additionally use equation (13) for linear velocities - Although this depends on how c is defined c th (theoretical) computed from any c ( M ) relation Alternatively use equation (18) with c dependence - No scaling because they are normalized |
Haloes can be reshaped by the ratio of mass enclosed at a radius r from the halo centre in each case. A scaled particle originally at r ′ should be moved to r ′′ , given by lar spectral shape to that of the Millennium Simulation. The historical popularity of τ CDMmodels is explained in MP14.
$$r ^ { \prime \prime } & = f ^ { \prime \prime - 1 } [ f ^ { \prime } ( r ^ { \prime } ) ] \, & ( 2 3 ) \quad \text{orde}$$
where f -1 indicates the inverse function. f ′′ will be the value of f in the target cosmology whereas f ′ will be the value in the original cosmology after it has been scaled . Particle positions relative to the CM can then be reassigned via
$$y ^ { \prime \prime } & = \frac { r ^ { \prime \prime } } { r ^ { \prime } } y ^ { \prime } \, & ( 2 4 ) & \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
y ^ { \prime \prime } & = \frac { r ^ { \prime \prime } } { r ^ { \prime } } y ^ { \prime } \, & ( 2 4 ) & \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
y & = \text{plitu} \\ & \quad \text{to cc}$$
so that they end up with the correct interior distribution for haloes in the new cosmology, but allowing the haloes to maintain any asphericity they may have had. This also means that the haloes retain any dispersion in structure and environmental dependence that they may have had in the original simulation. This requires the input of a c M ( ) relation, and as before we adopt the relations of Bullock et al. (2001). For 'real' haloes it is possible for particles to be outside the official virial radius (calculated from equation 16) for haloes defined with a friends-of-friends (FOF) algorithm, which is used in this work. Using equation (24) is still possible in this case but f > 1 for these particles.
Halo particle velocities, u , relative to the CM velocity, can be reassigned via
$$u ^ { \prime \prime } & = \frac { \sigma _ { v } ^ { \prime \prime } } { \sigma _ { v } ^ { \prime } } u ^ { \prime } \, & ( 2 5 ) & \quad z. \ In \\ & \quad t i o n s$$
with σ v taken from either equation (19) or (18). This should make the halo velocity dispersion appropriate for the new cosmology whilst maintaining the dispersion and environmental dependence in the dispersion relation for haloes of a given mass.
## 3 SIMULATIONS
We now show the operation of halo rescaling in redshift space, using the same matched set of simulations, with corresponding halo catalogues, as in MP14. The simulation parameters are given in Table 2. The target ΛCDM simulation is a WMAP1-style cosmology (Spergel et al. 2003) run with the same transfer function as that of the Millennium Simulation (Springel et al. 2005) which was generated using CMBFAST (Seljak & Zaldarriaga 1996). The input simulation τ CDM is a flat matter-only simulation run with a DEFW transfer function (Davis et al. 1985) tuned to have a simi-
We take this extreme approach (no Λ or baryons) in order to see how well rescaling can work over any range of models of conceivable interest. For smaller variations in cosmological parameters, we may expect that the results will be correspondingly more accurate. Additionally; any practical application of this method would scale to a variety of cosmological models from a single parent simulation of high σ 8 , such that it explored a large range of fluctuation amplitudes. In this case scalings to standard models are likely to come from the Ω m /similarequal 1 regime of the parent simulation anyway, even if it is ΛCDM.
Initial conditions were generated at z i = 199, using a glass initial load of 512 3 particles, created using the N-GenIC code. The simulations were run using the cosmological N -body code Gadget-2 of Springel (2005). As in MP14, we used the same phases for the Fourier modes in the target and original simulations, so that the approximate and exact target halo fields can be compared visually, and not purely at the level of power spectra. This also allows us to analyse the results of the rescaling without the added complication of cosmic variance.
Initially, we ran the original simulation to z = 0 in a box of size L ; but having computed the best scaling parameters ( s , z ) to match the target simulation, we then re-ran the original simulation to generate an output at exactly redshift z . In practice one would need to interpolate particle positions between simulation outputs near to redshift z if one was interested in particles, or constrain the scaling redshift to be one of a set of z (close to the best fit) for which one already had an output. This would be necessary in the case of halo catalogues because halo mergers mean that it is not straightforward to interpolate haloes between catalogues at different epochs. We also ran a simulation of the target cosmology to z ′ = 0 in a box of size L ′ = sL , using the same mode phases and amplitudes for the displacement fields so that structures in the target and scaled simulations could be compared directly without the added complication of cosmic variance.
Halo catalogues were generated with the public FoF code www-hpcc.astro.washington.edu/tools/fof.html , using a linking length of b = 0 2 times . the mean interparticle separation. No attempt was made to reject unbound particles.
For our simulations the best-fit scaling parameters are summarised in Table 3. The match to the mass function produced by this scaling was shown in MP14 and is good
Table 2. Cosmological parameters for our simulations. As a target we use a ΛCDM model with a WMAP1 type cosmology and as an input model we simulate a CDM-only model with a DEFW (Davis et al. 1985) spectrum having a similar spectral shape (Γ = 0 21) . to that of the ΛCDM model but lacking a BAO feature. Each simulation was run with N 3 = 512 3 particles, gravitational forces were softened at 20 h -1 kpc and initial conditions generated using N-GenIC on an initial glass load at a starting redshift z i = 199.
| Simulation | L | Ω m | Ω Λ | Ω b | h | σ 8 | n s | Γ |
|--------------|---------------|-------|-------|-------|------|-------|-------|------|
| ΛCDM | 780 h - 1 Mpc | 0.25 | 0.75 | 0.045 | 0.73 | 0.9 | 1 | - |
| τ CDM | 500 h - 1 Mpc | 1 | 0 | - | 0.5 | 0.8 | 1 | 0.21 |
Table 3. Best fit scaling parameters s and z ′ for scaling between our original τ CDM model and our target ΛCDM model together with the mass range of haloes in the original simulation ( M 1 → M 2 ) and the non-linear wave number in the target cosmology k nl .
| Original | Target | z | z ′ s | M 1 | M 2 | s m | k ′ nl | b eff |
|------------|----------|------|---------|-----------------------------------|-----------------------------------|-------|----------------|---------|
| τ CDM | ΛCDM | 0.22 | 0 1.56 | 2 . 58 × 10 13 h - 1 M /circledot | 7 . 19 × 10 15 h - 1 M /circledot | 0.95 | 0 . 15 h Mpc - | 1.57 |
only to the 10% level across all masses, despite σ R ( ) being matched to within 1% across all scales relevant to these haloes (Fig. 1 of MP14). A similar level of disagreement in the measured mass function was found in AW10 (their Fig. 7) in converting between WMAP1 and WMAP3 cosmologies and plausibly reflects non-universality of the mass function.
$$\underset { \text{these} } { \mathbb { m } { e } } \right ) \, \text{be-} \quad \Delta _ { \ell } ^ { 2 } ( k ) = \frac { 2 \ell + 1 } { 2 } \int _ { - 1 } ^ { 1 } P _ { \ell } ( \mu ) \, \Delta _ { s } ^ { 2 } ( k, \mu ) \, \mathrm d \mu \,, \quad \, \quad ( 2 9 ) \\ \text{nt in} \quad \text{where $P_{\ell}(\nu)$ are the Legendre nolynomials} \cdot P_{\ell}(\nu)=1$ and}$$
## 4 POWER SPECTRA
## 4.1 Estimation
In order to measure power spectra in redshift space, the particles are moved to their redshift-space positions. We use the distant-observer approximation, and move particles to redshift space positions s , via
$$s = x + \frac { v \cdot \hat { x } _ { i } } { a H ( a ) } \hat { x } _ { i } & & ( 2 6 ) & & \text{in} \quad \text{$\quad$}$$
where x i is arbitrarily chosen as one of the three coordinates and v is the peculiar velocity. The density field is then created using nearest grid point (NGP) binning of the particles onto an m 3 mesh; the effect of this in Fourier space is allowed for by dividing by the normalized transform of a cubic cell,
$$J ( { \mathbf k } ) = \sin c \left ( \theta _ { x } \right ) \sin c \left ( \theta _ { y } \right ) \sin c \left ( \theta _ { z } \right ) \,, \quad \quad \ \ ( 2 7 ) \quad \ \ w e \, s l$$
where θ i = k L/ m i 2 .
We compute the 2D anisotropic power spectrum of both haloes and particles by binning linearly in µ = cos θ , where θ is the angle of the wavevector to the chosen line of sight (LOS), and logarithmically in k , between the box scale mode and the Nyquist mode. We use the dimensionless redshiftspace power spectrum, ∆ ( 2 k, µ ), defined as the contribution to the redshift space variance per logarithmic interval in k and linear interval in µ
$$& \sigma _ { s } ^ { 2 } = \int _ { - 1 } ^ { 1 } d \mu \int _ { 0 } ^ { \infty } d \ln k \, \Delta _ { s } ^ { 2 } ( k, \mu ) \,. & \quad ( 2 8 ) \quad \text{ ifec} \\ & \cap \dots \, \ \lambda \, 2 / 2 \,. \, \cdots \, \ \lambda \, \dots \, \dots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \end{array}$$
Once ∆ ( 2 k, µ ) has been computed from the particle or halo distribution we additionally compute the monopole ( /lscript = 0) and quadrupole ( /lscript = 2) moments of the full 2D distribution (e.g. Cole et al. 1994). These are given by where P /lscript ( µ ) are the Legendre polynomials; P 0 ( µ ) = 1 and P 2 ( µ ) = (3 µ 2 -1) / 2. All odd moments vanish due to the symmetry ∆ ( 2 s k, µ ) = ∆ ( 2 s k, -µ ). Almost all the redshiftspace signal resides in the quadrupole, and we will concentrate on this. In order to compute the multipoles of ∆ 2 s ( k, µ ) we fit a model 'monopole + quadrupole' to ∆ ( 2 s k, µ ). This is necessary because this function is sampled sparsely at low k due to the finite periodic geometry of the simulation box, so that equation (29) is not well approximated by a discrete sum. Due to the orthogonality of the multipoles it is not necessary to fit a model with all (infinite) Legendre polynomials included. We always plot ∆ ( 2 2 k /k ) 1 5 . linearly in this work because at low k (box modes) the quadrupole is very noisy and can be negative, despite the Kaiser expectation.
In making estimates of power spectra from a distribution of discrete objects, it is normal to subtract shot noise:
$$\intertext { a t e s } \underset { \text{cre-} } { \ e r } ^ { 2 } ( k ) \to \Delta ^ { 2 } ( k ) - 4 \pi \left ( \frac { k } { 2 \pi } \right ) ^ { 3 } \frac { L ^ { 3 } } { N ^ { 3 } } \,, \quad \quad ( 3 0 ) \\ \text{ticles} \quad \text{where $N^{3}$ is the total number of objects under study. This}$$
where N 3 is the total number of objects under study. This correction is only important at small scales, and the shot noise correction has been shown to be accurate even for wellevolved glass initial conditions (Smith et al. 2003). However, we should note that this correction is inappropriate if we are purely concerned with haloes, rather than the larger number of galaxies or mass particles reconstituted from them. The shot-noise correction attempts to remove a small-scale randomness, but the locations of the haloes have a physical significance: they are the density field. In the halo model, the small-scale nonlinear correlations are of the form of shot noise owing to the finite numbers of haloes, softened via convolution with the halo profile. The halo-only plots in MP14 were in error in this respect, since shot noise was subtracted. But in practice the effect was small and in any case does not affect the differential comparison of models where the mass functions have been adjusted to be identical.
## 4.2 Models
Kaiser (1987) showed that the linear theory result for the redshift space galaxy power spectrum, ∆ 2 s g , , is given by
µ
Figure 1. The redshift-space power spectrum of the full mass distribution as a function of k and µ for the rescaled simulation (left), where the AW10 method has been applied, and in addition we have restructured the internal physical and velocity structure of haloes found in the rescaled particle distribution (Section 2.2.3). This is compared to the spectrum measured in the target simulation (right). Differences in the spectra are difficult to identify visually and residuals are therefore shown in Fig. 2.
$$\Delta _ { s, g } ^ { 2 } ( k, \mu ) = ( b + f _ { g } \mu ^ { 2 } ) ^ { 2 } \Delta _ { m } ^ { 2 } ( k ) = ( 1 + \beta \mu ^ { 2 } ) ^ { 2 } \Delta _ { g } ^ { 2 } ( k ) \, \ \ ( 3 1 ) \quad \text{displ}$$
where ∆ 2 m ( k ) and ∆ ( 2 g k ) the real space matter and galaxy power spectra and β = f /b g . The growth rate features in this equation because the linear velocity field depends on the growth rate in the ZA (equation 4); this gives a universal apparent displacement to all tracers, even though their effective real-space displacements are biased in general. Expanding ∆ 2 s g , ( k, µ ) in terms of Legendre polynomials (Cole et al. 1994), the monopole and quadrupole are given by
$$\stackrel { \text{-\vee } } { \Delta _ { 0 } ^ { 2 } ( k ) } & = \left ( 1 + \frac { 2 } { 3 } \beta + \frac { 1 } { 5 } \beta ^ { 2 } \right ) \Delta ^ { 2 } ( k ) \, \quad \ \underset { \text{-\substack { (3)} \\ \Delta _ { 2 } ^ { 2 } ( k ) } } { \Delta _ { 2 } ^ { 2 } ( k ) } & = \left ( \frac { 4 } { 3 } \beta + \frac { 4 } { 7 } \beta ^ { 2 } \right ) \Delta ^ { 2 } ( k ) \, \quad \ \underset { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack { \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{ \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{-\substack \text{ }
- \cdot \........................................................................................................................................................................................................
- + \cdot \...................................................................................................................................................................................................$$
and their ratio is a single function of β :
$$G ( k ) \equiv \frac { \Delta _ { 2 } ^ { 2 } ( k ) } { \Delta _ { 0 } ^ { 2 } ( k ) } \simeq \frac { 1 + \frac { 2 } { 3 } \beta + \frac { 1 } { 5 } \beta ^ { 2 } } { \frac { 4 } { 3 } \beta + \frac { 4 } { 7 } \beta ^ { 2 } } \. \quad \quad \quad ( 3 3 ) \quad \quad \quad 5 \ 1$$
It is infeasible to predict b with sufficient precision to convert this to the desired growth rate, but the apparent amplitude of real-space clustering can be inferred accurately, so the pure dark-matter combination f σ g 8 can be measured.
Equation (31) is only valid on linear scales and breaks down more quickly than the linear approximation for the real-space power spectrum due to velocities becoming nonlinear before densities (Scoccimarro 2004). More complicated expressions for the redshift-space power exist in the literature (e.g. Scoccimarro 2004; Taruya et al. 2010) and some prescriptions are compared by de la Torre & Guzzo (2012). These models all attempt to deal with two related issues: (1) deviations of the real-space power spectrum from linear theory; (2) additional sources of anisotropy beyond the bare Kaiser model. The latter effects amount to more rapid damping of modes that lie close to the line of sight (e.g. Kwan et al. 2012), known under the generic title of 'fingersof-God' (FOG), since this damping amounts to a radial convolution in configuration space. A common simple model displaying both these effects is the 'Kaiser+Lorentzian' (Peacock & Dodds 1994), in which the nonlinear power spectrum is subject to the Kaiser anisotropy and a radial damping controlled by the pairwise dispersion σ :
$$\stackrel { \sim } { \text{esal} } } _ { \text{effect} ^ { - } } \quad \Delta _ { s } ^ { 2 } ( k, \mu ) = \frac { ( b + f _ { g } \mu ^ { 2 } ) ^ { 2 } \Delta _ { N L } ^ { 2 } ( k ) } { 1 + k ^ { 2 } \sigma ^ { 2 } \mu ^ { 2 } / 2 } \. \quad \quad \quad ( 3 4 )$$
In practice, there are two distinct sources of FOG effects: pure quasi-linear terms (which can be estimated e.g. via the ZA), plus the internal velocity dispersion of haloes. In our modelling, the first kind of FOG should be included directly, since we work with a dynamically realistic halo catalogue. But the internal velocity dispersions are an important contributors to the overall effect, and we need to understand the effect on these of internal halo rescaling.
## 5 RESULTS OF RESCALING
## 5.1 Enhanced AW10 approach
We now show a number of examples of the effect of rescaling on the redshift-space power spectrum. We begin with the original AW10 method applied to a full particle distribution, with additional restructuring of internal halo properties as in Section 2.2.3. The full redshift space power spectrum before and after scaling is shown in Fig. 1 where differences in the power between the rescaled and target simulation are difficult to see by eye. Residuals are therefore shown in in Fig. 2 together with the monopole and quadrupole moments. The monopole is very well reproduced (at the 1% level up to k = 0 1 . h Mpc -1 ), including good recovery of the BAO feature from the large-scale displacement step, as in the case of real space. Deviations are seen at smaller scales if no changes are made to the internal structure of haloes. Improvements are gained either by restructuring the haloes or by replacing them entirely. To reconstitute haloes of a given mass we assigned halo virial radii according to equation (16), took
Figure 2. Redshift-space results for a rescaled particle distribution, using the original AW10 method supplemented by various alterations of halo internal structure. The upper two panels show the monopole (left) and quadrupole (right) power spectra with the arrow showing the non-linear scale. The monopole is accurate to 3% to k = 1 h Mpc -1 if the restructuring method is used, including good reproduction of the BAO feature via the displacement field correction. The quadrupole is less well reproduced with errors at the 5% level at large scales, but clearly applying the displacement field step of the method improves the match, and restructuring extends the match to quasi-linear scales ( k /similarequal 0 2 . h Mpc -1 ). At non-linear scales it is regurgitation that performs better. The lower four panels show the fractional residuals for the full redshift-space power spectrum at each stage of the rescaling process when compared to the target ΛCDM simulation. The middle left panel shows the scaling in redshift and size and the middle right panel shows the addition of the displacement field step. The bottom left panel then shows the effect of removing haloes and regurgitating haloes with theoretical velocity dispersions back into the parent particle distribution; the bottom right panel shows the effect of restructuring haloes. Regurgitation and restructuring haloes perform slightly different in the non-linear portions of redshift space. It is clear from this plot that the good agreement of the restructuring monopole is partially due to cancellations of errors across the full redshift-space plane.
concentrations from the Bullock et al. (2001) relations and assigned Gaussian velocity dispersions from equation (19). Haloes were created in an aspherical manner, as described in Section 2, although we found that the asphericity mattered very little at the level of the redshift-space power.
In contrast to the case of real space, shown in Fig. 8 of MP14, restructuring haloes performs better than regurgitation when compared in redshift space. Looking at the ratios to the target in the full redshift-space power information, shown in Fig. 2, one can see that regurgitation performs better for perpendicular modes that are unaffected by redshift distortions, but that restructuring performs slightly better across all modes, accounting for the slightly better overall prediction for the monopole. The eventual monopole in the restructured case is good to 3% up to k = 1 h Mpc -1 , whereas the regurgitated monopole is good to 5%. The quadrupole is noisier and 5% deviations can be seen up to the non-linear scale where the fractional error blows up as the quadrupole changes sign. However, it is still evident that the displacement field step of the method provides a better quadrupole at large scales.
## 5.2 Rescaling halo catalogues
We now turn to the results of rescaling the halo catalogues directly. Fig. 3 compares the redshift-space power of the rescaled halo distribution to that of the target haloes. As in real space, the original AW10 method fails to handle the BAOfeature correctly until haloes are given mass-dependent additional displacements. But the error is smaller in redshift space, because the additional velocity field does not depend on halo mass. The agreement in the monopole power is around 5% for the full halo sample out to k = 1 h Mpc -1 , but contains a slant which is shot noise due to the numbers of haloes not being perfectly matched (recall that the mass functions still differ at the 10% level post scaling). Note that although the non-linear scales have not been modified here the match is still reasonable due to the lack of an internal velocity induced FOG in the halo distribution. However, the rescaled quadrupole is significantly less accurate than the monopole, with residuals at the 10% level. Moreover, there is some suggestion of a remaining large-scale residual that correlates with the BAO signal, of amplitude about 5% and there are some large deviations at large scales. Additionally the displacement field step of the method seems to improve matters only marginally.
## 5.3 Reconstituted haloes
We now show the results of reconstituting haloes by modifying the positions and velocities of the halo particles. It is illuminating to test this aspect of our approach directly in an unscaled simulation (the L = 780 h -1 Mpc ΛCDM simulation discussed in Section 3). We show the redshiftspace monopole, quadrupole and the full 2D redshift-space power spectrum in Fig. 4. In both figures we compare results using reconstituted halo particles to the 'exact' results based on the particles from which the halo catalogue was constructed. This tests our assumptions about both the halo internal physical and velocity structure. Haloes were reconstructed with NFW profiles (equation 15) with concentrations from Bullock et al. (2001) and velocity dispersion was assumed to be Gaussian with a standard deviation given by equation (19). The agreement in the monopole is almost perfect (within 1% up to k = 1 h Mpc -1 ) if catalogued velocity dispersions are used, but deviations occur when using a theoretically predicted σ v , which predicts too high a power, suggesting that σ v is too low and FOG are not strong enough. This can be seen via a direct comparison of predicted and catalogued dispersions, which show the predicted value to be low by a factor of 1 07. According to equation . (16), such a factor would arise if the virialized overdensity threshold was raised to about ∆ v /similarequal 300. We note that an increased ∆ v is consistent with spherical model predictions for ΛCDM models (e.g. Bryan & Norman 1998) but is incompatible with our choice of b = 0 2 as a linking length. . We have not pursued the possibility of changing ∆ v as we do not want to introduce free parameters that may have an unknown dependence on cosmology or simulation resolution (e.g. Power & Knebe 2006; Smith et al. 2014).
Finally in Fig. 5 the redshift-space power of particles in reconstituted haloes is compared to that of particles in haloes in the target cosmology. The agreement is good at linear scales once a biased displacement field has been used. At non-linear scales the agreement is less good if one uses the theoretical dispersion relation in equation (19) compared to using a rescaled version of the catalogued halo velocity dispersion; in the later case the eventual agreement for the monopole is impressive and is at the 3% level across all scales shown. Although the quadrupole has a larger error, this still remains within 5% in regions where it is not passing through zero. In 2D redshift space one can see that large fractional errors in the catalogued dispersion case are concentrated at high k and high µ ; these contribute less to the monopole because the full power is damped here and the monopole is a simple average. Using theoretical dispersions perform worse across the entirety of redshift space, even at surprisingly large scales for high µ . This plot, combined with the tests in Fig. 4 gives hope that the theoretical dispersion relation could be modified slightly to produce better results. We have not pursued this here as our aim was to see how far one could get without resorting to tuning additional parameters.
## 5.4 Recovery of the growth rate
A realistic goal of forthcoming galaxy redshift surveys will be to measure the growth rate of cosmic structure with ∼ 1% accuracy (e.g. DESI , Euclid ). Can such a precision be attained via simulations that have undergone rescaling? Fig. 6 shows the recovered G k ( ) = ∆ ( 2 2 k / ) ∆ ( ) ratio for: the full 2 0 k rescaled particle distribution; the rescaled halo distribution; the distribution of particles in haloes reconstituted from halo catalogues.
At large scales for the case of the full matter distribution all approaches are accurate to around 5% out to k = 0 1 . h Mpc -1 . If haloes are restructured this accuracy is extended out to k = 0 2 . h Mpc -1 . For the case of both the haloes and halo particles large deviations are seen at the largest scales which remedy themselves to the 7% level for haloes, and 5% level for halo particles, out to the non-linear scale.
In each plot the linear theory G lin is shown (equation 33); for the full matter distribution G lin /similarequal 0 55. . and for the haloes, we expect G lin /similarequal 0 37 . or 0 28 . for respectively
Figure 3. The upper two panels show the halo redshift-space monopole (left) and quadrupole (right) power at each stage of the scaling with the arrow denoting the non-linear scale. The monopole is recovered at the few % level across all scales, but the match surprisingly degrades somewhat at large scales. As in MP14 applying a biased displacement field ( zsd ) to haloes outperforms the original AW10 method (AW10). The slant is shot noise due to differing halo numbers, which arises due to the slightly imperfect match for the halo mass function. The quadrupole is only recovered at the 10% level and contains large residuals that are only very slightly remedied by the displacement field correction. The lower panels show the fractional residuals of the 2D redshift-space power spectrum of the original to target simulations after scaling in box size and redshift (bottom left; equivalent to zs curve) and after additionally applying a biased displacement field correction, and a velocity field correction, to halo positions (bottom right; equivalent to zsd curve). The match is improved by applying the displacement and velocity field correction in the final stage of the method. The match is reasonable at small scales because there are no strongly non-linear FOG effects in the halo distribution.
the equal-weighted and mass-weighted cases (taking into account the effective bias discussed in Section 2). Fig. 6 indicates that the rescaled simulations are consistent with these growth rates but that large deviations are seen around the box scale, although in the case of haloes a simulation containing larger scales (thus more linear modes) would help to support this conclusion.
a simple proxy for a galaxy catalogue. Here, the last panel in Fig. 6 shows that usefully precise results to wavelengths large enough to extract most of the cosmological information ( k /similarequal 0 3 . h Mpc -1 ) would require improved treatment of the internal halo velocity dispersion, with the rescaled values and pure theoretical values bracketing the truth.
At smaller scales, the results are most satisfactory in the case of the halo distribution, where the error in the quadrupole-to-monopole ratio is relatively flat, at typically around 4%. There are occasional spikes to higher values, but these are unlikely to disrupt the recovery of the growth rate at this typical error rate. Conversely, the results for the overall mass are less impressive, with errors exceeding 5% for a broad band of wavenumbers beyond the nonlinear scale. Fortunately, this case is not of practical interest, whereas we are most concerned with particles in haloes as
In a practical survey analysis one could marginalize over velocity dispersion nuisance parameters, but this would diminish the statistical power of the highk data. Some encouragement comes from the fact that the non-linear portion of G k ( ) for haloes is particularly well matched, which is related to the fact that they lack a strongly non-linear FOG. It is plausible that future surveys may attempt to target only halo central galaxies in order to mitigate the effect of FOG in this way, or to use other weighting schemes (e.g. Seljak et al. 2009).
In any case, the example illustrated here undoubtedly
Figure 4. A comparison of the power spectra of particles in haloes reconstituted from a halo catalogue, compared to the original simulation particles in the haloes. The upper panels show the monopole (left) and quadrupole (right) power while the lower two panels show the fractional residuals for the full redshift-space power spectrum as function of k and µ in the case when haloes are assigned velocity dispersions using equation (19; left; equivalent to 'theory' curve above) as compared with the measured velocity dispersions (right; equivalent to 'catalogue' curve above). The error in the quadrupole blows up around k = 0 1 . h Mpc -1 where the quadrupole power changes sign. Using the catalogued value of the dispersions produces more accurate results, with the reconstituted monopole being almost perfectly reproduced (within /similarequal 1%) up to k = 1 h Mpc -1 . The lower panels show that the error from using the predicted dispersion differs across redshift space and extends to larger scales than that from using the catalogued dispersions. The monopole from using catalogued dispersions is recovered well because errors here are concentrated in the highµ -highk region of the plane where the absolute magnitude of the power is low due to FOG. This means that the monopole is relatively unaffected but this is less true of the quadrupole, which differences in µ . The match is perfect at large scales in all cases because the haloes are placed perfectly accurately by construction.
makes greater demands on the method than would be encountered in most practical applications, where the range of parameter space to be spanned would consist of models that are closer to ΛCDM. The above errors would undoubtedly decline in proportion to the distance between models in parameter space. This in itself could create new problems: if the degree of rescaling is very small, then there will be little chance of finding an output at an input redshift z suitable to match a desired target z ′ . Instead, we may have to accept that the original simulation defines a grid in z ′ for the target cosmology, over which we have no choice. But provided the outputs are reasonably well spaced, this need not be a problem: construction of a light-cone mock galaxy sample from HOD galaxies populating the snapshots would proceed in much the same way, independent of the exact set of z ′ values.
## 6 SUMMARY AND CONCLUSIONS
In this paper we have considered the rescaling of simulated dark-matter haloes in order to create catalogues of haloes that are characteristic of an altered cosmological model, focusing on the behavior of the method in redshift space. This extends our previous work in real space (Mead & Peacock 2014; MP14), which was based on the original proposal of Angulo & White (2010; AW10). MP14 demonstrated that rescaling of simulations could be made to work in real space directly from halo catalogues, without needing to manipu-
Figure 5. The upper two panels show the monopole (left) and quadrupole (right) power spectra of particles in reconstituted haloes compared to particles in haloes in the target simulation. If the original AW10 method is used (AW10) and haloes are reconstituted with catalogued halo velocity dispersions a large residual BAO feature is seen up to the non-linear scale (arrow), which is removed if a biased displacement field is used as in the other two curves. In these two cases better results are obtained for the monopole and quadrupole if rescaled catalogued dispersions are used (catalogue) compared to theoretical dispersions (theory). However, the quadrupole is relatively insensitive to the method used in the scaling and shows a larger error which remains at the 5% level on linear scales. The fractional residuals of the full redshift-space power spectrum is shown in the lower two panels from using biased displacements together with theoretical velocity dispersions (left, equivalent to 'theory' curves above) and rescaled catalogued dispersions (right, equivalent to 'catalogue' curves above). Errors in the catalogued case are concentrated at high k and high µ , where there is less overall power, resulting in a better recovery of the monopole and quadrupole.
late the full particle data. In addition to the resulting advantage in speed and storage, this approach also allows improvements over AW10 in terms of the dependence of bias on halo mass and the small-scale clustering, which is affected by internal changes to halo structure.
For these reasons, it is clearly of interest to see if the MP14 approach also works acceptably in redshift space, allowing it to be used as the basis for a complete approach to the construction of simulated surveys. We have therefore carried out the extreme test of rescaling a halo catalogue generated from a matter-only τ CDM simulation into that of a more standard ΛCDM model. In redshift space the MP14 method works well at the level of the monopole, as in the original AW10 case. For the full particle distribution the redshift-space monopole was recovered at the 2% level up to k = 0 1 . h Mpc -1 and to 3% to smaller scales if halo internal properties are also manipulated. Rescaling also worked well when applied to haloes in redshift space, although the improvements gained from using a biased displacement field are less marked because redshift space mixes in the unbiased velocity field. The monopole agrees at the few % level out to k = 1 h Mpc -1 , although the numerical results show some deviations at the very largest scales investigated and the origin of these is not known. For reconstituted haloes, the monopole power spectrum was recovered at the 1% level up to k = 0 1 . h Mpc -1 if a biased displacement field is used and the agreement is at the 3% level up to k = 1 h Mpc -1 if catalogued dispersions are also rescaled.
The quadrupole to monopole ratio, G k ( ), is of especial interest because it can be used in the linear regime to infer the growth rate of cosmic structure. We found that this could be recovered at the typical level of around 5% for haloes themselves, even out to wavenumbers as high as k = 1 h Mpc -1 . This is most encouraging given how far our
Figure 6. The value of G k ( ) = ∆ ( 2 2 k / ) ∆ ( ) recovered from the simulations before and after scaling. In each case the non-linear scale is 2 0 k shown with an arrow. The top left panel shows the case of the full particle distribution where methods of regurgitation and restructuring have been used to alter halo interiors. Here the residual difference is small across the full range of scales, and is at the 5% level on linear scales, with the non-linear tail being reproduced at the 5% level by all methods, in the region where the quadrupole is not passing through zero. The full range of scales is best matched by the restructuring method. The top right panel shows the case of haloes themselves, whereas the bottom panel shows the case of reconstituted halo particles with both theoretical and catalogued velocity dispersions used in the reconstitution, in both cases a comparison is made with the original AW10 method, which performs well here because the residual BAO is effectively divided out and with it the error due to a slightly incorrect b M ( ) relation. For haloes the value of G k ( ) is reproduced at the 10% level across all scales and when the halo particles are considered the match is at the 5% level when the quadrupole is not passing through zero and the non-linear tail is fairly well recovered in this case.
input τ CDM model is from anything close to ΛCDM, and it suggests that the required 1% precision should readily be obtained for models that are allowed by existing cosmological constraints.
But for particles in haloes (an example of a galaxy HOD, albeit an unrealistic one), the highk precision in G k ( ) is considerably poorer, reflecting the difficulty in achieving an appropriate level for the internal halo velocity dispersions using simple scaling recipes. Again this problem will diminish (at least in the case of rescaling 'observed' dispersions) as the input model approaches closer to the target. There are several routes by which the modelling of the halo-particle power spectrum may be improved - either via more sophisticated dynamical modelling of the halo velocity dispersion, or via weighting schemes that suppress this component of the FOG. We therefore see the current results as giving encouragement that halo rescaling will be a practically useful modelling method in the era when data on redshift-space distortions reaches the 1% level.
## ACKNOWLEDGEMENTS
AJM acknowledges the support of an STFC studentship and support from the European Research Council under the EC FP7 grant number 240185.
## REFERENCES
Anderson L. et al., 2014, MNRAS, 439, 83 Angulo R. E., White S. D. M., 2010, MNRAS, 405, 143 Bryan G. L., Norman M. L., 1998, ApJ, 495, 80 Bullock J. S., Kolatt T. S., Sigad Y., Somerville R. S., Kravtsov A. V., Klypin A. A., Primack J. R., Dekel A., 2001, MNRAS, 321, 559 Cole S., Fisher K. B., Weinberg D. H., 1994, MNRAS, 267, 785 Cooray A., Sheth R., 2002, Physics Reports, 372, 1 Davis M., Efstathiou G., Frenk C. S., White S. D. M., 1985,
ApJ, 292, 371
## 14 A. J. Mead and J. A. Peacock
de la Torre S., Guzzo L., 2012, MNRAS, 427, 327 de la Torre S. et al., 2013, A&A, 557, A54
Dooley G. A., Griffen B. F., Zukin P., Ji A. P., Vogelsberger
M., Hernquist L. E., Frebel A., 2014, ApJ, 786, 50
Eisenstein D. J., Seo H.-J., Sirko E., Spergel D. N., 2007, APJ, 664, 675
Eke V. R., Cole S., Frenk C. S., 1996, MNRAS
Eke V. R., Navarro J. F., Steinmetz M., 2001, ApJ, 554, 114 Guo Q., White S., Angulo R. E., Henriques B., Lemson G., Boylan-Kolchin M., Thomas P., Short C., 2013, MNRAS, 428, 1351
Kaiser N., 1987, MNRAS, 227, 1
Kazantzidis S., Magorrian J., Moore B., 2004, ApJ, 601, 37 Kwan J., Lewis G. F., Linder E. V., 2012, ApJ, 748, 78
Lacey C., Cole S., 1993, MNRAS, 262, 627
Li B., Efstathiou G., 2012, MNRAS, 421, 1431
Luki´ c Z., Heitmann K., Habib S., Bashinsky S., Ricker P. M., 2007, ApJ, 671, 1160
Mead A. J., Peacock J. A., 2014, MNRAS, 440, 1233
Merritt D., Graham A. W., Moore B., Diemand J., Terzi´ c
B., 2006, AJ, 132, 2685
Merritt D., Navarro J. F., Ludlow A., Jenkins A., 2005, ApJL, 624, L85
Moore B., Quinn T., Governato F., Stadel J., Lake G., 1999, MNRAS, 310, 1147
Navarro J. F., Frenk C. S., White S. D. M., 1997, ApJ, 490, 493
Neto A. F. et al., 2007, MNRAS, 381, 1450
Padmanabhan N., Xu X., Eisenstein D. J., Scalzo R., Cuesta
- A. J., Mehta K. T., Kazin E., 2012, MNRAS, 427, 2132
Peacock J. A., Dodds S. J., 1994, MNRAS, 267, 1020
Peacock J. A., Smith R. E., 2000, MNRAS, 318, 1144
Percival W. J., 2005, A&A, 443, 819
Power C., Knebe A., 2006, MNRAS, 370, 691
Press W. H., Schechter P., 1974, ApJ, 187, 425
Rasera Y., Alimi J.-M., Courtin J., Roy F., Corasaniti P.S., F¨ uzfa A., Boucher V., 2010, in American Institute of Physics Conference Series, Vol. 1241, American Institute of Physics Conference Series, Alimi J.-M., Fu¨zfa o A., eds., pp. 1134-1139
Reed D. S., Bower R., Frenk C. S., Jenkins A., Theuns T., 2007, MNRAS, 374, 2
Ruiz A. N., Padilla N. D., Dom´ ınguez M. J., Cora S. A., 2011, MNRAS, 418, 2422
Samushia L. et al., 2013, MNRAS, 429, 1514
Schmidt F., Lima M., Oyaizu H., Hu W., 2009, Phys. Rev.
## D, 79, 083518
Scoccimarro R., 2004, Phys. Rev. D, 70, 083007
Seljak U., 2000, MNRAS, 318, 203
Seljak U., Hamaus N., Desjacques V., 2009, Physical Review Letters, 103, 091303
Seljak U., Zaldarriaga M., 1996, ApJ, 469, 437
Sheth R. K., Mo H. J., Tormen G., 2001, MNRAS, 323, 1
Sheth R. K., Tormen G., 1999, MNRAS, 308, 119
Simha V., Cole S., 2013, MNRAS, 436, 1142
Smith R. E., Desjacques V., Marian L., 2011, Phys. Rev. D, 83, 043526
Smith R. E. et al., 2003, MNRAS, 341, 1311
Smith R. E., Reed D. S., Potter D., Marian L., Crocce M.,
Moore B., 2014, MNRAS, 440, 249
Spergel D. N. et al., 2003, ApJs, 148, 175
Springel V., 2005, MNRAS, 364, 1105
Taruya A., Nishimichi T., Saito S., 2010, Phys. Rev. D, 82,
Taylor A., Joachimi B., Kitching T., 2013, MNRAS, 432,
Tinker J., Kravtsov A. V., Klypin A., Abazajian K., Warren M., Yepes G., Gottl¨ber S., Holz D. E., 2008, ApJ, 688, o
Wojtak R., /suppress Lokas E. L., Gottl¨ber S., Mamon G. A., 2005, o
Springel V. et al., 2005, Nature, 435, 629 063522 1928 709 MNRAS, 361, L1 Zel'dovich Y. B., 1970, A&A, 5, 84 Zheng Z. et al., 2005, ApJ, 633, 791
|
10.1093/mnras/stu1964
|
[
"Alexander Mead",
"John Peacock"
] |
2014-08-05T17:44:26+00:00
|
2014-11-24T19:49:22+00:00
|
[
"astro-ph.CO"
] |
Remapping simulated halo catalogues in redshift space
|
We discuss the extension to redshift space of a rescaling algorithm, designed
to alter the effective cosmology of a pre-existing simulated particle
distribution or catalogue of dark matter haloes. The rescaling approach was
initially developed by Angulo & White and was adapted and applied to halo
catalogues in real space in our previous work. This algorithm requires no
information other than the initial and target cosmological parameters, and it
contains no tuned parameters. It is shown here that the rescaling method also
works well in redshift space, and that the rescaled simulations can reproduce
the growth rate of cosmological density fluctuations appropriate for the target
cosmology. Even when rescaling a grossly non-standard model with Lambda=0 and
zero baryons, the redshift-space power spectrum of standard LCDM can be
reproduced to about 5% error for k<0.2h Mpc^-1. The ratio of
quadrupole-to-monopole power spectra remains correct to the same tolerance up
to k=1h Mpc^-1, provided that the input halo catalogue contains measured
internal velocity dispersions.
|
1408.1050v1
|
## THE ASTROSPHERE OF THE ASYMPTOTIC GIANT BRANCH STAR CIT6
Raghvendra Sahai
Jet Propulsion Laboratory, MS 183-900, California Institute of Technology, Pasadena, CA 91109
[email protected] and
Galen P. Mack-Crane
Department of Physics, Occidental College, Los Angeles, CA 90041
## ABSTRACT
We have discovered two extended half-ring structures in a far-ultraviolet image taken with the GALEX satellite of the well-known mass-losing carbon star CIT6 (RWLMi). The northern (southern) ring is brighter (fainter) with a diameter of ∼ 15 ′ ( ∼ 18 ). ′ These structures most likely represent the astrosphere resulting from the shock interaction of CIT 6's molecular wind with the Warm Interstellar Medium, as it moves through the latter. These data provide a direct estimate of the size of CIT 6's circumstellar envelope that is a factor ∼ 20 larger than previous estimates based on CO millimeter-wave line data. We find that CIT6 has been undergoing heavy mass-loss for at least 93,000 yr and the total envelope mass is 0 29 . M /circledot or larger, assuming a constant mass-loss rate of 3 2 . × 10 -6 M /circledot yr -1 . Assuming that the shock front has reached a steadystate and CIT6's motion relative to the ISM is in the sky-plane, we measure the termination-shock standoff distance directly from the image and find that CIT6 is moving at a speed of about /greaterorsimilar 39 (0.17 cm -3 /n ISM ) 1 2 / km s -1 through the interstellar medium around it. However, comparisons with published numerical simulations and analytical modelling shows that CIT 6's forward shock (the northern ring) departs from the parabolic shape expected in steady-state. We discuss several possible explanations for this departure.
Subject headings: stars: AGB and post-AGB, stars: mass-loss, stars: individual (CIT6), circumstellar matter, reflection nebulae
## 1. Introduction
The carbon-rich AGB star CIT6 (RWLMi) is probably the most well-studied carbon star after IRC+10216 that is known to be experiencing heavy mass-loss during its Asymptotic Giant Branch (AGB) evolution. Such stars eject large quantities of processed material, enriched with carbon manufactured in their interiors as a result of 3α nucleosynthesis, into the interstellar medium via extensive dusty molecular winds that operate during the AGB phase. At 400 pc, CIT 6 is somewhat more distant than IRC+10216, and has been extensively observed from radio to optical wavelengths, with a variety of imaging and spectroscopic techniques.
The central star is a long-period variable with a period of about 640 days (Alksnis 1995), a bolometric luminosity of about 10 4 L /circledot , and an average mass-loss rate of 3 2 . × 10 -6 M /circledot yr -1 (Zhang et al. 2009) resulting in a large circumstellar envelope (CSE) expanding at about 18 km s -1 . HST imaging at optical and near-infrared wavelengths reveals the presence of a small, roughly bipolar nebula, suggesting that the object is transitioning into the pre-planetary nebula phase (Schmidt et al. 2002). These authors also found the presence of faint, diffuse arcs 1 ′′ -4 ′′ from the central star, and suggested that the primary star has a main-sequence companion of spectral type A-F at a separation greater than 40 AU. Recently, Claussen et al. (2011) discovered the presence of multiple, partial circumstellar ring structures in CIT 6 at even larger distances from the center (up to ∼ 8 ), ′′ from their mapping of HC N J=4-3 emission using the VLA. These arc structures have been interpreted 3 and modeled as a 3-dimensional spiral-shock structure induced in the CSE due to the presence of a binary companion (Kim et al. 2013).
The full extent of CIT 6's CSE has been traced most sensitively in CO J=1-0 emission, and has a half-power diameter of 35 ′′ (Neri et al. 1998). But since the outer extent of CO emission is limited by the photodissociation of this species by the interstellar ultraviolet radiation to ∼ 2 × 10 17 cm, direct evidence for AGB mass-loss in CIT6 is limited to a relatively short (expansion) time-scale of about ∼ 3500 yr. Thus the total amount of matter
ejected into the ISM by CIT 6, which depends linearly on the envelope's outer extent if the mass-loss has been constant, remains unknown, and likely vastly underestimated.
In this paper, we report deep GALEX images that trace the CIT6 CSE to an outer radius that is more than an order of magnitude larger than the above estimate, and likely represents the full duration of the current heavy mass-loss. The outer edge of the CSE has become visible as a result of its interaction with the ISM as CIT 6 moves through the latter. We report our analysis of the shape, size and structure of the CSE-ISM interaction, and compare our results to those expected from theoretical models and numerical simulations. We determine CIT6's motion relative to the local ISM and provide new lower limits for the duration of heavy mass-loss and the total mass of ejecta in this object.
## 2. Observations & Results
We retrieved pipeline-calibrated FUV and NUV images of CIT6 from the GALEX archive; the bandpass (angular resolution) is 1344-1786 ˚ (4 5 ) and 1771-2831 ˚ (6 0 ), reA . ′′ A . ′′ spectively, and the pixel size is 1 5 . ′′ × 1 5 . ′′ (Morrissey et al. 2005). The data were taken on 2009 Feb 01, each with an exposure time of 30824 sec. In Fig. 1a, we show an FUV image of the star and its CSE, and in Fig. 1b, the corresponding NUV image. Field stars in the FUV image have been removed using a customised IDL routine which replaces a small region covering each star's PSF with a tile of random noise representative of the surrounding sky. The sky noise was sampled separately at the four corners of each tile and linearly interpolated throughout, so as to preserve gradients in the local sky background to first order.
A magnified view of the relatively bright FUV nebula seen around CIT 6 is shown in Fig. 2. Bright nebulosity can be seen in the center of the FUV image, around the location of CIT6's central star. In addition, the image shows two bright, extended (size ∼ 14 ) half-ring ′ structures, with diameters of about 15 ′ and 18 . ′ No detectable counterpart to these ring structures is found in the NUV image (Fig. 1b).
Although the northern ring appears to be roughly circular around the central star's location, closer inspection supported by examination of radial intensity cuts at different position angles (Fig. 3) shows it be flattened in the northerly direction. On the southern
side, the ring is not as prominently limb-brightened as in the north, mostly appearing as a bright edge which is roughly circular but at a larger radius from the central star than the northern ring (Fig. 4).
## 3. An Astrosphere around CIT6
In a large imaging survey at 70 and 160 µ m with the Herschel Space Observatory that revealed bow shocks for ∼ 40%, and detached rings for ∼ 20%, of a sample of 78 evolved stars (AGB stars and red supergiants), Cox et al. (2012: hereafter Cetal12) did not find any wind-ISM interaction structure around CIT 6. However, they did find 'eye' shaped wind-ISM interaction structures ('two elliptical non-concentric arcs at opposing sides of the central source, both have a covering angle of /lessorsimilar 180 ') around seven AGB stars that resemble ◦ the northern and southern FUV ring structures in CIT 6. Given the strong indirect evidence that CIT6's central star is a binary, it is interesting that Cetal12 find that 5 of their 7 'eye' objects show evidence of binarity . 1 Cetal12 stated that their data were not adequate to either confirm or exclude a connection between binarity and the 'eyes' morphology. With the inclusion of CIT 6 to this list, the evidence for such a connection is strengthened.
A plausible interpretation of the FUV emission ring structures is that they represent the interaction of the expanding CSE of CIT 6 with the local ISM. We consider alternative interpretations later, but find them less likely ( § 6). The shorter radial distance and greater brightness of the northern ring from the star, compared to that of the southern one, implies that the star is moving roughly northwards through the local ISM, producing a strong shock front at the northern outer edge of its CSE.
The northern FUV ring around CIT6 thus represents the astrosheath with the outer edge of this ring corresponding to the astropause, and the inner edge to the termination shock (see Fig. 2d, Ueta 2008). The region interior to the latter consists of the unshocked, freely-streaming stellar wind; the innermost part of this region is seen due to the scattering of ambient Galactic starlight from dust in the wind.
1 these 5 include 2 potential binaries and 3 visual binaries: see Cetal12's Table 1 for details and references
A similar large astrosphere was found around IRC+10216 in GALEX images by Sahai & Chronopoulos (2010: SC10). Assuming, as for IRC+10216's astrosheath (SC10), that the FUV emission mechanism is due to collisional excitation of H 2 by hot electrons in shocked gas, which produces no detectable counterpart in the NUV band (see Martin et al. 2007), the non-detection of the ring structure in the NUV is not surprising. We note that the brightest region of the astrosheath has an excess FUV intensity (over the background) of about 0 013 . µ Jypix -1 , whereas the 3 σ noise in the NUV in this region is 0 036 . µ Jypix -1 . Assuming the relative FUV-to-NUV brightness for the astrosheath in CIT 6 is the same as in IRC+10216 (about 6, SC10), the CIT 6 NUV image, even with smoothing to reduce the noise, lacks the sensitivity needed to detect a counterpart to the FUV emission
We follow a similar procedure as described in SC10 to analyze CIT 6's astrosphere. We measured the northern ring's radial offset in different directions from the central star, using radial intensity cuts at different position angles. Since the emission from the ring (the astrosheath) is rather faint, we averaged the intensity over seven 23 ◦ wedges spanning the northern limb. These cuts (Fig. 3) show that the radius of the astrosheath varies systematically, reaching a minimum roughly in the northward direction, and implying that CIT 6 motion through the local ISM is indeed northward.
We have fit a model radial intensity curve derived from a limb-brightened spherical shell to the FUV radial brightness profiles, assuming the surface brightness to be proportional to the column density, and extracted the astrosheath's inner and outer radii ( R 1 and R c , using the nomenclature in Fig. 1 of Weaver et al. 1977). We assumed a two-piece inverse-square density profile in our model, one for 120 ′′ /lessorsimilar r < R 1 , and the other for R > r > R c 1 , with a jump in density at r = R 1 .
From a 46 ◦ wedge centered at PA = 2 ◦ which encloses the symmetry axis (Fig. 5a), we find R 1 = 373 ′′ and R c = 396 . These values of ′′ R 1 and R c are not sensitive to the assumed density profiles within these two regions, as they are largely determined by the radial location of the intensity peak and the radial width of the steeply-falling intensity curve just beyond this peak (e.g., see model fit in Fig. 5a).
We cannot derive absolute values of the densities from our modelling since the proportionality factor between the brightness and the column density is purely phenomenological;
furthermore, since the emission mechanisms in the two regions are different, the value of the derived density jump is not physical.
We note the systematic presence of a 'shoulder' in the FUV radial intensity cuts, extending to about 50 ′′ beyond the outer edge of the astropause ( r = R c ) in the four cuts nearest to the symmetry axis. Although comparably bright features are also seen at larger radii beyond this shoulder, their location varies from cut to cut. It is plausible that these features have the same origin as other patchy emission regions that are present in many other parts of the full field-of-view (see Fig. 1a), and are likely due to scattered light from dust in the ISM, unrelated to CIT 6. Our tentative conclusion is that the shoulder emission is due to a coherent structure that lies just outside the astropause, i.e., r > R c - this structure can be seen marginally in Fig. 5b, and probably represents the bow-shock interface separating the shocked and unshocked ISM. A similar structure was found by SC10 in IRC+10216.
The post-shock temperature in the bow-shock region is expected to be high, about (3 / 16 k ) ¯ µV 2 ∗ ∼ 2 × 10 K (assuming a strong shock, where the stellar velocity relative to the 4 ISM, V ∗ = 39km s -1 , § 4), where ¯ µ ∼ 10 -24 g is the mean mass per particle for fully ionized gas. The emission in this region is thus most likely dominated by the two-photon continuous emission of H (Spitzer & Greenstein 1951).
## 4. CIT6's Motion through the ISM, Mass-Loss Duration and Circumstellar Mass
We estimate the star's velocity V ∗ through the surrounding ISM using the relationship between l 1 , the distance of the termination shock from the star along the astropause's symmetry axis (i.e., the termination-shock standoff distance), and V ∗ (km s -1 ) = 10 V ∗ , 6 (Eqn. 1 of van Buren & McCray 1988 ): 2
$$l _ { 1 } ( \text{cm} ) = 1. 7 4 \times 1 0 ^ { 1 9 } \, ( \dot { M } _ { *, - 6 } V _ { w, 8 } ) ^ { 1 / 2 } \, \left ( \bar { \mu } _ { H } \, n _ { \text{ISM} } \right ) ^ { - 1 / 2 } V _ { *, 6 } ^ { - 1 }$$
where ˙ M ∗ -, 6 is the stellar mass-loss rate in units of 10 -6 M /circledot yr -1 , V w, 8 is the wind velocity
2 there is a missing minus sign in the exponent of ¯ µ H in their equation, which we have corrected below
in units of 10 3 km s -1 , ¯ µ H is the dimensionless mean molecular mass per H atom, and n ISM is the ISM number density in cm -3 .
Given the strong asymmetry between the northern and southern hemispheres, we first make the simplifying assumption that the astropause's symmetry axis lies in the sky-plane, i.e., the inclination angle, φ = 90 . ◦ We find l 1 = R D 1 = 2 2 . × 10 18 cm, using the value of R 1 = 373 ′′ derived earlier, and the distance D=400 pc. Substituting this value of l 1 in Eqn. 1, with ˙ M ∗ -, 6 = 3 2, . V w, 8 = 0 018, and ¯ . µ H = 1 33 (for an 89/11 mixture of H/He), we . find V ∗ = 39 ( n ISM / 0 17) . -1 2 / km s -1 . Our choice of n ISM is discussed in § 4.1. The value of V ∗ (i) does not depend on the (uncertain) distance, D , to CIT6, since both l 1 and ˙ M 1 2 / ∗ -, 6 scale linearly with D , and (ii) depends only weakly on the uncertain value of the ISM density at CIT6's location. The inclination angle φ may be significantly smaller than 90 ◦ - we discuss this in § 4.3.
## 4.1. Density of the ISM around CIT 6
We estimate the ISM number density near CIT6 based on the star's location in the Galaxy as follows. First we determine the the density of neutral hydrogen, H i around CIT6. We approximate the H i disk scale height h z ( R ) using the relation (Kalberla & Kerp 2009)
$$h _ { z } ( R ) = h _ { 0 } e ^ { ( R - R _ { \odot } ) / R _ { 0 } }, \text{ for } 5 < R < 3 5 \, \text{kpc},$$
where R is the Galactocentric radius, and R /circledot is the distance of the Sun from the Galactic center, and h 0 = 0 15 kpc, . R 0 = 9 8 kpc. . Taking R /circledot = 8 33 kpc (Gillessen et al 2009), and . using CIT6's galactic coordinates ( l = 197 ◦ . 7147, b = +55 ◦ . 9642, its distance from the Sun of 0 4 kpc, we find . R = 8 44 kpc, which implies . h z (8 44) = 0 152 kpc. . . The midplane density of H i (i.e., at z = 0) is found from the relation (Kalberla & Kerp 2009)
$$n _ { \text{H1} } \sim n _ { 0 } e ^ { - ( R - R _ { \odot } ) / R _ { n } } \text{ for } 7 < R < 3 5 \, \text{kpc},$$
where n 0 = 0 9 cm . -3 is the density at R = 0 and R n = 3 15 kpc. . Combining the midplane H density at CIT6's Galactocentric radius, i n H i ∼ 0 87, . with the fractional H i density of 0.038 (given by e -ln (2)( z/h z ) 2 ) at CIT6's height above the galactic plane ( z = 0 33 kpc), .
we find a relatively low value for the H i density, n H i ∼ 0 033. . Thus it is likely that the ISM surrounding CIT6 is ionized and that CIT 6 is embedded in the Warm Ionized Medium (WIM), specifically its thick-disk that extends more than a kpc above and below the Galactic plane. The volume-filling factor of the WIM in the midplane in this region is f = 0 04 . ± 0 01, . which grows exponentially with z over the 0 -1 4 kpc range as . e | z /H | f , where the scale height H f ∼ 0 7 kpc. (Gaensler et al. . 2008). In the neighborhood of CIT 6, f ( z = 0 33 kpc) = 0 065. . . Because the volume-filling factor is small, n ISM is likely less than n typ , where n typ is the typical internal electron density for clouds in the thick-disk component of the WIM. In the midplane, n typ = 0 34 . ± 0 06 cm . -3 and decays with z as e -| | z /H N , where the scale height H N = 0 5 kpc . (Gaensler et al. 2008).
In the neighborhood of CIT 6, n typ ( z = 0 33 kpc) = 0 17 cm . . -3 . Assuming that CIT6 is embedded in a WIM cloud, n ISM = 0 17 cm . -3 , which is the value that we use for our estimate of V ∗ . Our value of n ISM is larger than an estimate by Cetal12, who find n ISM ≈ 0 04 cm . -3 , assuming a representative global average of the density given by n H ( z ) = 2 e ( -| | z / 100 pc ) , although acknowledging that the structure of the ISM entails fluctuations in the density (and filling factor) on all spatial scales.
## 4.2. Duration of Mass Loss and Circumstellar Mass
The FUV emission traces the AGB stellar wind in CIT 6 to a much larger distance from the star than that derived from previous measurements of the CO J=1-0 and 2-1 emission (Neri et al. 1998). These authors take the CSE outer radius to be that at which CO is photodissociated by the interstellar radiation field. Since the photodissociation radius depends on ˙ M 0 58 . and Neri et al. (1998) use ˙ M =7 2 . × 10 -6 M /circledot yr -1 , we scale their value of this radius for our mass-loss rate ˙ M =3 2 . × 10 -6 M /circledot yr -1 to derive a photodissociation radius of 1 7 . × 10 17 cm. This estimate of the CSE outer radius implies a mass-loss duration of 3500 yr.
We use the astropause size to substantially revise (upwards) the above estimate of the duration, P , of heavy mass-loss in CIT 6. We take the radius of the termination shock in the direction orthogonal to the symmetry axis (432 ′′ or 2 6 . × 10 18 cm at D = 400pc) as a measure of the outer radius to which the unshocked wind has expanded, since the radial
extent in that direction is independent of the inclination angle.
We estimate P by deriving expansion time-scales ( P , P u s ) for the unshocked and shocked wind regions separately; P u = 45 240yr from the ratio of the termination shock radius to , V w , and P s = 47 950 yr from the ratio of the astrosheath width (38 ) to an average velocity , ′′ for this region, ¯ V s = 1 50 km s . -1 . We take ¯ V s = V / s 2, where V s = V w ( γ -1) / γ ( + 1) = V / w 6 = 3 00 km s . -1 , is the velocity in the astrosheath just beyond the termination shock, with γ = 7 5 for diatomic gas, and assuming the latter to be adiabatic. / The actual value of V s should be less than the adiabatic value, since the astrosheath appears to have cooled to some degree - the astrosheath's width of 38 , ′′ or 2 3 . × 10 17 cm, derived from fitting the radial intensity cut in a direction orthogonal to the symmetry axis, is smaller than the adiabatic value, ≈ 0 47 . l 1 = 1 0 . × 10 18 cm (Eqn. 2, Van Buren & McCray 1988) by factor of about 4. Furthermore, once a complete balance has been established between the ram pressure of the stellar wind and that of the ISM, the leading edge of the astropause (i.e., the forward shock structure) remains a fixed distance ahead of the moving star (Weaver et al. 1977). Hence, if the shock interaction of CIT 6's wind with the ISM has reached equilibrium, P = P u + P s = 93 190 yr is a lower limit, in which case CIT 6 has been undergoing mass-loss , for at least 93,000 years, and the total CSE mass is > 0 29 . M /circledot .
Model fits to IRAS far-infrared scan data have been used to derive dust shell sizes for a large sample of AGB stars, including CIT 6 by Young et al. (1995: Yetal95) - for the latter, they estimated an outer radius of 7 . ′ However, this result must be regarded with caution as (a) inspection of the data and model fits by Yetal95 for CIT 6 (see their Fig 7, panels D(i) and D(ii)), as well as many other stars in their sample, shows the presence of large negative and positive fluctuations in the derived intensity profile of the extended dust-shell emission, (b) the Herschel PACS imaging data in the Cetal12 study did not reveal an extended dust shell around CIT6, and (c) a comparison of the outer radii of the dust shells for a sample of AGB stars in Yetal95 with those reported to have bow-shock (Class I), eye (Class II), or ring (Class III) structures in Cetal12, shows that the IRAS-based value is significantly larger (i.e., by factors as large as ∼ 2 -5).
It is possible that the shells inferred from the IRAS data represent circumstellar structures that lie beyond the dust structures found with Herschel, but were not detected by the latter because of inadequate sensitivity. For example, in the case of Y CVn, an extended dust
shell is clearly seen in ISO maps (Izumiura et al. 1996) with about the same size as inferred by Yetal95, but only very faintly in the Cetal12 study. However, for CRL2688, where Speck et al. (2000) reported the presence of two dust shells from ISO scan data, Spitzer imaging by Do et al. (2007) showed no such shells at a level well below the intensities expected from the ISO results. We believe that a detailed re-investigation is needed to resolve such discrepancies in results related to the presence of extended dust shells as derived from IRAS and ISO, and from the more recent Spitzer and Herschel missions (utilizing modern large-format detector arrays).
If, however, the much larger circumstellar shells inferred by Yetal95 are real, then their presence poses a problem for each such object common to the Cetal12 study in which the (smaller) dust shell represents a bow-shock structure due to the wind-ISM interaction resulting from the object's motion through the ISM (based on its morphology, i.e., Class I and possibly Class II). In order to resolve this problem, one would require the ISM to be streaming through 'holes' in the outer IRAS-detected shell in order to interact with the inner circumstellar shell in these objects - numerical simulations would be needed to determine if this is a plausible scenario.
## 4.3. Forward Shock Shape and Inclination
Has CIT6's forward shock structure reached a steady state? The shape of this structure is expected to change with time, being initially circular and becoming increasingly parabolic until it reaches its steady state morphology (Weaver et al. 1977), given by the analytic solution of Wilkin (1996). We define η (0 / 90) as the ratio between the radius of the termination shock along the shock's symmetry axis to that along an orthogonal direction. Employing Eqn. 9 in Wilkin (1996), the analytic value is η (0 / 90) = 1 / 3 1 2 / = 0 577, and is reached . after ∼ 40 000 yr in simulations of astrospheres around AGB stars that are applicable to , CIT6 (see Fig. 11 of Mohamed et al. 2012 [MML12]). However, from CIT 6's observed shock structure, we find η (0 / 90) = 373 / 432 = 0 86, assuming that the inclination angle is, . φ = 90 . ◦
The discrepancy between the observed and expected steady-state value of η (0 / 90) in CIT6 is not unprecedented - a similar discrepancy exists for the forward shock structures seen in IRC+10216 and the red supergiant Betelgeuse, where this parameter is ∼ 0 8 (SC10, .
Decin et al. 2012).
In the MML12 simulations, η = 0 86 corresponds to an age of only . ∼ 5 000 years. , We found earlier that CIT 6 has been undergoing mass loss for a period at least as long as P u = 45 240yr, and possibly much longer. , If it has been interacting with the ISM during most of this period it is likely to have reached a steady state, in which case the discrepancy between the observed and expected values of η (0 / 90) would imply that the star's direction of motion of the star is not in the plane of the sky.
Mac Low et al.'s (1991: MLetal91) paraboloidal shock models for various inclination angles (their Fig. 5) show that as φ becomes smaller , the ratio of the radial distance between the star and the apex of the projected emission paraboloid, to the (unprojected) standoff distance, becomes larger . A visual comparison of CIT 6's FUV emission morphology (Fig. 1) with the surface brightness contours in Fig. 5 of MLetal91, indicate that φ is small, ∼ 30 , ◦ and our measured value of R 1 is larger than l 1 by a significant factor. We estimate that this factor may be as large as ∼ 1 5 . -2, by comparing the length of the unprojected standoff distance vector, to its apparent value as estimated from the distance between the star and the apex defined by the emission contours for φ = 30 ◦ and 90 ◦ in Fig. 5 of MLetal91. Thus the value of V ∗ may be as high as ∼ 70 km s -1 .
The value of φ depends on the orientation of the relative motion vector between CIT 6 and the WIM cloud with which it is interacting. Thus, if both CIT 6 and the WIM cloud have similar radial velocities, then the relative motion vector would lie in the sky plane, i.e., φ ∼ 90 . ◦ CIT6's radial velocity is V lsr = -1 5 km s . -1 (Olofsson et al. 1993). For the WIM, Haffner et al. (2003) conclude, using data from the Wisconsin H α (WHAM) survey that the H α emitting gas at high galactic latitudes is clearly biased towards negative V lsr values; inspection of the longitude-velocity plots for b=50 ◦ and 60 ◦ in Fig. 9 of their paper shows relatively strong emission extending to V lsr values of -50 to -75 km s -1 at CIT'6's longitude. Thus, it is quite possible that the relative motion vector between CIT 6 and the local WIM has a significant radial component, implying that φ is significantly less than 90 ◦ .
Alternatively, the density of the ISM around CIT 6 is much lower than few × 0 1 cm . -3 , or its mass-loss rate was significantly higher in the distant past (since the time required to reach a steady state varies as ( n ISM / ˙ M ) 1 2 / , Weaver et al. 1977). It is also possible
that CIT6 has only recently entered the higher-density region with which it is interacting - this alternative finds some support in the relatively smooth shock structure observed in CIT6, as hydrodynamical instabilities take time to develop. For example, MML12 attribute the smoothness of the shock structure of Betelgeuse to the bow-shock interaction being too young for instabilities to have developed (they constrain the age to < 30 000 yr). , A note of caution here is in order - the production of instabilities in different simulation studies varies significantly, both in terms of the time-scales over which they are produced and in their size and structure, as noted by Decin et al (2012) while comparing their simulations of the bow-shock around the red supergiant Betelguese with those of MML12 and Cetal12.
Other explanations for a smooth shock structure are also possible, including the suppression of large scale instabilities due to a warm ISM (Decin et al 2012) (which may be appropriate for CIT 6), and the presence of a magnetic field in the direction of motion (van Marle et al. 2014).
## 4.4. Comparison with IRC+10216
It is instructive to compare the astrosphere of CIT 6 with that of IRC+10216, the only other carbon-rich star that has been found to show such a structure in GALEX images (SC10). SC10 noted that IRC+10216's forward shock structure is relatively smooth and does not show the large-scale instabilities expected in such interactions (Blondin & Koerwer 1998), and CIT6 is similar in this respect. While the absence of large-scale instabilities for CIT6's shock may be explained as a result of it being surrounded by a warm ISM, this explanation is not applicable for IRC+10216, which (as we show below) is most likely immersed in neutral (and cooler) gas.
Given IRC+10216's Galactocentric radius R = 8 36 kpc (derived from its galactic co-. ordinates l = 221 ◦ . 4466, b = +45 ◦ . 0604, and D=120pc), we find the H i scale height to be h z = 0 15 kpc, giving a fractional density of 0 80 at its height above the Galactic plane of . . z = 85pc. With a mid-plane density of 0.89 cm -3 at R = 8 36 kpc, the density of H . i near IRC+10216 is then ∼ 0 72 cm . -3 , implying that the ISM near IRC+10216 is probably largely neutral (supporting SC10's choice of n ISM = 1cm -3 ), thus significantly cooler than the WIM.
The morphologies of the astrotail regions of CIT6 and IRC+10216 differ markedly, with CIT6 showing a relatively smooth arc structure, whereas IRC+10216 shows prominent vortices. SC10 attribute these as resulting from the vortex shedding seen in numerical simulations of AGB wind-ISM interactions that occurs at very long interaction time-scales ( ∼ 500,000 yr) (Wareing et al. 2007). Thus it appears that the period of wind-ISM interaction in IRC+10216 is much longer than that of CIT 6.
## 5. Comparison with 3D Numerical Simulations
We compare our observations of CIT 6 with the results of 3D simulations of the bowshock around Betelgeuse ( α Ori) by MML12, since there is reasonable similarity between the physical parameters relevant to the shock structure employed in these models and those applicable to CIT 6. MML12's model D has ˙ M ∗ = 3 1 . × 10 -6 M /circledot yr -1 , V w = 17km s -1 , V ∗ = 72 5 km s . -1 and n ISM = 0 3. The density structure for model D (rightmost panel of Fig. 7 in . MML12) appears quite similar to our image of CIT 6 - specifically, the density plot shows the presence of a high-density semi-circular structure in the direction of motion (leftwards in this figure), and a roughly semi-circular 'edge' structure in the opposite direction. The latter is at a larger radial distance from the central star than the former, similar to what we observe in CIT6, and shows a density contrast of a factor 10 or more relative to the immediate environment beyond it (seen in red/brown hues). We note that the parabolicshaped structure seen in yellow in MML12's Figure 7 represents high-temperature ( > 10 4 K) ionized gas in the bow-shock, which at best is only seen very faintly, and only around the forward shock in our observations - hence its absence in the tail region of the observed astrosphere of CIT 6 is most likely due to inadequate sensitivity. The southern ring, like the northern one, is only seen in FUV emission and not in the NUV, suggesting that the emission mechanism for the former is the same as the latter, namely, collisional excitation of H 2 by hot electrons.
MML12 find that the gas in the 'bow shock arc (where the gas is strongly decelerated)', i.e., the astrosheath in our description, behaves nearly isothermally, consistent with our finding for CIT 6 that the average astrosheath width is significantly smaller than the expected adiabatic value ( § 4.2). But as noted earlier, in the simulation the stellar wind-ISM interaction
has been ongoing for 32,000 yr and is close to steady state with η (0 / 90) ∼ 0 6, whereas the . observed value of η (0 / 90) is significantly higher.
## 6. Alternative Models for the FUV Rings around CIT 6
We now consider alternatives to the astrosphere model for the origin of the FUV emission rings around CIT6. Given the presence of the partial arcs seen by Claussen et al. (2011) and Schmidt et al. (2002) within ∼ 8 ′′ of the central star, believed to be created by a spiral shock due to the presence of a central binary, one possibility is that the much more distant FUV emission rings also have the same origin. In this model, the radial pitch of the spiral, i.e., the separation between successive windings is equal to P V exp , where P is the binary period and V exp is the expansion velocity of the CSE (e.g., Mauron & Huggins 2006). Hence, assuming V exp does not change substantially over the timescales corresponding to the rings, the current separation of the windings ( ∼ 2 6), . ′′ provides an upper limit to the separation in the past (because the period P of the binary increases as the central star loses mass via its wind). Clearly, given the much larger observed difference in the average radius of the northern and southern rings ( ∼ 200 ), these cannot represent successive windings, even if ′′ we allowed for a V exp that was larger in the past by a factor 10 over its current value.
Another possible model is one in which the northern and southern rings represent two discrete mass-loss episodes in CIT 6's distant past. However, such a model would then require that each of these episodic ejections was confined to only a hemispherical region - with either both ejections occcuring at different times with roughly the same velocity, or at the same time with different velocities (or some combination of these). Such a mass-ejection history would be unprecedented for an AGB star, and would not fit into the standard model of roughly spherical AGB mass-loss (i.e., driven by radiation pressure on dust grains). It also does not provide an explanation for the systematic increase in the radius of the northern ring around its symmetry axis. We note that although the so-called detached shell phenomenon observed in a few carbon-rich stars does produce large single shell structures (e.g., the Class III objects in the Cetal12 survey), these shells, believed to result from a relatively short episode of enhanced mass-loss rate (and possibly wind velocity as well) initiated by a Heshell flash (e.g., Olofsson et al. 1990, Steffen & Sch¨nberner 2000), are circular and complete o
and thus morphologically different from CIT 6's FUV emission rings.
## 7. Concluding Remarks
In summary, the GALEX images of CIT6 show an unusually detailed picture of the interaction of the wind from a carbon-rich AGB star with the ISM due to its motion in the latter. This interaction process has been studied using UV and far-infrared observations for other AGB stars (e.g., SC10, Cetal12 and references therein) and a red supergiant ( α Ori: Ueta et al. 2008, Decin et al. 2012, le Bertre et al. 2012), as well as HI 21 cm line observations (Matthews et al. 2013, and references therein). The far-infrared imaging observations with Herschel by Cetal12 have discovered the largest number of astrospheres by far, clearly tracing the detailed geometry of the shock front in the wind-ISM interaction for a large sample of AGB stars. The HI observations reveal, for a smaller sample, that the CSEs have been significantly influenced by the wind-ISM interaction, but since the angular resolution of these observations is relatively low ( ∼ 1 ), ′ it is difficult to spatially separate the detailed shape and structure of the astropause from the HI emission of the CSE as a whole. However, kinematical information extracted from the observed HI line profiles provides support for deceleration of the freely expanding CSE by the ISM.
Including CIT 6, there are now three examples of AGB stars (IRC+10216, Mira, and CIT6) and one RSG ( α Ori) that show astrospheres in FUV emission. An exhaustive search of the GALEX image archive in order to determine if there are additional such examples would be useful, but is outside the scope of this paper. However, we speculate that such a search would not be very succesful given the very long exposure times that were needed for the GALEX imaging to reveal the relatively faint FUV emission from the wind-ISM interaction in these objects (e.g., 30,824 s for CIT 6; 8783 s for IRC+10216; 48,915 s for α Ori.)
Our study provides valuable new data for understanding the CSE-ISM interaction, and shows that there are specific features in the resulting shock structures that are not adequately reproduced by current numerical simulations. These features include the relatively high value of the parameter η (0 / 90) compared to the steady-state value, and the relative smoothness of the shock structure implying the general lack of large-scale instabilities. Given that such discrepancies are seen in other objects as well, new modelling that is focussed on
addressing these will help improve our understanding of this interaction, and our ability to make inferences about the unprecedented long history of AGB mass-loss revealed by these interactions. For example, one avenue that merits detailed investigation is when the windISM interaction involves an encounter of the star's CSE with a relatively high density cloud in an inhomogeneous ISM, so that the timescale for the interaction is substantially smaller that the mass-ejection timescale.
We would like to thank an anonymous referee for detailed comments that have helped us improve the discussion in this paper. RS's contribution to the research described in this publication was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with NASA. RS thanks NASA for financial support via a GALEX GO and ADAP award. GMC thanks JPL for a NASA Student Independent Research Internship (SIRI).
## REFERENCES
Alksnis, A. 1995, Baltic Astronomy, 4, 79 Blondin, J. M., & Koerwer, J. F. 1998, New Astronomy, 3, 571 Claussen, M. J., Sjouwerman, L. O., Rupen, M. P., et al. 2011, ApJ, 739, L5 Cox, N. L. J., Kerschbaum, F., van Marle, A.-J., et al. 2012, A&A, 537, A35 Decin, L., Cox, N. L. J., Royer, P., et al. 2012, A&A, 548, A113 Do, T., Morris, M., Sahai, R., & Stapelfeldt, K. 2007, AJ, 134, 1419 Gaensler, B. M., Madsen, G. J., Chatterjee, S., & Mao, S. A. 2008, PASA, 25, 184 Gillessen, S., Eisenhauer, F., Trippe, S., et al. 2009, ApJ, 692, 1075 Haffner, L. M., Reynolds, R. J., Tufte, S. L., et al. 2003, ApJS, 149, 405 Izumiura, H., Hashimoto, O., Kawara, K., Yamamura, I., & Waters, L. B. F. M. 1996, A&A,
315, L221
Kalberla, P. M. W., & Kerp, J. 2009, ARA&A, 47, 27 Kim, H., Hsieh, I.-T., Liu, S.-Y., & Taam, R. E. 2013, ApJ, 776, 86 Le Bertre, T., Matthews, L. D., G´rard, E., & Libert, Y. 2012, MNRAS, 422, 3433 e Mac Low, M.-M., van Buren, D., Wood, D. O. S., & Churchwell, E. 1991, ApJ, 369, 395 Martin, D. C., et al. 2007, Nature, 448, 780 Matthews, L. D., Le Bertre, T., G´rard, E., & Johnson, M. C. 2013, AJ, 145, 97 e Mauron, N., & Huggins, P. J. 2006, A&A, 452, 257 Mohamed, S., Mackey, J., & Langer, N. 2012, A&A, 541, A1 Morrissey, P., et al. 2005, ApJ, 619, L7 Neri, R., Kahane, C., Lucas, R., Bujarrabal, V., & Loup, C. 1998, A&AS, 130, 1 Olofsson, H., Carlstrom, U., Eriksson, K., Gustafsson, B., & Willson, L. A. 1990, A&A, 230, L13 Olofsson, H., Eriksson, K., Gustafsson, B., & Carlstrom, U. 1993, ApJS, 87, 267 Sahai, R., & Chronopoulos, C. K. 2010, ApJ, 711, L53 Schmidt, G. D., Hines, D. C., & Swift, S. 2002, ApJ, 576, 429 Speck, A. K., Meixner, M., & Knapp, G. R. 2000, ApJ, 545, L145 Spitzer, L., Jr., & Greenstein, J. L. 1951, ApJ, 114, 407 Steffen, M., & Sch¨nberner, D. 2000, A&A, 357, 180 o Ueta, T. 2008, ApJ, 687, L33 Ueta, T., et al. 2008, PASJ, 60, 407 van Buren, D., & McCray, R. 1988, ApJ, 329, L93 van Marle, A. J., Decin, L., & Meliani, Z. 2014, A&A, 561, A152
```
<_SQL_>
```
Young, K., Phillips, T. G., & Knapp, G. R. 1993, ApJS, 86, 517
This preprint was prepared with the AAS LT E X macros v5.2. A
0.00550
Fig. 1.- (a) FUV GALEX image of CIT6 (the circular field-of-view (FOV) has a diameter of 72.8' × 72.8'); the image was boxcar-smoothed using a 3 × 3 pixel box, and displayed using a linear stretch (in false color). Units in the colorbar are cps/pixel, implying a flux of 107 µ Jy per 1 5 . ′′ × 1 5 . ′′ pixel (b) NUV GALEX image of CIT6 (same FOV as in a), displayed using a linear stretch (in false color). Units in the colorbar are cps/pixel, implying a flux of 35 3 . µ Jy per 1 5 . ′′ × 1 5 . ′′ pixel
Fig. 2.- As in Fig. 1(a), but showing a magnified view of the FUV emission structures around CIT6. The panel size is 24.75' × 24.75'. The location of the central star is marked with a *.
Fig. 3.- Cuts (solid blue curves) of the radial intensity averaged over each of seven 23 ◦ angular wedges spanning the northen ring seen in the FUV emission nebulosity around CIT6. The intensity cuts have been shifted vertically by constant offsets, indicated by black dashed lines, for clarity. The cut PA s are noted above the cuts on the image. Intensities are displayed in units of 2 00 . × 10 -4 cps/pixel (0 0214 . µ Jy/pixel).
Fig. 4.- As in Fig. 3, but for angular wedges spanning the southern ring seen in the FUV emission nebulosity around CIT 6.
Fig. 5.- (a) Cut of radial intensity (blue) averaged over a 46 ◦ degree wedge centered at 2 ◦ in the FUV. A model fit to the intensity is shown in red, along with the approximate locations of the termination shock ( R 1 ) and the astropause ( R c ). FUV intensity units are as in Fig. 3. (b) Detail of the FUV emission, rotated from its original orientation (as in Fig 1). Box size is 10 25 . ′ × 17 5 ; smoothing, stretch, and units as in Fig 1. . ′
|
10.1088/0004-6256/148/4/74
|
[
"Raghvendra Sahai",
"Galen P. Mack-Crane"
] |
2014-08-05T17:50:02+00:00
|
2014-08-05T17:50:02+00:00
|
[
"astro-ph.SR"
] |
The astrosphere of the Asymptotic Giant Branch star CIT 6
|
We have discovered two extended half-ring structures in a far-ultraviolet
image taken with the GALEX satellite of the well-known mass-losing carbon star
CIT 6 (RW LMi). The northern (southern) ring is brighter (fainter) with a
diameter of about 15 arcmin (18 arcmin). These structures most likely represent
the astrosphere resulting from the shock interaction of CIT 6's molecular wind
with the Warm Interstellar Medium, as it moves through the latter. These data
provide a direct estimate of the size of CIT 6's circumstellar envelope that is
a factor ~20 larger than previous estimates based on CO millimeter-wave line
data. We find that CIT 6 has been undergoing heavy mass-loss for at least
93,000 yr and the total envelope mass is 0.29 Msun or larger, assuming a
constant mass-loss rate of 3.2 x 10^(-6) Msun/yr. Assuming that the shock front
has reached a steady-state and CIT 6's motion relative to the ISM is in the
sky-plane, we measure the termination-shock standoff distance directly from the
image and find that CIT 6 is moving at a speed of about 39 (0.17
cm^(-3)/n_ISM)^(1/2) km/s through the interstellar medium around it. However,
comparisons with published numerical simulations and analytical modelling shows
that CIT 6's forward shock (the northern ring) departs from the parabolic shape
expected in steady-state. We discuss several possible explanations for this
departure.
|
1408.1051v4
|
## SUBCRITICAL CONTACT SURGERIES AND THE TOPOLOGY OF SYMPLECTIC FILLINGS
PAOLO GHIGGINI, KLAUS NIEDERKR ¨ GER, AND CHRIS WENDL U
Abstract. By a result of Eliashberg, every symplectic filling of a threedimensional contact connected sum is obtained by performing a boundary connected sum on another symplectic filling. We prove a partial generalization of this result for subcritical contact surgeries in higher dimensions: given any contact manifold that arises from another contact manifold by subcritical surgery, its belt sphere is null-bordant in the oriented bordism group Ω SO ∗ ( W ) of any symplectically aspherical filling W , and in dimension five, it will even be nullhomotopic. More generally, if the filling is not aspherical but is semipositive, then the belt sphere will be trivial in H ∗ ( W ). Using the same methods, we show that the contact connected sum decomposition for tight contact structures in dimension three does not extend to higher dimensions: in particular, we exhibit connected sums of manifolds of dimension at least five with Stein fillable contact structures that do not arise as contact connected sums. The proofs are based on holomorphic disk-filling techniques, with families of Legendrian open books (so-called ' Lob s') as boundary conditions.
## 1. Introduction
1.1. The main result and corollaries. The idea of constructing contact manifolds as boundaries of symplectic 2 n -manifolds by attaching handles of index at most n goes back to Eliashberg [Eli90b] and Weinstein [Wei91]. In this context, a special role is played by subcritical handles, i.e. handles with index strictly less than n . One well-known result on this topic concerns subcritical Stein fillings, which are known to be flexible in the sense that their symplectic geometry is determined by homotopy theory, see [CE12]. There are also known restrictions on the topological types of subcritical fillings, e.g. by results of M.-L. Yau [Yau04] and Oancea-Viterbo [OV12, Prop. 5.7], the homology of a subcritical filling with vanishing first Chern class is uniquely determined by its contact boundary; this result can be viewed as a partial generalization of the Eliashberg-Floer-McDuff theorem [McD91] classifying symplectically aspherical fillings of standard contact spheres up to diffeomorphism. In dimension three, there is a much stronger result due to Eliashberg [Eli90a; CE12]: in this case every subcritical surgery is a connected sum, and the result states that if ( M ,ξ ′ ′ ) is a closed contact 3-manifold obtained from another (possibly disconnected) contact manifold ( M,ξ ) by a connected sum, then every symplectic filling of ( M ,ξ ′ ′ ) is obtained by attaching a Weinstein 1-handle to a symplectic filling of ( M,ξ ). This implies that symplectic fillings of subcritically fillable contact 3-manifolds are unique up to symplectic deformation equivalence and blowup-in particular, their Stein fillings are unique up to symplectomorphism.
The present paper was motivated by the goal of generalizing Eliashberg's connected sum result to higher dimensions. The natural question in this setting is the following:
Question 1.1. Given a closed contact manifold ( M ,ξ ′ ′ ) that is obtained from another contact manifold ( M,ξ ) by subcritical contact surgery, is every (symplectically aspherical) filling of ( M ,ξ ′ ′ ) obtained by attaching a subcritical Weinstein handle to a symplectic filling of ( M,ξ ) ?
Classifying fillings up to symplectomorphism as suggested in this question would be far too ambitious in higher dimensions, e.g. the strongest result known so far, the Eliashberg-Floer-McDuff theorem, is essentially a classification of fillings up to homotopy type (the h -cobordism theorem then improves it to a classification up to diffeomorphism). Our objective in this paper will therefore be to understand the main homotopy-theoretic obstruction to an affirmative answer to Question 1.1.
To state the main result, let us first recall some basic notions. Given an oriented (2 n -1)-dimensional manifold M , a (positive, co-oriented) contact structure on M is a hyperplane distribution of the form ξ = ker α , where the contact form α is a smooth 1-form satisfying α ∧ ( dα ) n -1 > 0, and the co-orientation of ξ is determined by α > 0. In this paper, contact structures will always be assumed to be both positive (with respect to a given orientation on M ) and co-oriented, and all contact forms will be assumed compatible with the co-orientation. A compact symplectic 2 n -manifold ( W,ω ) with oriented boundary M = ∂W carrying a contact structure ξ is called a strong symplectic filling of ( M,ξ ) if ξ admits a contact form λ that extends to a primitive of ω on a neighborhood of ∂W . It is equivalent to say that the boundary is symplectically convex , as the vector field ω -dual to λ is then a Liouville vector field pointing transversely outward at ∂W . More generally, we say that ( W,ω ) is a weak symplectic filling of ( M,ξ ) if ξ is the bundle of complex tangencies for some ω -tame almost complex structure near ∂W that makes the boundary pseudoconvex (see [MNW13]).
Recall that if ( M,ξ ) contains a ( k -1)-dimensional isotropic sphere S k -1 at with trivial normal bundle, then one can perform a contact surgery of index k on ( M,ξ ) by attaching to ( -ε, 0] × M a handle of the form D k × D 2 n -k along a neighborhood of S k -1 at . The new contact manifold ( M ,ξ ′ ′ ) then contains the (2 n - -k 1)dimensional coisotropic sphere S 2 n - -k 1 belt = { 0 } × ∂ D 2 n -k , which we call the belt sphere of the surgery. This surgery operation was first introduced by Weinstein [Wei91], and we will give a more precise description of it in Section 3. A Weinstein handle yields a symplectic cobordism that can be attached to any weak filling ( W,ω ) of ( M,ξ ) for which ω is exact along S k -1 at ; the result is a weak filling of ( M ,ξ ′ ′ ) in which the belt sphere is necessarily contractible (Figure 1).
Our main result is the following.
Theorem 1.2. Suppose ( M ,ξ ′ ′ ) is a closed contact manifold of dimension 2 n -1 ≥ 3 that has been obtained from a manifold ( M,ξ ) by a contact surgery of index k ≤ n -1 , with belt sphere S 2 n - -k 1 belt ⊂ M ′ . Assume ( W ,ω ′ ′ ) is a weak symplectic filling of ( M ,ξ ′ ′ ) .
- (a) If ( W ,ω ′ ′ ) is semipositive, then the belt sphere represents the trivial homology class in H 2 n - -k 1 ( W ′ ; Z ) .
- (b) If ( W ,ω ′ ′ ) is symplectically aspherical, then the belt sphere represents the trivial element in the oriented bordism group Ω SO 2 n - -k 1 ( W ′ ) . If additionally either (1) M ′ is 5 -dimensional, or (2) M ′ is 7 -dimensional and k = 3 , then S 2 n - -k 1 belt is contractible in W ′ , that is, it represents the trivial class in π 2 n - -k 1 ( W ′ ) .
Figure 1. The belt sphere bounds an embedded disk inside the handle.
We now state two related results that follow via the same techniques.
Recall that in dimension three, convex surface theory gives rise to a contact prime decomposition theorem, implying e.g. that every tight contact structure on a closed 3-manifold of the form M 0 # M 1 arises as a contact connected sum of tight contact structures on M 0 and M 1 (see [Col97], or [Gei08, § 4.12] for more details). Some evidence against a higher-dimensional generalization of this result appeared in the recent work of Bowden, Crowley and Stipsicz [BCS15], providing also a negative answer to a topological version of Question 1.1: namely, there exist pairs of closed oriented manifolds M ,M 0 1 such that M 0 # M 1 admits a Stein fillable contact structure but M 0 and M 1 do not. This did not imply an actual answer to Question 1.1, however, as it was unclear whether the contact structures on M 0 # M 1 in the examples of [BCS15] could actually be contact connected sums, i.e. whether they arise from contact structures ξ 0 on M 0 and ξ 1 on M 1 by performing index 1 contact surgery. The following result gives a negative answer to the latter question, and shows that there is no hope of extending the contact prime decomposition theorem to higher dimensions. The theorem applies in particular whenever M is an almost contact S n -1 -bundle over S n that is not a homotopy sphere, so for instance M could be S n -1 × S n or-as in [BCS15]-the unit cotangent bundle of a sphere.
Theorem 1.3. Suppose M is a closed oriented manifold of dimension 2 n -1 ≥ 5 that is not a homotopy sphere but admits an almost contact structure Ξ and a Morse function with unique local maxima and minima and otherwise critical points of index n -1 and n only. Then M #( -M ) admits a Stein fillable contact structure that is homotopic to the almost contact structure Ξ#Ξ but is not isotopic to ξ 1 # ξ 2 for any contact structures ξ 1 and ξ 2 on M and -M respectively. 1
Note that the contact structures in the above statement are necessarily tight in the sense of Borman-Eliashberg-Murphy [BEM15]; this follows from Stein fillability, using [Nie06] and the observation in [BEM15] that any overtwisted contact structure is also PS-overtwisted.
The holomorphic disk techniques developed in this article can also be used as in the work of Hofer [Hof93] to prove the Weinstein conjecture for a wide class of contact manifolds obtained by subcritical surgery. The following theorem, proved in Section 7, is related to the well-known result that every subcritically Stein fillable contact form admits a contractible Reeb orbit. (A similar result specifically for index 1 surgeries appeared recently in [GZ16].)
1 Given an oriented manifold M with almost contact structure Ξ, we denote by -M the same manifold with reversed orientation, and let Ξ denote the almost contact structure on -M obtained
by inverting the co-orientation of Ξ.
Theorem 1.4. Assume ( M ,ξ ′ ′ ) is the result of performing a contact surgery of index k ≤ n -1 on a closed contact manifold ( M,ξ ) of dimension 2 n -1 ≥ 3 , with belt sphere S 2 n - -k 1 belt ⊂ M ′ , and suppose at least one of the following conditions holds:
glyph[negationslash]
- (1) [ S 2 n - -k 1 belt ] = 0 in Ω SO 2 n - -k 1 ( M ′ ) ;
glyph[negationslash]
- (2) [ S 2 n - -k 1 belt ] = 0 in π 2 n - -k 1 ( M ′ ) and either dim M ′ = 5 or dim M ′ = 7 with k = 3 ;
- (3) dim M ′ = 5 and ( M ,ξ ′ ′ ) is a contact connected sum ( M ,ξ 0 0 ) # ( M ,ξ 1 1 ) with the following two properties:
- (a) Neither M 0 nor M 1 is homeomorphic to a sphere;
- (b) If M 0 and M 1 are both rational homology spheres, then either both are not simply connected or at least one of them has infinite fundamental group.
Then every contact form for ξ admits a contractible Reeb orbit.
Before discussing the proofs, some further remarks about the main theorem are in order.
Remark 1.5 . We do not know whether the dimensional restriction for the contractibility result in part (b) of Theorem 1.2 is essential, but given the wide range of known contact geometric phenomena that can happen only in sufficiently high dimensions, we consider it plausible that the contractibility statement could be false without some restriction of this type (thus implying a definitively negative answer to Question 1.1 in general). It is apparent in any case that our method will not work in all dimensions, as the improvement from 'null-bordant' to 'nullhomotopic' involves subtle topological difficulties that increase with the dimension; see the beginning of Section 5 for more discussion of this.
Remark 1.6 . It is clear that nothing like Theorem 1.2 can be true for critical surgeries in general, i.e. the case k = n . There are obvious counterexamples already in dimension three, as any Legendrian knot L ⊂ ( M ,ξ ′ ′ ) can be viewed as the belt sphere arising from a critical contact surgery-take ( M,ξ ) in this case to be the result of a Legendrian (+1)-surgery along L . It is certainly not true in general that arbitrary Legendrian knots are nullhomologous in every filling of ( M,ξ )!
Remark 1.7 . The semipositivity assumption in part (a) of Theorem 1.2 is there for technical reasons and could presumably be lifted using more advanced technology (e.g. polyfolds, see [HWZ11]). In contrast, symplectic asphericity in part (b) is a geometrically meaningful condition that, while not needed for Eliashberg's threedimensional version of this result, cannot generally be removed in higher dimensions; see Example 1.8 below. The answer to Question 1.1 thus becomes negative without this assumption.
Example 1.8 . The blowup of the total space of the rank 2 holomorphic vector bundle O -( 2) ⊕ O over C P 1 at the zero section can be viewed as a (not symplectically aspherical) weak filling of a subcritically Stein fillable contact manifold ( M ,ξ ′ ′ ) containing a belt sphere that is homotopically nontrivial in the filling. This is a special case of the following construction, which gives examples with subcritical handles of any even index k = 2 m ≥ 2 in any dimension 2 n ≥ 2 k +2 ≥ 6.
Choose integers m,glyph[lscript] ≥ 1 and set n = 2 m + , and suppose (Σ glyph[lscript] , σ ) is a 2 m -dimensional closed symplectic manifold. Then consider the 2 n -dimensional Weinstein manifold T ∗ Σ × C glyph[lscript] , i.e. the glyph[lscript] -fold stabilization of T ∗ Σ with its standard Weinstein structure, and denote its ideal contact boundary by ( M ,ξ ′ ′ ). Any Morse function on Σ gives rise to a Weinstein handle decomposition of T ∗ Σ × C glyph[lscript] , such that the function's maximum q ∈ Σ corresponds to an ( n -glyph[lscript] )-handle whose belt sphere S n + glyph[lscript] -1 belt is isotopic to the unit sphere in T ∗ q Σ × C glyph[lscript] . Let Σ ⊂ T ∗ Σ denote the zero section, so Σ ×{ } 0 is an isotropic submanifold in T ∗ Σ × C glyph[lscript] , and denote by π : T ∗ Σ × C glyph[lscript] → Σ ×{ } 0 the obvious projection. Then for any ε > 0 sufficiently small, adding επ σ ∗ to the natural exact symplectic form on T ∗ Σ × C glyph[lscript] gives a weak filling of ( M ,ξ ′ ′ ) with Σ ×{ } 0 as a symplectic submanifold. We can then blow up along this submanifold, as explained in [MS98, Section 7.1]. This produces a new weak filling ( W ,ω ′ ′ ) of ( M ,ξ ′ ′ ), in which the belt sphere S n + glyph[lscript] -1 belt ⊂ M ′ is nullhomologous but homotopically nontrivial: indeed, every fiber T ∗ q Σ × C glyph[lscript] has now been replaced by its blowup at the point (0 , 0), which can be viewed as the tautological line bundle over C P m glyph[lscript] + -1 , so the bundle projection sends S n + glyph[lscript] -1 belt to a generator of π 2( m glyph[lscript] + ) -1 ( C P m glyph[lscript] + -1 ) = ∼ Z .
The special case with Σ = S 2 and glyph[lscript] = 1 gives the construction described at the beginning of this example, because the total space of O -( 2) is a deformation of T ∗ S 2 .
The following represents another easy application of Theorem 1.2.
Example 1.9 . Suppose ( M ,ξ 1 1 ) is a contact 5-manifold obtained by a subcritical surgery of index 2 on a sphere ( S 5 , ξ ), where ξ is any contact structure. Then M 1 is diffeomorphic to either S 2 × S 3 or S 2 ˜ × S 3 , i.e. the trivial or nontrivial 3-sphere bundle over the 2-sphere. Indeed, closed loops in S 5 are automatically unknotted, and the possible framings of the surgery are classified by the elements of π 1 ( SO(3) ) ∼ = Z 2 . If ( W,ω ) is any symplectically aspherical weak filling of ( M ,ξ 1 1 ), then by Theorem 1.2, the fiber { p } × S 3 is a contractible 3-sphere in W .
Now take ( M ,ξ 2 2 ) to be the unit cotangent bundle of the 3-sphere or, more generally, any contact manifold supported by a contact open book with page T ∗ S 2 and monodromy isotopic to a 2 k -fold product of Dehn twists for some integer k ≥ 1. Then M 2 will be diffeomorphic to S 2 × S 3 , but ( M ,ξ 2 2 ) admits a Stein filling that contracts to a bouquet of 2 k -1 three-dimensional spheres (see e.g. [KN05]). We conclude that whenever ( M ,ξ 1 1 ) admits a symplectically aspherical weak filling, it is not contactomorphic to ( M ,ξ 2 2 ). This implies for instance that the contact structures induced on the ideal contact boundaries of T ∗ S 3 and T ∗ S 2 × C (cf. Remark 1.7) are not isomorphic.
There are presumably other ways to distinguish ξ 1 and ξ 2 in many cases, e.g. using Symplectic Homology, but the technique described above is much more topological.
1.2. Idea of the proof. Our proof of Theorem 1.2 is based on a higher-dimensional analogue of the disk-filling methods underlying Eliashberg's result in dimension three [Eli90a]. Such methods work whenever one can find a suitable submanifold to serve as a boundary condition for holomorphic disks, and the natural object to consider in this case is known as a Legendrian open book or ' Lob '. Let us recall the definition, which is due originally to the second author and Rechtman [NR11].
Definition 1.10. A Lob in a (2 n -1)-dimensional contact manifold ( M,ξ ) is a closed n -dimensional submanifold L ⊂ M equipped with an open book decomposition π : L \ B → S 1 whose binding B ⊂ L is an ( n -2)-dimensional isotropic submanifold of ( M,ξ ), and whose pages π -1 ( ∗ ) are each Legendrian submanifolds of ( M,ξ ).
glyph[negationslash]
The simplest interesting example of a Lob occurs at the center of the 'neck' in any 3-dimensional contact connected sum: here we find a 2-sphere S ⊂ M on which the characteristic foliation ξ ∩ TS traces an S 1 -family of longitudes connecting the north and south poles, so we can regard the longitudes as pages of an open book with the poles as binding. Such spheres were used as totally real boundary conditions for holomorphic disks in [BG83; Gro85; Eli90a], and similarly in Hofer's proof of the Weinstein conjecture [Hof93] for contact 3-manifolds ( M,ξ ) with π 2 ( M ) = 0.
In higher dimensions, a Lob L ⊂ ( M,ξ ) with binding B ⊂ L similarly defines a totally real submanifold { 0 } × ( L \ B ) in the symplectization R × M of ( M,ξ ), and thus serves as a natural boundary condition for pseudoholomorphic disks. Moreover,
Figure 2. A schematic picture of the Bishop family around the binding of a Lob L .
for a suitably 'standard' choice of almost complex structure near the binding, a Lob always gives rise to a canonical family of holomorphic disks near { 0 }× B whose boundaries foliate a neighborhood of B in L \ B (see Figure 2). This is the so-called Bishop family of holomorphic disks, and it has the useful property that no other holomorphic curve can enter the region occupied by the Bishop disks from outside. For a unified treatment of the essential analysis for Bishop disks with boundary on a Lob , see [Nie13].
As in the 3-dimensional case, we will see that the belt sphere of a surgery of index n -1 on a contact (2 n -1)-manifold is also naturally a Lob , so there is again a natural moduli space of holomorphic disks that fill the belt sphere, implying that it is nullhomologous. For surgeries of index k < n -1, the belt sphere has dimension 2 n -k -1 > n , and thus cannot be a Lob , but we will show that after a suitable deformation, the belt sphere can be viewed as a parametrized family of Lob s, giving rise to a well-behaved moduli space of disks with moving boundary condition. It should now be clear why this method cannot work for critical surgeries: the belt sphere in this case has dimension n -1, so it is too small to define a totally real boundary condition.
The construction of the family of Lob s foliating a general subcritical belt sphere is somewhat less than straightforward: as we will see in Section 3, the standard model for a contact form after surgery does not lend itself well to this construction, but a natural family of Lob s can be found after deforming to a different model of the belt sphere as piecewise smooth boundary of a poly-disk.
Let us now discuss the topological reasons why the family of Lob s foliating S 2 n - -k 1 belt ⊂ ( M ,ξ ′ ′ ) in Theorem 1.2 places constraints on the filling ( W ,ω ′ ′ ). We focus for now on the case where ( W ,ω ′ ′ ) is symplectically aspherical, which rules out bubbling of holomorphic spheres. For a suitable choice of tame almost complex structure J on ( W ,ω ′ ′ ), the Bishop families associated to S 2 n - -k 1 belt generate a compactified moduli space M of J -holomorphic disks in W ′ with one marked point, whose boundaries are mapped to S 2 n - -k 1 belt . In light of the marked point, this moduli space is necessarily diffeomorphic to a manifold with boundary and corners of the form
$$\overline { \Sigma } \times \mathbb { D } ^ { 2 } \,,$$
where Σ is a smooth, compact, connected and oriented manifold with boundary and corners, whose boundary is a sphere. Furthermore, the natural evaluation map
$$\text{ev} \colon ( \overline { \mathcal { M } }, \partial \overline { \mathcal { M } } ) \rightarrow ( W ^ { \prime }, S _ { \text{belt} } ^ { 2 n - k - 1 } )$$
is smooth, and its restriction
$$e v | _ { \partial \overline { \mathcal { M } } } \colon \partial \overline { \mathcal { M } } \rightarrow S _ { \text{belt} } ^ { 2 n - k - 1 }$$
is a continuous map of degree ± 1. The latter follows readily from the properties of the Bishop family, which provide a nonempty open subset U ⊂ W ′ that intersects S 2 n - -k 1 belt and is the diffeomorphic image of ev -1 ( U ) ⊂ M .
This description of the evaluation map ev: M→ W ′ already implies the homological part of Theorem 1.2, i.e. that [ S 2 n - -k 1 belt ] = 0 ∈ H 2 n - -k 1 ( W ′ ). To deduce stronger constraints, we will apply two further techniques. The first consists in performing surgery on the moduli space M to simplify its topology and suitably extending the evaluation map in order to prove [ S 2 n - -k 1 belt ] = 0 ∈ π 2 n - -k 1 ( W ′ ). This method works when the dimension of M is not too large. The second method is relevant in particular to the case k = 0 of the main result, as well as to Theorem 1.3, and is based on the following topological lemma.
Lemma 1.11. Let X,Y be smooth orientable compact manifolds with boundary and corners such that ∂Y is homeomorphic to a sphere and dim X + 2 = dim Y ≥ 3 . Write X ′ = X × D 2 , and assume that
$$f \colon ( X ^ { \prime }, \partial X ^ { \prime } ) \rightarrow ( Y, \partial Y )$$
is a continuous map that is smooth on the interior of X ′ , and for which we find an open subset U ⊂ ˚ Y such that f | f -1 ( U ) : f -1 ( U ) → U is a diffeomorphism.
Then Y is contractible.
While it will not be essential to most of our arguments, note that the h -cobordism theorem implies:
Corollary 1.12. If dim Y ≥ 5 in the lemma, then Y is diffeomorphic to a ball.
The k = 0 case of Theorem 1.2 is the case where ( M ,ξ ′ ′ ) is the standard contact sphere and the belt sphere is the entirety of M ′ . In this setting, applying the above lemma to the evaluation map ev: ( M = Σ × D 2 , ∂ M → ) ( W , ′ S 2 n -1 ) implies that W ′ must be diffeomorphic to a ball, hence this reproves the Eliashberg-FloerMcDuff theorem. We will explain this argument in more detail in § 2, including the proof of the lemma (see Lemma 2.3). In another context, we will also apply the lemma in § 6 to demonstrate that the contact structures arising on the boundaries
of certain Stein domains which look topologically like connected sums cannot arise from index 1 contact surgery, thus proving Theorem 1.3.
Here is a brief outline of the paper. In Section 2, we provide a foretaste of the methods in the rest of the paper by using them to give an alternative proof of the Eliashberg-Floer-McDuff theorem. Section 3 then explains the general case of the family of Lob s associated to a subcritical belt sphere. In Section 4, we define the relevant moduli space of holomorphic disks and establish its basic properties, leading to the proof of the homological part of Theorem 1.2. Section 5 then improves this to a homotopical statement in cases where the moduli space has sufficiently low dimension. Finally, in § 6 and § 7 respectively we prove Theorems 1.3 and 1.4 on contact connected sums and contractible Reeb orbits. The paper concludes with a brief appendix addressing the technical question of orientability for our moduli space of holomorphic disks.
## Acknowledgments
The authors would like to thank Yasha Eliashberg, Fran¸ cois Laudenbach, Patrick Massot, Emmy Murphy, Otto van Koert and Andr´ as Stipsicz for useful conversations. They are also very grateful to the referees for careful reading and suggesting many improvements.
The first and second authors were partially supported during this project by ANR grant ANR-10-JCJC 0102 . The third author is partially supported by a Royal Society University Research Fellowship and by EPSRC grant EP/K011588/1 . Additionally, all three authors gratefully acknowledge the support of the European Science Foundation's 'CAST' research network.
## 2. The Eliashberg-Floer-McDuff theorem revisited
In this section we modify slightly the proof of the Eliashberg-Floer-McDuff theorem [McD91, Theorem 1.5] in order to illustrate the methods that will be applied in the rest of the article. The original argument worked by capping off the symplectic filling and then sweeping through it with a moduli space of holomorphic spheres. Our version will be the same in many respects, but has more in common with the 3-dimensional argument of Eliashberg in [Eli90a]: instead of spheres, we use holomorphic disks attached to a family of Lob s.
Theorem 2.1 (Eliashberg-Floer-McDuff) . Let S 2 n -1 ⊂ C n be the unit sphere with its standard contact structure ξ 0 given by the complex tangencies to the sphere, that is,
$$\xi _ { 0 } = T \mathbb { S } ^ { 2 n - 1 } \cap \left ( i \cdot T \mathbb { S } ^ { 2 n - 1 } \right ) \,.$$
Every symplectically aspherical filling of ( S 2 n -1 , ξ 0 ) is diffeomorphic to the (2 n ) -ball.
Let z = x + i y = ( x 1 + iy 1 , . . . , x n + iy n ) be the coordinates of C n . The function f : C n → [0 , ∞ ) given by
$$f ( \mathbf z ) = \sum _ { j = 1 } ^ { n } ( x _ { j } ^ { 2 } + y _ { j } ^ { 2 } )$$
is plurisubharmonic, and the unit sphere is the boundary of the ball
$$\mathbb { D } ^ { 2 n } = \left \{ \mathbf z \in \mathbb { C } ^ { n } \, \Big | \ f ( \mathbf z ) \leq 1 \right \} \,.$$
The main geometric ingredient needed for our proof is a foliation of S 2 n -1 (minus some singular subset) by a family of Lob s, but this idea does not seem to work when applied directly to the unit sphere f -1 (1). Instead, we will deform the sphere to a different shape, which does contain a suitable family of Lob s that will suffice for our purposes. Define two functions g A , g B : C n → [0 , ∞ ) by
$$g _ { A } ( \mathbf z ) & = y _ { 1 } ^ { 2 } + \dots + y _ { n - 1 } ^ { 2 } \\ g _ { B } ( \mathbf z ) & = x _ { 1 } ^ { 2 } + \dots + x _ { n - 1 } ^ { 2 } + x _ { n } ^ { 2 } + y _ { n } ^ { 2 } \,.$$
Note that g B is strictly plurisubharmonic and g A is weakly plurisubharmonic as
$$- d d ^ { c } g _ { A } & = 2 \, d x _ { 1 } \wedge d y _ { 1 } + \dots + 2 \, d x _ { n - 1 } \wedge d y _ { n - 1 } \\ - d d ^ { c } g _ { B } & = 2 \, d x _ { 1 } \wedge d y _ { 1 } + \dots + 2 \, d x _ { n - 1 } \wedge d y _ { n - 1 } + 4 \, d x _ { n } \wedge d y _ { n } \.$$
We will now consider the subset (see Fig. 3)
$$\widehat { \mathbb { D } } ^ { 2 n } = \left \{ \mathbf z \in \mathbb { C } ^ { n } \, \Big | \ g _ { B } ( \mathbf z ) \leq 1 \right \} \cap \left \{ \mathbf z \in \mathbb { C } ^ { n } \, \Big | \ g _ { A } ( \mathbf z ) \leq 1 \right \} \,.$$
Up to reordering the coordinates, ̂ D 2 n is a bi-disk D n +1 × D n -1 ⊂ R 2 n , which clearly contains the unit ball. Its boundary is not a smooth manifold, but is nonetheless homeomorphic to the unit sphere. It decomposes as
$$\partial \widehat { \mathbb { D } } ^ { 2 n } \cong \mathbb { D } ^ { n + 1 } \times \left ( \partial \mathbb { D } ^ { n - 1 } \right ) \cup \left ( \partial \mathbb { D } ^ { n + 1 } \right ) \times \mathbb { D } ^ { n - 1 } = S _ { A } \cup S _ { B } \,,$$
where we have used the notation
$$S _ { A } \coloneqq \left \{ g _ { A } = 1 \right \} \cap \partial \widehat { \mathbb { D } } ^ { 2 n } \cong \mathbb { D } ^ { n + 1 } \times \mathbb { S } ^ { n - 2 }$$
and
$$S _ { B } \coloneqq \left \{ g _ { B } = 1 \right \} \cap \partial \widehat { \mathbb { D } } ^ { 2 n } \cong \mathbb { S } ^ { n } \times \mathbb { D } ^ { n - 1 } \,.$$
Figure 3. We find a family of Lob s by deforming the sphere to the boundary of a bi-disk. One of the two parts of the boundary, which we denote by S B , will then be foliated by Lob s.
Let now ( W,ω ) be a symplectically aspherical filling of ( S 2 n -1 , ξ 0 ) . If it is only a weak filling, we can extend it by attaching a symplectic collar to obtain a strong symplectic filling of the sphere [MNW13, Remark 2.11] because ω | T S 2 n -1 is exact. This filling is diffeomorphic to the initial one, and it is also still symplectically aspherical, because any 2-sphere can just be pushed by a homotopy entirely into the
old symplectic filling. After rescaling the symplectic form, the extended symplectic manifold will be a strong symplectic filling of the unit sphere.
Remove now the interior ˚ D 2 n of the unit ball from ̂ D 2 n , and glue ̂ D 2 n \ ˚ D 2 n symplectically onto the filling W . Denote this new symplectic manifold by ( ̂ ̂ W,ω ). Clearly ̂ W is homeomorphic to W . Using holomorphic disks, we will show as in the original paper by McDuff that ̂ W is contractible, so that the h -cobordism theorem [Sma62; Mil65] implies that W must be diffeomorphic to D 2 n whenever 2 n -1 ≥ 5.
To study ̂ W using holomorphic curves, choose first an almost complex structure J on ̂ W that is tamed by ̂ ω and that agrees on a small neighborhood of ∂ ̂ D 2 n in ̂ W with the standard complex structure i on C n . The holomorphic curves we are interested in are attached to a family of Lob s, which we will introduce now. Let Ψ: S n × D n -1 → S B ⊂ ∂W ̂ be the embedding into the boundary of ̂ W given by
↦
$$\left ( ( a _ { 1 }, a _ { 2 }, \dots, a _ { n + 1 } ) ; ( b _ { 1 }, \dots, b _ { n - 1 } ) \right ) \mapsto \left ( a _ { 1 } + i b _ { 1 }, \dots, a _ { n - 1 } + i b _ { n - 1 }, a _ { n } + i a _ { n + 1 } \right ).$$
The image of Ψ lies in S B ⊂ ∂W ̂ , and the J -complex tangencies on the corresponding part of ∂W ̂ are the kernel of the 1-form
$$- d ^ { c } g _ { B } = 2 x _ { 1 } \, d y _ { 1 } + \dots + 2 x _ { n - 1 } \, d y _ { n - 1 } + 2 \left ( x _ { n } \, d y _ { n } - y _ { n } \, d x _ { n } \right ) \,.$$
We obtain for the pull-back
$$\Psi ^ { * } \left ( - d ^ { c } g _ { B } \right ) = 2 a _ { 1 } \, d b _ { 1 } + \dots + 2 a _ { n - 1 } \, d b _ { n - 1 } + 2 \left ( a _ { n } \, d a _ { n + 1 } - a _ { n + 1 } \, d a _ { n } \right )$$
so that the restriction of Ψ ∗ ( -d g c B ) to each of the spheres S n × { ( b 1 , . . . , b n -1 ) = const } gives
$$2 \left ( a _ { n } \, d a _ { n + 1 } - a _ { n + 1 } \, d a _ { n } \right ) \,.$$
This means that the projection
↦
$$\left ( ( a _ { 1 }, a _ { 2 }, \dots, a _ { n + 1 } ) ; ( b _ { 1 }, \dots, b _ { n - 1 } ) \right ) \mapsto \arg ( a _ { n } + i a _ { n + 1 } ) \in \mathbb { S } ^ { 1 }$$
defines for each ( b 1 , . . . , b n -1 ) a Lob with the ( n -1)-ball as pages and trivial monodromy. From now on we denote the points in D n -1 by b = ( b 1 , . . . , b n -1 ), and write for the Lob
$$L _ { \mathbf b } = \Psi ( \mathbb { S } ^ { n } \times \{ \mathbf b \} ),$$
and B b for its binding.
For the technical details of the following part, we refer to Section 4.3. We will study the space
$$\widetilde { \mathcal { M } } _ { * } = \left \{ ( \mathbf b, u, z _ { 0 } ) \, \Big | \, \mathbf b \in \mathbb { D } ^ { n - 1 }, \, u \colon ( \mathbb { D } ^ { 2 }, \partial \mathbb { D } ^ { 2 } ) \to \left ( \widehat { W }, L _ { \mathbf b } \right ), \, z _ { 0 } \in \mathbb { D } ^ { 2 } \right \}$$
of nonconstant holomorphic maps from a disk, equipped with one marked point z 0 , and with boundary sent into one of the Lob s L b . Additionally, we require that u is homotopic to a Bishop disk as an element in π 2 ( ̂ W,L b \ B b ) , and we denote the corresponding subset by ˜ M . Next we divide ˜ M by the action of the group Aut( D 2 ) of biholomorphic transformations on D 2 , where ϕ ∈ Aut( D 2 ) acts on ˜ M via
$$\varphi \cdot ( { \mathbf b }, u, z _ { 0 } ) = \left ( { \mathbf b }, u \circ \varphi ^ { - 1 }, \varphi ( z _ { 0 } ) \right ).$$
We denote the moduli space ˜ M / Aut( D 2 ) by M . Note that for every class [ b , u, z ] in M , we can fix a unique representative ( b , u 0 , z 0 ) by choosing a parametrization
of u such that
$$( 2. 1 ) \quad \ \ u ( z ) \in \begin{cases} \text{the $0$ degree page of the Lob,} \quad \text{ if $z=1$,} \\ \text{the $\pi/2$ degree page of the Lob,} \quad \text{if $z=1$.} \end{cases}$$
A corollary of this is that the moduli space M (before the compactification, see below) is a trivial disk bundle over the space of unmarked disks. This is the key fact that will allow us to 'push' the topology of W into its boundary (which is the geometric analogue of the algebraic argument given in [McD91] and [OV12]).
Next, we need to understand the compactification of M . Note first that typical holomorphic disks are surrounded by a neighborhood of other typical holomorphic disks, that is, they represent smooth points of the interior of the moduli space M . With 'typical', we mean smooth holomorphic disks whose interior points are mapped to the interior of ̂ W , and whose boundary sits on a Lob L b that is not a boundary Lob , i.e. for which ‖ b ‖ < 1, and such that the disk does not touch the binding B b of the Lob .
Let us now consider the remaining cases. The boundary of ̂ W consists of S A ⊂ { g A = 1 } and S B ⊂ { g B = 1 } , which are weakly and strongly plurisubharmonic hypersurfaces respectively. A disk touching S B with one of its interior points will automatically be constant. If the disk touches S A instead, then it needs to be entirely contained in this hypersurface, and in particular its boundary will lie on a Lob with ‖ b ‖ = 1; below we will explain how to understand the disks in this second case explicitly.
For every Lob L b , there is a certain neighborhood of its binding B b that is only intersected by Bishop disks. Since there is exactly one disk meeting every point of this neighborhood, that is, the evaluation map
↦
$$\text{ev} \colon \mathcal { M } \rightarrow \dot { W }, \, [ \mathbf b, u, z _ { 0 } ] \mapsto u ( z _ { 0 } )$$
restricts close to B b to a diffeomorphism, it follows that the compactification M contains disks that collapse to a point in the binding. In [NR11] it was shown that adding these constant disks to M , corresponds to adding points which lie on the smooth boundary of the compactification M .
Before understanding the bubbling, we will discuss disks whose boundary lies in a Lob L b ⊂ S A .
Lemma 2.2. Suppose u ∈ M maps ∂ D 2 to a Lob L b such that ‖ b ‖ = 1 . Then the image of u is completely contained in S A , and moreover, it is obtained by the intersection of a complex line parallel to the z n -plane with S A .
Proof. Parametrize the disk by polar coordinates re iφ . By acting on the coordinates z , . . . , z 1 n -1 with a matrix in SO n ( -1) (regarded as an element of SU( n -1) with real entries), we can assume without loss of generality that the Lob L b corresponds to the parameter b = (1 0 , , . . . , 0), as the functions g A and g B are invariant under such an action. In particular it follows that the y 1 -coordinate of u has its maximum on the boundary of u . The x 1 -coordinate of u | ∂ D 2 is bounded, and hence there is an angle e iφ 0 at which the derivative
$$\frac { d } { d \phi } \Big | _ { \phi = \phi _ { 0 } } x _ { 1 } \Big ( u ( e ^ { i \phi } ) \Big ) = 0$$
is zero. Complex multiplication gives i · ∂ r = ∂ φ , hence
$$d y _ { 1 } ( D u \cdot \partial _ { r } ) = d y _ { 1 } ( D u \cdot ( - i \cdot \partial _ { \phi } ) ) = - d y _ { 1 } ( i \cdot D u \cdot \partial _ { \phi } ) = - d x _ { 1 } ( D u \cdot \partial _ { \phi } ) = 0 \.$$
It follows that the outward derivative of the y 1 -coordinate vanishes at the point e iφ 0 ∈ D 2 , so that according to the boundary point lemma, y 1 must equal the constant 1 on the whole disk, and as a consequence u lies entirely in S A .
The y 2 -to y n -1 -coordinates are all 0 on the boundary of the disk, and hence by the maximum principle, they need to be both maximal and minimal on all of the disk. With the Cauchy-Riemann equation we obtain that the x 1 -up to x n -1 -coordinates of u need all to be constant on u (for more details read Section 4.2). glyph[square]
As explained in Section 4.3, no bubbling can occur under our assumptions, and hence M will be a compact manifold with boundary and corners (the boundary is smooth everywhere with the exception of the disks corresponding to the edges of ̂ W ). Moreover, the moduli space is orientable (see Appendix A) and the evaluation map
↦
$$\text{ev} \colon ( \overline { \mathcal { M } }, \partial \overline { \mathcal { M } } ) \rightarrow ( \widehat { W }, \partial \widehat { W } ), \, [ \text{b}, u, z _ { 0 } ] \mapsto u ( z _ { 0 } )$$
is a degree 1 map, that is, it maps the fundamental class [ M ∈ ] H 2 n ( M M , ∂ ; Z ) onto the fundamental class [ ̂ W ] ∈ H 2 n ( ̂ W,∂W ̂ ; Z ) . We are therefore in a position to apply the following topological result, which was stated as Lemma 1.11 in the introduction.
Lemma 2.3. Let X,Y be smooth orientable compact manifolds with boundary and corners such that ∂Y is homeomorphic to a sphere and dim X + 2 = dim Y ≥ 3 . Write X ′ = X × D 2 , and assume that
$$f \colon ( X ^ { \prime }, \partial X ^ { \prime } ) \rightarrow ( Y, \partial Y )$$
is a continuous map that is smooth on the interior of X ′ , and for which we find an open subset U ⊂ ˚ Y such that f | f -1 ( U ) : f -1 ( U ) → U is a diffeomorphism.
Then Y is contractible.
Proof. Note that by Whitehead's theorem, it suffices to show that Y is weakly contractible, that is, π j ( Y ) = 0 for all j > 0. Using Hurewicz's theorem we will show instead that Y is simply connected and satisfies H Y j ( ; Z ) = 0 for all j > 0.
- (i) We will first consider the fundamental group of Y . Choose the base point p 0 ∈ U ⊂ ˚ . Y Let γ be a smooth, embedded loop representing a class in π 1 ( Y, p 0 ) that lies in the interior of Y . After a perturbation, we can assume that γ is transverse to the map f , so f -1 ( γ ) will be a finite collection of loops Γ 0 , . . . , Γ N in X ′ .
There is one loop, say Γ 0 , that is mapped to γ with degree one. The reason for this is that γ runs through U , where f is a diffeomorphism. Then the loop f ◦ Γ 0 is homotopic to γ , and thus represents the same class in π 1 ( Y, p 0 ) .
Using the fact that X ′ is diffeomorphic to a trivial disk bundle, we may shift Γ 0 into the boundary of X ′ (just by moving it inside the D 2 -factor). In particular, this shows that [ γ ] = [ f ◦ Γ ] can be represented by a loop that lives in the boundary 0 of Y , and is thus contractible.
- (ii) Next we need to compute the homology of Y . It is easy to see that the image of
$$f _ { * } \colon H _ { * } ( X ^ { \prime } ; \mathbb { Z } ) \to H _ { * } ( Y ; \mathbb { Z } )$$
is trivial. Indeed, all homology groups H k ( X ′ ; Z ) with k ≥ dim X are trivial, so that we only need to study k < dim X < dim ∂Y . Let A ∈ H k ( Y ; Z ) be a homology
class that lies in the image of f ∗ so that there is a B ∈ H k ( X ′ ; Z ) with A = f B ∗ . Since X ′ = X × D 2 where D 2 is contractible, B can be represented by a cycle in X ×{ } p for any point p ∈ D 2 ; in particular we are free to choose p ∈ ∂ D 2 , hence B is represented by a cycle in ∂X ′ . This implies that the class A is homologous to a cycle in the sphere ∂Y , which shows that A must be trivial.
We will now show that f ∗ : H k ( X ′ ; Z ) → H k ( Y ; Z ) is surjective for every k ≤ 1 2 dim Y . Assume for now that k < 1 2 dim Y and that we already have shown for every l < k that H Y l ( ; Z ) = 0. From (i), we see that k ≥ 2. By Hurewicz's theorem we know that H k ( Y ; Z ) is isomorphic to π k ( Y, p 0 ) so that we can represent any homology class in H k ( Y ; Z ) by a map
$$s \colon \mathbb { S } ^ { k } \rightarrow Y \.$$
We additionally assume again that the base point p 0 lies in U . After a generic perturbation (see [Hir94, Theorem II.2.12]), s will be an embedding that is transverse to f , and we find a closed (possibly disconnected) smooth submanifold f -1 ( s ) ⊂ X ′ . As in step (i), there is a unique connected component S of f -1 ( s ) which passes through the base point p 0 . Clearly we have a map S → S k of degree one that makes the following diagram commute:
/
/
GLYPH<31>
GLYPH<31>
~
~
As we wanted to show, it follows that f ( S ) represents the same class in H k ( Y ; Z ) as s , so that f ∗ : H k ( X ′ ; Z ) → H k ( Y ; Z ) is surjective, and as a consequence H k ( Y ; Z ) is trivial.
Let us now study the case k = 1 2 dim Y . Again we may represent any element in H k ( Y ; Z ) by a map
$$s \colon \mathbb { S } ^ { k } \rightarrow Y$$
that goes through p 0 ∈ U . Perturbing s , we may assume that it is an immersion with transverse self-intersections, and also that it is transverse to f (position the double-points at regular values of f ).
The map F := ( f, s ): X ′ × S k → Y × Y is transverse to the diagonal glyph[triangle] Y ⊂ Y × Y , and it follows that F -1 ( glyph[triangle] Y ) is a closed smooth submanifold of X ′ × S k . Let S be the unique component of F -1 ( glyph[triangle] Y ) that is mapped by F to ( p , p 0 0 ). By definition we have f ◦ Π 1 | S = s ◦ Π 2 | S , where Π 1 : X ′ × S k → X ′ and Π 2 : X ′ × S k → S k are the canonical projections. Furthermore Π 2 is a degree one map.
This ends the proof, because [ f ◦ Π 1 | S ] = [ s ] ∈ H k ( Y ; Z ).
(iii) Above we have seen that H k ( Y ; Z ) is trivial for all 1 ≤ k ≤ 1 2 dim Y . Using Poincar´ e-Lefschetz duality, H dim Y -k ( Y, ∂Y ; Z ) is isomorphic to H k ( Y ; Z ) and thus it also vanishes (for 1 ≤ k ≤ 1 2 dim Y ). The long exact sequence of the pair then implies that H dim Y -k ( Y ; Z ) = { 0 } , and finally the universal coefficient theorem tells us that H dim Y -k ( Y ; Z ) = { 0 } , so that Y has trivial homology. glyph[square]
To conclude the proof of Theorem 2.1, simply apply the lemma and Corollary 1.12 to the moduli space and its evaluation map found above. It follows that ̂ W is diffeomorphic to a ball.
## 3. Weinstein handles and contact surgeries
In this section, we will give a precise description of Weinstein handle attachment, and then show how to deform the contact structure near the belt sphere in order to find a suitable family of Lob s.
glyph[negationslash]
Assume ( W,ω ) is a weak symplectic filling of a (2 n -1)-dimensional contact manifold ( M,ξ ), choose an isotropic sphere S k -1 at ↪ → ( M,ξ ) with trivial normal bundle νS k -1 at , and suppose that ω | TS k -1 at is exact (this is obviously always the case if k = 3). After deforming the symplectic structure in a small neighborhood of the boundary using [MNW13, Remark 2.11], we can find a Liouville vector field X L on a neighborhood U ⊂ W close to S k -1 at that points transversely out of W , and such that
$$\alpha | _ { U \cap M } = ( \iota _ { X _ { L } } \omega ) | _ { T ( U \cap M ) }$$
is a contact form for ξ .
Topologically, the handle attachment can be described as follows. Choose a trivialization of the normal bundle νS k -1 at in M , identifying a tubular neighborhood with
$$T ^ { * } \mathbb { S } ^ { k - 1 } \times \mathbb { R } ^ { 2 ( n - k ) + 1 } \cong \mathbb { S } ^ { k - 1 } \times \mathbb { R } ^ { 2 n - k } \,,$$
and let the k -handle H k be the poly-disk
$$H _ { k } \coloneqq \mathbb { D } ^ { k } \times \mathbb { D } ^ { 2 n - k } \,.$$
The boundary of H k can be written as the union
$$\partial H _ { k } \coloneqq \mathbb { S } ^ { k - 1 } \times \mathbb { D } ^ { 2 n - k } \cup \mathbb { D } ^ { k } \times \mathbb { S } ^ { 2 n - k - 1 } \,.$$
Using the obvious homeomorphism, we can 'glue' H k along the subset S k -1 × D 2 n -k ⊂ ∂H k to the neighborhood of S k -1 at using the trivialization chosen above. Denote the new manifold by
$$W ^ { \prime } = W \cup _ { \nu S _ { \text{at} } ^ { k - 1 } } H _ { k } \,.$$
Note that the gluing operation depends on the trivialization chosen for the normal bundle of S k -1 at .
The boundary M ′ = ∂W ′ is obtained from the old contact manifold M by removing the neighborhood of S k -1 at , and gluing in the free boundary component of the handle, that is,
$$M ^ { \prime } = \left ( M \, \sum \nu S _ { \text{at} } ^ { k - 1 } \right ) \cup \left ( \mathbb { D } ^ { k } \times \mathbb { S } ^ { 2 n - k - 1 } \right ).$$
This operation changing M to M ′ is called a contact surgery of index k along S k -1 at . We will recall below how the natural symplectic structure on W ′ with weakly contact-type boundary M ′ is defined. The belt sphere S 2 n - -k 1 belt of the handle H k is the 'cosphere to the gluing sphere',
$$S _ { \text{belt} } ^ { 2 n - k - 1 } \coloneqq \{ 0 \} \times \mathbb { S } ^ { 2 n - k - 1 } \subset \partial H _ { k } \,.$$
Note that contact surgery can also be defined as an operation on contact manifolds without assuming that they are symplectically fillable: one only need regard ( M,ξ ) as the contact boundary of a piece of its symplectization ( ( -ε, 0] × M,d e α ( t ) ) . Topologically, a surgery of index k is the operation of removing a small neighborhood of the isotropic sphere S k -1 at from the contact manifold ( M,ξ ) and gluing into the cavity a standard patch that is diffeomorphic to D k × S 2 n - -k 1 .
3.1. Attaching Weinstein handles. We now define a symplectic model for a subcritical Weinstein handle of index k in dimension 2 n . For this, split C n into C k × C m × C with n = k + m +1, and write the coordinates on C k as
$$\mathbf z ^ { - } = \mathbf x ^ { - } + i \mathbf y ^ { - } \coloneqq \left ( x _ { 1 } ^ { - } + i y _ { 1 } ^ { - }, \dots, x _ { k } ^ { - } + i y _ { k } ^ { - } \right ),$$
the ones on C m as
$$\mathbf z ^ { + } = \mathbf x ^ { + } + i \mathbf y ^ { + } \coloneqq \left ( x _ { 1 } ^ { + } + i y _ { 1 } ^ { + }, \dots, x _ { m } ^ { + } + i y _ { m } ^ { + } \right ),$$
and the ones on C as
$$z ^ { \circ } = x ^ { \circ } + i \, y ^ { \circ } \,.$$
The coordinate z ◦ behaves like any other of the coordinates in z + , and in the usual descriptions of the handle attachment, it is not distinguished from z + . (For a critical handle attachment k = n , thus there are no z + -or z ◦ -coordinates.) The reason why we have introduced this more complicated notation is to prepare for the deformation we will perform in the next section, in which the z ◦ -coordinate will play a particular role.
As a model for the handle, take
$$H _ { r } \coloneqq \mathbb { D } ^ { k } \times \mathbb { D } _ { r } ^ { 2 n - k } \subset \mathbb { C } ^ { n }$$
where the first disk corresponds to the y --coordinates, the second disk to the ( x -, z + , z ◦ )-coordinates, and r > 0 is a constant, i.e. H r is the intersection of the two subsets
$$\left \{ \left \| y ^ { - } \right \| ^ { 2 } \leq 1 \right \} \, \cap \, \left \{ \left \| x ^ { - } \right \| ^ { 2 } + \left \| z ^ { + } \right \| ^ { 2 } + \left | z ^ { \circ } \right | ^ { 2 } \leq r ^ { 2 } \right \} \,.$$
We denote
$$\partial _ { - } H _ { r } = \partial \mathbb { D } ^ { k } \times \mathbb { D } _ { r } ^ { 2 n - k }, \ \partial _ { + } H _ { r } = \mathbb { D } ^ { k } \times \partial \mathbb { D } _ { r } ^ { 2 n - k } \,.$$
The core of ∂ -H r is the ( k -1)-sphere
$$S _ { - } \coloneqq \left \{ \left \| y ^ { - } \right \| ^ { 2 } = 1, \, x ^ { - } = 0, \, z ^ { + } = 0, \, z ^ { \circ } = 0 \right \}$$
which will be identified with the attachment sphere in a contact manifold; the core of ∂ + H r is the ( n + m )-sphere (note that n + m = 2 n -k -1)
$$S _ { + } \coloneqq \left \{ y ^ { - } = 0, \, \left \| x ^ { - } \right \| ^ { 2 } + \left \| z ^ { + } \right \| ^ { 2 } + \left | z ^ { \circ } \right | ^ { 2 } = r ^ { 2 } \right \} \,.$$
Choose on C n the symplectic form
$$\omega = 2 \sum _ { r = 1 } ^ { k } d x _ { r } ^ { - } \wedge d y _ { r } ^ { - } + 4 \sum _ { s = 1 } ^ { m } d x _ { s } ^ { + } \wedge d y _ { s } ^ { + } + 4 \, d x ^ { \circ } \wedge d y ^ { \circ } \,.$$
It admits the Liouville form
$$\lambda = 2 \sum _ { r = 1 } ^ { k } ( 2 x _ { r } ^ { - } d y _ { r } ^ { - } + y _ { r } ^ { - } d x _ { r } ^ { - } ) + 2 \sum _ { s = 1 } ^ { m } ( x _ { s } ^ { + } d y _ { s } ^ { + } - y _ { s } ^ { + } d x _ { s } ^ { + } ) + 2 \left ( x ^ { \circ } d y ^ { \circ } - y ^ { \circ } d x ^ { \circ } \right )$$
that is associated to the Liouville vector field
$$k$$
$$X _ { L } = \sum _ { r = 1 } ^ { k } & \left ( 2 x _ { r } ^ { - } \, \frac { \partial } { \partial x _ { r } ^ { - } } - y _ { r } ^ { - } \, \frac { \partial } { \partial y _ { r } ^ { - } } \right ) \\ & \quad + \frac { 1 } { 2 } \sum _ { s = 1 } ^ { m } & \left ( x _ { s } ^ { + } \, \frac { \partial } { \partial x _ { s } ^ { + } } + y _ { s } ^ { + } \, \frac { \partial } { \partial y _ { s } ^ { + } } \right ) + \frac { 1 } { 2 } \left ( x ^ { \circ } \, \frac { \partial } { \partial x ^ { \circ } } + y ^ { \circ } \, \frac { \partial } { \partial y ^ { \circ } } \right ).$$
The field X L points outward through ∂ + H r and inward at ∂ -H r , so that both ∂ + H r and ∂ -H r are contact type hypersurfaces with the corresponding coorientations. The core S -⊂ ∂ -H r is an isotropic sphere with trivial conformal symplectic normal bundle.
Figure 4. The handle can be glued onto a contact manifold.
Let now ( M,ξ ) be a given contact manifold and let S k -1 at ⊂ M be a ( k -1)dimensional isotropic sphere with trivial conformal symplectic normal bundle that will serve as the attaching sphere of the k -handle H r . Fixing r > 0 small enough, ∂ -H r , endowed with the contact structure induced by λ , is contactomorphic to a neighborhood N ( S k -1 at ) ⊂ ( M,ξ ) of S k -1 at . We choose a contact form α for ξ on M such that α | N ( S k -1 at ) can be glued to λ | T ∂ ( -H r ) and define the Liouville manifold
$$\left ( W _ { 0 }, \lambda _ { 0 } \right ) \coloneqq \left ( \left ( - \varepsilon, 0 \right ) \times M, e ^ { t } \alpha \right ) \cup _ { \partial _ { - } H _ { r } } \left ( H _ { r }, \lambda \right ).$$
The positive boundary of W 0 (denoted ∂ + W 0 ) has two smooth faces M \ N ( S k -1 at ) and ∂ + H r , meeting along a corner which is the image of the corner ∂ D k × ∂ D 2 n -k r in ∂H r , see Fig. 4. Fix a small neighborhood U of the corner, and choose a smooth hypersurface M ′ that matches ∂ + W 0 outside of U , and is transverse to X L in U . Denote the induced contact structure on M ′ by ξ ′ = TM ′ ∩ ker λ 0 .
Note that the constant r > 0 can be made arbitrarily small, without changing the isotopy class of the contact structure on M ′ ; we can shrink the size of the handle continuously (including the smoothing) which allows us to apply Gray stability to obtain an isotopy with support in the model neighborhood.
The belt sphere S 2 n - -k 1 belt = S n + m belt of the k -handle is the core S + of ∂ + H r .
3.2. Families of Lob s on a deformed subcritical handle. To find the desired family of Lob s, we will now modify the contact structure in a neighborhood of the belt sphere in two steps. The first deformation is borrowed from the recent article [GZ16]; it replaces a technically more complicated method that was used in an earlier version of this paper.
Consider again a 'thin' handle H r = D k × D 2 n -k r ⊂ C n with r glyph[lessmuch] 1 as used above. Suppose that the rounding of the corners has been performed for values of ‖ y -‖ in the interval [1 -ε, 1]. The part of ∂ + H r outside the smoothing region lies in the level set { f = r 2 } of the function
$$f ( z ^ { - }, z ^ { + }, z ^ { \circ } ) = \left \| { \mathbf x } ^ { - } \right \| ^ { 2 } + \left \| { \mathbf z } ^ { + } \right \| ^ { 2 } + \left | z ^ { \circ } \right | ^ { 2 }.$$
Wewould like to modify the Liouville field on a neighborhood of the belt sphere so that the induced contact structure coincides with the field of complex hyperplanes on the boundary. For this, add the Hamiltonian vector field X H of a function H : C n → R to X L , since then ˆ X L := X L + X H will still be a Liouville field.
Let ρ : [0 , ∞ → -) [ 1 0] be a smooth function that is equal to , -1 on the interval [0 , √ 1 -2 ], equal to 0 on the interval [ ε √ 1 -ε, 1] and that increases monotonically in between. Define the Hamiltonian function
$$H ( z ^ { - }, z ^ { + }, z ^ { \circ } ) \coloneqq 2 \left \langle { \mathbf x } ^ { - }, { \mathbf y } ^ { - } \right \rangle \rho \left ( \left \| { \mathbf y } ^ { - } \right \| ^ { 2 } \right ).$$
The Hamiltonian vector field corresponding to H is
$$X _ { H } = - \sum _ { r = 1 } ^ { k } y _ { r } ^ { - } \rho ( \| y ^ { - } \| ^ { 2 } ) \, \frac { \partial } { \partial y _ { r } ^ { - } } + \sum _ { r = 1 } ^ { k } \left ( 2 \, \langle x ^ { - }, y ^ { - } \rangle \, y _ { r } ^ { - } \rho ^ { \prime } ( \| y ^ { - } \| ^ { 2 } ) + x _ { r } ^ { - } \rho ( \| y ^ { - } \| ^ { 2 } ) \right ) \, \frac { \partial } { \partial x _ { r } ^ { - } } \.$$
The vector field ˆ X L agrees outside the support of ρ with X L , and it is everywhere transverse to ∂ + H r as can be seen from
$$\dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \\ \mathcal { L } _ { \hat { X } _ { L } } f & = \mathcal { L } _ { X _ { L } } f + \mathcal { L } _ { X _ { H } } f \\ & = 4 \left \| \mathbf x ^ { - } \right \| ^ { 2 } + \left \| \mathbf z ^ { + } \right \| ^ { 2 } + \left | z ^ { \circ } \right | ^ { 2 } + \mathcal { L } _ { X _ { H } } \left \| \mathbf x ^ { - } \right \| ^ { 2 } \\ & = \left ( 4 + 2 \, \rho \right ) \left \| \mathbf x ^ { - } \right \| ^ { 2 } + \left \| \mathbf z ^ { + } \right \| ^ { 2 } + \left | z ^ { \circ } \right | ^ { 2 } + 4 \left \langle \mathbf x ^ { - }, \mathbf y ^ { - } \right \rangle ^ { 2 } \rho ^ { \prime } > 0 \,,$$
because ρ ≥ -1, and ρ ′ ≥ 0.
It follows that λ and ˆ := λ ι ˆ X L ω induce isotopic contact structures on M ′ .
The contact structure on the domain ∂ + H r ∩ {‖ y -‖ 2 ≤ 1 -2 ε } is the kernel of the Liouville form
$$\hat { \lambda } = \lambda + d H = 2 \, \sum _ { r = 1 } ^ { k } x _ { r } ^ { - } \, d y _ { r } ^ { - } + 2 \, \sum _ { s = 1 } ^ { m } \left ( x _ { s } ^ { + } \, d y _ { s } ^ { + } - y _ { s } ^ { + } \, d x _ { s } ^ { + } \right ) + 2 \, ( x ^ { \circ } \, d y ^ { \circ } - y ^ { \circ } \, d x ^ { \circ } ) \.$$
Remark 3.1 . This first deformation shows that the surgered manifold contains a neighborhood of the belt sphere that is contactomorphic to a cylinder { f = r 2 } ∩ { ‖ y -‖ 2 ≤ 1 / 2 } ⊂ C n with r arbitrarily small and a contact structure given as kernel of ˆ . λ Note that ˆ λ on the domain under consideration is equal to the differential -d f c , i.e. the contact structure on our domain coincides with the complex tangencies. This is the key fact that we will exploit in the second deformation below.
To continue, we consider the setting of Theorem 1.2, in which ( M ,ξ ′ ′ ) was a fillable contact manifold obtained by subcritical surgery. Since this will be the main object of study from now on, it will be convenient to simplify the notation, hence we assume (unlike in the statement of Theorem 1.2) that ( M,ξ ) is a closed contact manifold of dimension 2 n -1 that has been obtained by a surgery of index k ∈ { 1 , . . . , n - } 1 from another contact manifold, and let ( W,ω ) be a weak symplectic filling of ( M,ξ ). Since the theorem in dimension three already follows from the much stronger result of Eliashberg [Eli90a], we are free to assume n ≥ 3. The belt sphere then has dimension 2 n -k -1 ≥ n ≥ 3, hence the restriction of ω to S 2 n - -k 1 belt is automatically exact. It follows (using [MNW13, Remark 2.11]) that ω can be deformed in a collar neighborhood of ∂W so that an outwardly transverse
Liouville vector field exists near S 2 n - -k 1 belt , and we are therefore free to pretend in the following discussion that ( W,ω ) is a strong filling of ( M,ξ ). In particular, we may assume that the symplectic structure on a collar neighborhood close to the belt sphere looks like the symplectic structure on the boundary of the model of the handle, and we may identify both.
Let f : C n → [0 , ∞ ) be again the plurisubharmonic function
$$f ( z ^ { - }, z ^ { + }, z ^ { \circ } ) = \left \| { \mathbf x } ^ { - } \right \| ^ { 2 } + \left \| { \mathbf z } ^ { + } \right \| ^ { 2 } + \left | z ^ { \circ } \right | ^ { 2 }.$$
By the explanations above (see Remark 3.1), the belt sphere S n + m belt has a neighborhood U M ⊂ M that is contactomorphic to the cylinder
$$C _ { r } \coloneqq \left \{ ( z ^ { - }, z ^ { + }, z ^ { \circ } ) \in \mathbb { C } ^ { n } \, \left | \ f ( z ^ { - }, z ^ { + }, z ^ { \circ } ) = r ^ { 2 }, \, \left \| y ^ { - } \right \| ^ { 2 } < 1 / 2 \right \}$$
for arbitrarily small r glyph[lessmuch] 1 with contact structure ˆ given as the kernel of the Liouξ ville form ˆ = λ -d f c | TC r , and since the cylinder is a level set of f , this also means that ˆ are the complex tangencies of ξ C r . We denote by U W a small neighborhood of U M in W that is symplectomorphic to the subset
$$\left \{ ( z ^ { - }, z ^ { + }, z ^ { \circ } ) \in \mathbb { C } ^ { n } \, \left | \ f ( z ^ { - }, z ^ { + }, z ^ { \circ } ) \in ( r ^ { 2 } - \delta, r ^ { 2 } ], \, \left \| y ^ { - } \right \| ^ { 2 } < 1 / 2 \right \}$$
with symplectic form ω = -dd f c .
Using the embedding of U W into our model, we can extend the symplectic filling ( W,ω ) by attaching the following compact symplectic subdomain of C n : replace f by G := max { g A , g B } which is obtained as the maximum of the two functions:
$$g _ { A } ( { \mathbf z } ^ { - }, { \mathbf z } ^ { + }, z ^ { \circ } ) \coloneqq \left \| { \mathbf y } ^ { + } \right \| ^ { 2 }$$
and
$$g _ { B } ( \mathbf z ^ { - }, \mathbf z ^ { + }, z ^ { \circ } ) \coloneqq \left \| x ^ { - } \right \| ^ { 2 } + \left \| \mathbf x ^ { + } \right \| ^ { 2 } + \psi \left ( \left \| y ^ { - } \right \| \right ) \cdot \left \| \mathbf y ^ { + } \right \| ^ { 2 } + \left | z ^ { \circ } \right | ^ { 2 } \,,$$
where ψ is a cut-off function that vanishes close to 0, and increases until it reaches 1 close to ‖ y -‖ = 1 / √ 2.
Clearly g A is a weakly plurisubharmonic function. The function g B is strictly plurisubharmonic on a neighborhood of { y + = 0 } because the last term of
$$- d d ^ { c } g _ { B } = 2 \sum _ { r = 1 } ^ { k } d x _ { r } ^ { - } \wedge d y _ { r } ^ { - } + 2 \sum _ { s = 1 } ^ { m } d x _ { s } ^ { + } \wedge d y _ { s } ^ { + } + 4 \, d x ^ { \circ } \wedge d y ^ { \circ } - d d ^ { c } \left ( \psi ( \| y ^ { - } \| ) \cdot \| y ^ { + } \| ^ { 2 } \right )$$
simplifies along this subset to
$$& \text{smpimes along usuosec} \, \omega \\ & - d d ^ { c } \Big ( \psi ( \| y ^ { - } \| ) \cdot \| y ^ { + } \| ^ { 2 } \Big ) = - \psi ( \| y ^ { - } \| ) \, d d ^ { c } \| y ^ { + } \| ^ { 2 } - d \| y ^ { + } \| ^ { 2 } \wedge d ^ { c } \psi ( \| y ^ { - } \| ) \\ & & - d \psi ( \| y ^ { - } \| ) \wedge d ^ { c } \| y ^ { + } \| ^ { 2 } - \| y ^ { + } \| ^ { 2 } \, d d ^ { c } \psi ( \| y ^ { - } \| ) \\ & & = 2 \psi ( \| y ^ { - } \| ) \, \sum _ { s = 1 } ^ { m } d x _ { s } ^ { + } \wedge d y _ { s } ^ { + } \,,$$
which is weakly plurisubharmonic. This implies that if the chosen handle C r is thin enough, that is if r > 0 has been chosen sufficiently small, then g B will be strictly plurisubharmonic on its neighborhood.
For large values of ‖ y -‖ 2 , the cut-off function is equal to 1 and g B agrees with f . Since it also dominates g A , the level set { G = r 2 } glues smoothly to the given
contact manifold, and it bounds a symplectic manifold
$$\widehat { W } = W \cup \left \{ ( z ^ { - }, z ^ { + }, z ^ { \circ } ) \in \mathbb { C } ^ { n } \ | \ f ( z ^ { - }, z ^ { + }, z ^ { \circ } ) \geq r ^ { 2 } \text{ and } G ( z ^ { - }, z ^ { + }, z ^ { \circ } ) \leq r ^ { 2 } \right \}$$
obtained from the given symplectic filling W by attaching to it the symplectic domain lying in our model between the level sets { f = r 2 } and { G = r 2 } , see Figures 5 and 6. Note that the boundaries of W and ̂ ) W are continuously isotopic. We also write W model for the subdomain U W ∪ ( ̂ W \ W that lies entirely in C n .
We decompose the boundary of ̂ W into three domains, which we denote by M reg , M A and M B . Here M A and M B are the parts of ∂W ̂ that lie in the level set { g A = r 2 } or { g B = r 2 } respectively, and satisfy ‖ y -‖ 2 < 1 / 2; M reg is the remaining part of the boundary of ̂ W , i.e. the part that is disjoint from the boundary of the deformed handle.
The boundary of the handle contains a deformation of the belt sphere, which we will still write as S n + m belt = { y -= 0 } even though it has edges. The cut-off function ψ vanishes on a neighborhood of S n + m belt , so that G simplifies to
$$G ( z ^ { - }, z ^ { + }, z ^ { \circ } ) \coloneqq \max \left \{ \left \| y ^ { + } \right \| ^ { 2 }, \left \| x ^ { - } \right \| ^ { 2 } + \left \| x ^ { + } \right \| ^ { 2 } + \left | z ^ { \circ } \right | ^ { 2 } \right \}.$$
It follows that S n + m belt is the boundary of the poly-disk
$$\left \{ \| \mathbf y ^ { + } \| ^ { 2 } \leq r ^ { 2 }, \, \mathbf y ^ { - } = 0 \right \} \cap \left \{ \| \mathbf x ^ { - } \| ^ { 2 } + \| \mathbf x ^ { + } \| ^ { 2 } + | z ^ { \circ } | ^ { 2 } \leq r ^ { 2 }, \mathbf y ^ { - } = 0 \right \} \cong \mathbb { D } ^ { m } \times \mathbb { D } ^ { n + 1 } \,.$$
Figure 5. We need to deform the handle to find a suitable family of Lob s in the belt sphere. The new handle will have edges. The green area in the picture represents the boundary M A , the yellow one is the boundary M B , and the grey part is M reg . The belt sphere S A ∪ S B corresponding to the deformed handle also has edges. The part S B is foliated by Lob s.
We can decompose the boundary of the poly-disk D m × D n +1 as a union of two smooth parts
$$\partial \left ( \mathbb { D } ^ { m } \times \mathbb { D } ^ { n + 1 } \right ) = \mathbb { S } ^ { m - 1 } \times \mathbb { D } ^ { n + 1 } \cup \mathbb { D } ^ { m } \times \mathbb { S } ^ { n } \,.$$
We will denote the first part of the belt sphere by
$$S _ { A } \coloneqq \left \{ y ^ { - } = 0, \, \| y ^ { + } \| ^ { 2 } = r ^ { 2 }, \, \| x ^ { - } \| ^ { 2 } + \| x ^ { + } \| ^ { 2 } + | z ^ { \circ } | ^ { 2 } \leq r ^ { 2 } \right \} \cong \mathbb { S } ^ { m - 1 } \times \mathbb { D } ^ { n + 1 } \,,$$
Figure 6. The deformed handle differs from the original one by the attachment of a symplectic cobordism. This cobordism can also be added to any other symplectic filling of the surgered contact manifold. The figure shows a cut through this cobordism.
but for now, we will be mostly interested in the second part
$$S _ { B } \coloneqq \left \{ y ^ { - } = 0, \, \| y ^ { + } \| ^ { 2 } \leq r ^ { 2 }, \, \| x ^ { - } \| ^ { 2 } + \| x ^ { + } \| ^ { 2 } + | z ^ { \circ } | ^ { 2 } = r ^ { 2 } \right \} \cong \mathbb { D } ^ { m } \times \mathbb { S } ^ { n } \,.$$
It lies in the i -convex hypersurface M B := { g B = r 2 } , whose complex tangencies are the kernel of the 1-form -( d g c B )∣ ∣ TM B . Close to { y -= 0 } , we compute
$$- d ^ { c } g _ { B } = 2 \, \sum _ { r = 1 } ^ { k } x _ { r } ^ { - } \, d y _ { r } ^ { - } + 2 \, \sum _ { s = 1 } ^ { m } x _ { s } ^ { + } \, d y _ { s } ^ { + } + 2 \left ( x ^ { \circ } \, d y ^ { \circ } - y ^ { \circ } \, d x ^ { \circ } \right )$$
which simplifies on S B ⊂ { y -= 0 } further to
$$- d ^ { c } g _ { B } = 2 \, \sum _ { s = 1 } ^ { m } x _ { s } ^ { + } \, d y _ { s } ^ { + } + 2 \left ( x ^ { \circ } \, d y ^ { \circ } - y ^ { \circ } \, d x ^ { \circ } \right ) \,.$$
The submanifold S B ∼ = D m × S n can be foliated by the n -spheres with constant y + -value. For every fixed value of y + = b + ∈ D m r we write the corresponding leaf as
$$L _ { \mathbf b ^ { + } } = \left \{ ( \mathbf y ^ { - }, \mathbf y ^ { + } ) = ( 0, \mathbf b ^ { + } ), \, \left \| \mathbf x ^ { - } \right \| ^ { 2 } + \left \| \mathbf x ^ { + } \right \| ^ { 2 } + \left | z ^ { \circ } \right | ^ { 2 } = r ^ { 2 } \right \} \,.$$
The restriction of -d g c B to each of these spheres is
$$2 \left ( x ^ { \circ } \, d y ^ { \circ } - y ^ { \circ } \, d x ^ { \circ } \right )$$
so that L b + is actually a spherical Lob in the (strictly) i -convex level set M B . The binding B b + of the Lob L b + is given by the set of points where z ◦ vanishes, i.e. B b + ∼ = S n -2 ; the pages of the open book are the fibers of the map
↦
$$\vartheta \colon L _ { \mathbf b ^ { + } } \, \langle \, B _ { \mathbf b ^ { + } } \to \mathbb { S } ^ { 1 }, \, ( \mathbf z ^ { - }, \mathbf z ^ { + }, z ^ { \circ } ) \mapsto \frac { z ^ { \circ } } { | z ^ { \circ } | } \.$$
In the following sections we will study holomorphic disks that each have boundary on one of the Lob s L b + .
## 4. The space of holomorphic disks attached to the belt sphere
We will now construct the moduli space of pseudoholomorphic disks needed for the proof of Theorem 1.2 and show that it is a smooth manifold with boundary. We continue with the setup and notation used in § 3.2 above.
Assumptions 4.1 . Choose an almost complex structure J on ̂ W with the following properties:
- · J is tamed by ω ;
- · J agrees on W model with the standard complex structure i ;
- · the unmodified domain M reg in ∂W is J -convex, and its J -complex tangencies agree with ξ .
4.1. The top stratum of the moduli space. We define ˜ ( M ̂ W,S B ; J ) as the moduli space of 'parametrized' curves ( b + , u, z 0 ) , where:
- (i) b + ∈ D m r is a point in the m -disk parametrizing the Lob s L b + ⊂ S B described in the previous section;
- (ii) u : ( D 2 , ∂ D 2 ) → ( ̂ W,L b + \ B b + ) is a J -holomorphic map which is trivial in π 2 ( ̂ W,L b + ); and
- (iii) z 0 is a marked point in the closed unit disk D 2 .
Additionally we require that
$$\left ( \vartheta \circ u \right ) \Big | _ { \partial \mathbb { D } ^ { 2 } } \colon \mathbb { S } ^ { 1 } \rightarrow \mathbb { S } ^ { 1 }$$
is a degree 1 map, that is, the boundary of each disk makes one turn around the binding of the open book.
Since L b + lies in the strictly convex hypersurface M B , the map ( ϑ ◦ u )∣ ∣ ∂ D 2 is a diffeomorphism, i.e. the disk intersects every page of the Lob L b + precisely once; see [Nie13, Corollary II.1.11]. It is very easy to deduce from this that the disks in ˜ ( M ̂ W,S B ; J ) are somewhere injective near the boundary. However, since we are not really free to choose the almost complex structure near the boundary, we need more to achieve transversality. We say that a J -holomorphic disk u is simple if it is somewhere injective in an open dense set of D 2 . For closed holomorphic curves, simple means not multiply covered, but for disks the situation is more complicated; see [Laz00]. However our situation is not the most general one, and it is possible to adapt the arguments of [MS04, Proposition 2.5.1] to prove that all disks in ˜ ( M ̂ W,S B ; J ) are simple.
## Lemma 4.2. Every disk u ∈ M ˜ ( ̂ W,S B ; J ) is simple.
Proof. We know u : D 2 → W is embedded near the boundary. Let X denote the set of points z ∈ D 2 such that either of the following is true:
- (i) D u z = 0, or
- (ii) There exists a different point z ′ ∈ D 2 such that u restricted to disjoint neighborhoods of z and z ′ has an isolated intersection u z ( ) = u z ( ′ ).
(Recall that either u z ( ) = u z ( ′ ) is an isolated intersection, or there exists neighbourhoods of z and z ′ with the same image.)
Standard local results plus the fact that u is embedded at ∂ D 2 tell us that X is a finite set of interior points. The image u ( D 2 \ X ) is then a smoothly embedded J -holomorphic submanifold of W ; in particular, it is a Riemann surface Σ with ˙
connected boundary and finitely many punctures. The inclusion of Σ into ˙ W is then a J -holomorphic embedding, and it extends over the punctures to a J -holomorphic map v : Σ → W , which is not necessarily an embedding but has only finitely many critical points and self-intersections.
At this point we don't know the topology of Σ, except that it has connected boundary. But the original map u , restricted to D 2 \ X , defines a holomorphic map to Σ, which then extends by removal of singularities to a holomorphic map ˙ ϕ : D 2 → Σ such that u = v ◦ ϕ . Given the properties of u at the boundary, ϕ must restrict to a diffeomorphism ∂ D 2 → ∂ Σ, and it maps interior to interior. So it has degree one, and is therefore biholomorphic. glyph[square]
Lemma 4.2 will allow us to use the following transversality result.
Proposition 4.3. Let ( W,J ) be an almost complex manifold, and let D m ε × L ⊂ W be a submanifold for which every slice L x := { x } × L is a totally real submanifold. For generic choices of J satisfying Assumptions 4.1, the following holds. Suppose u 0 : ( D 2 , ∂ D 2 ) → ( W,L 0 ) is any J -holomorphic map such that
- · the interior points of u 0 do not touch the boundary of W ;
- · the boundary of u 0 lies in the interior of L 0 ;
- · the disk u 0 is simple .
Let ˜ M be the space of all J -holomorphic maps
$$u \colon \left ( \mathbb { D } ^ { 2 }, \partial \mathbb { D } ^ { 2 } \right ) \rightarrow \left ( W, L _ { \mathbf x } \right )$$
for all x ∈ D m ε .
Then the space of solutions in ˜ M close to u 0 forms a smooth ball that has u 0 as its center and whose dimension is
$$\dim \mathcal { M } = \frac { 1 } { 2 } \dim W + \mu \left ( u _ { 0 } ^ { * } T W, u _ { 0 } ^ { * } T L _ { 0 } \right ) + m$$
where µ u TW,u TL ( ∗ 0 ∗ 0 0 ) denotes the Maslov index of the disk u 0 .
Proof. The result is standard if m = 0, in which case 1 2 dim W + µ u TW,u TL ( ∗ 0 ∗ 0 0 ) is the Fredholm index of the linearized Cauchy-Riemann operator on a suitable Banach space of sections of u TW ∗ 0 with totally real boundary condition; see [MS04, Section 3.2]. For m > 0, the linearized problem is the same as that of the m = 0 case, but with an extra m -dimensional space of smooth sections added to the domain in order to allow for the moving boundary condition, cf. [Wen05, § 4.5]. Thus the Fredholm index becomes larger by m . Given the corresponding enlargement of the nonlinear configuration space, the proof of transversality for generic J works as in the standard case by defining a suitable universal moduli space and applying the Sard-Smale theorem, see e.g. [MS04, Chapter 3]. glyph[square]
Recall from Section 3.2 that the family of Lob s is parametrized by a disk D m r of some fixed radius r glyph[lessmuch] 1. We define
$$\widetilde { \mathcal { M } _ { \text{int} } } ( \widehat { W }, S _ { B } ; J ) = \left \{ \left ( \mathbf b ^ { + }, u, z _ { 0 } \right ) \in \widetilde { \mathcal { M } } ( \widehat { W }, S _ { B } ; J ) \ | \ \| \mathbf b ^ { + } \| < r \right \} \,.$$
Since g B is plurisubharmonic and g A is weakly plurisubharmonic, this subspace consists of J -holomorphic disks that map the interior of D 2 to the interior of ̂ W ; see also Proposition 4.8.
Corollary 4.4. The subspace ˜ M int ( ̂ W,S B ; J ) of the parametrized moduli space is a smooth manifold with boundary, and its dimension is 2 n -k +3 , where k is the index of the surgery. Its boundary consists of triples ( b + , u, z 0 ) with z 0 ∈ ∂ D 2 .
Note that triples ( b + , u, z 0 ) with ‖ b + ‖ = r do not belong to ˜ M int ( ̂ W,S B ; J ) , and therefore are not points of its boundary.
Proof. By definition the elements of ˜ M int ( ̂ W,S B ; J ) satisfy the hypotheses of Proposition 4.3, and therefore every J -holomorphic disk u ∈ M ˜ int ( ̂ W,S B ; J ) has an open neighbourhood which is diffeomorphic to a ball of dimension 1 2 dim ̂ W + µ u TW,u TL ( ∗ ̂ ∗ b + ) + m +2 - the presence of the marked point adds 2 to the index. A simple computation shows that µ u TW,u TL ( ∗ ̂ ∗ b + ) = 2 for all ( b + , u, z 0 ), so that ˜ M int ( ̂ W,S B ; J ) is a smooth manifold with boundary of dimension
$$\dim \widetilde { \mathcal { M } _ { \text{int} } } ( \widehat { W }, S _ { B } ; J ) = n + m + 4 \.$$
(The boundary points are those with the mark point in ∂ D 2 .) Since m = n -k -1, we obtain the desired formula for the dimension. glyph[square]
In the next subsection we will analyze what happens when ‖ b + ‖ = r , and we will also show that ˜ ( M ̂ W,S B ; J ) is non-empty. To consider geometric disks instead of parametrized ones, we divide ˜ ( M ̂ W,S B ; J ) by the group of biholomorphic reparametrizations of D 2 ⊂ C . We define the moduli space of 'unparametrized' curves:
$$\mathcal { M } ( \widehat { W }, S _ { B } ; J ) = \widetilde { \mathcal { M } } ( \widehat { W }, S _ { B } ; J ) / \sim \,,$$
where ( b + , u, z 0 ) ∼ ( ˜ b + , u, z ˜ ˜ 0 ) if and only if b + = ˜ b + and there exists a transformation ϕ ∈ Aut( D 2 ) with u = ˜ u ◦ ϕ -1 and z 0 = ϕ z ( ˜ 0 ). The action of the reparametrization group Aut( D 2 ) preserves ˜ M int ( ̂ W,S B ; J ) ; we denote its quotient by M int ( ̂ W,S B ; J ) .
Proposition 4.5. The subspace M int ( ̂ W,S B ; J ) of the moduli space is a smooth manifold with boundary of dimension 2 n -k .
Proof. The map ϑ : S B \{ z ◦ = 0 } → S 1 is globally defined for all Lob s in the family. Therefore we can define the subset
$$\widetilde { \mathcal { M } } _ { 0 } ( \widehat { W }, S _ { B } ; J ) \subset \widetilde { \mathcal { M } _ { \text{int} } } ( \widehat { W }, S _ { B } ; J )$$
consisting of triples ( b + , u, z 0 ) such that
$$\vartheta ( u ( z ) ) = \begin{cases} 1 & \text{if $z=1$}, \\ i & \text{if $z=i$}, \\ - 1 & \text{if $z=-1$}. \end{cases}$$
We know that ˜ ( ̂ M 0 W,S B ; J ) is a submanifold of ˜ M int ( ̂ W,S B ; J ) because ϑ ◦ u | ∂ D 2 is a diffeomorphism and the biholomorphism group of the disk is triply transitive on ∂ D 2 . Then the subset ˜ ( ̂ M 0 W,S B ; J ) provides a global slice for the action of Aut( D 2 ) on ˜ M int ( ̂ W,S B ; J ) . glyph[square]
4.2. The Bishop disks. In this section, we want to study a certain class of disks in M ( ̂ W,S B ; J ) that lie entirely in the model neighborhood W model and that can be described explicitly.
A Bishop disk is a disk that we obtain by intersecting a z ◦ -plane in C n with constant ( z -, z + )-coordinates with the model neighborhood W model . A possible way to parametrize it is as a map
$$u \colon \left ( \mathbb { D } ^ { 2 }, \partial \mathbb { D } ^ { 2 } \right ) \rightarrow \left ( W, L _ { \mathbf b ^ { + } } \right ) \,,$$
with constant coordinates ( y -, y + ) = ( 0 b , + ), constant x -and x + -coordinates, so we write
$$u ( z ) = \left ( { \mathbf x } ^ { - } ; { \mathbf x } ^ { + } + i { \mathbf b } ^ { + } ; C z \right ) \,,$$
where C = √ r 2 -‖ x -‖ 2 -‖ x + 2 ‖ . The Bishop disks are the buds from which the moduli space will grow, and it is therefore important to establish that they are Fredholm regular, meaning that their linearized Cauchy-Riemann operators are surjective. This is ensured by the following 'automatic' transversality lemma (see [Nie13, Section III.1.3]).
Lemma 4.6. Let u : D 2 → ̂ W be a Bishop disk with image in W model and boundary mapped to the Lob L b + . Then its linearized Cauchy-Riemann operator D ( b + ,u ) , defined on suitable Banach space completions with totally real boundary condition determined by the Lob L b + , is surjective.
Corollary 4.7. The triples ( b + , u, z 0 ) where u is a Bishop disk with image in W model are regular points of the moduli space ˜ ( M ̂ W,S B ; J ) .
Proof. The relevant linearized operator is the same as D ( b + ,u ) in Lemma 4.6, except that the moving boundary condition satisfied by J -holomorphic maps in ˜ ( M ̂ W,S B ; J ) means that this domain must be enlarged by some finite-dimensional space of smooth sections, allowing the boundary to move to different Lob s in the family (see Appendix A for more details). The target of the operator remains the same, so surjectivity of D ( b + ,u ) in Lemma 4.6 immediately implies surjectivity on the enlarged domain. glyph[square]
The rest of this subsection will be concerned with the proof that the Bishop disks are the only holomorphic disks in W model .
Proposition 4.8. If a holomorphic disk
$$u \colon \left ( \mathbb { D } ^ { 2 }, \partial \mathbb { D } ^ { 2 } \right ) \rightarrow \left ( \widehat { W }, L _ { \mathfrak { b } ^ { + } } \right )$$
touches the boundary of ̂ W at an interior point of D 2 , then either it is constant or it is a multiple cover of a Bishop disk that is completely contained in S A ⊂ M A ∩ S n + m belt .
Proof. Let z 0 ∈ ˚ D 2 be a point in the interior of the disk at which u touches M A , M B , or M reg . We will obtain the desired statement by using the maximum principle; we only need to be a bit more careful compared with the standard situation, because the boundary of ̂ W is defined piecewise as a union of level sets of different plurisubharmonic functions.
Assume first that u z ( 0 ) touches M B . The function g B is not defined on the whole symplectic filling, but we may nonetheless assume that g B exists on a small
neighborhood of u z ( 0 ), hence we find an open subset U ⊂ ˚ D 2 containing z 0 such that
$$( g _ { B } \circ u ) | _ { U } \colon U \to \mathbb { R }$$
is a plurisubharmonic function having a maximum at z 0 . It follows from the maximum principle that g B ◦ u | U is constant, and due to strong convexity it even follows that the holomorphic map u | U itself must be constant. This implies that the open set U chosen above can in fact be extended to the whole disk, and u will be a constant disk.
Note that this argument also remains valid if u z ( 0 ) lies in the edge where M A and M B meet. The disk lies in the model locally in the domain with g B ≤ r 2 , and thus g B ◦ u | U still has a local maximum at z 0 ∈ ˚ , as used previously. D 2 Similarly, the argument can be used verbatim for disks that touch M reg , and this implies in fact that there are no disks at all touching M reg at interior points, because a constant disk must lie in L b + ⊂ S n + m belt , which is disjoint from M reg .
Let us now assume that the disk u touches the hypersurface M A at z 0 . Again, we find an open subset U ⊂ ˚ D 2 containing z 0 for which
$$( g _ { A } \circ u ) | _ { U } \colon U \to \mathbb { R }$$
is defined and has a maximum at z 0 . By weak plurisubharmonicity, this function must be constant.
Now it is easy to see that we can choose U to be the whole disk D 2 , because by continuity, the image of every point z ∈ U lies in ∂W ̂ . If z is an interior point of the disk, and if u z ( ) is an interior point of M A , i.e. it does not lie in M A ∩ M B , then we can extend U to a larger open domain that contains z in its interior. If z is an interior point but u z ( ) does lie in M A ∩ M B , then we know by the first part of the proof that u must be a constant map. In both cases the whole disk lies in M A .
It remains to see that a nonconstant holomorphic disk lying in M A must be a Bishop disk (or a multiple cover). We know that all coordinate functions are harmonic, and hence each of them must attain both its maximum and its minimum at a point on the boundary of the disk. The boundary of u lies in L b + ⊂ { y -= 0 } , and hence it follows that all of the y --coordinates vanish on the disk. From the Cauchy-Riemann equation, we then see that the x --coordinates of the disk will be constant.
Similarly, the y + -coordinates of the disk must all be equal to b + , because L b + ⊂ { y + = b + } , and again by the Cauchy-Riemann equation also the x + -coordinates will be constant.
The only nonconstant coordinate functions of the disk are the z ◦ -coordinate, and they span a round disk. glyph[square]
Recall that
$$B _ { \mathbf b ^ { + } } = L _ { \mathbf b ^ { + } } \cap \{ z ^ { \circ } = 0 \}$$
is the binding of the Lob L b + .
Proposition 4.9. There exists an open subset V ⊂ W model , containing B b + for every b + ∈ D m r , such that every holomorphic disk
$$u \colon ( \mathbb { D } ^ { 2 }, \partial \mathbb { D } ^ { 2 } ) \rightarrow ( \widehat { W }, L _ { \mathfrak { b } ^ { + } } )$$
in ˜ ( M ̂ W,S B ; J ) intersecting V must be a Bishop disk up to reparametrization.
## Proof. Note that
$$h ( z ^ { - }, z ^ { + }, z ^ { \circ } ) = \left \| x ^ { - } \right \| ^ { 2 } - \frac { 1 } { 2 } \left \| y ^ { - } \right \| ^ { 2 } + \left \| x ^ { + } \right \| ^ { 2 } \right \| \right \| ^ { 2 }$$
is a weakly plurisubharmonic function on W model . Its value on the binding B b + is r 2 , and it decreases along the Lob . If we choose a sufficiently small ε > 0, we can make sure that V := h -1 ( ( r 2 -ε, + ∞ ) ) ∩ W model is an open neighborhood of B b + with V ⊂ W model . It follows in fact from g B ≤ r 2 and h > r 2 -ε that
$$g _ { B } ( z ^ { - }, z ^ { + }, z ^ { \circ } ) - h ( z ^ { - }, z ^ { + }, z ^ { \circ } ) = \frac { 1 } { 2 } \left \| y ^ { - } \right \| ^ { 2 } + \psi ( \left \| y ^ { - } \right \| ) \cdot \left \| y ^ { + } \right \| ^ { 2 } + \left | z ^ { \circ } \right | ^ { 2 } < \varepsilon \,,$$
so that both the y -and the z ◦ -coordinates are small in V , and in particular we can assume that ψ = 0 on V . On the other hand,
$$\left \| { \mathbf x } ^ { - } \right \| ^ { 2 } + \left \| { \mathbf x } ^ { + } \right \| ^ { 2 } > r ^ { 2 } - \varepsilon + \frac { 1 } { 2 } \left \| { \mathbf y } ^ { - } \right \| ^ { 2 } \geq r ^ { 2 } - \varepsilon \\ \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots$$
implies that every point in V lies in an arbitrarily small neighborhood of S B .
Let now u : ( D 2 , ∂ D 2 ) → ( ̂ W,L b + ) be a holomorphic disk whose image intersects V . Assume that h ◦ u is not constant: then we can choose by Sard's theorem a slightly smaller number ε ′ < ε for which r 2 -ε ′ will be a regular value of h ◦ u , so that the subdomain
$$G \coloneqq \left \{ z \in \mathbb { D } ^ { 2 } \, \Big | \ ( h \circ u ) ( z ) \geq r ^ { 2 } - \varepsilon ^ { \prime } \right \}$$
is compact and has piecewise smooth boundary, which we denote by
$$\partial G = \partial _ { + } G \cup \partial _ { - } G \,,$$
where ∂ + G = G ∩ ∂ D 2 lies in the boundary of the unit disk, and ∂ -G lies in the interior of the unit disk. Denote the restriction
$$u | _ { G } \colon G \to W _ { \text{model} }$$
by u G .
By the maximum principle, it follows that the maximum of h ◦ u G on each component of G must lie on the boundary of that component. Clearly then the boundary of every component of G must intersect ∂ + G , because otherwise h ◦ u G would have an interior maximum on that component, so it would be equal to r 2 -ε ′ , but this contradicts the assumption that r 2 -ε ′ is a regular value.
It follows then that every component of G must intersect ∂ D 2 , and since h ◦ u G is minimal along ∂ -G , the maximum of h ◦ u G must lie at a point z 0 ∈ ∂ + G ⊂ ∂ D 2 . By the boundary point lemma, a version of the maximum principle at the boundary (see for example [Nie13, Theorem II.1.3]), the derivative of h ◦ u G at z 0 in the outward radial direction must be strictly positive. We choose polar coordinates ( r, ϕ ) on D 2 . Using the fact that u is J -holomorphic, we can write
$$\partial _ { r } ( h \circ u ) & = d h ( D u \cdot \partial _ { r } ) = d h ( D u \cdot ( - i \cdot \partial _ { \varphi } ) ) \\ & = - d h ( i \cdot D u \cdot \partial _ { \varphi } ) = - d ^ { c } h ( D u \cdot \partial _ { \varphi } )$$
but note that
$$- d ^ { c } h = \sum _ { r = 1 } ^ { k } \left ( 2 x _ { r } ^ { - } \, d y _ { r } ^ { - } + \, y _ { r } ^ { - } \, d x _ { r } ^ { - } \right ) + 2 \, \sum _ { s = 1 } ^ { m } x _ { s } ^ { + } \, d y _ { s } ^ { + } \,.$$
We obtain -d h c · Du · ∂ ϕ = 0 along the whole boundary of the disk, because the boundary of u lies in the Lob L b + , which is a subset of { y -= 0 y , + = b + } . It
follows that ∂ r ( h ◦ u ) = -d h Du c ( · ∂ ϕ ) vanishes at z 0 , and by the boundary point lemma, the disk must be contained in one of the level sets of h , so in particular it lies in V ⊂ W model .
The rest of the statement follows from standard arguments. All of the coordinate functions on W model are harmonic, hence they must attain their maxima and minima on the boundary of the disk. Since y -= 0 along ∂ D 2 , the y --coordinates of u are zero on the whole disk, and using the Cauchy-Riemann equation, we see that the x --coordinates must be constant on the disk. Similar arguments work for y + and x + , and we finally conclude that u must be a Bishop disk. glyph[square]
Proposition 4.10. Let L b + be a Lob that lies in the hypersurface M A , i.e. b + ∈ D m r has been chosen such that ‖ b + ‖ = r . Then up to parametrization, every holomorphic disk u in ˜ ( M ̂ W,S B ; J ) whose boundary lies in L b + is a Bishop disk.
Proof. Note that L b + lies in the level set of the weakly plurisubharmonic function g A : ̂ W → R . It suffices to prove that the image of u has to lie entirely in M A ⊂ W model , as this already implies the desired statement by Proposition 4.8.
Since the whole boundary u ∂ ( D 2 ) lies in M A , we can find a closed annulus G ⊂ D 2 having ∂ D 2 as one of its boundary components such that
$$( g _ { A } \circ u ) | _ { G } \colon G \to \mathbb { R }$$
is defined and everywhere weakly plurisubharmonic, and it takes its maximum along ∂ D 2 ⊂ G .
Assume first that the disk u is tangent to M A at one of its boundary points. We can apply the boundary point lemma around this point (see again [Nie13, Theorem II.1.3]) to deduce that ( g A ◦ u )∣ ∣ G has to be constant on all of G . In particular this implies that u G ( ) lies in M A , and u touches M A also with one of its interior points. Proposition 4.8 then implies that u is either constant or one of the Bishop disks.
Conversely suppose that u is everywhere transverse to M A , meaning that ∂ r ( g A ◦ u ) ( z ) is strictly positive for every z ∈ ∂ D 2 . The restriction u | G is a J -holomorphic map whose image lies in W model ; moreover, g A ◦ u ≡ r 2 on ∂ D 2 and g A ◦ u < r 2 on the inner boundary of G . Introduce on G the polar coordinates z = ρe iϕ .
Note that along ∂ D 2 , all of the y + -coordinates are constant in the ϕ -direction, because the boundary of the disk lies in the Lob L b + . Multiplying the complex coordinates z + by a suitable SO( m )-matrix (the standard complex structure i and the functions g A , g B are invariant under such a multiplication), we may assume that b + = ( r, 0 , . . . , 0). It follows that the y + 1 -coordinate of u | G has its maximum on ∂ D 2 . Note now that the x + 1 -coordinate of u | S 1 is bounded, and hence it necessarily must take a maximum at some point e iϕ 0 ∈ S 1 = ∂ D 2 , so that
$$\frac { d } { d \varphi } \Big | _ { \varphi = \varphi _ { 0 } } x _ { 1 } ^ { + } \Big ( u ( e ^ { i \varphi } ) \Big ) = 0 \.$$
Again, we can use complex multiplication to see i · ∂ ρ = ∂ ϕ , hence
$$d y _ { 1 } ^ { + } ( D u \cdot \partial _ { \rho } ) = d y _ { 1 } ^ { + } ( D u \cdot ( - i \cdot \partial _ { \varphi } ) ) = - d y _ { 1 } ^ { + } ( i \cdot D u \cdot \partial _ { \varphi } ) = - d x _ { 1 } ^ { + } ( D u \cdot \partial _ { \varphi } )$$
and in particular the radial derivative of y + 1 vanishes at e iϕ 0 , so that by the boundary point lemma, y + 1 must be constant on all of G .
Using the fact that r 2 = ∣ ∣ y + 2 1 ∣ ∣ ≤ g A ( z -, z + , z ◦ ) ≤ r 2 everywhere on G , we deduce that all of u G ( ) lies in M A . In particular, u touches M A at an interior point, which allows us to conclude the proof by applying Proposition 4.8. glyph[square]
We end this subsection with a description of the global topology of the moduli spaces ˜ ( M ̂ W,S B ; J ) and M ( ̂ W,S B ; J ) .
Proposition 4.11. The parametrized moduli space ˜ ( M ̂ W,S B ; J ) is a smooth (2 n -k + 3) -dimensional manifold with boundary and corners. Its boundary has two smooth strata, one corresponding to elements ( b + , u, z 0 ) with ‖ b + ‖ = r , and the other corresponding to elements ( b + , u, z 0 ) with | z 0 | = 1 .
The moduli space M ( ̂ W,S B ; J ) is a smooth (2 n -k ) -dimensional manifold with boundary and corners, which decomposes as a product
$$\mathcal { M } ( \widehat { W }, S _ { B } ; J ) = \Sigma \times \mathbb { D } ^ { 2 }$$
where Σ is a (non-compact) manifold with boundary.
Proof. Let ( b + , u, z 0 ) be an element of ˜ M int ( ̂ W,S B ; J ) . Since ˜ M int ( ̂ W,S B ; J ) is open in ˜ ( M ̂ W,S B ; J ) , it follows from Proposition 4.8 that the image of u does not touch ∂W ̂ with any interior point if ‖ b + ‖ < r , and that ( b + , u, z 0 ) has a neighborhood in ˜ ( M ̂ W,S B ; J ) which is diffeomorphic to an open ball ˚ D 2 n -k +3 if | z 0 | < 1, or to a half-ball ˚ D 2 n -k +1 × { z ∈ C ∣ ∣ Im z ≥ 0 } if | z 0 | = 1.
Now we consider the elements ( b + , u, z 0 ) of ˜ ( M ̂ W,S B ; J ) such that the image of u touches ∂W ̂ at an interior point. Again by Proposition 4.8 this implies ‖ b + ‖ = . r We know by Proposition 4.10 that u will be a Bishop disk up to reparametrization. Since Bishop disks are regular by Corollary 4.7, a neighborhood of ( b + , u, z ) in ˜ ( M ̂ W,S B ; J ) looks like the neighborhood of a boundary point of an ( n + m +2)manifold. A priori the boundary of the unparametrized moduli space may contain one more stratum consisting of triples ( b + , u, z 0 ) such that the image of u touches the binding B b + of the Lob L b + at a boundary point. However, in this case, the image of u would have to intersect the neighborhood V from Proposition 4.9, and therefore u would be a Bishop disk. Since Bishop disks which intersect the binding are constant and, by definition, do not belong to ˜ ( M ̂ W,S B ; J ) , the possibility that the image of u touches B b + cannot occur. (We will see later that those constant disks must be added to the compactification of the moduli space.)
By definition, the space ˜ ( M ̂ W,S B ; J ) of parametrized disks with a marked point is a trivial disk bundle. The moduli space M ( ̂ W,S B ; J ) is then a disk bundle over the moduli space of the same holomorphic curves without the marked point, and the projection map is simply the map that forgets the marked point.
In our case though, it is even true that M ( ̂ W,S B ; J ) is a trivial disk bundle. Recall that the map ϑ : S B \ { z ◦ = 0 } → S 1 is globally defined for all Lob s in the family. Hence every equivalence class [ b + , u, z ] in M ( ̂ W,S B ; J ) has a unique representative ( b + , u 0 , z 0 ), defined by fixing a parametrization of u such that
$$\vartheta ( u ( z ) ) = \begin{cases} 1 & \text{if $z=1$}, \\ i & \text{if $z=i$}, \\ - 1 & \text{if $z=-1$}, \end{cases}$$
as we did in the proof of Proposition 4.5. This choice of parametrization gives a global slice for the action of Aut( D 2 ) on ˜ ( M ̂ W,S B ; J ) , which identifies M ( ̂ W,S B ; J ) with a subset of ˜ ( M ̂ ( W,S B ; J ) . Then M ( ̂ W,S B ; J ) is a trivial disk bundle because the same is true of ˜ ̂ M W,S B ; J ) . glyph[square]
4.3. Topology of the compactified moduli space. In the previous sections, we introduced the moduli space we want to use, and we showed that all the disks intersecting certain domains of the model neighborhood W model must be Bishop disks. Our aim in this section is to study the topology of the natural compactification of that moduli space. The compactification of M ( ̂ W,S B ; J ) involves two phenomena: (1) Gromov convergence to stable nodal holomorphic disks (see e.g. [Fra08; FZ15]), and (2) degeneration to constant maps in the binding of a Lob . In order to accommodate the latter without losing the extra disk-bundle structure provided by the marked point, we shall (as in Proposition 4.5) replace M ( ̂ W,S B ; J ) by the space ˜ ( ̂ M 0 W,S B ; J ) of parametrized curves that satisfy the condition (4.1). This introduces a hint of extra book-keeping into the following statement, but the reader should keep in mind that the space we are actually interested in is always M ( ̂ W,S B ; J ) .
Proposition 4.12. Any sequence ( b + j , u j , z j ) ∈ M ˜ ( ̂ W,S B ; J ) satisfying the condition (4.1) has a subsequence that converges to a unique configuration of one of the following types:
- (1) An element of the moduli space ( b + ∞ , u ∞ , z ∞ ) ∈ M ˜ ( ̂ W,S B ; J ) , still satisfying (4.1) ;
- (2) A triple ( b + ∞ , p ∞ , z ∞ ) , where b + ∞ ∈ D m , p ∞ represents the constant map at the point p ∞ ∈ B b + ∞ and z ∞ ∈ D 2 ; or
- (3) A triple ( b + ∞ , t ∞ , z ∞ ) , where b + ∞ ∈ D m , t ∞ is a stable nodal holomorphic disk with boundary on L b + ∞ ⊂ S n + m belt , consisting of a single nonconstant disk with a tree of sphere bubbles attached, and z ∞ is a marked point on the domain of t ∞ .
Convergence in cases (1) and (2) is in the C ∞ -topology, and in case (3) it is in the sense of Gromov. If W is symplectically aspherical, then the third case does not occur.
Proof. Since the parameters b + j belong to the closed ball D m r , we can extract a first subsequence from ( b + j , u j , z j ) for which the parameters converge to some limit b + ∞ ∈ D m r . For simplicity, we still denote this subsequence by ( b + j , u j , z j ).
The usual statement of Gromov's compactness theorem for holomorphic disks (see [Fra08, § 4]) applies to sequences of unparametrized curves with fixed numbers of interior marked points and/or boundary marked points. Thus in order to apply the theorem to ( b + j , u j , z j ) ∈ M ˜ ( ̂ W,S B ; J ) , it will be convenient to observe that parametrized curves satisfying (4.1) can be identified in a canonical way with unparametrized stable nodal J -holomorphic disks carrying one interior marked point (corresponding to z j ) and three extra boundary marked points (corresponding to the points 1 , i, -1), where the latter are required to satisfy incidence conditions under the evaluation map. In this picture, smooth (i.e. non-nodal) unparametrized curves with extra boundary marked points correspond to triples ( b + j , u j , z j ) with
z j ∈ D 2 \ ∂ D 2 , and triples with z j ∈ ∂ D 2 are identified with nodal curves that consist of a nonconstant disk u j attached by a node at z j to a single constant ('ghost') disk on which the interior marked point lives.
With this identification understood, suppose the maps u j have images bounded away from the binding B b + ∞ . Then after taking a subsequence, we can assume by Gromov compactness that the corresponding sequence of unparametrized stable curves with extra boundary marked points converges in the Gromov topology to a smooth or nodal J -holomorphic disk. Note that each of the unparametrized curves has a unique parametrization for which the (ordered) set of boundary marked points is (1 , i, -1), thus if the Gromov limit is smooth, then this means z j converges to an interior point of D 2 and u j converges in C ∞ to a smooth J -holomorphic disk u ∞ . Similarly, if the nodal limit consists only of one nonconstant J -holomorphic disk u ∞ and one ghost disk containing the interior marked point, then this means that u j converges in C ∞ to u ∞ while z j converges to a point in ∂ D 2 . In all other cases, u j can be viewed as converging to a bubble tree which may include both spheres and disks, while z j converges to an interior or boundary point on one of the components.
Suppose now that the sequence ( u , z j j ) converges to a bubble tree ( t ∞ , z ∞ ). We will show that t ∞ does not contain any nonconstant disk bubble. (Since the boundary marked points are always mapped to distinct points in the image, stability then implies that with the exception of the cases interpreted above as smooth limits, t ∞ contains no disk bubbles at all.) Suppose on the contrary that the sequence u j bubbles a nonconstant disk v with boundary on the Lob L b + ∞ . The points 1 , -1 , i divide ∂ D 2 into three segments, one of which is necessarily disjoint from the bubbling region. The fact that for each j the function ϑ ◦ u j | ∂ D 2 is a diffeomorphism then implies that ϑ ◦ v | ∂ D 2 is not surjective. Thus the boundary of v is somewhere tangent a page of the Lob L b + ∞ , but it follows from a standard argument using the boundary point lemma [Nie13, Theorem II.1.3] that the disk v cannot exist. We conclude that t ∞ is a bubble tree containing only holomorphic spheres. This is case (3).
If there is a sequence w j ∈ D 2 such that u j ( w j ) approaches the binding, then the maps u j are Bishop disks for j large enough because, sooner or later, the images of the u j will intersect the domain V as in Proposition 4.9 nontrivially. This implies that the limit u ∞ is the constant map at a point p ∞ ∈ B b + ∞ , and we have case (2). glyph[square]
Using the natural identification of M ( ̂ W,S B ; J ) with the space of parametrized curves satisfying (4.1), we can now compactify M ( ̂ W,S B ; J ) by adding the limiting configurations described in Proposition 4.12. We will denote this compactified moduli space by M ( ̂ W,S B ; J ) . Its 'boundary'
$$\partial \overline { \mathcal { M } } ( \widehat { W }, S _ { B } ; J ) \subset \overline { \mathcal { M } } ( \widehat { W }, S _ { B } ; J )$$
can be defined naturally as the set of equivalence classes [( b + , u, z )] for which either b + ∈ ∂ D 2 r , z ∈ ∂ D 2 (including cases where the domain of u contains sphere bubbles), or u is a constant map into the binding of a Lob . The compactification can also be decomposed naturally into two disjoint pieces,
$$\overline { \mathcal { M } } ( \widehat { W }, S _ { B } ; J ) = \mathcal { M } _ { s m o o t h } \cup \mathcal { M } _ { b u b b l e },$$
defined as the subsets consisting of non-nodal and nodal curves respectively. We define
$$\partial \mathcal { M } _ { \text{smooth} } \coloneqq \partial \overline { \mathcal { M } } ( \widehat { W }, S _ { B } ; J ) \cap \mathcal { M } _ { \text{smooth} } \,.$$
The next proposition describes the topology of the compactified moduli space. We refer to [MS04, § 6.5] for general facts about pseudocycles, and [Sch99; Zin08] for the fact that pseudocycles up to bordism can be identified with integral homology classes.
Proposition 4.13. Let ( W,ω ) be a symplectic filling of a contact (2 n -1) -manifold ( M,ξ ) . Suppose that ( M,ξ ) has been obtained by a surgery of index k ≤ n -1 from another contact manifold, and that S n + m belt is the corresponding belt sphere (with n = k + m + 1 ). Deform W as described in Section 3.2 to ( ̂ W,ω ) , and let M ( ̂ W,S B ; J ) be the compactification of the moduli space M ( ̂ W,S B ; J ) of disks attached to S B ⊂ S n + m belt . Then ev: M ( ̂ W,S B ; J ) → ̂ W extends to a continuous map
$$\overline { \text{ev} } \colon \left ( \overline { \mathcal { M } } ( \widehat { W }, S _ { B } ; J ), \partial \overline { \mathcal { M } } ( \widehat { W }, S _ { B } ; J ) \right ) \to \left ( \widehat { W }, S _ { \text{belt} } ^ { n + m } \right ).$$
Moreover:
- (a) If ( W,ω ) is semipositive, then
$$\ e v | _ { \partial \mathcal { M } _ { s m o o t h } } \colon \partial \mathcal { M } _ { s m o o t h } \to \dot { W }$$
defines an ( n + m ) -dimensional pseudocycle in ̂ W representing the homology class ± [ S n + m belt ] ∈ H n + m ( ̂ W ; Z ) , and
$$\text{ev} | _ { \mathcal { M } _ { \text{smooth} } } \colon \mathcal { M } _ { \text{smooth} } \rightarrow \tilde { W }$$
defines a bordism of the above pseudocycle to the empty ( n + m ) -dimensional pseudocycle in ̂ W .
- (b) If ( W,ω ) is symplectically aspherical, then M ( ̂ W,S B ; J ) is homeomorphic to a manifold with boundary and corners of the form
$$\overline { \Sigma } \times \mathbb { D } ^ { 2 } \,,$$
where Σ is a smooth, compact, connected and oriented ( n + m -1) -manifold with boundary and corners, whose boundary is homeomorphic to S n + m -2 . Furthermore,
$$\text{ev} | _ { \partial \overline { \mathcal { M } } \left ( \widehat { W }, S _ { B } ; J \right ) } \colon \partial \overline { \mathcal { M } } \left ( \widehat { W }, S _ { B } ; J \right ) \rightarrow S _ { \text{belt} } ^ { n + m }$$
is a map of degree ± 1 .
Proof. Let us describe the natural topology on M ( ̂ W,S B ; J ) , using again the identification of M ( ̂ W,S B ; J ) with the slice in ˜ ( M ̂ W,S B ; J ) defined via the conditions (4.1). The boundary of the uncompactified space M ( ̂ W,S B ; J ) consists of holomorphic disks whose marked points lie in ∂ D 2 , together with Bishop disks with boundary on a Lob L b + with b + ∈ ∂ D 2 r . Proposition 4.12, provides a description of the two additional limit objects we need to consider. If ( b + , p ∞ , z ) is one of the constant disks appearing in case (2) of Proposition 4.12, then it follows from Proposition 4.9 that it is surrounded only by Bishop disks. Using the parametrization of the Bishop disks given at the beginning of Section 4.2-in this description the constant disks are obtained by choosing C = 0-we can add the constant disks
( b + , p ∞ , z ) to the chosen slice in ˜ ( M ̂ W,S B ; J ) , and give it a smooth structure that agrees with the one induced by C ∞ -convergence of maps. Attaching the constant disks in this way corresponds to adding boundary points to the global slice. Defining a smooth structure on M ( ̂ W,S B ; J ) in this way, it is straightforward to see that the evaluation map extends smoothly to the constant disks in M ( ̂ W,S B ; J ) .
The other singular points we need to consider in M ( ̂ W,S B ; J ) are bubble trees, each consisting of one holomorphic disk and several holomorphic spheres. If ( W,ω ) is symplectically aspherical as in case (b), it does not contain any holomorphic spheres, and hence no bubbles can appear. In this case, the compactified moduli space M ( ̂ W,S B ; J ) will be diffeomorphic to
## Σ × D 2
according to Propositon 4.11, where Σ is a smooth compact manifold with boundary and corners. If we are in case (a), then bubbling of spheres may occur, but standard index counting arguments using the semipositivity assumption imply that such bubbling is a 'codimension 2 phenomenon'. The restriction of ev to ∂ M smooth is then a pseudocycle, and the restriction to M smooth is a bordism of this pseudocycle to the trivial one.
In the absence of bubbling, M ( ̂ W,S B ; J ) is a trivial disk bundle over the compact base manifold Σ, whose boundary consists of Bishop disks sitting on boundary Lob s and/or collapsing into the binding. The boundary Lob s are parametrized by b + ∈ ∂ D m r ∼ = S m -1 , and there is precisely one (unparametrized) Bishop disk going through every point of the page of the Legendrian open book of L b + , hence we conclude that the first disks can be parametrized by S m -1 × D n -1 . The binding of a Lob L b + , on the other hand, is diffeomorphic to S n -2 , and since there is a D m r -worth of Lob s, we conclude that M ( ̂ W,S B ; J ) contains a family of constant disks that is parametrized by D m r × S n -2 . These two parts meet at their boundaries and form a (topological) manifold homeomorphic to S n + m -2 , as claimed.
Finally, we observe that the restriction of ev to ∂ M ( ̂ W,S B ; J ) is always bijective on some subset consisting of Bishop disks, so it is a map of degree ± 1 onto S n + m belt whenever ∂ M ( ̂ W,S B ; J ) is a topological manifold. More generally, this implies that the pseudocycle ev | ∂ M smooth : ∂ M smooth → S n + m belt represents a generator of H n + m ( S n + m belt ; Z ) whenever it is well defined.
The orientability of M ( ̂ W,S B ; J ) is shown in Appendix A. glyph[square]
Proof of Theorem 1.2. The proof of statement (a) follows directly from part (a) of Proposition 4.13, using the natural identification between singular homology classes and bordism classes of pseudocycles, see [Sch99; Zin08].
Statement (b) can be obtained by using part (b) of Proposition 4.13. Since M ( ̂ W,S B ; J ) is diffeomorphic to a trivial disk bundle, we can apply Proposition 5.1 in the general situation, or Propositions 5.2 and 5.3 respectively, when n + m = 3 or n + m = 4. This implies that S n + m belt is the trivial element in the oriented bordism group Ω SO n + m ( W ), and it is even contractible in W , if n + m = 3 or 4 as claimed. glyph[square]
## 5. Surgery on moduli spaces
If bubbling can be ruled out in the proof of Theorem 1.2 given in the previous section, we can use topological results to conclude that the belt sphere is not only nullhomologous but null-bordant in Ω SO ∗ ( ̂ W ), and in cases where dim M ( ̂ W,S B ; J ) ≤ 5, it is even null-homotopic. The idea in both cases is to attach handles to the moduli space and extend the evaluation map accordingly so that we obtain a new space together with a map into ̂ W which will be topologically simpler than the original moduli space. Our argument for this will make essential use of the fact that the moduli space is naturally a trivial disk bundle.
Note that if Σ is a compact oriented k -manifold with boundary, then using handle attachments to turn Σ × D 2 into a ball cannot succeed unless Σ × D 2 admits an embedding into R k +2 , which cannot always be guaranteed, i.e. in general there are topological obstructions to the applicability of this technique to obtain contractibility of the belt sphere. We will show that these can be overcome if dim Σ ≤ 3.
Proposition 5.1. Let W be a compact manifold possibly with boundary, and let S ⊂ W be an embedded ( k + 1) -sphere. Assume that Σ is a compact connected orientable k -manifold with non-empty boundary. Let X be Σ × D 2 , and let
$$f \colon ( X, \partial X ) \rightarrow ( W, S )$$
be a continuous map, whose restriction to the boundary
$$f | _ { \partial X } \colon \partial X \rightarrow S$$
is of degree 1 . Then it follows that S is null-bordant in Ω SO k +1 ( W ) .
Proof. We will assume k ≥ 2, since for k ≤ 1, Σ is either a point or a closed interval. The manifold Σ has a handle decomposition that consists of a k -disk with finitely many handles of index < k attached to it (attached successively in order of their indices). Since the product of a k -dimensional handle with D 2 is a ( k +2)-dimensional handle of the same index, X = Σ × D 2 is built up by attaching ( k +2)-dimensional handles of index < k . For each q = 0 , . . . , k -1, let X ( q ) ⊂ X denote the union of all the handles up to index q , so X = X ( k -1) .
For every 1-handle in X (1) we can find a closed curve in the boundary of X (1) by pushing the core of the handle into ∂X (1) , and connecting the end points with a path in ∂X (1) that does not intersect any other 1-handle. This is possible, because X (0) consists of a unique 0-handle that in particular is connected. These curves intersect the belt sphere of the corresponding handle exactly once. Moreover since dim ∂X (1) = k +1, there is enough space to assume that the loops corresponding to different 1-handles are disjoint from each other and also disjoint from any of the attaching circles of the 2-handles needed to obtain X (2) . The loops thus embed into ∂X (2) , and we can repeat this reasoning to see that they also embed into ∂X (3) and so on up to ∂X .
Standard Morse theory implies that a q -handle can be canceled out by attaching a ( q +1)-handle along an embedded q -sphere that intersects the belt sphere of the q -handle exactly once. It is thus possible to convert X (1) into a ball by attaching 2handles, each corresponding to one of the 1-handles. Since the loops also embed into ∂X , we may equally well attach the 2-handles to X , obtaining in this way a compact connected orientable ( k +2)-manifold X ′ that admits a handle decomposition with exactly one 0-handle, and no handles of index 1 or of index ≥ k .
The map f : X → W can be extended to the newly added 2-handles because ∂X , and hence also the attaching curves, are mapped to S , which is simply connected. We can thus construct a continuous map
$$f ^ { \prime } \colon ( X ^ { \prime }, \partial X ^ { \prime } ) \to ( W, S ) \,.$$
The restriction of f ′ to the boundary is still of degree 1, because the image of the 2-handles can be assumed to be a thin set in the ( k +1)-sphere S .
To finish the proof, consider the double of X ′ , obtained by gluing a copy of X ′ with reversed orientation to itself along its boundary. One can decompose the double into handles such that for each q -handle in the original X ′ , there is a corresponding ( k +2 -q )-handle in the second copy (think of the two copies of X ′ as carrying Morse functions f and c -f for some constant c ∈ R ). Our handle decomposition of the double therefore has exactly one handle of index k + 2 and none of index k +1. Let ̂ X denote the result of removing the ( k +2)-handle. Then ̂ X is a compact connected orientable ( k +2)-manifold obtained from X ′ by attaching additional handles of various indices 3 , . . . , k , and ∂X ̂ ∼ = S k +1 . Since π q ( S ) = 0 for all q = 2 , . . . , k , the map f ′ extends from X ′ to a continuous map
$$\widehat { f } \colon ( \widehat { X }, \partial \widehat { X } ) \rightarrow ( W, S ) \,,$$
mapping all of the additional handles into S . Note in particular that the restriction of ̂ f to ∂X ̂ is a degree 1 map to S , because all the handles added to X ′ are of index lower than dim S .
The restriction ̂ ∣ f ∣ ∣ ∂X ̂ is homotopic to a diffeomorphism between the two ( k +1)spheres ∂X ̂ and S , and it follows that S is null-bordant in Ω SO k +1 ( W ). glyph[square]
Proposition 5.2. Let W be a compact manifold possibly with boundary, and let S ⊂ W be an embedded 3 -sphere. Assume that Σ is a compact connected orientable surface with non-empty boundary. Let X be Σ × D 2 , and let
$$f \colon ( X, \partial X ) \rightarrow ( W, S )$$
be a continuous map, whose restriction to the boundary
$$f | _ { \partial X } \colon \partial X \rightarrow S$$
is of degree 1 . Then it follows that S is contractible in W .
Proof. The proof is a special case of the argument used for Proposition 5.1. The difference is that X consists only of a 4-disk and 1-handles, thus after passing to X ′ by attaching 2-handles, it already follows that X ′ is diffeomorphic to a 4-disk D 4 . We then obtain a continuous map
$$f ^ { \prime } \colon \left ( \mathbb { D } ^ { 4 }, \partial \mathbb { D } ^ { 4 } \right ) \rightarrow \left ( W, S \right )$$
whose restriction to the boundary is still of degree 1, and is therefore homotopic to a homeomorphism ∂ D 4 → S . glyph[square]
Proposition 5.3. Let W be a compact manifold, possibly with boundary, and let S ⊂ W be an embedded 4 -sphere. Let Σ be a compact connected orientable 3 -manifold with non-empty boundary ∂ Σ = ∼ S 2 . Assume that X is Σ × D 2 , and that
$$f \colon ( X, \partial X ) \rightarrow ( W, S )$$
is a continuous map, whose restriction to the boundary
$$f | _ { \partial X } \colon \partial X \rightarrow S$$
is of degree 1 . Then it follows that S is contractible in W .
Proof. The manifold Σ is an orientable 3-manifold minus a ball. It admits a handle decomposition given by a 3-ball with 1- and 2-handles attached, and we may assume that Σ has been obtained by attaching first the 1-handles and then the 2-handles. As in the proof of Proposition 5.1, it follows that the manifold X is built up by first attaching 1- and then 2-handles to a 5-ball D 5 . Denote by X (1) the intermediate space consisting only of the 5-ball and the 1-handles.
It is easy to find for every 1-handle an embedded loop in ∂X (1) that intersects the belt sphere of the handle exactly once. For dimensional reasons, these loops will be generically disjoint from each other, but they will also generically not intersect any of the attaching circles of the 2-handles. We can cancel all 1-handles of X (1) by attaching 2-handles along the chosen loops. The chosen loops also embed into ∂X , hence we can also kill all 1-handles by attaching 2-handles to X . Note that we could get rid of the 1-handles without choosing a particular framing when attaching the 2-handles; however X is parallelizable (as is any oriented 3-manifold), and using Lemma 5.4 below, we attach the 2-handles in such a way that the resulting manifold is also parallelizable. Since the image f ( ∂X ) lies in the sphere S , we can extend f to the additional 2-handles without changing the degree of f | ∂X .
After the previous step, we will assume that X is a 5-manifold with trivial tangent bundle that has been obtained by gluing 2-handles to the 5-ball. Every embedding of S 1 into S 4 = ∂ D 5 is isotopic to a standard one for dimensional reasons (see [Hae62]), and it follows that the 2-handles are all attached along unknots. Note also that these unknots are unlinked since we may shrink the first unknot into an arbitrarily small ball, so that the other loops will bound embedded disks that are disjoint from this ball.
We may therefore assume that X is the boundary sum of a finite collection of 5-manifolds, each consisting of a 5-ball with a single 2-handle attached along an unknot. The only invariant of each such manifold is the framing of the 2-handle. It is given by a loop in SO(3), which means there are only two choices, corresponding to the two elements of π 1 (SO(3)). In fact, each of these manifolds is diffeomorphic to either the trivial rank 3 bundle over S 2 or the twisted one, S 2 ˜ × D 3 . The total space of the twisted one is not parallelizable: it suffices to study T ( S 2 ˜ × D 3 )∣ ∣ S 2 ˜ ×{ } 0 which is obtained by clutching two copies of C ⊕ R 3 over two disks together. The gluing map is e 2 iφ ⊕ ψ , where ψ is the nontrivial loop in π 1 (SO(3)), but since this is the nontrivial element of π 1 (SO(5)), the bundle is not trivial. It follows that X is the boundary connected sum of copies of S 2 × D 3 .
We can then also kill the 2-handles by attaching 3-handles, and the map f extends to this new manifold. This proves that S is homotopically trivial in the filling. glyph[square]
The following lemma was used above in the proof of Proposition 5.3.
Lemma 5.4. Let X be a compact parallelizable n -manifold with boundary, and let γ be an embedded loop in ∂X . Assume n ≥ 5 . Then one can choose a framing of γ such that the manifold
$$X \cup _ { \gamma } H _ { 2 }$$
obtained by attaching a 2 -handle H 2 along γ is also parallelizable.
Proof. A framing of γ is an oriented trivialization of the normal bundle ν γ ( ) of γ in ∂X . Given one framing, any other one can be obtained by multiplying the first one in each fiber with a matrix in GL + ( n -2), i.e. the second framing can be represented with respect to the first one by a map S 1 → GL ( + n -2). We are only interested in framings up to homotopy, hence it follows that all framings are classified by π 1 ( GL ( + n -2) ) , and since GL + ( n -2) glyph[similarequal] SO( n -2), there are only two possible choices.
Choose now a trivialization of TX . Such a trivialization allows us to identify TX | γ with S 1 × R n . Any other trivialization of TX | γ can be represented with respect to the first one by a map S 1 → GL ( ), that is, up to homotopy there are also + n only two trivializations of TX | γ , corresponding to the elements of π 1 (SO( n )). In particular, any framing ( e , . . . , e 1 n -2 ) of ν γ ( ) extends to a basis ( f , f 1 2 , e 1 , . . . , e n -2 ) of TX | γ , where the vector fields f 1 and f 2 are given by
$$f _ { 1 } & = \vec { n } \, \cos \phi - \dot { \gamma } \, \sin \phi \\ f _ { 2 } & = \vec { n } \, \sin \phi + \dot { \gamma } \, \cos \phi$$
where φ parametrizes γ . Here glyph[vector] n denotes the outward normal vector field to the boundary ∂X , and ˙ γ is the tangent vector field to the loop γ . If this basis is not homotopic to the given trivialization of TX | γ , it suffices to choose instead
$$\left ( f _ { 1 }, f _ { 2 }, e _ { 1 } \cos \phi - e _ { 2 } \sin \phi, e _ { 1 } \sin \phi + e _ { 2 } \cos \phi, e _ { 3 }, \dots, e _ { n - 2 } \right ) \,,$$
which corresponds to the second framing of γ , but also to the other homotopy class of possible trivializations of TX | γ . It is thus possible to homotope the trivialization of TX into one that coincides close to γ with ( f , f 1 2 , e 1 , . . . , e n -2 ), where ( e , . . . , e 1 n -2 ) is a framing of γ .
On H 2 = D 2 × D n -2 with coordinates ( x, y ; z ) ∈ D 2 × D n -2 , the attaching circle { x 2 + y 2 = 1 , z = 0 } has the obvious framing ( ∂ , ∂ u v , ∂ z 1 , . . . , ∂ z n -2 ) . If we glue H 2 to X with the chosen framing, then the trivialization ( f , f 1 2 , e 1 , . . . , e n -2 ) extends to ( ∂ , ∂ u v , ∂ z 1 , . . . , ∂ z n -2 ) , so the manifold X ∪ γ H 2 has trivial tangent bundle, as desired. glyph[square]
## 6. Contact structures that are not contact connected sums
In this section we prove Theorem 1.3. The construction we are going to use is inspired by a similar one in [BCS15], though we do not need the full strength of that paper.
Let M be a closed (2 n -1)-dimensional manifold that admits an almost contact structure and that has a handle decomposition with a single handle of index 0, a single one of index 2 n -1, and otherwise only handles of indices n -1 and n . We assume also that M is not a homotopy sphere, which by the Hurewicz theorem implies that it must have nontrivial homology in dimension n -1 or n . Possible examples include the unit cotangent bundle of S n , and S n -1 × S n ; the first carries a canonical contact structure, and the second is easily seen to be almost contact since it is stably parallelizable.
Remove a small open disk D from M and denote the resulting manifold by M ∗ . The product manifold W = M ∗ × -[ 1 1] is compact and has boundary and corners, , and after smoothing, its boundary
$$\partial W = M ^ { * } \times \{ - 1 \} \cup \left ( \partial M ^ { * } \times [ - 1, 1 ] \right ) \cup M ^ { * } \times \{ + 1 \}$$
is diffeomorphic to M #( -M ).
Now we can proceed with the proof of Theorem 1.3.
Proof of Theorem 1.3. By assumption, the manifold M ∗ admits a Morse function with outward pointing gradient at the boundary and critical points of index at most n , and the same is therefore true of W . Moreover, any almost contact structure Ξ on M induces an almost complex structure on W , thus by a well-known theorem of Eliashberg [Eli90b], there is a Stein structure whose complex structure is homotopic to the given almost complex structure. The boundary ∂W ∼ = M #( -M ) inherits from this Stein structure a contact structure ξ which is homotopic to the almost contact structure Ξ # Ξ. Note that the belt sphere of the connected sum (i.e. the center of the 'neck' in M #( -M )) is
$$S \coloneqq \partial M ^ { * } \times \{ 0 \} \subset \partial W \.$$
Arguing by contradiction, suppose now that ξ 1 and ξ 2 are positive contact structures on M and -M respectively such that ξ 1 # ξ 2 is isotopic to ξ . Then after a deformation of the Stein structure near ∂W and hence an isotopy of ξ , we can assume ξ in a neighborhood of S is contactomorphic to the contact structure on a neighborhood of the belt sphere of an index one Weinstein handle. According to Proposition 4.13 there is a compact (2 n -1)-dimensional moduli space M∼ × = Σ D 2 and an evaluation map ev: ( M M → , ∂ ) ( W,S ) such that
- (1) ev | ∂ M : ∂ M→ S has degree one, and
- (2) Σ is a compact orientable (2 n -3)-manifold with non-empty boundary.
Moreover, ev is a diffeomorphism on some open subset.
Consider the projection p : W = M ∗ × -[ 1 1] , → M ∗ , which maps S to ∂M ∗ , and denote by f : M→ M ∗ the composition f = p ◦ ev. It is easy to check that f : M → M ∗ now satisfies the conditions of Lemma 1.11. This implies that M ∗ has vanishing homology in positive degrees, and is thus a contradiction. glyph[square]
On the other hand, note that there is no homotopical obstruction to decomposing ( M #( -M ,ξ ) ) , because ξ is, by construction, homotopic to Ξ # Ξ.
Remark 6.1 . A similar argument can be used to find examples of Stein fillable contact structures that are homotopic (through almost contact structures) but not isotopic to contact structures obtained via subcritical surgery of arbitrary index k = 1 , . . . , n -1. The above is the k = 1 case of this result.
## 7. The Weinstein conjecture for subcritical surgeries
We will now prove Theorem 1.4, the existence of contractible Reeb orbits for certain contact manifolds ( M ,ξ ′ ′ ) obtained by subcritical surgery.
Under either of the first two conditions stated in the theorem, the proof is a trivial modification of the proof of Theorem 1.2, following [Hof93]. Suppose α is the contact form for which we'd like to find a contractible Reeb orbit, and let α ′ denote a second contact form that matches the one given in our Weinstein surgery model near the belt sphere S 2 n - -k 1 belt . After rescaling α , we can find an exact symplectic structure on R × M ′ that matches d e α ( t ) on ( -∞ -, 1] × M ′ and d e α ( t ′ ) on [ -1 2 / , ∞ × ) M ′ . We then choose a compatible almost complex structure and, as in Theorem 1.2, study the moduli space of holomorphic disks in R × M ′ with boundary in the Lob s obtained by deformation from { 0 } × S 2 n - -k 1 belt ⊂ R × M ′ . If
α admits no contractible Reeb orbits, then bubbling is impossible, so the proof of Theorem 1.2 shows that S 2 n - -k 1 belt will be null-bordant in R × M ′ , and thus also in M ′ . If n = 3, or n = 4 with k = 3, it also shows that S 2 n - -k 1 belt is trivial in π 2 n - -k 1 ( R × M ′ ) = π 2 n - -k 1 ( M ′ ).
It remains to handle the third condition in Theorem 1.4, which specifically concerns contact connected sums in dimension five. The above argument shows that in this situation, if there is no contractible Reeb orbit, then the belt sphere must be nullhomotopic. But the following theorem of Ruberman [Rub97] says that this can only happen in the cases excluded by the third condition.
Theorem 7.1 (Ruberman) . Let M be a closed oriented manifold, and suppose S is an embedded codimension 1 sphere that is nullhomotopic. Then either S is the boundary of a homotopy-ball embedded in M , or M is the connected sum
$$M = N _ { 0 } \# N _ { 1 }$$
of two rational homology spheres N 0 and N 1 , one of which is simply connected, while the other has finite fundamental group.
The proof of Theorem 1.4 is thus complete.
## Appendix A. Orientability of the moduli spaces
In this appendix we prove that the moduli spaces used in this paper are orientable. Let us fix some notation which will be used in the proof. Fix a real number p > 2. Let B denote the space of pairs ( b + , u ) where:
- · b + ∈ D m r with ‖ b + ‖ < r ;
- · u : D 2 → W is a map of class W 1 ,p such that u ∂ ( D 2 ) ⊂ L b + \ B b + where L b + denotes the Lob indexed by b + and B b + its binding; and
- · ϑ ◦ u | ∂ D 2 has degree one, where ϑ : L b + \ B b + → S 1 is the fibration of the Lob .
Of course the information about b + is already contained in u , and b + only serves for book-keeping. We denote by B b + the fibers of the projection p : B → D m r , i.e. p ( b + , u ) = b + . Then B b + consists of the maps u ∈ B such that u ∂ ( D 2 ) ⊂ L b + \ B b + .
The linearized Cauchy-Riemann operator at ( b + , u ) ∈ B (or to be more precise, the vertical differential of the nonlinear Cauchy-Riemann operator, as defined in [MS04]), will be denoted ˜ D ( b + ,u ) . Recall that this depends on a choice of connection on W , though it is independent of this choice whenever u is J -holomorphic. We define
$$\det \widetilde { D } _ { ( \mathbf b ^ { + }, u ) } = \Lambda ^ { t o p } \ker \widetilde { D } _ { ( \mathbf b ^ { + }, u ) } \otimes \Lambda ^ { t o p } ( \text{ocker} \, \widetilde { D } _ { ( \mathbf b ^ { + }, u ) } ) ^ { * } \,.$$
The determinant bundle D → B is the real rank-one bundle whose fiber at ( b + , u ) is det ˜ D ( b + ,u ) . In order to prove that the moduli space ˜ M int ( ̂ W,S B ; J ) is orientable, it suffices to show that D → B is trivial.
To better understand the determinant bundle we take a closer look at the linearized Cauchy-Riemann operator. The tangent space T ( b + ,u ) B consists of sections ξ ∈ W 1 ,p ( u TW ∗ ̂ ) such that, for all z ∈ ∂ D 2 , they satisfy ξ z ( ) ∈ T u z ( ) S n + m belt and moreover the projection of ξ z ( ) to T b + D m r is independent of z ∈ ∂ D 2 . It contains the subspace T u B b + ⊂ T ( b + ,u ) B which is defined as
$$T _ { u } \mathcal { B } _ { \mathbf b ^ { + } } = \left \{ \xi \in W ^ { 1, p } ( u ^ { * } T \widehat { W } ) \ | \ \xi ( z ) \in T _ { u ( z ) } L _ { \mathbf b ^ { + } } \, \forall z \in \partial \mathbb { D } ^ { 2 } \right \}$$
and therefore we can identify
$$T _ { ( \mathfrak { b } ^ { + }, u ) } \mathcal { B } \cong T _ { u } \mathcal { B } _ { \mathfrak { b } ^ { + } } \oplus T _ { \mathfrak { b } ^ { + } } \mathbb { D } _ { r } ^ { m } \,.$$
The tangent spaces T u B b + are the fibers of a vector bundles over B which we will denote T vert B and the decomposition (A.1) globalizes to a bundle isomorphism
$$T \mathcal { B } \cong T ^ { \text{vert} } \mathcal { B } \oplus \mathfrak { p } ^ { * } T \mathbb { D } _ { r } ^ { m } \.$$
Although the above isomorphism is not canonical, its homotopy class is.
We denote the restriction of ˜ D ( b + ,u ) to T u B b + by D ( b + ,u ) . If we write the elements of T ( b + ,u ) B as pairs ( ξ, v ) ∈ T u B b + ⊕ T b + D m r using the identification in Equation (A.1), we can decompose D + as
˜ ( b ,u )
$$D _ { ( \mathbf b ^ { + }, u ) } ( \xi, v ) = D _ { ( \mathbf b ^ { + }, u ) } ( \xi ) + K _ { ( \mathbf b ^ { + }, u ) } ( v ) \.$$
The operator D ( b + ,u ) is the linearization at u of the nonlinear Cauchy-Riemann operator defined on B b + , and therefore it is a linear Cauchy-Riemann type operator.
Let D → B ′ the real line bundle whose fiber at ( b + , u ) is det D ( b + ,u ) . Since the determinant line bundles of homotopic families of Fredholm operators are isomorphic, we obtain an isomorphism
$$\mathcal { D } \cong \mathcal { D } ^ { \prime } \otimes \mathfrak { p } ^ { * } \Lambda ^ { m } T \mathbb { D } _ { r } ^ { m }$$
by homotoping the operators K ( b + ,u ) to the zero operator via a linear homotopy. Note that this defines a homotopy of families of Fredholm operators because the operators K ( b + ,u ) are defined on a finite dimensional space. By the isomorphism (A.2), the triviality of D is equivalent to the triviality of D ′ , so from now on we will concentrate on this second bundle.
↦
Triviality of a rank-one real line bundle can be checked on loops. Thus let ( b + · , u · ): S 1 → B be a loop in B , i.e. θ → ( b + θ , u θ ). From a different point of view we have a map ˜: u S 1 × D 2 → W defined as ˜( u θ, z ) = u θ ( z ). We define the vector bundle T vert S n + m belt such that T vert p S n + m belt = T L p b + ( p ) , where L b + ( p ) denotes the Lob containing p . We define a complex vector bundle E → S 1 × D 2 by E = ˜ u TW ∗ ⊕ C and a real vector subbundle F → S 1 × ∂ D 2 of E | S 1 × ∂ D 2 by F = ( ˜ u | S 1 × ∂ D 2 ) ∗ T vert S n + m belt ⊕ R . Here C and R denote the trivial complex and real line bundle, respectively.
We denote by E θ the restriction of E to { } × θ D 2 , by F θ the restriction of F to { } × θ D 2 and by Γ( E , F θ θ ) the sections of E θ which take values in F θ along ∂ D 2 . Similarly, let Γ( C R , ) denote the sections of the trivial line bundle C over D 2 with real values at ∂ D 2 , and D 0 the standard Cauchy-Riemann operator acting on Γ( C R , ). We consider the family of linear Cauchy-Riemann type operators D + θ = D ′ u θ ⊕ D 0 acting on Γ( E , F θ θ ). This family gives rise to a determinant line bundle D + → S 1 . Since det D 0 = R , we have D + ∼ = u ∗ · D ′ . Therefore studying the orientability of D + is equivalent to studying the orientability of the moduli space.
Being spheres, the Lob s are stably parallelizable and the parameter space D m r is contractible, so T vert S n + m belt is stably trivial. An orthonormal trivialization of T vert S n + m belt ⊕ R can be pulled back to an orthonormal trivialization ν 0 of F . We can also regard ν 0 as a unitary trivialization of E | S 1 × ∂ D 2 , because F is a Lagrangian subbundle of E | S 1 × ∂ D 2 . However ν 0 does not extend to a unitary trivialization of E . In fact it does not extend to the meridian disks of S 1 × D 2 , because ( E , F θ θ ) has Maslov index two. We choose a map A ∂ : D 2 → U n ( ) such that the trivialization ν defined as ν θ, z ( ) = A z ( ) -1 ν 0 ( θ, z ) extends to a trivialization of E over any meridian
disk. (Of course the new trivialization is no longer an orthonormal trivialization of F .) We extend ν to a trivialization of E on a regular neighborhood of ( S 1 × ∂ D 2 ) ∪ ( { θ 0 } × D 2 ) for a fixed θ 0 ∈ S 1 . The complement of this neighborhood in S 2 × D 2 is a ball. We can extend ν inside this ball because π 2 ( U n ( )) = 0.
↦
Then ν defines an isomorphism ( E,F ) = ( ∼ C n , F ′ ), where F ′ ( θ,z ) = A z ( ) R n . The operators D + θ become D + 0 + a θ , where D + 0 is the standard Cauchy-Riemann operator on Γ( C n , F ′ ) and a θ ∈ Ω 0 1 , ( T D 2 , E ). Since a θ belongs to a contractible space, the loop θ → D + θ can be continuously deformed to a constant loop. Then D ′ is a trivial line bundle. This ends the proof of the orientability of the moduli space ˜ M int ( ̂ W,S B ; J ).
## References
[BCS15] J. Bowden, D. Crowley, and A. Stipsicz. 'The topology of Stein fillable manifolds in high dimensions II'. In: Geom. Topol. 19 (2015), 29953030 (electronic). doi : 10.2140/gt.2015.19.2995 .
[BEM15] S. Borman, Y. Eliashberg, and E. Murphy. 'Existence and classification of overtwisted contact structures in all dimensions'. In: Acta Math. 215.2 (2015), pp. 281-361. doi : 10.1007/s11511-016-0134-4 .
[BG83] E. Bedford and B. Gaveau. 'Envelopes of holomorphy of certain 2spheres in C 2 '. In: Amer. J. Math. 105.4 (1983), pp. 975-1009. doi : 10.2307/2374301 .
[CE12] K. Cieliebak and Y. Eliashberg. From Stein to Weinstein and back . Vol. 59. American Mathematical Society Colloquium Publications. Symplectic geometry of affine complex manifolds. American Mathematical Society, Providence, RI, 2012, pp. xii+364.
[Col97] V. Colin. 'Chirurgies d'indice un et isotopies de sph`res dans les vari´t´s e e e de contact tendues'. In: C. R. Acad. Sci. Paris S´ er. I Math. 324.6 (1997), pp. 659-663. doi : 10.1016/S0764-4442(97)86985-6 .
[Eli90a]
Y. Eliashberg. 'Filling by holomorphic discs and its applications'. In: Geometry of low-dimensional manifolds, 2 (Durham, 1989) . Vol. 151. London Math. Soc. Lecture Note Ser. Cambridge: Cambridge Univ. Press, 1990, pp. 45-67. doi : 10.1017/CBO9780511629341.006 .
[Eli90b] Y. Eliashberg. 'Topological characterization of Stein manifolds of dimension > 2'. In: Internat. J. Math. 1.1 (1990), pp. 29-46. doi : 10. 1142/S0129167X90000034 .
[Fra08] U. Frauenfelder. 'Gromov convergence of pseudoholomorphic disks'. In: J. Fixed Point Theory Appl. 3.2 (2008), pp. 215-271. doi : 10.1007/ s11784-008-0078-1 .
[FZ15]
- U. Frauenfelder and K. Zehmisch. 'Gromov compactness for holomorphic discs with totally real boundary conditions'. In: Journal of Fixed Point Theory and Applications (2015), pp. 1-20. doi : 10.1007/s11784015-0229-0 .
[Gei08] H. Geiges. An introduction to contact topology . Vol. 109. Cambridge Studies in Advanced Mathematics. Cambridge: Cambridge University Press, 2008, pp. xvi+440. doi : 10.1017/CBO9780511611438 .
[Gro85] M. Gromov. 'Pseudo holomorphic curves in symplectic manifolds'. In: Invent. Math. 82 (1985), pp. 307-347. doi : 10.1007/BF01388806 .
| [GZ16] | H. Geiges and K. Zehmisch. 'The Weinstein Conjecture for Connected Sums'. In: Int. Math. Res. Not. IMRN 2016.2 (2016), pp. 325-342. doi : 10.1093/imrn/rnv124 . |
|----------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [Hae62] | A. Haefliger. 'Plongements diff´rentiables e dans le domaine stable'. In: Comment. Math. Helv. 37 (1962/1963), pp. 155-176. doi : 10.1007/ BF02566970 . |
| [Hir94] | M. Hirsch. Differential topology . Vol. 33. Graduate Texts in Mathemat- ics. Corrected reprint of the 1976 original. Springer-Verlag, New York, 1994, pp. x+222. doi : 10.1007/978-1-4684-9449-5 . |
| [Hof93] | H. Hofer. 'Pseudoholomorphic curves in symplectizations with appli- cations to the Weinstein conjecture in dimension three'. In: Invent. Math. 114.3 (1993), pp. 515-563. doi : 10.1007/BF01232679 . |
| [HWZ11] | H. Hofer, K. Wysocki, and E. Zehnder. 'Applications of polyfold theory I: Gromov-Witten theory'. preprint. 2011. |
| [KN05] | O. van Koert and K. Niederkr¨ger. u 'Open book decompositions for contact structures on Brieskorn manifolds'. In: Proc. Amer. Math. Soc. 133.12 (2005), pp. 3679-3686. doi : 10.1090/S0002-9939-05-07944-X . |
| [Laz00] | L. Lazzarini. 'Existence of a somewhere injective pseudo-holomorphic disc'. In: Geom. Funct. Anal. 10.4 (2000), pp. 829-862. doi : 10.1007/ PL00001640 . |
| [McD91] | D. McDuff. 'Symplectic manifolds with contact type boundaries'. In: Invent. Math. 103.3 (1991), pp. 651-671. doi : 10.1007/BF01239530 . |
| [Mil65] | J. Milnor. Lectures on the h -cobordism theorem . Notes by L. Sieben- mann and J. Sondow. Princeton, N.J.: Princeton University Press, 1965, pp. v+116. |
| [MNW13] | P. Massot, K. Niederkr¨ger, u and C. Wendl. 'Weak and strong fillabil- ity of higher dimensional contact manifolds'. In: Invent. Math. 192.2 (2013), pp. 287-373. doi : 10.1007/s00222-012-0412-5 . |
| [MS04] | D. McDuff and D. Salamon. J -holomorphic curves and symplectic topol- ogy . Colloquium Publications. American Mathematical Society 52. Prov- idence, RI: American Mathematical Society (AMS)., 2004. |
| [MS98] | D. McDuff and D. Salamon. Introduction to symplectic topology. 2nd ed. Oxford Mathematical Monographs. New York, NY: Oxford University Press., 1998. |
| [Nie06] | K. Niederkr¨ger. u 'The plastikstufe - a generalization of the overtwisted disk to higher dimensions'. In: Algebr. Geom. Topol. 6 (2006), pp. 2473- 2508. doi : 10.2140/agt.2006.6.2473 . |
| [Nie13] | K. Niederkr¨ger. u 'On fillability of contact manifolds'. Habilitation ` a diriger des recherches. oai:tel.archives-ouvertes.fr:tel-00922320: Univer- sit´ e Paul Sabatier - Toulouse III, 2013. url : http://tel.archives- ouvertes.fr/tel-00922320 . |
| [NR11] | K. Niederkr¨ger u and A. Rechtman. 'The Weinstein conjecture in the presence of submanifolds having a Legendrian foliation'. In: J. Topol. Anal. 3.4 (2011), pp. 405-421. doi : 10.1142/S1793525311000684 . |
| [OV12] | A. Oancea and C. Viterbo. 'On the topology of fillings of contact mani- folds and applications'. In: Comment. Math. Helv. 87.1 (2012), pp. 41- 69. doi : 10.4171/CMH/248 . |
42
## REFERENCES
[Rub97] D. Ruberman. 'Null-homotopic embedded spheres of codimension one'. In: Tight and taut submanifolds (Berkeley, CA, 1994) . Vol. 32. Math. Sci. Res. Inst. Publ. Cambridge: Cambridge Univ. Press, 1997, pp. 229-
232.
[Sch99]
M. Schwarz. 'Equivalences for Morse homology'. In: Geometry and topology in dynamics (Winston-Salem, NC, 1998/San Antonio, TX, 1999) . Vol. 246. Contemp. Math. Amer. Math. Soc., Providence, RI, 1999, pp. 197-216. doi : 10.1090/conm/246/03785 .
S. Smale. 'On the structure of manifolds'. In: Amer. J. Math. 84 (1962), pp. 387-399. doi : .
[Sma62] 10.2307/2372978
[Wei91]
A. Weinstein. 'Contact surgery and symplectic handlebodies'. In: Hokkaido Math. J. 20.2 (1991), pp. 241-251. doi : 10.14492/hokmj/1381413841 .
[Wen05]
[Yau04]
C. Wendl. Finite energy foliations and surgery on transverse links . Thesis (Ph.D.)-New York University. ProQuest LLC, Ann Arbor, MI, 2005, p. 411. url : http://search.proquest.com/docview/305467259 .
Mei-Lin Yau. 'Cylindrical contact homology of subcritical Stein-fillable contact manifolds'. In: Geom. Topol. 8 (2004), pp. 1243-1280. doi : 10.2140/gt.2004.8.1243 .
[Zin08]
A. Zinger. 'Pseudocycles and integral homology'. In: Trans. Amer. Math. Soc. 360.5 (2008), pp. 2741-2765. doi : 10.1090/S0002-994707-04440-6 .
- (P. Ghiggini) Laboratoire de Math´ ematiques Jean Leray, BP 92208, 2, Rue de la Houssini`re, F-44322 Nantes Cedex 03, FRANCE e
E-mail address : [email protected]
- (K. Niederkr¨ger) u Alfr´ ed R´nyi Institute of Mathematics, Hungarian Academy of Scie ences, POB 127, H-1364 Budapest, HUNGARY
- (K. Niederkr¨ger) u Institut de math´matiques de Toulouse, Universit´ Paul Sabatier e e Toulouse III, 118 route de Narbonne, F-31062 Toulouse Cedex 9, FRANCE
E-mail address : [email protected]
- (C. Wendl) Department of Mathematics, University College London, Gower Street, London WC1E 6BT, UNITED KINGDOM
E-mail address : [email protected]
|
10.5802/jep.31
|
[
"Paolo Ghiggini",
"Klaus Niederkrüger",
"Chris Wendl"
] |
2014-08-05T17:55:49+00:00
|
2016-02-28T00:44:57+00:00
|
[
"math.SG",
"57R17 (Primary), 53D10, 32Q65, 57R65 (Secondary)"
] |
Subcritical contact surgeries and the topology of symplectic fillings
|
By a result of Eliashberg, every symplectic filling of a three-dimensional
contact connected sum is obtained by performing a boundary connected sum on
another symplectic filling. We prove a partial generalization of this result
for subcritical contact surgeries in higher dimensions: given any contact
manifold that arises from another contact manifold by subcritical surgery, its
belt sphere is null-bordant in the oriented bordism group $\Omega SO^*(W)$ of
any symplectically aspherical filling $W$, and in dimension five, it will even
be nullhomotopic. More generally, if the filling is not aspherical but is
semipositive, then the belt sphere will be trivial in $H^*(W)$. Using the same
methods, we show that the contact connected sum decomposition for tight contact
structures in dimension three does not extend to higher dimensions: in
particular, we exhibit connected sums of manifolds of dimension at least five
with Stein fillable contact structures that do not arise as contact connected
sums. The proofs are based on holomorphic disk-filling techniques, with
families of Legendrian open books (so-called "Lobs") as boundary conditions.
|
1408.1052v1
|
## Optimal path selection in Graded network using Artificial Bee Colony algorithm with Agent enabled Information
## Kavitha Sooda
T. R. Gopalakrishnan Nair
Advanced Networking Research Group, RIIC, DSI and Asst. Professor, Dept. of CSE Nitte Meenakshi Institute of Technology Bangalore - 560064, India
e-mail: [email protected]
Abstract -In this paper we propose a network aware approach for routing in graded network using Artificial Bee Colony (ABC) algorithm. ABC has been used as a good search process for optimality exploitation and exploration. The paper shows how ABC approach has been utilized for determining the optimal path based on bandwidth availability of the link and how it outperformed non graded network while deriving the optimal path. The selection of the nodes is based on the direction of the destination node also. This would help in narrowing down the number of nodes participating in routing. Here an agent system governs the collection of QoS parameters of the nodes. Also a quadrant is synthesized with centre as the source node. Based on the information of which quadrant the destination belongs, a search is performed. Among the many searches observed by the onlooker bees the best path is selected based on which onlooker bee comes back to source with information of the optimal path. The simulation result shows that the path convergence in graded network with ABC was 30% faster than non-graded ABC.
Keywords-Network Awareness; Graded Network; Bio-inspired algorithm; ABC; Optimal path
## I. INTRODUCTION
In the recent past, Swarm intelligence has been applied for studying collective behavior of known population. It has become a research interest to many research scientists of related fields in recent years. The Internet in present scenario has confronted with problems like security, scalability, performance due to rapid increase in the end-users and new service demands. Internet test-beds are required to support reliable end-to-end research for federation and management of Internet. Here routing and forwarding has become the core problem for delivery of data from one node to another. Current routing strategy whether it is wired or wireless routing protocols, are based on some predetermined calculation or prior knowledge of the connection between the nodes. This information was simply based on the request packets sent. Generally for any intelligent action [1] to take place a collective behavior of the topology needs to be studied prior to
ARAMCO Endowed Chair- Technology, PMU, KSA Advanced Networking Research Group VP, Research, Research and Industry Incubation Center (RIIC), Dayananda Sagar Institutions, Bangalore - 560078, India e-mail: [email protected], www.trgnair.org
the path determination process. Also there exists a need for self organizing the topology when changes occur. This has led to the introduction of Swarm intelligence, a nature inspired process to handle this kind of situation [2-3]. Here the agents locally collect information based on some predefined parameter [4]. Later they all communicate with each other globally to arrive at the desired result. Because of its self adaptive property, swarm intelligence has gained importance in many field of research.
The remainder of the paper is organized as follows: Section 2 outlines the overview of Network Awareness and Graded network. Section 3 deals with ABC algorithm. Section 4 presents the details of the mathematical model. Section 5 consists of performance results. Finally, the last section concludes the paper.
## II. NETWORK AWARENESS AND GRADED NETWORK
## A. Network Awareness
Network awareness [5] is much dependent on querying the distributed system state space such as the scenario of Internet. It deals with supplementing the knowledge of network characteristics while deciding on the network locations where the query operators are executed. There exists a need for querying widely in distributed system that enables intelligent data acquisition from the nodes which can participate in network routing. Existing querying approaches are suitable for small-scale systems and it fails when the network grows. The network awareness approach helps in one way the implementation of querying process in a large distributed environment. In general, the network operating principle is largely based on the characteristics like topology and bandwidth. In this paper we incorporate the network aware management strategy. It utilizes the network characteristics to identify better nodes where the operator can approach or query the node to identify suitable situation. The network characteristic is dealt based on the QoS metric such as the bandwidth. The algorithm, initially assume that any given node is a feasible candidate for obtaining the optimal path the forward push of messages. Decision is taken on nodes, based on the query operator which will bifurcate the nodes in to feasible the non-feasible nodes. Next the neighbor of the previous feasible node is considered and the process is repeated until the destination node is reached. Here the network aware algorithm is based on exploration of the neighbor nodes as illustrated in Fig. 1.
Figure 1. Network Awareness Routing
Thus the environment awareness [5] can extract the nodes which are feasible to propagate the data in an efficient manner. This leads to an improved performance due to reduction in the search space and provides feasible path for the data to reach the desired destination in an efficient manner.
Figure 2. Network Awareness algorithm
Here in this paper, the network awareness algorithm i.e, level-1 approach of grading algorithm is applied on a physical network where the nodes which participate in routing are available as in [4]. Now a quadrant is designed where the quad is drawn with source as the centre point as shown in Fig. 2. Now the nodes which belong to the destination quadrant will participate in routing. Thus the optimal path is determined assuming sufficient nodes are availble for access and selection in the selected quadrant of interest.
## B. Graded network
The authors in [4] have discussed much on the graded network. The Grade value estimation method for implementing intelligent routing in the autonomic network is the core focus of research and is the thrust. Grade is like an index; it is made available everywhere and routing will much depend on it. It helps building the quality of the router, which is in fact triggered by the knowledge of the environment. The router must be an intelligent entity because it performs different operations using the environment information. It depends on input, output, load and resource availability. The graded network is setup with six QoS parameters. These are considered at two levels; level-1 and level-2. Level-1 is applied region-wise and the goal is to achieve favorable routing based on selected attributes. The values obtained from Level-1 must be able to eliminate the non-production node. Non-production nodes are those that come in the pitfall of congestion and possess less resource availability. These nodes must be identified by the algorithm that defines the gradient from most non-productive to productive nodes in a homogeneous network. This is identified by assigning a grade value from -3 to +3, i.e. it signifies the productivity value of the node. At this point, we are able to calibrate the routing process region-wise. The algorithm requires proactive decision making on the output obtained. This is because many paths exist to reach the destination, and we have to choose the most optimal path. Once the gradient value has been calculated, it can be made available as pervasive information packets to all the other nodes to obtain the optimal path constitution.
Now the graded function, i.e., Level-2 , considers the calculated values as its input of level-1 . The output of this function defines the route availability for the set of nodes considered. This shall be calculated for all the available paths leading towards the destination node. The mean value of the gradient in the grade function shows the success level of operation of the network which is obtained by the ABC algorithm.
## III. BEE COLONY ALGORITHMS
In recent years swarm intelligence has been the focus of interest in research. Swarm intelligence is about developing an algorithm or a distributed problem solving device inspired by the collective behavior of social insect colonies and other animals. Swarm in general refers to any restrained collection of interacting agents or individuals. The classical examples which depict the behavior of swarm are bees swarming around their hive. Similar architecture which mimic the swarming approach are the ant colony, immune system, bird flocking and the fish schooling [6].
## A. Artificial Bee Colony(ABC) for network routing
Artificial Bee colony algorithm is developed by Karaboga in 2005 [7], which were motivated by the intelligent behavior of the bees. It is an optimization tool, where search is performed in a distributed environment. Here the artificial bee fly find the food source based on experience or their nest mates and adjusts their position. Some artificial bees (scouts) choose the food source randomly without experience. If the nectar amount of the new food source is better, then the bees memorize this position and forget the previous food source position. Thus ABC performs a local search with onlooker and employed bees, with global search methods managed by the onlooker and scout bees, attempting to balance exploration and exploitation process.
In ABC, the search space is exploited by the employed bees and onlooker bees and the scout bee does the exploration. Here the recruitment rate represents the measure of how quickly the swarm locates and exploits new food source. The survival of the bees here depends on how quickly they are able to discover optimal or feasible solutions. In Artificial Bee colony algorithm the initial food source is equal to the number of employed bees.
In this paper we apply these principles to solve routing problems optimally. They are used in the model to test and develop forward path progress. The nectar collected as food from the neighboring nodes is used as the parameter in equivalence to the bandwidth available at the forward point at the evaluation time point of the virtual network under consideration. When the messages are in different state of dispatch requirement, general algorithm restore to geographical connectivity mainly table driven algorithm. In a highly exploded model of network upcoming future limitations of getting stuck with geographical information need to be avoided and the message need to be distributed based on the intelligence of the router and the information collected from the environment such as available bandwidth at very forward node angular positioning of forward vector, QoS parameters like jitter, delay, etc. In the model the employed bees are randomly assigned to the food source for exploitation. In every iteration, every employed bee determines a food source in the neighborhood and evaluates its nectar amount using fitness function. The i th food source position is represented by N ϵ i (n1 , n2 , ......n ). e Here ϵ represents the feasibly connectable nodes. F(n ) i determines the total nectar amount possibly accessible from Ni . After watching the dance of the employed bees, the onlooker bee goes to the region of food source at N i with probability p defined as, i
$$P _ { i } = & \frac { F ( N _ { i } ) } { \Sigma _ { k = 1 } ^ { ( } F ( n _ { i } ) } } \quad \quad \quad ( 1 ) \quad \text{$val$} \quad \text{$old$} \quad \text{$\text{$th$} }$$
where fs is the total number of food sources. The numerator represents the nectar amount at node N and denominator i represents the total nectar amount accumulated from current node to all its neighbours. Here the dance represents the amount of free available bandwidth. The more is the bandwidth availability, the more the bees are flying towards the node.
To derive the intermediate paths, let 𝐹 𝑚𝑎𝑥 1 be the first neighbour of the current node . This function derives the filtered maximum value among all the neighbours from the current node we are looking from. Here = symbol represents the updates on intermediate paths obtained from the current node. The onlooker finds the neighborhood food source of Ni and is represented using following equation where the + symbol represent the establishment of the intermediate path from current node to the node where maximum bandwidth to the destination direction is available ,
$$N _ { i } ( n _ { i } + 1 ) = C _ { i } ( n _ { i } ) + F _ { \max 1 } ^ { \ w } \left ( C _ { i } ( n _ { i } ) \right ) \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdclots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \colon$$
The next intermediate node selection takes place by the set of onlooker bees. The onlooker bees' number depends on the number of neighbours each node posses which is represented by w. The selection of the neighbour from the current intermediate node is represented by F max 1 w ( Ci ni ). The current intermediate node is represented by Ci ni . Fmax 1 w returns the node address which would have the highest bandwidth among all the neighbors of the current intermediate node. The result of this is concatenated with current node to determine the next intermediate path. Now C i ni is updated as,
$$\sigma \text{int at} \quad C _ { i } ( n _ { i } ) = N _ { i } ( n _ { i } + 1 )$$
This process is repeated until destination is reached or no link exists further to explore.
In this paper, when the system is setup a random activity shall start occupying certain bandwidth of the link. Here packets sent are of fixed size, i.e, 250 KB. Each node has 1MB of buffer space where the incoming packets are stored. The time taken to send one packet is one second from each node. The bandwidth utility on each link is monitored. A decision shall be taken whether to consider the link or not based on a threshold bandwidth. If the link has bandwidth above a threshold value then the link can participate in routing of the packets. This selection is applied after the quadrant structure is obtained.
Here the routers selected are the ones which are above the threshold bandwidth and the messages can transfer easily. The network awareness here is about identifying the nodes which satisfy QoS metric. This makes the routing decision to take place easily.
After the neighbours have been determined the fitness value is calculated. If this new fitness value is better than the old value then the bees move to this new food source otherwise it remains in the same node. When all employed bees have finished the search for food source it shares this information with the onlooker bees. Now onlooker bees select these food sources with a probability, p as given by (1) which i represents the best bandwidth on the link between N (t) and i Ni(t+1).
Pseudo code on application of ABC which is applied for network routing is as follows:
Initialize the nodes and bandwidth randomly.
The employed bees are initially the number of neighbors to source
## REPEAT
- The employed bees are moved onto their paths and fitness is calculated.
- The employed bees report to the onlooker bees about the information collection.
- Now selection is made where onlooker bees should move which is based on maximum bandwidth.
- If good node (means no link with sufficient bandwidth) not obtained scout bees (different quadrant search gets initiated) are used for searching.
## Memorize the best node found so far.
## UNTIL (termination criteria satisfied)
Fig. 3 illustrate the flow chart of the implementation of the ABC algorithm for network routing. Here the populations of solutions are randomly initialized and evaluated to test to see whether it is above the threshold bandwidth value. The employed bees initially will be initialized to the number of neighbors of the source [8-9].
Figure: 3 ABC Algorithm for network routing
Here ‗C' represents the cycle or the count for the performance of the ABC algorithm and M c represents the maximum cycle i.e., the total number of times ABC repeats.
The bee search for food source is to maximize the ratio E/T. Here E represents the amount of nectar food discovered and T represents the time spent for foraging. As far network is concern T is not permitted above a certain level for communication efficiency. And also foraging will take place not at very instant but at a selected interval of time which is again determined by the stability of the network in that particular region.
## B. Fitness function
The simulation topology is generated randomly. Here relaxation is imposed on the number of onlooker bees. The number of onlooker bees is equal to the number of neighbors to the source node who belong to the quadrant of the destination node. Here the fitness value is calculated on nodes which only belong to the quadrant. The nodes which satisfy the threshold bandwidth values are the ones which get selected and participate in routing which indicates the fitness value. Further the authors have planned to implement multi-objective parameters solvable by the fitness function [10].
## IV. MATHEMATICAL MODEL
Here we have proposed random network model with channel level stochastic Markov process for realizing the link capacity. Here every cell is considered to be a Jackson network [11]. Here links are interconnected by set of n links, L={1,2,3, …,n}, where every link has the same capacity, l ϵ L. let the flow arrival rate between any two nodes, say s and e be represented by λ se . Here we assume that each flow transmits packet at fixed unit rate with mean µ -1 , along the route it is assigned.
$$\begin{smallmatrix} \dots \dots \dots \\ \text{The traffic intensity of the link can be calculated as,} \\ \text{Traffic Intensity=\frac{Ps. Tk(t)}^{Abs} \\ \text{Lara\ Dc\ indirectac\ } tba\ method\ ciso\ } A b\ indirectac\ tba \end{smallmatrix}$$
Here Ps indicates the packet size, Ab indicates the available bandwidth and the traffic load on the link l ϵ L at time t is denoted by T(t). It represents the total number of l flows currently routed across it. The flow arrival rate to link l is denoted by γ . This flow represents a part of the flow arrival l between s and e that is routed to link l. Here the flow departure rate is proportional to the current link load T (t) and l is given by T(t) * µ. The packet size for the simulation is fixed l as well as the available bandwidth on the link.
The updates on the link is done every 30*t sec, which would represent the delay involved in building a grade network. Here the model is developed based on the given information on traffic load and flow rate. Here we assume that the traffic arrivals during [0, t] are routed based on the link states T(0) available at time 0. The link state T (t) tracks to the l l following differential equations which derives the total traffic arrival rate γ to link l. l
$$\frac { d \, T I ( t ) } { \stackrel { d t } { \longrightarrow } } = ( \gamma _ { 1. } T _ { 1 } ( t ) \\ \stackrel { d t } { \longrightarrow } \cdot \cdot \cdot \cdot \cdot$$
$$\dots \dots \dots \dots \dots \dots \dots \dots \\ \frac { d T | ( t ) } { d t } = ( \gamma _ { 1 } \cdot T _ { 1 } ( t ) * \mu ) & & ( 5 ) \\ \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \text{$i$} \, \end{cases}$$
where Tl(0) is the link state information at time interval [0,t]. This value is subtracted from link capacity to obtain the remaining fraction of the bandwidth available.
$$B _ { a } = L _ { c } \text{-} T _ { l } ( 0 ) ; & & ( 7 ) \quad \text{red}$$
Using Eq (6) and Eq (7) we can determine the traffic intensity from Eq. (4). Equation (7) helps is the calculation of whether congestion has occurred or not.
The model is designed based on the hexagon structure with nodes and link generated at random. In each of the cell structure, the jobs arrive from outside which are Poisson in process with rate α >0. Each arrival is independently routed to node j with probability p j ≥0 and 𝑘 𝑃 𝑖𝑗 = 1 𝑗 =1 . Here k represents the total number of nodes that a packet which has arrived from node i and can go through. Upon service completion at node i, a job may go to another node j with probability pij or leave the cell with probability qi = 1𝑘 𝑃𝑖𝑗 𝑗 =1 . This is similar to the Jackson open network and mapped to Eq. (6), for calculating the load on link l.
Thus the determination of the link capacity helps to identify where congestion has occurred or not. The agent here is used for aiding the grade network with relevant information when required. For instance at the first level of grading the parameters which are looked for are Bandwidth availability, congestion level check, delay, node density and resource allocation. All these five parameters are taken care by the agent. For instance the bandwidth availability is determined by Eq. (7) which in turn helps to check whether congestion has occurred or not. Delay is obtained by Eq. (4). Network node density is number of packets arriving to the node and resource allocation is randomly assigned.
We then apply local algorithm at level-2 trying to achieve global optimization, which is referred to as constrained optimization. Here we applying the probabilistic reasoning approach to derive optimal path which is based on agent information. At the second level of the grade approach we try to achieve maximum bandwidth for routing using the following maximization technique,
$$\text{Maximize } \sum _ { j \in N ( i ) } ^ { n } l _ { i j } \\ \text{Subject to } \sum _ { j \in N ( i ) } ^ { n } P ( l _ { j } ) \ = 1 \\ \text{Over } 0 \leq l _ { j } \leq 1 \text{ for all } j$$
Here lij represent the link capacity from node i to node j . We try to get the link which possesses maximum bandwidth. This is arrived by P(l ) which represent the probabilistic j bandwidth value for a given node where its neighboring node are of high bandwidth.
## V. SIMULATION RESULTS
The initial configuration of simulation model consists of a randomly distributed set of nodes, in a geographical area. They are capable of getting connected through links which could be created on demand. The user enters the source and destination node for which a quadrant is drawn with source as the centre. With this structure, depending on the destination node selection of the path is carried out. The quadrant structure reduces the search space nearly to one-fourth depending on the
TABLE I. RESULT ANALYSIS
| Total nodes | ABC with quality grading of nodes (A2) | ABC with quality grading of nodes (A2) | ABC without quality grading of nodes (A1) | ABC without quality grading of nodes (A1) |
|---------------|------------------------------------------|------------------------------------------|---------------------------------------------|---------------------------------------------|
| | Number of nodes selected | Route length | Number of nodes selected | Route length |
| 15 | 10 | 3 | 15 | 5 |
| 16 | 12 | 4 | 16 | 4 |
| 32 | 18 | 5 | 32 | 5 |
| 64 | 38 | 5 | 64 | 6 |
| 128 | 61 | 5 | 128 | 6 |
source location. Here the agent evaluates all the nodes in the quadrant where destination node belongs and performs level-1 grading. The level-2 grading is performed using ABC algorithm. Table. 1 is the comparative result of graded ABC and non-graded ABC. It shows that the route length obtained by grade ABC is less than route length obtained by non-graded ABC. The column two of the table shows how the search space has reduced due to the application of the quadrant design in graded ABC.
Here for implementing ABC, the number of onlooker bees is chosen based on the number of neighbors the source has. The employed bees now search for the best nectar among the available neighbors for each one of them. This search is continued until the destination is reached. Now among the different paths which exist between the source and destination one complete path is selected based on the bandwidth estimation of the links. Since here the link selected needs to be above a threshold value, there is less likely scenario for congestion to occur. Table. 2 shows the computation time obtained after applying ABC on graded and non-graded surface.
TABLE II. COMPUTATION TIME
| A2-Computation Time | A1- Computation Time |
|-----------------------|------------------------|
| 3 | 5 |
| 4 | 6 |
| 5 | 7 |
| 5.5 | 7.5 |
| 6 | 8 |
Figure 4, depicts the graph for the throughput and traffic intensity determined in both graded and non graded network. It is seen that the traffic intensity in grade ABC is less than nongraded network. Also the throughput was more or less better in graded ABC than in non-graded ABC.
The paper here presented the optimal path determination using the bio inspired artificial bee colony approach for graded network. The approach showed a good performance in terms of convergence rate as the path search took place only in one fourth of the search area. A quality path with sufficient bandwidth was determined. Here we have shown that without a pre-defined message path we are able to obtain the paths. The routers are able to dynamically select the forward nodes and propagate the messages successfully to the destination. The result showed the traffic intensity was low in graded ABC as against to non-graded ABC.
Figure 4. Graph plot for realizing through and traffic intensity for graded ABC and non-graded ABC
However the simulation has its own limitation. Currently only few hundred nodes were used to test and evaluate the principles of the model. Future work involves expanding the simulation to higher number of nodes realizing near real time network scenarios.
The simulation setup shall be modified to a zone based structure where there shall be a zone coordinator which will be responsible improving scalability and for achieving parallelism.
## REFERENCES
- [1] T. R. Gopalakrishnan Nair, Kavitha Sooda, Abhijith, -Transformation of networks using cognitive approaches‖, JRI, Vol.1, No. 1, December, 2008, pp.7-14.
- [2] H.F. Wedde, M. Farooq, Y. Zhang, Beehive: an efficient fault-tolerant routing algorithm inspired by honeybee behavior, ant colony, optimization and swarm intelligence, in: 4th International Workshop, ANTS 2004, Brussels, Belgium, September 5-8.
- [3] J. Kennedy, R.C. Eberhart, -Particle Swarm Optimization‖, IEEE International Conference on Neural Networks, vol. 4, 1995, pp. 19421948.
- [4] T. R. Gopalakrishnan Nair, Kavitha Sooda, -Particle Swarm Optimization for Realizing Intelligent Routing in Networks with Quality Grading‖, WiCOM, 7 th IEEE International Conference on Wireless Communications, Networking and Computing, Wuhan, China, 23-25 th September, pp. 1-4, 2011.
- [5] Yanif Ahmad Ugur Cetintemel -Network-Aware Query Processing for Stream-based Applications‖, Proceedings of the 30th VLDB Conference, Toronto, Canada, 2004.
- [6] V. Tereshko, Reaction-diffusion model of a honeybee colony's foraging behaviour, in: M. Schoenauer (Ed.), Parallel Problem Solving from Nature VI, Lecture Notes in Computer Science, vol. 1917, SpringerVerlag, Berlin, 2000, pp. 807-816.
- [7] D. Karaboga, An idea based on honeybee swarm for numerical optimization, Technical Report TR06, Erciyes University, Engineering Faculty, Computer Engineering Department, 2005.
- [8] B. Basturk, D. Karaboga, An artificial bee colony (ABC) algorithm for numeric function optimization, in: IEEE Swarm Intelligence Symposium 2006, Indianapolis, Indiana, USA, May 2006.
- [9] D. Karaboga, B. Basturk, On the performance of artificial bee colony (ABC) algorithm, Applied Soft Computing 8 (1) (2008), pp.687-697.
- [10] T. R. Gopalakrishnan Nair, Kavitha Sooda- A Novel Adaptive Routing through Fitness Function Estimation Technique with multiple QoS parameters Compliance‖, 4 t h International Conference on Information Processing, ICIP-2010, Bangalore, India, 6th - 8 t h August 2010, pp. 489-491.
- [11] J. Medhi, Stochastic Models in Queueing theory , Elsevier, Academic press, 2006.
|
10.1109/HIS.2012.6421356
|
[
"Kavitha Sooda",
"T. R. Gopalakrishnan Nair"
] |
2014-08-05T18:02:25+00:00
|
2014-08-05T18:02:25+00:00
|
[
"cs.NI"
] |
Optimal path selection in Graded network using Artificial Bee Colony algorithm with Agent enabled Information
|
In this paper we propose a network aware approach for routing in graded
network using Artificial Bee Colony (ABC) algorithm. ABC has been used as a
good search process for optimality exploitation and exploration. The paper
shows how ABC approach has been utilized for determining the optimal path based
on bandwidth availability of the link and how it outperformed non graded
network while deriving the optimal path. The selection of the nodes is based on
the direction of the destination node also. This would help in narrowing down
the number of nodes participating in routing. Here an agent system governs the
collection of QoS parameters of the nodes. Also a quadrant is synthesized with
centre as the source node. Based on the information of which quadrant the
destination belongs, a search is performed. Among the many searches observed by
the onlooker bees the best path is selected based on which onlooker bee comes
back to source with information of the optimal path. The simulation result
shows that the path convergence in graded network with ABC was 30% faster than
non-graded ABC.
|
1408.1053v1
|
## Dynamical Stability of Terrestrial Planets in the Binary α Centauri System
## E. Andrade-Ines /star and T.A. Michtchenko
Instituto de Astronomia, Geof´ ısica e Ciˆncias Atmosf´ricas, USP, Rua do Mat˜o 1226, 05508-090 S˜o Paulo, Brazil e e a a
6 August 2014
## ABSTRACT
In this paper, we investigate whether hypothetical Earth-like planets have high probability of remaining on stable orbits inside the habitable zones around the stars A and B of α Centauri, for lengths of time compatible with the evolution of life. We introduce a stability criterion based on the solution of the restricted three-body problem and apply it to the α Centauri system. In this way, we determine the regions of the short-term stability of the satellite-type (S-type) planetary orbits, in both planar and three-dimensional cases. We also study the long-term stability of hypothetical planets through the dynamical mapping of the habitable zones of the stars. The topology of the maps is analyzed using the semi-analytical secular Hamiltonian model and possible processes responsible for long-lasting instabilities are identified. We verify that the planetary motion inside the habitable zones is regular, regardless of high eccentricities, for inclinations smaller than 40 ◦ . We show that the variation of the orbital distance of the planet located in the habitable zones of the binary is comparable to that of Earth, if the planet is close to the Mode I stationary solution. This result brings positive expectations for finding habitable planets in binary stars.
Key words: celestial mechanics; stars: binaries: general; planets and satellites: dynamical evolution and stability; planets and satellites: individual: α Centauri;
## 1 INTRODUCTION
About 50% of the known main-sequence-stars belong to the class of binary or multiple stellar systems (Abt 1979; Duquennoy & Mayor 1991; Raghavan et al. 2010). Due to inherent difficulties in monitoring the radial velocities of multi-star systems, these systems have been mostly ignored by researchers performing searches for extra-solar planets. However, in last years, due to improving precision and advances of detection methods, binaries have become increasingly interesting targets of planet searches, and several research programs looking for planets in binary systems are in progress.
that more planets around the star B may exist, particularly inside the Habitable Zone (HZ) (Guedes et al. 2008; Th´ ebault et al. 2009). Although these planets are yet to be discovered in binary systems, recent works have shown that terrestrial planets can form and have a stable orbit around a star of a binary in a close S-type system, depending on the binary orbital and physical parameters (Haghighipour 2006; Haghighipour & Raymond 2007; Quintana et al. 2007; Eggl et al. 2013; Kaltenegger & Haghighipour 2013).
The α Centauri binary is of special interest. Being the closest stellar system to the Sun, as well as composed of two stars very similar to the Sun, it has been always a major target for extrasolar planet surveys, although it was not until recently that a putative planet has been found orbiting around α Centauri B with a period of 3.236 days (Dumusque et al. 2012; for an opposing view see Hatzes 2013). Furthermore, studies of the planet formation suggest
/star E-mail: [email protected]
The high mass of the disturbing stellar companion and the frequent high eccentricity/inclination of its relative orbit make the perturbation theories based on classical Lagrangian expansions inapplicable in this case. A possible approximation is the expansion of the disturbing function using the Legendre polynomials (Heppenheimer 1978; Ford et al. 2000; Laskar & Bou´ e 2010). However, although this development avoids the constraints due to the high eccentricities and inclinations, its convergence is very slow, allowing its application only in the case of hierarchical systems (Nelson 2000; Boss 2006; Haghighipour 2006; Th´ ebault et al. 2009; Eggl et al. 2013).
An alternative to overcome the problem is the extension of a perturbation theory up to second order in the disturber
mass. This was done in Giuppone et al. (2011), where the developed model was applied to the γ Cephei binary system. Although the results obtained have shown an improvement over the first-order model, the authors have emphasized that the model was not workable due to its high complexity; thus they preferred the use of empiric correction terms, specific for the γ Cephei configuration.
In this work, we perform a study of the stability of the planetary motion in binary systems combining the analytical and numerical approaches. For this, the stability concept is separated into the short-term stability and the long-term stability. The short-term stability is based on the Hill criterion, which does not require any kind of expansion and averaging of the disturbing function. In such a way, it does not introduce the constrains on neither the mass of the disturbing body nor the eccentricity/inclination of its orbit.
The Hill criterion is a powerful approach, which has been used for over a century (Hill 1878); an in-depth review of its applications can be found in Szebehely (1984). Analytical approaches based on the Hill stability concept have been developed in Marchal & Bozis (1982), Gladman (1993), among others, also as some empirical relations between stable orbits and the physical parameters of the binary (Graziani & Black 1981; Black 1982; Pendleton & Black 1983; Rabl & Dvorak 1988; Holman & Wiegert 1999). In our paper, we develop a variation of the Hill approach, referring to it as Maximum Apocentric Distance (or MAcDcriterion), and apply it to the α Centauri stars A and B. To simplify the applications of the method, we present the explicit formula for the maximal apocentric distances as a function of the physical and orbital parameters of the stellar binary.
The study of the long-term stability of a binary-planet system (of the order of the lifetime of the stars) is done for the S-type (satellite type) planetary motion, when the planet revolves around one of the stellar components (Dvorak et al. 1986). The planet motion is modeled semi-analytically in the frame of the secular three-body problem (Michtchenko et al. 2006) and applied to the α Centauri hypothetical Earth-like planets placed in the HZs of the stars A and B. The topology of the phase space is investigated, without any restriction on the magnitude of the eccentricities and inclinations of the planets. The boundaries of the secular stability of the system, as a function of the masses and the orbital parameters of the planets, are then obtained. Notwithstanding the averaging being done at first order in masses, the relatively large separation between the stars allows a good agreement of the analytical results with those obtained numerically. The direct numerical investigations are done through the dynamical mapping of the regions of habitability of both stars in the α Centauri system.
The physical and orbital parameters of the α Centaury binary used in this paper were taken from Pourbaix et al. (2002). They are the masses of the components A and B, m A = 1 105 . ± 0 0070 . M /circledot and m B = 0 934 . ± 0 0061 . M /circledot , respectively, and the eccentricity e = 0 518 . ± 0 0013 and the . semimajor axis a = 17 57 . ± 0 022 . arcsec (or a = 23 4 . ± 0 03 AU, if the parallax . of 746 8 . ± 1 2 . mas is adopted) of the relative orbit.
This paper is structured as in the following. In the next section, we study the short-term stability of the system, introducing a stability criterion and applying it to both the planar and 3D cases for planets on S-type orbits. In Section 3 we study the long-term stability, applying the semianalytical 3D model and comparing the results to those obtained through dynamical mapping. In Section 4, we investigate whether the planets remain inside the HZ around the central star, regardless of large amplitudes of short-term oscillations provoked by the stellar companion. Finally, the summary is given in Section 5.
## 2 SHORT-TERM DYNAMICAL STABILITY
In this section, we introduce the stability criterion for S-type planetary orbits in binary star systems based on the concept of the Hill stability (Marchal & Bozis 1982; Gladman 1993, among others) and apply it to both the planar and 3D motions of the fictitious planets in the α Centaury system.
## 2.1 Planar model
Our approach is based on the classical solution of the restricted three-body problem and, in particular, on the Zero Velocity Surface method. Zero velocity surfaces are levels of the S /vector ( r )-function defined by the position of the planet with respect to the two stars (Marchal & Bozis 1982). The position vectors are defined with respect to the center of mass of the system in the non-inertial rotating reference frame, in which the two stars always lie along the x -axis.
Let us consider a system composed by two stars, 1 and 2, and a planet p . The S /vector ( r )- function, also known as Jacobi integral or relative energy, can be written as
$$\tanetri { e \, \text{for} } _ { \text{len} } \quad S ( \vec { r } _ { p } ) = \sum _ { i } \frac { m _ { i } } { M } \left [ \frac { ( \vec { r } _ { p } - \vec { r } _ { i } ) ^ { 2 } } { d ^ { 2 } } + \frac { 2 d } { | \vec { r } _ { p } - \vec { r } _ { i } | } \right ], \quad \quad ( 1 )$$
where m i and /vector r i ( i = 1 2) are the stellar , masses and the position vectors and /vector r p is the position vector of the planet; M is the total mass of the system and d = | /vector r 1 -/vector r 2 | . Figure 1 shows the positions of the stars m 1 and m 2 separated by the distance d . In the rotating frame, the function S /vector ( r p ) allows five equilibrium solutions (three co-linear and two triangular Lagrangian points), all located in the plane of the motion of the stars, which will be chosen as a reference plane. For our purpose, the co-linear solutions, and, particularly, L 1 solution, are of special interest. These solutions are located on the x -axis and their co-ordinates can be obtained searching for singular points on the surface (1), where the vectors /vector r are just replaced by scalar x -component of the corresponding body.
It should be noted that the expression (1) was developed in the uniformly rotating reference frame, when two stars move on circular orbits around the baricenter of the system and their mutual distance d is unchanged. To apply it to the case when the stars have elliptic orbits, it is reasonable to adopt a minimal value for d , when the stars are at the pericenter of their relative orbit. In this configuration, the perturbations on the S-like planetary orbit are maximal and possible instabilities may appear.
Figure 1 shows the projection of one zero-velocity surface on the reference plane (red curve); the surface was calculated using the orbital elements of the α Centauri binary and the minimal distance between the stars. This level holds the
Figure 1. Zero velocity curve corresponding to the Lagrangian L 1 solution (red curve), obtained for the stars at the pericenter of their relative orbit. The circumferences of stability, of radii R 1 cr around m 1 , and R 2 cr around m 2 , are shown by blue and green curves, respectively.
Lagrangian solution L 1 , which appears as a singular saddlelike point on the x -axis. The same level intersects the x -axis at two other points; the positions of these points, x i cr , can be calculated from the condition
$$S ( x _ { c r } ^ { i } ) - S ( x _ { L _ { 1 } } ) = 0, \quad i = 1, 2. \quad \quad ( 2 ) \quad \text{satisf} \quad \text{satisf} \quad \text{satisf} \quad \text{satisf} \quad \text{satisf} \quad \text{satisf}$$
We assume that these intersections define a sphere of stability around each star. By definition, the sphere of stability is centered on the corresponding star and its radius is given by R i cr = | x i cr -x i | , where x i cr is the abscissa of the closest to the star intersection, while the abscissa of the star is x i . The projections of the spheres of stability on the reference plane are shown in Figure 1, as a blue circumference centered on the star 1 and a green circumference centered on the star 2.
It is worth noting that Equation (2), which is just a polynomial function of x i cr , can be easily solved numerically. For the sake of simplicity, we present here approximate expressions that allow us to obtain the radii. The expressions are obtained applying the method of least squares. Without loss of generality, we assume that the star 1 is a central star, while the star 2 is a perturber. In the case when the stellar mass ratio µ = m /m 2 1 belongs to the interval 0 1 . /lessorequalslant µ /lessorequalslant 1 0, . the radius of the sphere of stability is given as
$$& R _ { c r } / a \approx 0. 6 6 8 2 3 - 0. 6 3 7 4 0 e - 0. 7 4 5 4 9 \mu + \begin{pmatrix} & \text{sphe} \\ & + 0. 4 5 4 9 6 e \mu + 1. 0 4 9 2 \mu ^ { 2 } - 0. 2 3 1 7 9 e \mu ^ { 2 } + & ( 3 ) & \text{is wc} \\ & - 0. 8 7 7 2 2 \mu ^ { 3 } + 0. 3 1 5 4 1 \mu ^ { 4 }, & \text{long} \\ & - 0. 8 7 7 2 2 \mu ^ { 3 } + 0. 3 1 5 4 1 \mu ^ { 4 }, \end{pmatrix}$$
where a and e are the semimajor axis and the eccentricity of the stellar relative orbit, respectively. For the stellar mass ratio µ = m /m 2 1 from the interval 1 0 . < µ /lessorequalslant 10 0, . the radius of the sphere is given as
$$& R _ { c r } / a \approx 0. 4 5 2 6 5 - 0. 4 1 9 2 1 e - 0. 0 7 0 7 5 4 \mu + \begin{pmatrix} & \text{In tt} \\ + 0. 0 3 9 6 1 7 e \mu + 0. 0 1 0 8 6 5 \mu ^ { 2 } - 2. 1 3 9 4 \times 1 0 ^ { - 3 } e \mu ^ { 2 } + \end{pmatrix} & \text{$(4)$} & \text{stabi} \\ & - 9. 3 7 2 9 \times 1 0 ^ { - 4 } \mu ^ { 3 } + 3. 3 8 8 6 \times 1 0 ^ { - 5 } \mu ^ { 4 }. & \text{$bina} \\ & \text{$\mathbb{m.m.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.t.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s. s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.r.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.n.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.g.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.S.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.c.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.ns.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.h.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.l.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.x.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.e.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.o.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.s.
s$$
Both approximations have errors smaller than 10% of the exact solutions, for 0 /lessorequalslant e /lessorequalslant 0 9. The error drops to less . than 4%, considering the interval of eccentricities 0 /lessorequalslant e /lessorequalslant 0 7. .
Now we introduce our criterion of stability. It states that
Figure 2. Dynamical Power Spectrum: the proper oscillation frequencies as functions of the semimajor axis of the planet orbiting the star α Centauri A, in logarithmic scale. The top axis is the planetary semimajor axis in unit of the semimajor axis of the binary relative orbit. The frequency unit is 1 /yr .
the stable motion of the planet orbiting one of the stars must satisfy, at any instant, the following condition
$$\text{re of} \quad a _ { p } ( 1 + e _ { p } ) < R _ { c r },$$
where a p and e p are the planetary osculating semimajor axis and eccentricity, with respect to the corresponding central star, and R cr is the radius of the sphere of stability of this star. The interpretation of the expression (5) is simple: since the orbital evolution of the planet is confined to the inside of the circumference of radius R cr , even at the most distant position from the central star (at apocenter), the stability of this motion is guaranteed by the deficit of the relative energy (1) needed to cross the zero-velocity level holding the L 1 solution. It is worth noting that, due to the restricted threebody approximation, the size of the domains of stability is defined by solely the physical and orbital parameters of the stars.
Hereafter we will refer to our criterion as Maximum Apocentric Distance or MAcD-criterion. It assures that planetary orbits that exceed the critical distance with respect to the central star (equivalent to the radius of the sphere of stability of this star) will be unstable. However, it is worth stressing that this condition does not guarantee the long-term (of order of the lifetime of the stars) stability of the planet motion. The secular dynamics of the system and its stability will be discussed in Section 3.
## 2.2 Application to the α Centauri AB System
In this section, we apply the MAcD-criterion to analyze the stability of planetary S-type orbits in the α Centauri AB binary.
The first step is to determine the radii of the spheres of stability of the stars A and B. In the case when the star A is a central star and the star B is a perturber, the stellar mass ratio is µ = m /m B A = 0 849 . < 1 0 and the expression (3) . must be used. In the case when the central star is B and the perturbing star is A, µ = m /m A B = 1 178 . > 1 0 . and the
expression (4) is applied. As a result, for given values of the semimajor axis and eccentricity of the relative stellar orbit, a = 23 4 AU and . e = 0 518, we obtain the critical distances . for the stars A and B as R A cr = 4 7 AU and . R B cr = 4 4 AU, . respectively. The spheres of stability of the stars A and B are shown in Figure 1, where m 1 = m A and m 2 = m B .
In the following step, we numerically integrate the system, in order to determine the stability of the planetary orbit defined by the condition (5). In this stage, the choice of the suitable time-length is important: it should be neither too large to raise the computational costs, nor too small to lose information on the proper oscillations in the planetary motion. These oscillations are usually separated in the short-period terms, associated with the Keplerian motions of the planet and the perturbing star, and the long-period secular terms, associated with the precession of the planetary orbit. Eventually, depending on the ratio of the mean motions of the planet and the perturbing star, there are also mean-motion resonance terms whose periods are intermediate lying between the short and the long periods; the resonant conditions will not be studied in this paper.
Although the MAcD-criterion is applied to assess the short-term stability of the planet motion, the long-period secular terms are still important. This is due to the fact that, during the precession, the pericenter line of the planetary orbit becomes, at some instant, aligned with the pericenter line of the orbit of the binary. If, at this instant, the perturbing star is at the pericenter of its relative orbit, the short-period perturbations become more intense and instabilities may appear. Thus, the choice of the integration timespan must account for the proper secular period of the planetary motion.
To determine the secular period, we apply the Dynamical Power Spectrum method (Michtchenko et al. 2002; Ferraz-Mello et al. 2005). This method is based on the Fourier decomposition of oscillations of the orbital elements and provides an information on the proper frequencies of the system under study and their evolution with changing parameters. In our study, we have chosen the semimajor axis of the planet as a parameter, which varied in the range between 0.5 AU and 3.5 AU. The proper frequencies of the S-type planet orbiting α Centauri A obtained for each value of planetary semimajor axis over 10 5 years, are presented in Figure 2. We can clearly observe the presence of the three distinct frequency bands associated to three degreesof-freedom of the problem. The first band, located in the high-frequency interval -2 0 . < log( f ) < 0, is composed of the orbital frequency of the planet, f p , and its linear combinations with other proper frequencies. According to the third Kepler's law, this frequency decreases with decreasing a p . The second band is confined to the very close neighborhood of log( f ) = -1 9 . and corresponds to the orbital frequency of the binary relative orbit, f star .
The third band is located in the low-frequency interval -5 < log( f ) < -3 and corresponds to the secular frequency, f ∆ /pi1 . This frequency tends to increase with increasing a p , due to increasing perturbations on the planetary orbit from the stellar companion. The overlap of the bands occurs at a p > 2 5 AU, introducing the long-term instabilities in the . planetary motion. The dense vertical lines visible already at a p > 2 0 AU in Figure 2, are associated with the mean-.
motion resonances in this region. Finally, for a p > 3 0 AU, . mostly all orbits are disrupted during 10 5 years.
From Figure 2, we obtain that the secular period of the planet in the α Centauri system varies approximately from 10 3 years up to 10 4 years. Assuming the integration timespan within this range, we can now numerically integrate the planetary path, in order to verify the stability of the system defined by the condition (5).
The quality of the prediction obtained with the MAcDcriterion can be verified by constructing dynamical maps around each star of the α Centauri binary. For this, we use a grid of 384 × 100 initial conditions on the ( a , e p p ) representative plane and a terrestrial planet of mass m p = 5 × 10 -5 M /circledot . Without loss of generality, we fix the initial values of the mean longitudes at λ A = λ B = 0 ◦ and of the longitudes of pericenter at ∆ /pi1 = /pi1 A,B -/pi1 p = 0 ◦ and 180 . Each ini-◦ tial condition is numerically integrated using two different timespans: one is equal to 10 3 years ( ∼ 12 5 orbital periods . of the binary) and the second is equal to 10 4 years ( ∼ 125 orbital periods of the binary), which cover, respectively, ∼ 1 and ∼ 10 secular periods at a p ≈ 3 AU.
The solutions obtained are then analyzed applying the MAcD-criterion, in order to calculate the maximal apocentric distances that the planet reaches over the assumed timespan. The values obtained are presented using the grey scale code in Figure 3, where the domains of the strongly unstable motion characterized by collisions and escape of the planet are dashed. On the planes we also plot, by the red curves, the levels whose values correspond to the radii of the spheres of stability, obtained using Equations 3 and 4 as 4 7 AU, for the star A, and 4 4 AU, for the star B. . .
As expected, the apocentric distances increase with the increasing planetary semimajor axis and decay with the increasing eccentricity. In the domains of the regular motion, the levels are smooth functions of the orbital elements of the planet. This behaviour is changed when the initial conditions approach the domains of the unstable motion, whose boundaries are defined, according to our MAcD-criterion, by the critical distances shown by the red curves in Figure 3. In this case, the critical levels show erratic behaviour, characteristic of chaotic orbits. It is worth noting the good agreement between the positions of the critical levels (red curves) and the boundaries of the domains of instabilities provided by purely numerical simulations. This confirms the quality of our criterion predictions. Moreover, the comparison of the topology of the dynamical maps obtained over 1 secular period of the planets (top row in Figure 3) and those obtained over 10 secular periods (bottom row), shows no significant differences. This means that the time-length of 1 secular period is a good choice for the application of the MAcD-criterion.
## 2.3 Comparison with the Holman & Wiegert (1999) criterion
A similar criterion was proposed in Holman & Wiegert (1999): it states that, for planetary motion around one of the stars to be stable, the initial value of the semimajor axis of the planet must obey a relationship a p < F cr ( a, e, µ ), where F cr is a function of the physical and orbital characteristics of the stars. By means of a series of numerical integrations, the authors have found an empirical expression for F cr as
Figure 3. Dynamic maps around α Centauri A (left panels) and B (right panels): short-term stability in the planar case. The grey scale code represents the maximal (apocentric) distance of the planet from the central star during 10 3 yr (top panels) and 10 4 yr (bottom panels) of evolution; the bars relate the grey tones to the values of the apocentric distances (in astronomical units). The positive and negative values along the x -axis indicate that the initial values of ∆ /pi1 are 0 ◦ or 180 ◦ , respectively. The hatched regions correspond to the ejected from the system planets. The red curves correspond to levels of the radii of the stability spheres ( R A r c = 4 7AU and . R B r c = 4 4 AU). The blue curves show the critical distances given by the Holman & Wiegert (1999) criterion of stability, while the cyan . curves show the boundaries of the corresponding grey areas (see text for details).
$$\begin{array} { c } F _ { c r } ( a, e ) / a = ( 0. 4 6 4 \pm 0. 0 0 6 ) + ( - 0. 3 8 0 \pm 0. 0 1 0 ) \mu + \\ + ( - 0. 6 3 1 \pm 0. 0 3 4 ) e + ( 0. 5 8 6 \pm 0. 0 6 1 ) \mu e + & & ( 6 ) & & \text{corr} \\ + ( 0. 1 5 0 \pm 0. 0 4 1 ) e ^ { 2 } + ( - 0. 1 9 8 \pm 0. 0 4 7 ) \mu e ^ { 2 }, & & \text{clear} \\ & & & & & & \text{time} \end{array}$$
where µ = m / m B ( A + m B ) and a and e are the semimajor axis and the eccentricity of the stellar relative orbit, respectively. The approximation is valid in the range 0 1 . /lessorequalslant µ /lessorequalslant 0 9, . for 0 0 . /lessorequalslant e /lessorequalslant 0 8, . with the error smaller than 11% (as claimed by the authors). In their approach, the authors assumed that the planet orbits were initially circular, which means that the results obtained using Equation (6) are valid only along the y -axis in Figure 3. For purpose of comparison with our criterion predictions, we re-write the Holman & Wiegert (1999) criterion as a p (1 + e p ) < F cr and apply it to the α Centauri binary.
The result is shown by the blue curves in Figure 3. The correlation with the MAcD-criterion is notable, but it is clear that the Holman & Wiegert (1999) criterion overestimates the instabilities in the planet motion, in this way, excluding from the study a significative portion of still stable configurations. It is worth noting that the Holman & Wiegert (1999) criterion (6), also as other empirical criteria (e.g., Rabl & Dvorak 1988, Pilat-Lohinger & Dvorak 2002, Fatuzzo et al. 2006, among others), introduces a concept of grey area in the parameter space of the problem, in which the state of a system changes from stability to instability (Haghighipour 2010, Chapter 11). The grey area is defined as the confidence interval which measures the reliability of the estimation obtained from the fitting procedure. Applying the formula (6) to the α Centauri system and us-
ing the traditional law of error propagation (Ku 1966), we obtain the confidence intervals for the stars A and B as F A cr = 2 80 . ± 0 64 AU and . F B cr = 2 54 . ± 0 69 AU, respec-. tively. The borders of the corresponding grey areas on the ( e p , a p )-plans are shown by the cyan-colored curves in Figure 3. However, it must be emphasized here that the grey area idea does not apply to the MAcD-criterion (5), which is founded on the basic concept of the Hill stability and cannot be interpreted in statistical terms. Due to short time intervals of numerical integrations, the method can still originate effects analogous to 'grey areas' (seen as red-colored zones in Figure 3), but these zones rapidly decrease with increasing integration timespans.
Also, it is worth noting here that the purposes of both methods are different. The purpose of the Holman & Wiegert (1999) criterion is a prompt estimation of the stability of one initially circular configuration of a putative planet in the binary, given the distance of the planet to the central star. Notwithstanding the facility in implementation (it does not require numerical integrations), this method still suffers on some limitations. One of these is due to assumption that the planet starts on zero-eccentricity orbit. Due to very strong gravitational perturbations produced by the stellar companion on the planetary motion, this assumption can scarcely be correct.
Another limitation is related to the detection of the planets in binary systems and consequent determination of their orbits through the best-fitting procedures. It seems unlikely that the sets of the osculating orbital elements of the planets provided will be compatible with the circular orbits needed in applying the Holman & Wiegert (1999) criterion. Our MAcD-criterion can overcome this limitation, since it is based on the implementation of numerical integrations which use the sets of the orbital parameters of the planets as input.
## 2.4 Extension to the 3D
The MAcD-criterion can be easily extended to the inclined to the reference plane orbits. Indeed, since the maximum distance of the planet from the central star in space is still its apocentric distance, we can apply the same condition of stability a p (1+ e p ) < R cr , where R cr is a radius of the sphere of stability given by Equations (3), for the star A, and (4), for the star B.
To test the stability of the inclined planetary orbits, we construct the dynamical maps, each with a 400 × 400 grid of initial conditions, on the ( a p , I p ) representative planes, with the planetary eccentricity fixed at e p = 0 05, . again with the planetary mass fixed at m p = 5 × 10 -5 M /circledot . Our previous study (Michtchenko et al. 2006) pointed out that the initial angular conditions of the secular angle ∆ /pi1 and the argument of pericenter 2 ω p can be fixed at either 0 ◦ or 180 ◦ , without loss of generality. The maps obtained are shown in Figure 4.
We can see that the boundaries of the stable motion are well defined by the MAcD-criterion (red levels), when the apocentric distances approach the critical values R A cr = 4 7 AU and . R B cr = 4 4 AU. The maximum values of . the planetary semimajor axis, for which the motions still remain stable, are ≈ 3 5 AU and . ≈ 3 AU, for the stars A and B, respectively. For smaller values of the semimajor axis, there are zones of regular motion for any values of the planetary inclination. The stable motion is robust at inclinations below ∼ 40 . Above this inclination, we can detect sporadic ◦ chaotic solutions (red dots) deeply inside the region of the stable motion. This is an indicator that the analysis of the short-term stability is not sufficient in this case and must be extended to a longer length-times. This will be done in the next section, where we will show that the instabilities at I p > 40 ◦ are originated by the Lidov-Kozai secular resonance.
The method described in this section provides a global view of the short-term stability of the system. It is conceptually simple, fast and easily implementable. It allows us to detect promptly the domains of regular motion, avoiding, in this way, long-term integrations and analysis of unstable orbits. However, due to the analysis of orbital behaviour of the planets over short timespans, this approach provides no information on the long-term (of order of the lifetime of the stars) stability of the system. The specific study on the stability of the planetary motion in the binary over the age of the system will be done in the next section.
## 3 LONG-TERM DYNAMICAL STABILITY
Under a 'long-term' stability we mean the stability of the system over the lifetime of the central star. One way to assess the long-term stability of the planet motion is direct Nbody numerical simulations, which, depending on the stellar types and ages, must be done over several billions of years. This approach suffers from several limitations and the most serious from these is extremely high computing costs.
An alternative is the dynamical mapping of the domain under study. This approach is also founded on purely numerical integrations, but the time intervals of the simulations are much shorter in this case, and, therefore, we are allowed to explore a large set of initial conditions. Analyzing the topological structure of dynamical maps, we can detect the mechanisms which could induce long-term instabilities in the planetary motion, and determine domains of their action.
The topology analysis is based on the analytical modeling of the general three-body problem briefly described in the following (for more details, see Michtchenko et al. 2006). In the Jacobian reference, the Hamiltonian which describes the motion of the planet ( m p ) around one of the stars ( m A ), with the second star ( m B ) as the disturber, is written as (Brouwer & Clemence 1961)
$$\underset { \text{Our} } { \text{gain} } \quad \mathcal { H } = \mathcal { H } _ { \text{Kep} } + \mathcal { H } _ { \text{Pert} },$$
where:
$$\underset { \alpha _ { p } ^ { \circ } } { \text{and} } \quad \mathcal { H } _ { \text{Kep} } = - \frac { G \, m _ { p } \, m _ { A } } { 2 \, a _ { p } } - \frac { G \, m _ { B } \, ( m _ { A } + m _ { p } ) } { 2 \, a _ { B } }, \quad \text{and}$$
$$\intertext { m o - } \intertext { \mathcal { H } } _ { \text{Per} } = - \frac { G \, m _ { p } \, m _ { B } } { r _ { p B } } - G \, m _ { A } \, m _ { B } \left ( \frac { 1 } { r _ { A B } } - \frac { 1 } { r _ { 2 } } \right ), \quad \ \ ( 9 )$$
and a i is the osculating semimajor axis of i -th orbit, /vector r B is the position vector of m B relative to the center of mass of m A and m p and /vector r iB is the distance between m i and m B ; G is the gravitational constant. Hereafter, the indices i = A,p,B stand for the central star, planet and perturbing star orbits, respectively.
Figure 4. Same as in Figure 3, except in the 3D case, with a p and I p being the semimajor axis and the inclination of the planet and the initial eccentricity e p fixed at 0 05. The positive and negative values on . x ( y )-axis indicate that the initial ∆ /pi1 (2 ω p ) are 0 ◦ or 180 ◦ , respectively.
Associated with the Keplerian part of the Hamiltonian, a set of mass-weighted Delaunay's elliptic variables is introduced as
$$\begin{array} { c c } M _ { i } = \text{mean anomaly}, & L _ { i } = m _ { i } ^ { \prime } \sqrt { \mu _ { i } a _ { i } }, & \text{ingly} \\ \omega _ { i } = \arg u n t { \text{ of epicenter}, \quad } G _ { i } = L _ { i } \sqrt { 1 - e _ { i } ^ { 2 } }, & ( 1 0 ) & \text{The} \\ \Omega _ { i } = \text{longitude of node}, & H _ { i } = L _ { i } \sqrt { 1 - e _ { i } ^ { 2 } } \cos I _ { i }, & \text{max} \\ \text{integ} \end{array}$$
where e i and I i are the eccentricities and inclinations, respectively; µ p = G m ( A + m p ), µ B = G m ( A + m p + m B ), m ′ p = m m / m p A ( A + m p ) and m ′ B = m B ( m A + m / m p ) ( A + m p + m B ).
$$\frac { \imath _ { i } ^ { \prime } \sqrt { \mu _ { i } a _ { i } } } { \imath \sqrt { 1 - \imath } } \\ \frac { \sum ^ { 1 } \infty } { \sim } \, \text{and in} \, \sum ^ { 1 } \infty$$
We assume that the planet and the perturbing star are not involved in any mean-motion resonance of low order. Thus we can perform numerically the averaging of the Hamiltonian (7) with respect to the mean anomalies of the planet and the disturber (Michtchenko & Malhotra 2004). The secular Hamiltonian is then defined by
$$\overline { \mathcal { H } } _ { s } = - \frac { 1 } { ( 2 \pi ) ^ { 2 } } \int _ { 0 } ^ { 2 \pi } \int _ { 0 } ^ { 2 \pi } R ( L _ { i }, G _ { i }, H _ { i }, M _ { i }, \omega _ { i }, \Omega _ { i } ) \, d M _ { p } d M _ { B }, ( 1 1 ) \quad \text{the corr} \quad \\ \quad \text{The clos}$$
where the Keplerian part is constant and therefore needs not be considered.
## 3.1 Planar case
The results are shown in Figure 5. The grey color scale is used to plot the spectral number: light regions correspond to small numbers and regular motions, while the increasingly dark tones indicate large numbers and unstable orbits. The hatched patterns correspond to the domains where the MAcD-criterion of stability is not satisfied over the chosen integration timespan (see Sect. 2.1). The blue curves show the families of the stable stationary solutions of the secular problem given by the Hamiltonian model (11), frequently referred to as Mode I of motion (Michtchenko & Ferraz-Mello 2001); these solutions are also known as forced eccentricities of the secular problem (Heppenheimer 1978; Giuppone et al. 2011). In the planar case, the secular problem is reduced to one degree of freedom and the secular solutions are obtained by plotting the energy and angular momentum levels on the representative plane (for a detailed description, see Michtchenko & Malhotra 2004). The secular planetary motion projected on the ( e , a p p )-plane is an oscillation around the corresponding equilibrium belonging to the blue curve. The closer the initial condition is to the Mode I solution, the smaller is the amplitude of the oscillation of its eccentricity. This fact explains the very stable, nearly harmonic, dynamics of the planet in the vicinity of the blue curve (light grey region), even at larger values of a p .
In this section, we construct the dynamical maps of coplanar motions, using a grid of 392 × 100 osculating initial conditions on the ( e p , a p )-plane and fixing the initial angular variables at M p = M A,B = 0 and ∆ /pi1 = 0. Each initial condition was integrated over 10 4 years ( ∼ 125 orbital periods of the binary or ∼ 10 secular periods, see Figure 2), with the planetary mass fixed at m p = 5 × 10 -5 M /circledot . The solution obtained was Fourier analyzed, in order to calculate the spectral number corresponding to the used initial condition. The spectral number defines the degree of chaos in the planetary motion: its small values correspond to regular motions, while its larger values correspond to unstable motions (Michtchenko et al. 2002; Ferraz-Mello et al. 2005).
However, it is worth stressing that the initial conditions used in construction of the dynamical maps are osculating orbital elements; thus, the comparison of the results of numerical integrations to those provided by the secular model (11) should be done carefully. Moreover, the mass of the disturbing body is comparable to the mass of the central star; this fact reduces the domain of the application of the firstorder in masses secular model (11). We plot in Figure 5, by the red curves, the families of the secular stationary solutions obtained through the purely numerical filtering procedure, as described in Giuppone et al. (2011). The deviation of the analytically obtained family from that obtained numerically (with no restriction on the perturbing mass) clearly occurs at larger semimajor axes, when the perturba-
Figure 5. Dynamic maps around α Centauri A (left panel) and B (right panel): long-term stability in the planar case. The grey scale code represents the spectral number (stability indicator) in logarithmic scale: the lighter tones correspond to regular motion, while the darker tones correspond to increasingly chaotic motion. The positive and negative values along the x -axis indicate that the initial values of ∆ /pi1 are 0 ◦ or 180 ◦ , respectively. The hatched patterns represent ejected or collided orbits. The blue curves are families of stationary solutions provided by the secular semi-analytical Hamiltonian (11), while the red curves are families obtained through numerical averaging (Giuppone et al. 2011). The right-hand axes on both planes are the planetary semimajor axis in unit of the semimajor axis of the binary ( a /a p B ).
tions due to the stellar companion become stronger. Since, in this region, the planetary motion is stable only in the very narrow vicinity of the Mode I, the correct determination of the family is essential in this case.
The two families match each other at small semimajor axes of the planet, up to ∼ 1 5 AU . (or a /a p B < 0 06 . ), for small eccentricities. The dynamical maps in Figure 5 show that the planetary dynamics in this case is very regular (white color), even at very-high eccentricities. It is worth noting that the HZs around the stars are located close to these regions (see Section 4).
bust and stable, even at large distances from the central star, this is a favorite location of several known planetary extra-solar systems (e.g., υ Andromedae c-d planet pair). Michtchenko & Rodr´ ıguez (2011) shows that, during migration, the planets are guided toward the stationary configurations, independently on the specific migration mechanism. Moreover, Giuppone et al. (2011) shows that the planet formation is preferential on the orbital configurations corresponding to the Mode I stationary solutions.
When the semimajor axis of the planetary orbit increases, the perturbations due to the stellar companion become stronger and the planetary motion increasingly nonharmonic (except in the vicinity of the Mode I of motion). The dark tones on the dynamical maps in Figure 5 confirm this fact. Several high order mean-motion resonances, from 13/1 to up 20/1, are dense in these regions; they appear as thin horizontal strips on the dynamical maps. The phenomenon can be also observed in the power dynamical spectrum in Figure 2, where the resonant initial conditions appear as vertically scattered dots, at a p > 2 AU. Since these resonances can destabilize the planetary motion, the longliving planets are not expected to be found in this region.
Finally, at relative distances above 0.14 (in units of the semimajor of the binary orbit), the planetary motion is strongly chaotic and the planets are ejected from the binary (hatched regions). The boundary of these domains match the boundaries of short-term instabilities showed in Figure 3.
It is worth to emphasize the features of the planet motion near the Mode I stationary solutions. Being ro-
## 3.2 Three-dimensional case
We construct the dynamical maps on the plane ( e p , I p ) of the initial osculating values of the planetary eccentricity and the inclination, with a terrestrial planet of mass m p = 5 × 10 -5 M /circledot . The spectral analysis method was applied on the 201 × 201-grid, with the spacings ∆ e p = 0 004 and ∆ . I p = 0 8 . ◦ . The initial semimajor axis of the planet was fixed at a p = 1 0 AU; for this value, the planet is distant from the . perturbing star and the first-order Hamiltonian model (11) can be used to investigate the topology of the dynamical maps.
The maps are shown in Figure 6. The positive and negative values on the x y ( )-axis stand for the initial values of the angle ∆ /pi1 p (2 ω p ) of 0 ◦ and 180 , ◦ respectively (see Michtchenko et al. 2006, for more details). The power dynamical spectrum in Figure 2 allows us to estimate the secular period of the planet at a p = 1AU as ∼ 10 4 years. This means that an appropriate timespan for simulations of the secular dynamics must be of order of at least 10 5 years. The chosen timespan of integration for our dynamical map was 500,000 years ( ∼ 6000 orbital periods of the binary), for
Figure 6. Same as in Figure 5, except the long-term stability in the 3D case, with e p and I p being the eccentricity and the inclination of the planet and the initial semimajor axis fixed at 1 0 AU. The positive and negative values on . x ( y )-axis indicate that the initial ∆ /pi1 (2 ω p ) are 0 ◦ or 180 ◦ , respectively. The blue and red curves represent the stable and unstable periodic solutions of the Hamiltonian (11), respectively.
each initial condition on the grid; the orbits disrupted during this time are shown by the hatched pattern in Figure 6. The surviving orbits are shown by the grey scale code, from light tones, for nearly periodic orbits, to darker tones, for increasingly unstable orbits.
the red curves. In this case, the secular angle ∆ /pi1 librates either around 0 or 180 ◦ .
The two color curves represent the periodic solutions obtained through the secular model (11). The blue curves are location the of the stable periodic solutions, while the red curves are locations of the unstable solution. The longliving motions are expected to be found near the blue curves and far from the red curves. Indeed, the analysis of the structure of the dynamical maps of the representative planes reveals that, for low inclinations I p < 40 , the maps are dom-◦ inated by a light grey background, which means that the low-inclination secular dynamics is regular.
In contrast, the vicinity of the red curves is dominated by the darker tones, indicating unstable motions. The instability is related to the existence of separatrices between different regimes of motion. Their domains, at high inclinations, show complex dynamical behaviour, with the presence of several regimes of resonant motion (the detailed description of these regimes of secular motion can be found in Michtchenko et al. 2006). The dominating behaviour is the e p -I p coupling, or Lidov-Kozai resonance, which destabilizes the nearly circular orbits at I p > 40 . For higher values ◦ of eccentricity, the planetary dynamics is characterized by the coupled large variation of the eccentricity and inclination and the libration of the angle ω p around ± 90 . ◦ The domains of the Lidov-Kozai resonance are located on the lower half-planes in Figure 6; the behaviour of the system is very regular in these regions.
The ∆ /pi1 secular resonance also exists in the highinclination regions; its domains are thin white strips along
It is worth emphasizing that we have detected a large portion of the chaotic orbits (black tones on the graphs in Figure 6), but only a small quantity of these orbits are actually ejected from the system (dashed regions). This is because the chaotic processes are slow in the secular system and the timespan of 500,000 years is not sufficient for detecting the planetary escapes. However, the escapes are imminent and, as a consequence, it will be impossible to detect the planets inside these regions of high instabilities.
## 4 DYNAMICAL EVOLUTION INSIDE THE HZ
Regardless of the robust long-term stability of the planetary motion in the large domains around the stars of α Centauri, the planetary evolution inside the HZs, should be analyzed, in order to verify whether the planet motion is confined inside the HZ. The concept of the HZ itself, even when introduced for single stars (e.g., Kasting et al. 1993), is very complex. In binary stars, the definition of the HZ is further complicated due to several additional aspects. One is the combined radiation of both stellar components on the terrestrial planet studied recently in several papers (e.g., Forgan 2012, Kaltenegger & Haghighipour 2013).
Another aspect concerns the large oscillations of the orbital elements of the planet, that could affect the conditions of habitability of the planet or even drive the planet far from the HZ. This behaviour was studied in Eggl et al. (2012, 2013), where the authors considered the averaged effects produced by the secondary star on the variation of the eccentricity of the planet, yet in the case of coplanar
Figure 7. Maps of the orbital distances of the planets around the star α Centauri A (left) and the star B (right). The grey scale represents levels of the distance amplitude oscillation of the planet for different initial values of eccentricity and inclinations. The initial semimajor axis is fixed at a p = 1 6 AU, for . α Centauri A, and a p = 0 9 AU, for . α Centauri B. The yellow, red, blue and green curves represent the width of the HZ given by Forgan (2012), Eggl et al. (2013) and Kaltenegger & Haghighipour (2013), narrow and empirical.
Table 1. The inner and outer boundaries of the HZs (in AU) and the HZs width in the α Centauri AB binary, calculated in Forgan (2012), Eggl et al.(2013) and Kaltenegger & Haghighipour (2013) (KH13) (conservative and optimistic definitions).
width of the HZs, in order to infer whether the planet remain within the HZ boundaries over 10,000 years ( ∼ 125 orbital periods of the binary).
| Authors | Star | Inner HZ | Outer HZ | Width |
|--------------------|--------|------------|------------|---------|
| Eggl et al. (2013) | A | 1.12 | 1.81 | 0.71 |
| KH13 conservative | A | 1.2 | 2.07 | 0.87 |
| KH13 optimistic | A | 0.92 | 2.18 | 1.3 |
| Forgan (2012) | B | 0.65 | 0.9 | 0.25 |
| Eggl et al. (2013) | B | 0.65 | 1.13 | 0.48 |
| KH13 conservative | B | 0.71 | 1.26 | 0.55 |
| KH13 optimistic | B | 0.54 | 1.32 | 0.78 |
and initially circular planetary orbits. In this work we explore the conditions for which the planets remain inside the HZ, in the case of eccentric and inclined orbits. For this, we use the different definitions of the HZs around the α Centauri stars given in the papers Forgan (2012), Eggl et al. (2013) and Kaltenegger & Haghighipour (2013)(see Table 1). It should be emphasized that these definitions are not equally weighted. In fact, the KH13 definition can be considered more realistic because it takes into account the presence of an atmosphere on a terrestrial or super-Earth planet. However, the detailed discussion on the HZ calculations is out of the scope of this paper.
For each definition from Table 1, we calculate the maximum and minimum distances of the planet from the central star, for different initial eccentricities and inclinations of the planet orbit and the initial semimajor axis fixed at a p = 1 6 AU, around . α Centauri A, and a p = 0 9 AU, around . α Centauri B; these values roughly correspond to the averaged positions of all HZs defined in Table 1.
The numerical test is done using numerical integrations of planetary orbits over 10,000 years. The values of the maximal and minimal distances obtained are compared to the
The difference between the maximal and minimal distances of the planet from the central star is plotted on the ( e p , I p )-planes in Figure 7, using the grey scale code: the difference is smaller in the light regions and larger in the dark regions. The regions in white tone correspond to the minimal possible deviations of the planet from the center of the HZ. The simple comparison with the graphs in Figure 6 allows us conclude that these regions are associated with the Mode I secular solutions whose location is given by blue curves in Figure 6.
An interesting feature of this analysis is that initially circular orbits exhibit significant deviations from the center of the HZ, that, for high inclinations, could result in the planets spending some time outside the HZs.
As discussed in Eggl et al. (2012), the planet remaining inside the HZ during its evolution is not the only requirement for the planet to host life, and it is neither an excluding factor. Even if the planet spends some time of its orbit outside the HZ, as long as the mean radiation received is still within the HZ limits, the planet could be considered habitable, although in a more restrict way. Nevertheless, it is expected that orbits with smaller oscillations are able more likely to sustain some kind of life.
It is also interesting to compare the orbital evolution of a terrestrial S-type planet in α Centauri AB with the evolution of the terrestrial planet in a planetary system, with a Jupiter-sized planet. The comparison is done using the configurations defined by the orbital parameters of α Centauri AB, except, in the case of the planetary system, the binary companion is replaced by a Jupiter-mass planet. The results are shown in Figure 8, where the variation of the orbital distance of the planet revolving around α Centauri B is plotted by the red curve, in the case when the perturber is the star α Centauri A (1 1 . M /circledot ), and by the black curve, when the
Figure 8. Time evolution of the planetary distance to α Centauri B when the perturber is α Centauri A (red) and a Jupiter-mass planet (blue).
perturber is a hypothetical Jupiter-mass planet (10 -3 M /circledot ). In both simulations, the same orbital and physical parameters of the planet and perturbing body were used, except the mass of the perturber. The initial semimajor axis and the eccentricity of the planet were fixed at a p = 0 9 AU (in-. side the HZ of the star) and e p = 0 016, respectively; note . that this value of the eccentricity corresponds to the actual eccentricity of Earth. The initial orbital elements of the perturber were the same of the relative orbit of the α Centauri binary.
We can observe in Figure 8 that the amplitudes of the oscillation of the orbital distance of the planet are of the same order of magnitude, despite that the perturbing masses differ in order 3. This result is consistent with the theoretical predictions of the restricted three-body models (e.g., Heppenheimer 1978) which state that the forced eccentricity of the test particle is independent on the masses of the large bodies (at least, in first order in masses). Therefore, the orbital distance oscillation of the planet in the binary system is similar to that of a terrestrial planet perturbed by a giant planet. The main difference is the timescale of the oscillation: it is of order of 10 4 years, in the binary system, and of order of 10 7 years, in the planetary system. However, this difference seems to be not important in the context of the Continuously Habitable Zone.
In addition, we have obtained that, in the case of the real Earth-Jupiter system, the forced eccentricity of the Earth's orbit is equal to 0 016; . in the case of a putative terrestrial planet orbiting α Centauri B, the forced eccentricity of the planetary orbit is equal to 0 032. Calculating . the variation of the orbital distances in both cases, we have found that its amplitude is ∼ 0 032 AU, for the Earth's or-. bit, and ∼ 0 07 AU, for the planet located near the Mode I . stationary solution of the HZ of α Centauri B, that is, both variations are of same order of amplitude. This result brings positive perspectives of finding habitable planets in binary star systems.
## 5 SUMMARY
In this work, we present a methodology to analyze the planetary dynamics of S-type orbits in binary star systems. We develop a short-term stability criterion (MAcD-criterion), which is based on the Hill concept of stability. Since this method does not require any kind of expansion and averaging of the disturbing function, there are no constrains on neither the mass of the disturbing body nor the eccentricity/inclination of its orbit. We apply our approach to the α Centauri binary system with a hypothetical Earth-like planet, orbiting one of the stars (S-type orbit), and show that the purposed criterion provides a global view of the stable and unstable domains in the space of the orbital elements. It is worth emphasizing that our approach is simple, fast and easily implementable. Its main purpose is to detect promptly the domains of regular motion, avoiding, in this way, long-term integrations and analysis of unstable orbits. It also allows us to quickly assess the stability of the best-fit solutions in the determination of the planetary orbits hosted by binary stars.
The long-term stability of the planetary orbits in the HZs of the α Centauri binary system is studied using the secular Hamiltonian approach. We show that the behaviour of the planets is well described by the first-order in masses model. The secular planetary motion is regular, even at high eccentricity domains of the HZs. There are no significant mean-motion resonances that could destabilize the secular dynamics. The stability of the planets is also detected through the 3D-analysis, at low and moderate inclinations of the planetary orbits ( < 40 ). For higher inclinations, the ◦ dominating regime of motion is the Lidov-Kozai resonance, which is separated from the purely secular regime of motion through the layers of strongly chaotic motion. Even though there are domains of very stable motion in the Lidov-Kozai resonance, the large excursions of the planetary eccentricities and inclinations seem to be unfavorable for the evolution of life inside the HZs. Beyond the HZ, where high-order mean-motion resonances take place, the first-order model fails. Thus, a second-order in masses model, as well as the modeling of the resonant behaviour, would be required to further study of this region.
Analyzing the dynamics of the planets evolving inside the HZs of the α Centaury stars, we obtain that the planetary orbits will remain within the boundaries of the HZs, if their eccentricities and inclinations are constrained to values close to the stationary solutions of the secular problem, the Mode I solutions. In this favorable scenario, the amplitude of the orbital distance variations are comparable to the amplitudes of the Earth's orbit, with the only difference being the timescales of the variations. The real effects of this on the planetary climate and geological conditions are still poorly known and it is expected that further interdisciplinary studies could bring to light considerable new information on this issue.
## 6 ACKNOWLEDGMENTS
This work was supported by the S˜o Paulo State Science a Foundation, FAPESP (grant 2010/01209-2), and the Brazilian National Research Council, CNPq. This work has made
use of the facilities of the Computation Center of the University of S˜ ao Paulo (LCCA-USP) and of the Laboratory of Astroinformatics (IAG/USP, NAT/Unicsul), whose purchase was made possible by the Brazilian agency FAPESP (grant 2009/54006-4) and the INCT-A. The authors are grateful to Dr. Haghighipour for numerous helpful suggestions/corrections on this paper.
## REFERENCES
Abt, H. A. 1979, Astronomical Journal, 84, 1591
- Barbieri, M., Marzari, F., & Scholl, H. 2002, Astronomy and Astrophysics, 396, 219
Black, D. C. 1982, Astronomical Journal, 87, 1333
- Boss, A. P. 2006, Astrophysical Journal, 641, 1148
- Brouwer, D., & Clemence, G. M. 1961, New York: Academic Press, 1961,
- Dumusque, X., Pepe, F., Lovis, C., et al. 2012, Nature, 491, 207
- Duquennoy, A., & Mayor, M. 1991, Astronomy and Astrophysics , 248, 485
- Dvorak, R., Froeschle, C., & Froeschle, C. 1986, Bulletin of the American Astronomical Society, 18, 842
- Eggl, S., Pilat-Lohinger, E., Funk, B., Georgakarakos, N., & Haghighipour, N. 2013, Monthly Notices of the Royal Astronomical Society, 428, 3104
- Eggl, S., Pilat-Lohinger, E., Georgakarakos, N., Gyergyovits, M., & Funk, B. 2012, Astrophysical Journal, 752, 74
Everhart, E. 1985, Dynamics of Comets: Their Origin and Evolution, Proceedings of IAU Colloq. 83, held in Rome, Italy, June 11-15, 1984. Edited by Andrea Carusi and Giovanni B. Valsecchi. Dordrecht: Reidel, Astrophysics and Space Science Library. Volume 115, 1985, p.185, 185 Fatuzzo, M., Adams, F. C., Gauvin, R., & Proszkow, E. M. 2006, PASP, 118, 1510
- Ferraz-Mello, S., Michtchenko, T. A., Beaug´, C., & Callee gari, N. 2005, Chaos and Stability in Planetary Systems, 683, 219
- Ford, E. B., Kozinsky, B., & Rasio, F. A. 2000, Astrophysical Journal, 535, 385
- Forgan, D. 2012, Monthly Notices of the Royal Astronomical Society, 422, 1241
- Giuppone, C. A., Leiva, A. M., Correa-Otto, J., & Beaug´, e C. 2011, Astronomy and Astrophysics , 530, A103 Gladman, B. 1993, Icarus, 106, 247
- Graziani, F., & Black, D. C. 1981, Astrophysical Journal, 251, 337
- Guedes, J. M., Rivera, E. J., Davis, E., et al. 2008, Astrophysical Journal, 679, 1582
- Haghighipour, N. 2006, Astrophysical Journal, 644, 543
- Haghighipour, N. 2010, Astrophysics and Space Science Library, 366
- Haghighipour, N., Dvorak, R., & Pilat-Lohinger, E. 2010, Astrophysics and Space Science Library, 366, 285
- Haghighipour, N., & Raymond, S. N. 2007, Astrophysical Journal, 666, 436
- Hatzes, A. P. 2013, Astrophysical Journal, 770, 133 Heppenheimer, T. A. 1978, Astronomy and Astrophysics , 65, 421
Hill, G. W., 1878, Am. J. Math. 1-5, 129
- Holman, M. J., & Wiegert, P. A. 1999, Astronomical Journal, 117, 621
- Kaltenegger, L., & Haghighipour, N. 2013, Astrophysical Journal, 777, 165
- Kasting, J. F., Whitmire, D. P., & Reynolds, R. T. 1993, Icarus, 101, 108
- Kervella, P., Th´venin, F., S´gransan, D., et al. 2003, Ase e tronomy and Astrophysics , 404, 1087
- Ku, H. H., Journal of Research of the National Bureau of Standards. Section C: Engineering and Instrumentation, 70C, No. 4, 263
- Laskar, J., & Bou´, G. 2010, Astronomy and Astrophysics e , 522, A60
- Marchal, C., & Bozis, G. 1982, Celestial Mechanics, 26, 311 Michtchenko, T. A., & Ferraz-Mello, S. 2001, Icarus, 149, 357
- Michtchenko, T. A., Lazzaro, D., Ferraz-Mello, S., & Roig, F. 2002, Icarus, 158, 343
- Michtchenko, T. A., Ferraz-Mello, S., & Beaug´, C. 2006, e Icarus, 181, 555
- Michtchenko, T. A., & Malhotra, R. 2004, Icarus, 168, 237 Michtchenko, T. A., & Rodr´ ıguez, A. 2011, Monthly Notices of the Royal Astronomical Society, 415, 2275
Nelson, A. F. 2000, Astrophysical Journal, 537, L65
- Pendleton, Y. J., & Black, D. C. 1983, Astronomical Journal, 88, 1415
- Pilat-Lohinger, E., & Dvorak, R. 2002, Celestial Mechanics and Dynamical Astronomy, 82, 143
- Pourbaix, D., Nidever, D., McCarthy, C., et al. 2002, Astronomy and Astrophysics, 386, 280
- Quintana, E. V., Lissauer, J. J., Chambers, J. E., & Duncan, M. J. 2002, Astrophysical Journal, 576, 982
- Quintana, E. V., Adams, F. C., Lissauer, J. J., & Chambers, J. E. 2007, Astrophysical Journal, 660, 807
- Rabl, G., & Dvorak, R. 1988, Astronomy and Astrophysics, 191, 385
- Raghavan, D., McAlister, H. A., Henry, T. J., et al. 2010, Astrophysical Journal Supplement, 190, 1
- Szebehely, V. 1984, Celestial Mechanics, 34, 49
- Th´ ebault, P., Marzari, F., & Scholl, H. 2008, Monthly Notices of the Royal Astronomical Society, 388, 1528
- Th´ ebault, P., Marzari, F., & Scholl, H. 2009, Monthly Notices of the Royal Astronomical Society, 393, L21
|
10.1093/mnras/stu1591
|
[
"Eduardo Andrade-Ines",
"Tatiana A. Michtchenko"
] |
2014-08-05T18:05:36+00:00
|
2014-08-05T18:05:36+00:00
|
[
"astro-ph.EP"
] |
Dynamical Stability of Terrestrial Planets in the Binary α Centauri System
|
In this paper, we investigate whether hypothetical Earth-like planets have
high probability of remaining on stable orbits inside the habitable zones
around the stars A and B of {\alpha} Centauri, for lengths of time compatible
with the evolution of life. We introduce a stability criterion based on the
solution of the restricted three-body problem and apply it to the {\alpha}
Centauri system. In this way, we determine the regions of the short-term
stability of the satellite-type (S-type) planetary orbits, in both planar and
three-dimensional cases. We also study the long-term stability of hypothetical
planets through the dynamical mapping of the habitable zones of the stars. The
topology of the maps is analyzed using the semi-analytical secular Hamiltonian
model and possible processes responsible for long-lasting instabilities are
identified. We verify that the planetary motion inside the habitable zones is
regular, regardless of high eccentricities, for inclinations smaller than
40{\deg}. We show that the variation of the orbital distance of the planet
located in the habitable zones of the binary is comparable to that of Earth, if
the planet is close to the Mode I stationary solution. This result brings
positive expectations for finding habitable planets in binary stars.
|
1408.1055v2
|
## Coherent Excitation Transfer in a Spin Chain of Three Rydberg Atoms
Daniel Barredo, Henning Labuhn, Sylvain Ravets, Thierry Lahaye, and Antoine Browaeys Laboratoire Charles Fabry, UMR 8501, Institut d'Optique, CNRS, Universit´ e Paris Sud 11, 2 avenue Augustin Fresnel, 91127 Palaiseau cedex, France
Charles S. Adams Joint Quantum Centre (JQC) Durham-Newcastle, Department of Physics, Durham University, Durham, DH1 3LE, United Kingdom
We study coherent excitation hopping in a spin chain realized using highly excited individually addressable Rydberg atoms. The dynamics are fully described in terms of an XY spin Hamiltonian with a long range resonant dipole-dipole coupling that scales as the inverse third power of the lattice spacing, C /R 3 3 . The experimental data demonstrate the importance of next neighbor interactions which are manifest as revivals in the excitation dynamics. The results suggest that arrays of Rydberg atoms are ideally suited to large scale, high-fidelity quantum simulation of spin dynamics.
Spin Hamiltonians, introduced in the early days of quantum mechanics to explain ferromagnetism, are widely used to study quantum magnetism [1]. Assemblies of interacting, localized spins are a paradigm of quantum many-body systems, where the interplay between interactions and geometryinduced frustration creates a wealth of intriguing quantum phases. Many other phenomena, such as coherent energy transfer, photochemistry or photosynthesis [2], can also be described using spin Hamiltonians. However, despite this fundamental significance, exact analytical solutions are known only for the simplest cases, and numerical simulations of strongly correlated spin systems are notoriously difficult.
For those reasons, quantum simulation of spin Hamiltonians by controllable systems raises great interest. Recently, various approaches were followed to simulate spin systems using tools of atomic physics [3], such as cold atoms [4-6] or polar molecules [7] in optical lattices, interacting via weak exchange or dipole-dipole interactions, or trapped ions with engineered effective interactions [8-10]. As compared to their condensed-matter counterparts, the spin couplings can be long range, which gives rise to new properties [11-14].
Rydberg atoms are a promising alternative platform for quantum simulation [15, 16]. In particular, they allow implementing various spin1 / 2 Hamiltonians on two-dimensional lattices with strong couplings, in the MHz range [17, 18]. Rydberg systems interacting through van der Waals interactions can be described by Ising-type Hamiltonians H = ∑ ij V ij σ σ z i z j where σ z is the z -Pauli matrix acting in the (pseudo-) spin Hilbert space, and V ij ∼ | r i -r j | -6 , where r i denotes the position of atom i [17-21]. On the other hand, spin-exchange, or XY , spin Hamiltonians of the form H = ∑ ij V ij ( σ + i σ -j + σ -i σ + j ) , where σ ± = σ x ± iσ y are spin-flip operators and V ij ∼ | r i -r j | -3 , can be realized by using two different Rydberg states, interacting directly via the resonant dipole-dipole interaction. However in this case, only incoherent transfer of excitations has been observed so far, due to the random atomic positions in the ensembles used in experiments [22-27].
In this Letter, we study the coherent dynamics of a spin excitation in a chain of three Rydberg atoms. The dipoledipole interaction between atoms is given by the XY Hamiltonian [28]
glyph[negationslash]
$$H = \frac { 1 } { 2 } \sum _ { i \neq j } \frac { C _ { 3 } } { R _ { i j } ^ { 3 } } \left ( \sigma _ { i } ^ { + } \sigma _ { j } ^ { - } + \sigma _ { i } ^ { - } \sigma _ { j } ^ { + } \right ), \quad \ \ ( 1 )$$
where R ij = | r i -r j | is the distance between atoms i and j . We calibrate the spin-spin coupling between two Rydberg atoms by investigating the temporal evolution of two Rydberg atoms prepared in the state |↑↓〉 , as a function of distance R between the atoms, up to R glyph[similarequal] 50 µ m. We then use three Rydberg atoms prepared in |↑↓↓〉 and study the propagation of the excitation through this minimalistic spin chain, observing the effect of long-range hopping of the excitation. The agreement between experimental data and the XY model without adjustable parameters validates our setup as a future quantum simulator for systems of many spins in arbitrary twodimensional arrays .
The experimental setup, shown in Fig. 1(a), is detailed in Ref. [30]. Briefly, we focus a red-detuned trapping beam with an aspheric lens into a magneto-optical trap of 87 Rb, to a waist glyph[similarequal] 1 µ m. Multiple traps at arbitrary distances are created by imprinting an appropriate phase on the trapping beam with a spatial light modulator [31]. Beccause of fast light-assisted collisions in the small trapping volume, at most one atom is present in each trap. The temperature of the trapped atoms is approximately 50 µ K. A 6 G magnetic field defines the quantization axis [32].
We encode the two spin states in the Rydberg states |↑〉 = ∣ ∣ 62 D 3 / 2 , m j = 3 2 / 〉 and |↓〉 = ∣ ∣ 63 P 1 / 2 , m j = 1 2 / 〉 [see Fig. 1(b)]. We trigger an experiment when an atom is detected in each trap. To prepare the atoms in a desired spin state, we first optically pump them in | g 〉 = 5 ∣ ∣ S 1 / 2 , F = 2 , m F = 2 〉 . We then switch off the traps to avoid inhomogeneous light shifts, and excite the atoms to |↑〉 = ∣ ∣ 62 D 3 / 2 , m J = 3 2 / 〉 via a two-photon transition (wavelengths 795 and 474 nm, with polarizations π and σ + , respectively), detuned from the intermediate state ∣ ∣ 5 P 1 / 2 , F = 2 , m F = 2 〉 by ∆ glyph[similarequal] 2 π × 740 MHz. From the |↑〉 state the atom can be transferred
E
FIG. 1. (color online). (a) Individual 87 Rb atoms in microtraps aligned along the quantization axis, defined by a B = 6 G magnetic field. (b) Excitation lasers couple the ground state | g 〉 = ∣ ∣ 5 S 1 / 2 , F = 2 , m F = 2 〉 and the Rydberg state |↑〉 = ∣ ∣ 62 D 3 / 2 , m j = 3 2 / 〉 with an effective Rabi frequency Ω opt . Microwaves couple |↑〉 to |↓〉 = 63 ∣ ∣ P 1 / 2 , m J = 1 2 / 〉 , with Rabi frequency Ω MW . (c) Microwave-driven Rabi oscillation of a single atom between |↑〉 and |↓〉 , yielding Ω MW = 2 π × 4 6 . MHz.
to |↓〉 = 63 ∣ ∣ P 1 / 2 , m J = 1 2 / 〉 using resonant microwaves at glyph[similarequal] 9 131 . GHz, emitted by an antenna outside the vacuum chamber.
To read out the state of an atom at the end of a sequence, we switch on the excitation lasers, coupling only |↑〉 back to the ground state. We then turn on the dipole traps to recapture ground-state atoms, while atoms in Rydberg states remain untrapped, and detect atoms in | g 〉 by fluorescence. Therefore if we detect an atom in its trap at the end of a sequence, we assume it was in |↑〉 , while a loss corresponds to the |↓〉 state. We reconstruct all the 2 N probabilities P i 1 ...i k ...i N of having i k atom in trap k , with i k = 0 or 1, for our N -trap system (with N = 1 2 , , or 3 ) by repeating the experiment typically 100 times. For instance for N = 3 , P 100 is the probability to recapture an atom in trap 1, while recapturing none in traps 2 and 3. The statistical error on the determination of the probabilities is below 5% . Figure 1(c) illustrates the coherent spin manipulation for a single atom, by showing Rabi oscillations between |↑〉 and |↓〉 : the probability P 1 to recapture the atom oscillates with a frequency Ω MW glyph[similarequal] 2 π × 4 6 . MHz. In 4 µ s , we induce more than 35 spin flips without observing noticeable damping.
We first use two atoms, aligned along the quantization axis, to directly measure the coupling between two spins as a function of their distance. The sequence is shown in Fig. 2(a). We illuminate atom 1 with an addressing beam [33] which induces a 20 MHz light shift, making it off resonant to the global Rydberg excitation. Atom 2 is excited to |↑〉 , and then transferred to |↓〉 using microwaves. Subsequently, atom 1 is optically excited to the |↑〉 state with the addressing beam switched off (atom 2 in |↓〉 is not affected by the Rydberg exci-
FIG. 2. (color online). (a) Sequence to observe spin exchange between two atoms. (b) Excitation hopping between states |↑↓〉 (blue disks) and |↓↑〉 (red disks) of two atoms separated by R = 30 µ m. Solid lines are sinusoidal fits, with frequency 2 E/h . (c) Interaction energy E (circles) versus R . Error bars are smaller than the symbols size. The line shows the theoretical prediction C /R 3 3 with C th 3 = 7965 MHz µ m 3 . The shaded area corresponds to our systematic 5% uncertainty in the calibration of R .
tation pulse). We let the system evolve for an adjustable time τ and read out the final state by deexciting |↑〉 back to | g 〉 . In the absence of experimental imperfections (see [28], section S.4), P 10 (respectively P 01 ) would give the population of |↑↓〉 (respectively |↓↑〉 ).
The evolution of P 10 ( τ ) and P 01 ( τ ) for two atoms prepared in |↑↓〉 separated by 30 µ m is shown in Fig. 2(b). The spin excitation oscillates back and forth between the two atoms, with a frequency 2 E/h ≈ 0 52 . MHz. The finite contrast is essentially due to spontaneous emission via the intermediate ∣ ∣ 5 P 1 / 2 〉 state during preparation and readout, which limits the oscillation amplitude to about 60% , and, to a lesser extent, to the onset of dipolar interactions during the second excitation pulse [28]. We then repeat the same experiment for several values of the distance R between the atoms, and observe spin-exchange oscillations for distances as large as 50 µ m. Figure 2(c) shows the measured interaction energies as a function of R , together with the expected C /R 3 3 behavior (solid line) for the theoretical value C th 3 = 7965 MHz µ m 3 of the C 3 coefficient, calculated from the dipole matrix elements 〈↑| ˆ d ± 1 |↓〉 [30, 34]. A power-law fit to the data (not shown) gives an exponent -2 93 . ± 0 20 . . Fixing the exponent to -3 gives C exp 3 = 7950 ± 130 MHz µ m 3 . The agreement
FIG. 3. (color online). Spin excitation transfer along a chain of three Rydberg atoms with nearest-neighbor separation of 20 µ m. (a) Theoretical dynamics for a system initially prepared in |↑↓↓〉 , and evolving under a Hamiltonian similar to (1), but with only nearest-neighbor interactions. (b) The same as (a), but for the the full Hamiltonian (1), including long-range interactions. (c) Experimental data (points) and prediction of the model taking into account experimental imperfections (see text), with no adjustable parameters. For perfect preparation and readout, the probabilities P ↑↓↓ (respectively P ↓↑↓ , P ↓↓↑ ) and P 100 (respectively P 010 , P 001 ) would coincide.
between data and theory is excellent.
We now extend the system to a three-spin chain, with a distance R = 20 µ mbetween the atoms. The sequence is similar to that in Fig. 2(a) for two atoms, except that we now use microwave transfer for atoms 2 and 3 to prepare | g ↓↓〉 . Here, the van der Waals interaction between the two atoms in |↑〉 is only ∼ 10 kHz for R = 20 µ m , and thus no blockade effect arises during excitation. We then excite atom 1 to prepare |↑↓↓〉 .
namics showing the effects of the long-range coupling. However, one notices differences with the ideal case: (i) the preparation is imperfect, as one starts with a significant population in |↓↑↓〉 , (ii) this, together with imperfect readout [28], reduces the overall amplitude of the oscillations, and (iii) the oscillations show some damping, which becomes significant for τ ≥ 4 µ s .
We first analyze theoretically the evolution of the system. Assuming that the initial state is | ψ (0) 〉 = |↑↓↓〉 , the dynamics induced by the XY Hamiltonian (1), which conserves the total magnetization ∑ i σ z i , occurs within the subspace spanned by {|↑↓↓〉 , |↓↑↓〉 , |↓↓↑〉} . Figures 3(a) and (b) show the calculated dynamics of the spin excitation, which moves back and forth between the extreme sites. Figure 3(a) corresponds to the case where only nearest-neighbor interactions are retained in (1). Periodic, fully contrasted oscillations at a frequency √ 2 C /R 3 3 are expected for the population of the extreme sites, while the population of |↓↑↓〉 oscillates twice as fast between 0 and 1 / 2 . In contrast, in Fig.3(b), the full Hamiltonian (1) is simulated, including the interaction between extreme sites. One observes a clear signature of this long-range coupling, as the dynamics now becomes aperiodic for the populations of |↑↓↓〉 and |↓↓↑〉 . The interplay of the couplings C /R 3 3 and C / 3 (8 R 3 ) between nearest- and nextnearest neighbors makes the eigenvalues of (1) incommensurate. The back-and-forth exchange of excitation is thus modulated by a slowly varying envelope due to the beating of these frequencies.
Figure 3(c) shows the experimental results for P 100 , P 010 , and P 001 (symbols). We observe qualitative agreement with Fig. 3(b), in particular the 'collapse and revival' in the dy-
Imperfect preparation and readout stem from the fact that, in addition to the spontaneous emission via the intermediate state during the optical pulses, the Rabi frequency for optical excitation ( glyph[similarequal] 5 3 . MHz) of atom 1 from | g 〉 to |↑〉 is not much higher than the interaction ( glyph[similarequal] 0 92 MHz . for R = 20 µ m ). Thus, during the excitation of atom 1, the spin excitation already has a significant probability to hop to atom 2. The damping essentially arises from the finite temperature of the atoms, which leads to changes in the interatomic distances, and thus in the couplings.
To go beyond this qualitative understanding of the limitations of our 'quantum simulator,' we add all known experimental imperfections to the XY model [28]. The result, shown by solid lines on Fig. 3(c) accurately reproduces the data with no free parameters. To obtain these curves, we simulate the full sequence, i.e., all three optical (de-)excitation pulses with or without the addressing beam, the microwave pulse, and evolution time, by solving the optical Bloch equations describing the dynamics of the internal states of the atoms, restricted to three states: | g 〉 , |↑〉 , and |↓〉 . Dissipation comes from both off resonant excitation of the intermediate ∣ ∣ 5 P 1 / 2 〉 state during the optical excitation pulse, and from the finite lifetimes of the Rydberg states (101 and 135 µ s for |↑〉 and |↓〉 , respectively [35]). The former effect is treated as an effective damping of the | g 〉 ↔ |↑〉 transition, present only during the
optical pulses, and with a damping rate chosen to match the damping of single-atom Rabi oscillations performed to calibrate the excitation Rabi frequency Ω opt [17].
We then account for the thermal motion of the atoms. A first consequence of the finite temperature ( T glyph[similarequal] 50 µ K ) is that at the beginning of the sequence, the atoms have random positions (the transverse rms extension of the thermal motion in each microtrap, of radial frequency 90 kHz, is about 120 nm) and random velocities ( 70 nm /µ s rms). During the sequence, the traps are switched off and the atoms are thus in free flight with their initial velocity. When solving the optical Bloch equations, we thus first draw the initial positions r 0 i and velocities v 0 i of each atom i according to a thermal distribution, and use time-dependent dipolar couplings C / 3 | ( r 0 i + v 0 i t ) -( r 0 j + v 0 j t ) | 3 in Eq. (1) [36]. We then average the results over 100 realizations. This yields a dephasing of the oscillations, resulting in a significant contrast reduction at long times.
A second effect of the temperature is that an atom has a small probability ε t ( ) to leave the trap region during the experiment. In this case, we mistakenly infer that it was in a Rydberg state at the end of the sequence. This leads to a small distortion of the measured populations P ijk ( i, j, k = 0 1 , ) [37], that we compute from the actual ones as described in [17]. We measure ε t ( ) (which increases with the duration t of the sequence, from ∼ 1% at t = 0 up to ∼ 20% for t = 7 µ s ) in a calibration experiment, and then use it to calculate the expected populations from the simulated ones [28].
Figure 4 shows how those two consequences of the finite temperature contribute to the observed damping in the dynamics of P 001 : both have sizable effects, but the dephasing due to fluctuations in the coupling dominates at long times. Reducing the atomic temperature using e.g. Raman cooling [38, 39] would render those effects negligible for our time scales, and allow the realization of a nearly ideal quantum simulator of spin dynamics.
In summary, we have measured the dynamics of a spin excitation in a minimal spin chain of three Rydberg atoms. The evolution of the system is accurately described by an XY Hamiltonian without any adjustable parameters. The obtained results are encouraging in view of scaling up the system to a larger number of spins. In particular, the residual motion of the atoms and the level of detection errors would already allow us to observe unambiguously the back-and-forth propagation of an excitation over a chain of ∼ 20 atoms [28]. However, so far, experiments with more than ∼ 5 atoms are hampered by the stochastic loading of the traps by single atoms [31]. In future work, we will thus explore various quasideterministic loading schemes that have been demonstrated at the level of a single [40, 41] or a few [42, 43] traps. Once this is achieved, our system will allow us to study the equivalent of an assembly of hard-core bosons on a 2D lattice with long-range, anisotropic hopping. We will also study dipolar interactions involving more than only two Rydberg states at an electrically tuned F¨rster resonance [44]. o Our system will be ideal to study exotic phases and frustration in quantum magnetism,
1
FIG. 4. (color online). Influence of the temperature on P 001 ( τ ) : simulated dynamics at zero temperature (black dashed line), and adding either only atom loss (green dotted line), or only atomic motion (blue solid line).
excitation hopping in complex networks [45, 46] or quantum walks with long-range hopping [47].
We acknowledge financial support by the EU [ERC Stg Grant ARENA, AQUTE Integrating Project, FET-Open Xtrack Project HAIRS, and EU Marie-Curie Program ITN COHERENCEFP7-PEOPLE-2010-ITN-265031 (H. L.)], and by R´ egion ˆ le-de-France I (LUMAT and Triangle de la Physique, projects LAGON and COLISCINA during the stay of C. S. A at LCF). C. S. A. also acknowledges support from the U. K. EPSRC and Durham University.
- [1] A. Auerbach, Interacting Electrons and Quantum Magnetism (Springer-Verlag, New York, 1994).
- [2] E. Collini, Chem. Soc. Rev 42 , 4932 (2013).
- [3] M. Lewenstein, A. Sanpera, and V. Ahufinger, Ultracold Atoms in Optical Lattices : Simulating Quantum Many-Body Systems (Oxford University Press, Oxford, U.K, 2012).
- [4] J. Simon, W. S. Bakr, R. Ma, M. E. Tai, P. M. Preiss, and M. Greiner, Nature (London) 472 , 307 (2011).
- [5] T. Fukuhara, P. Schauß, M. Endres, S. Hild, M. Cheneau, I. Bloch, and C. Gross, Nature (London) 502 , 76 (2013).
- [6] A. de Paz, A. Sharma, A. Chotia, E. Mar´ echal, J. H. Huckans, P. Pedri, L. Santos, O. Gorceix, L. Vernac, and B. LaburtheTolra, Phys. Rev. Lett. 111 , 185305 (2013).
- [7] B. Yan, S. A. Moses, B. Gadway, J. P. Covey, K. R. A. Hazzard, A. M. Rey, D. S. Jin, and J. Ye, Nature (London) 501 , 521 (2013).
- [8] K. Kim, M. S. Chang, S. Korenblit, R. Islam, E. E. Edwards, J. K. Freericks, G. D. Lin, L. M. Duan, and C. Monroe, Nature (London) 465 , 590 (2010).
- [9] P. Richerme, Z.-X. Gong, A. Lee, C. Senko, J. Smith, M. FossFeig, S. Michalakis, A. V. Gorshkov, and C. Monroe, Nature (London) 511 , 198 (2014).
- [10] P. Jurcevic, B. P. Lanyon, P. Hauke, C. Hempel, P. Zoller, R. Blatt, and C. F. Roos, Nature (London) 511 , 202 (2014).
- [11] P. Hauke, F. M. Cucchietti, A. M¨ uller-Hermes, M.-C. Ba˜uls, n J. I. Cirac, and M. Lewenstein, New J. Phys. 12 , 113037 (2010).
- [12] D. Peter, S. M¨ uller, S. Wessel, and H. P. B¨ uchler, Phys. Rev. Lett. 109 , 025303 (2012).
- [13] K. R. A. Hazzard, M. van den Worm, M. Foss-Feig, S. R. Manmana, E. Dalla Torre, T. Pfau, M. Kastner, and A. M. Rey, Phys. Rev A 90 , 063622 (2014).
- [14] M. Avellino, A. J. Fisher, and S. Bose, Phys. Rev A 74 , 012321 (2006).
- [15] M. Saffman, T. G. Walker, and K. Mølmer, Rev. Mod. Phys.
## 82 , 2313 (2010).
- [16] H. Weimer, M. M¨ uller, I. Lesanovsky, P. Zoller, and H.-P. B¨ uchler, Nat. Phys. 6 , 382 (2010).
- [17] D. Barredo, S. Ravets, H. Labuhn, L. B´ eguin, A. Vernier, F. Nogrette, T. Lahaye, and A. Browaeys, Phys. Rev. Lett. 112 , 183002 (2014).
- [18] P. Schauß, J. Zeiher, T. Fukuhara, S. Hild, M. Cheneau, T. Macr` ı, T. Pohl, I. Bloch, and C. Gross, arXiv:1404.0980.
- [19] H. Weimer, R. L¨ ow, T. Pfau, and H. P. B¨chler, Phys. Rev. Lett. u 101 , 250601 (2008).
- [20] T. Pohl, E. Demler, and M. D. Lukin, Phys. Rev. Lett. 104 , 043002 (2010).
- [21] I. Lesanovsky, Phys. Rev. Lett. 106 , 025301 (2011).
- [22] I. Mourachko, D. Comparat, F. de Tomasi, A. Fioretti, P. Nosbaum, V. M. Akulin, and P. Pillet, Phys. Rev. Lett. 80 , 253 (1998).
- [23] W. R. Anderson, J. R. Veale, and T. F. Gallagher, Phys. Rev. Lett. 80 , 249 (1998).
- [24] C. S. E. van Ditzhuijzen, A. F. Koenderink, J. V. Hern´ andez, F. Robicheaux, L. D. Noordam, and H. B. van Linden van den Heuvell, Phys. Rev. Lett. 100 , 243201 (2008).
- [25] G. G¨ unter, H. Schempp, M. Robert-de-Saint-Vincent, V. Gavryusev, S. Helmrich, C. S. Hofmann, S. Whitlock, and M. Weidem¨ uller, Science 342 , 954 (2013).
- [26] S. Bettelli, D. Maxwell, T. Fernholz, C. S. Adams, I. Lesanovsky, and C. Ates, Phys. Rev. A 88 , 043436 (2013).
- [27] D. Maxwell, D. J. Szwer, D. Paredes-Barato, H. Busche, J. D. Pritchard, A. Gauguet, K. J. Weatherill, M. P. A. Jones, and C. S. Adams, Phys. Rev. Lett. 110 , 103001 (2013).
- [28] See Supplemental Material, which includes Ref. [29], for extra details about the simulation of the spin chain dynamics.
- [29] C. Tuchendler, A. M. Lance, A. Browaeys, Y. R. P. Sortais, and P. Grangier, Phys. Rev. A 78 , 033425 (2008).
- [30] L. B´ eguin, A. Vernier, R. Chicireanu, T. Lahaye, and A. Browaeys, Phys. Rev. Lett. 110 , 263201 (2013).
- [31] F. Nogrette, H. Labuhn, S. Ravets, D. Barredo, L. B´ eguin, A. Vernier, T. Lahaye, and A. Browaeys, Phys. Rev. X 4 , 021034 (2014).
- [32] The stray electric field is below 5 mV cm / . Note that as (1) conserves the excitation number, the system is quite immune to decoherence due to global fluctuations in electric or magnetic fields.
- [33] H. Labuhn, S. Ravets, D. Barredo, L. B´ eguin, F. Nogrette, T. Lahaye, and A. Browaeys, Phys. Rev. A 90 , 023415 (2014).
- [34] A. Reinhard, T. C. Liebisch, B. Knuffman, and G. Raithel, Phys. Rev. A 75 , 032712 (2007).
- [35] I. I. Beterov, I. I. Ryabtsev, D. B. Tretyakov, and V. M. Entin, Phys. Rev. A 79 , 052504 (2009).
- [36] Here, we neglect mechanical effects due to the interactions between Rydberg atoms during the interaction time. Using a classical model we estimate that these induce a displacement in R below 30 nm for τ ≤ 7 µ s and are therefore negligible.
- [37] C. Shen and L.-M. Duan, New J. Phys. 14 , 053053 (2012).
- [38] A. M. Kaufman, B. J. Lester, and C. A. Regal, Phys. Rev. X 2 , 041014 (2012).
- [39] J. D. Thompson, T. G. Tiecke, A. S. Zibrov, V. Vuleti´ c, and M. D. Lukin, Phys. Rev. Lett. 110 , 133001 (2013).
- [40] T. Gr¨nzweig, A. Hilliard, M. McGovern, u and M. F. Andersen, Nat. Phys. 6 , 951 (2010).
- [41] M. Ebert, A. Gill, M. Gibbons, X. Zhang, M. Saffman and T. G. Walker, Phys. Rev. Lett. 112 , 043602 (2014).
- [42] Y. Miroshnychenko, W. Alt, I. Dotsenko, L. F¨ orster, M. Khudaverdyan, D. Meschede, D. Schrader and A. Rauschenbeutel, Nature (London) 442 , 151 (2006).
- [43] M. Schlosser, J. Kruse, C. Gierl, S. Teichmann, S. Tichelmann, and G. Birkl, New J. Phys. 14 , 123034 (2012).
- [44] S. Ravets, H. Labuhn, D. Barredo, L. B´ eguin, T. Lahaye, and A. Browaeys, Nat. Phys. 10 , 914 (2014).
- [45] S. W¨ uster, C. Ates, A. Eisfeld, and J. M. Rost, Phys. Rev. Lett. 105 , 053004 (2010).
- [46] S. Mostarda, F. Levi, D. Prada-Gracia, F. Mintert, and F. Rao, arXiv:1312.1833.
- [47] R. Cˆ ot´, A. Russell, E. E. Eyler, e and P. L. Gould, New J. Phys. 8 , 156 (2006)
## Supplemental Material: Coherent Excitation Transfer in a Spin Chain of Three Rydberg Atoms
## THE DIPOLE-DIPOLE INTERACTION AS AN XY HAMILTONIAN
Here we derive, for the sake of completeness, how the dipole-dipole interaction between two Rydberg atoms leads to the XY Hamiltonian (1) of the main text. We also emphasize the angular dependence of the C 3 coefficient, which is proportional to 1 -3 cos 2 θ .
The interaction between two neutral atoms separated by a distance R (large compared to the atom size, but small compared to the wavelength of the relevant transitions, such that the electrostatic limit applies) can be expressed, to leading order, through the dipole-dipole interaction:
$$V _ { \text{ddi} } = \frac { 1 } { 4 \pi \epsilon _ { 0 } } \frac { d _ { 1 } \cdot d _ { 2 } - 3 ( d _ { 1 } \cdot n ) ( d _ { 2 } \cdot n ) } { R ^ { 3 } },$$
where d i = ( d , d x y , d z ) is the electric dipole moment operator of atom i ( i = 1 2 , ), and n = R /R is the unit vector connecting the two atoms. We denote the quantization axis by z , and the angle between z and n by θ . In the spherical basis, it is convenient to use the spherical dipole operators:
$$\begin{cases} \begin{array} { c c } d _ { 0 } & = d _ { z } \\ d _ { + } & = - \left ( d _ { x } + i \, d _ { y } \right ) / \sqrt { 2 } \\ d _ { - } & = \left ( d _ { x } - i \, d _ { y } \right ) / \sqrt { 2 } \,. \end{array} \end{cases}$$
The operator d 0 conserves the magnetic quantum number m j , whereas the operators d ± change m j by one ( ∆ m j = ± 1 ). In the spherical basis, the dipole-dipole interaction can be written as:
$$\text{cal} & \text{sis, the alpha} \oplus \text{orization can be written as} \colon \\ V _ { d d i } = \frac { 1 } { 4 \pi \epsilon _ { 0 } } \frac { 1 } { R ^ { 3 } } \left [ \frac { 1 - 3 \cos ^ { 2 } \theta } { 2 } \left ( d _ { 1, + } d _ { 2, - } + d _ { 1, - } d _ { 2, + } + 2 d _ { 1, 0 } d _ { 2, 0 } \right ) \\ & \quad + \frac { 3 } { \sqrt { 2 } } \sin \theta \cos \theta \left ( d _ { 1, + } d _ { 2, 0 } - d _ { 1, - } d _ { 2, 0 } + d _ { 1, 0 } d _ { 2, + } - d _ { 1, 0 } d _ { 2, - } \right ) \\ & \quad - \frac { 3 } { 2 } \sin ^ { 2 } \theta \left ( d _ { 1, + } d _ { 2, + } + d _ { 1, - } d _ { 2, - } \right ) \right ].$$
The sum comprises three terms, with different angular dependence, that couple states where the total magnetic quantum number M = m (1) j + m (2) j changes, respectively, by ∆ M = 0 ∆ , M = ± 1 , and ∆ M = ± 2 .
We now restrict ourselves to only two states |↑〉 and |↓〉 of the form | n, L, J, m j 〉 and ∣ ∣ n , L , J ′ ′ ′ , m ′ j 〉 that fulfill the selection rules for the dipole operator d q ( q = 0 , ± 1 ), i.e. ∆ L = ± 1 , ∆ J = 0 , ± 1 , and ∆ m j = q (in the main text, they are ∣ ∣ 62 D 3 / 2 , m j = 3 2 / 〉 and ∣ ∣ 63 P 1 / 2 , m j = 1 2 / 〉 ). In the two-atom basis {|↑↑〉 , |↑↓〉 , |↓↑〉 , |↓↓〉} , V ddi couples only |↑↓〉 to |↓↑〉 , on the one hand, and |↑↑〉 together with |↓↓〉 on the other hand. In the latter case, the two pair states are separated in energy by several tens of GHz, and thus the effect of the dipolar coupling is negligible. In contrast, in the first case, the two pair states are always degenerate, and thus the dipolar coupling is resonant. Moreover, for those pair states, we always have ∆ M = 0 (since the two atoms just exchange their states), so only the first term in (S3) survives. Therefore, when restricted to the two pair states |↑↓〉 , |↓↑〉 , the interaction Hamiltonian takes the simple form:
$$V _ { d d i } = \frac { 1 } { 4 \pi \epsilon _ { 0 } } \frac { 1 - 3 \cos ^ { 2 } \theta } { 2 R ^ { 3 } } \left ( d _ { 1, + } d _ { 2, - } + d _ { 1, - } d _ { 2, + } + 2 d _ { 1, 0 } d _ { 2, 0 } \right )$$
or, in a matrix form in the basis { |↑↓〉 , |↓↑〉 } :
$$V _ { d d i } = \frac { 1 } { 4 \pi \epsilon _ { 0 } } \frac { 1 - 3 \cos ^ { 2 } \theta } { R ^ { 3 } } \left ( \begin{array} { c c } 0 & \tilde { C } _ { 3 } \\ \tilde { C } _ { 3 } & 0 \end{array} \right ).$$
There, ˜ C 3 ≡ 〈↑↓| d 1 + , d 2 , -+ d 1 , -d 2 + , +2 d 1 0 , d 2 0 , |↓↑〉 / 2 . Note that in this matrix element, only one of the three terms of the sum contributes, depending on the m j values of the |↑〉 and |↓〉 states [S1].
Now, in the context of spin Hamiltonians, it is desirable to rewrite the pseudo-spin operators in terms of the Pauli matrices. A first step towards this is then to denote the single-atom operators |↑〉 〈↓| as σ ↑↓ , and we end up with
$$V _ { d d i } = \frac { C _ { 3 } ( \theta ) } { R ^ { 3 } } \left ( \sigma _ { \uparrow \downarrow } ^ { ( 1 ) } \sigma _ { \downarrow \uparrow } ^ { ( 2 ) } + \text{ h. c.} \right )$$
where C 3 ( θ ) = ˜ (1 C 3 -3 cos 2 θ ) , and h c . . stands for hermitian conjugate. In Eq. (1) of the main text, to conform to the standard notations of the ladder operators σ ± = σ x ± iσ y defined from the Pauli matrices in the context of spin systems, we have written σ + (resp. σ -) for σ ↓↑ (resp. σ ↑↓ ). We emphasize here that the ± subscripts in this latter notation refer to the pseudo-spin space, and have nothing to do with the ± subscripts entering the definition of the spherical dipole (S2), which are related to the orientation of the dipoles with respect to the quantization axis in real space [S2].
## SIMULATION OF THE SPIN EXCITATION DYNAMICS IN A CHAIN
To investigate the origin of the reduced contrast and damping of the oscillations in our experiment, we solve the optical Bloch equations (OBEs) for the system of three atoms in the basis spanned by the states {| g 〉 , |↑〉 , |↓〉} . By including the ground state | g 〉 we can account for imperfections in the preparation of the initial configuration due to both spontaneous emission through the intermediate state, and the presence of the always resonant dipole-dipole interaction between the states |↑↓〉 and |↓↑〉 . We simulate the full experimental sequence, including the optical and microwave pulses to the Rydberg states, the time evolution under the XY Hamiltonian [(1) of the main text], and the final state readout. The total Hamiltonian of the system
$$H _ { t o t } = H _ { A \text{-} L } + H _ { \text{int} } \,,$$
is composed of two terms describing, respectively, the coupling of the atoms to the light and microwave fields, and the dipoledipole interaction. The first term, H A L -, reads:
$$H _ { A \ast L } = \hbar { \sum } _ { i } \frac { \Omega _ { L } ^ { ( i ) } } { 2 } \left ( \sigma _ { \uparrow g } ^ { ( i ) } + \sigma _ { g \uparrow } ^ { ( i ) } \right ) + \frac { \Omega _ { \text{MW} } } { 2 } \left ( \sigma _ { \downarrow \uparrow } ^ { ( i ) } + \sigma _ { \uparrow \downarrow } ^ { ( i ) } \right ) - \delta _ { L } ^ { ( i ) } \left ( \sigma _ { \uparrow \uparrow } ^ { ( i ) } + \sigma _ { \downarrow \downarrow } ^ { ( i ) } \right ) \,,$$
where σ ( ) i nn ′ = ∣ ∣ n ( ) i 〉 〈 n ′ ( ) i ∣ ∣ ( n, n ′ ∈ { g, ↑ ↓} , ) are the transition and projector operators for atom at site i , and the parameters Ω ( ) i L , and δ ( ) i L are the optical Rabi frequencies, and laser detunings for the transition | g 〉 ↔ |↑〉 respectively. These parameters are position-dependent to account for the slightly different Rabi frequencies and differential light shifts experienced by the atoms due to the gaussian profile of the excitation laser beams. For the microwave Rabi frequency Ω MW for the transition | d 〉 ↔ | p 〉 , we neglect any inhomogeneity of the microwave field at the local position of the atoms. Finally, the interaction term is given by the Hamiltonian:
glyph[negationslash]
$$H _ { \text{int} } = \frac { 1 } { 2 } \sum _ { i \neq j } \frac { C _ { 3 } } { R _ { i j } ( t ) ^ { 3 } } ( \sigma _ { \downarrow \uparrow } ^ { ( i ) } \sigma _ { \uparrow \downarrow } ^ { ( j ) } + \sigma _ { \uparrow \downarrow } ^ { ( i ) } \sigma _ { \downarrow \uparrow } ^ { ( j ) } ) \,.$$
To account for the thermal motion of the atoms during the experiment we introduce a time dependency in the interatomic distance R ij ( ) = t | ( r 0 i + v 0 i t ) -( r 0 j + v 0 j t ) | .
The Lindblad operator includes a sum over the decay channels for each atom,
$$L [ \rho ] = \frac { 1 } { 2 } \sum _ { i } \left ( \gamma _ { i } + \gamma _ { \uparrow } \right ) \left ( 2 \sigma _ { g \uparrow } ^ { ( i ) } \rho \sigma _ { \uparrow g } ^ { ( i ) } - \sigma _ { \uparrow \uparrow } ^ { ( i ) } \rho - \rho \sigma _ { \uparrow \uparrow } ^ { ( i ) } \right ) + \gamma _ { \downarrow } \left ( 2 \sigma _ { g \downarrow } ^ { ( i ) } \rho \sigma _ { \downarrow g } ^ { ( i ) } - \sigma _ { \downarrow \downarrow } ^ { ( i ) } \rho - \rho \sigma _ { \downarrow \downarrow } ^ { ( i ) } \right ) \,.$$
where γ i are the effective damping rates, and 1 /γ ↑ and 1 /γ ↓ are the lifetimes for the |↑〉 and |↓〉 Rydberg states. The dampings γ i (with 1 /γ i ∼ 1 µ s ) are mainly due to off-resonant spontaneous emission through the intermediate state ∣ ∣ 5 P 1 / 2 〉 and are only present during the optical pulses. For the lifetimes of the Rydberg states we use the effective values ( 1 /γ ↑ ∼ 101 µ s ; 1 /γ ↓ ∼ 135 µ s) calculated by Beterov et al. [Ref. 32 of the main text], as a single decay channel to the ground state | g 〉 . All the parameters, i.e., the single-atom Rabi frequencies Ω i , laser detunings δ i , and dampings γ i are measured independently by recording single-atom Rabi oscillations on each site of the array.
The final result of the simulation is obtained by averaging the solution of the OBEs over 100 realizations, starting from random atom positions r 0 i , r 0 j , with rms value σ r = √ k B T/mω 2 ⊥ glyph[similarequal] 120 nm (where k B is the Boltzmann constant and ω ⊥ ∼ 90 kHz is the measured radial frequency of the atoms of mass m in the trap), and velocities v 0 i , v 0 j (rms σ v = √ k B T/m glyph[similarequal] 70 nm/ µ s).
## READOUT DETECTION ERRORS
In our experiment, the outcome of every realization is a binary state: either the atom is recaptured or it has escaped the trapping region at the end of the sequence. If the atom is recaptured we assume it is in the ground state. On the contrary, a loss in
1.0
0
2
4
6
8
10
Figure S1. (a) Recapture probabilities P 111 (solid circles), P 110 + P 101 + P 011 (square symbols), P 100 + P 010 + P 001 (diamonds), and P 111 (open circles) for three atoms in a line as a function of the release trapping time t . The dashed line is the polynomial fit to the data from which we extract ε . Solid lines are the prediction of our loss model with the measured ε , which is shown in logarithmic scale in (b). The nonzero value of ε t ( ) for t → 0 is due to losses induced by background-gas collisions during the ∼ 200 ms duration of a full experimental sequence.
interpreted as the atom being in a Rydberg state ( |↑〉 or |↓〉 ), since Rydberg atoms are not trapped in our tweezers. However, due to the finite temperature ( glyph[similarequal] 50 µ K) of the atoms and collisions with the background gas, any atom has a probability ε = ( ) ε t to be lost during a sequence of total length t , independently of its internal state [Ref 20 of the main text]. Therefore, there exists a probability to falsely detect the state of the atom, which increases with the length of the sequence. For a single atom, the recapture probability P 1 is therefore P 1 = (1 -ε P ) g , where P g is the actual probability for the atom to be in the ground state. In the case of a loss, P 0 is related to the actual Rydberg state populations P ↑ and P ↓ by P 0 = (1 -ε )( P ↑ + P ↓ ) + ε . As an example, in the case of three atoms the relation between the observed recapture probability P 100 and the actual state populations P j 1 j 2 j 3 ( j k ∈ { g, ↑ ↓} , ) reads:
$$P _ { 1 0 0 } = ( 1 - \varepsilon ) [ P _ { g \uparrow \uparrow } + P _ { g \uparrow \downarrow } + P _ { g \downarrow \uparrow } + P _ { g \downarrow \downarrow } + \varepsilon ( P _ { g \uparrow g } + P _ { g \downarrow g } + P _ { g g \uparrow } + P _ { g g \downarrow } ) + \varepsilon ^ { 2 } P _ { g g g } ] \quad \text{(S11)}$$
To measure ε , we perform an experiment where we trap three ground state atoms in the line, switch the trap off for a variable duration t , and measure the populations at the end of the sequence.
We extract ε from a polynomial fit of the measured P 111 , where P 111 = (1 -ε ) 3 . We then evaluate the evolution of the rest of the recapture probabilities:
$$P _ { 0 1 1 } + P _ { 1 0 1 } + P _ { 1 1 0 } & = 3 \, \varepsilon ( 1 - \varepsilon ) ^ { 2 } \,, \\ P _ { 1 0 0 } + P _ { 0 1 0 } + P _ { 0 0 1 } & = 3 \, \varepsilon ^ { 2 } ( 1 - \varepsilon ) \,, \\ P _ { 0 0 0 } & = \varepsilon ^ { 3 } \,.$$
The comparison of the calculated recapture probabilities (solid lines in Fig. S1) with the experimental test supports the consistency of our loss model.
## ORIGIN OF THE FINITE CONTRAST OF THE OSCILLATIONS
In Fig. 2(b) of the main text, the measured contrast of the oscillations in the populations P 01 and P 10 is around 60% . This reduced contrast arises from two effects. The first one is imperfections in the preparation of the state |↑↓〉 mainly due to (i) inefficient optical pumping of the atoms in the state | g 〉 and (ii) spontaneous emission from the intermediate state ∣ ∣ 5 P 1 / 2 , F = 2 , m F = 2 〉 to the states ∣ ∣ 5 S 1 / 2 , F = 1 , m F = 1 〉 , ∣ ∣ 5 S 1 / 2 , F = 2 , m F = 1 〉 , and ∣ ∣ 5 S 1 / 2 , F = 2 , m F = 2 〉 . The second effect is that during the second and third optical pulses, of typical duration 100 ns, one cannot neglect totally the effect of the dipolar interactions between states |↑↓〉 and |↓↑〉 : the interaction-induced energy splitting between the two states is then 0.6 MHz, yielding a phase accumulation of ∼ 0 2 . rad, which is not totally negligible. The same effect also occurs for three atoms, as described in the main text. The full simulation (following the approach described in sections S.2 and S.3 above) of the two-atom system reproduces well the observed data.
## FULL DATASET FOR THE THREE-ATOM SPIN CHAIN DYNAMICS
The evolution of the system of three spins aligned along a chain coupled by long range dipole-dipole interactions is displayed in Fig. S2. Here we show, additionally, the probabilities not shown for clarity in Fig. 3 of the main text.
Solid lines are the result of the simulation through the OBEs as described in section S.1., where we also include the losserror correction (section S.2.). The atom loss effect is clearly visible in the probability P 000 at long interaction times. This effect further reduces the contrast of the oscillations for the populations P 100 , P 010 , and P 001 . The overall agreement of the simulations is very good, supporting the possibility to extend the experiments to larger two dimensional systems.
1
Figure S2. Spin excitation dynamics along a chain of three Rydberg atoms from a system initially prepared in |↑↓↓〉 . The complete set of experimental probabilities (symbols) and the result of the model (solid lines), including temperature effects and atom losses are shown.
## EFFECT OF THE ATOM TEMPERATURE IN THE DYNAMICS
In an attempt to investigate the limitations of the finite temperature of the atoms in the observed spin dynamics, especially in view of experiments with larger number of atoms, we simulate the evolution of the system at T = 10 µ K. The temperature of the atoms enters the simulation in (i) the rms extension of the random atomic positions and velocities in the time-dependent dipolar couplings, and (ii) the atom recapture probability ε t, T ( ) . To include (i) we perform Monte-Carlo simulations of the trajectories of the single atoms, taking into account the expected energy distribution of an atom in the trap [S3].
The result is shown in Fig. S3, where we compare the simulation for T = 10 µ K (solid lines) with the curves obtained at T = 50 µ K(dashed lines). In this case the damping of the oscillations are drastically reduced for times longer than ∼ 4 µ s. This suggests that a reduction of the temperature of the atoms by only one order of magnitude would be enough to make motional effects negligible for the timescales used in the experiment, so as to enable the nearly ideal quantum simulation of larger spin systems.
1
1
1
Figure S3. Simulated recapture probabilities P 100 , P 010 , and P 001 for a single atom temperature of T = 10 µ K(solid lines), and for T = 50 µ K (dashed lines).
## EXTENSION OF THE SYSTEM TO LONGER SPIN CHAINS
Considering the temperature of the atoms as the main source of dephasing in the excitation dynamics, we now explore its effect for a larger spin chain. We simulate the evolution of a system of N = 20 atoms in a line, separated by 20 µ m under
the Hamiltonian (1) (main text). In Fig. S4 we show the results for two different atom temperatures, T = 0 µ K (a), and for T = 50 µ K (b). The simulation assumes perfect preparation of the initial state | 1 ↑〉 = |↑↓ ... ↓〉 (imperfect preparation would reduce the overall amplitude of the oscillations, but would not induce any extra loss of coherence) and accounts for temperature effects through both (i) time-dependent distances R t ( ) appearing in the dipolar couplings and (ii) the finite recapture probability ε (as discussed in the previous sections). To account for (ii) we scale the obtained excitation probability by the factor [1 -ε t ( )] N -1 .
Dephasing in the dynamics is appreciable for interaction times longer than ∼ 4 µ s, but still allows for an unambiguous observation of the spin dynamics. This means that even with the finite temperature of the atoms in our tweezers, motional effects should not prevent the observation of coherent spin exchange in larger systems. Therefore, as stated in the conclusion of the main text, it is the stochastic character of the atom loading into the traps which is currently the main limitation for scalability, and we plan to explore several quasi-deterministic loading schemes that have already been demonstrated in other groups (refs. [37-40] of main text).
Figure S4. Simulated probabilities P i ↑ of finding atom i in the state |↑〉 after an interaction time τ for single atoms with a temperature T = 0 µ K (a), and T = 50 µ K (b). The system is initially prepared in the state P 1 ↑ .
[S1] For instance, for the states used in the main text, |↑〉 = 62 ∣ ∣ D 3 / 2 , m j = 3 2 / 〉 and |↓〉 = 62 ∣ ∣ P 1 / 2 , m j = 1 2 / 〉 , only the term d 1 + , d 2 , -has a nonzero contribution. If one used for instance the states |↑〉 = ∣ ∣ nD 3 / 2 , m j = 3 2 / 〉 and |↓〉 = ∣ ∣ n P ′ 3 / 2 , m j = 3 2 / 〉 , then only the term 2 d 1 0 , d 2 0 , would contribute.
[S2] One should not be confused by the notations, and believe that for instance the σ + σ -term arises from the d + d -term of the dipole dipole interaction: the σ + σ -term comes only from the fact that the resonant dipole-dipole interaction exchanges the two spin states, and this can arise even from a d d 0 0 term.
[S3] C. Tuchendler, A. M. Lance, A. Browaeys, Y. R. P. Sortais, and P. Grangier, Phys. Rev. A 78 , 033425 (2008).
|
10.1103/PhysRevLett.114.113002
|
[
"Daniel Barredo",
"Henning Labuhn",
"Sylvain Ravets",
"Thierry Lahaye",
"Antoine Browaeys",
"Charles S. Adams"
] |
2014-08-05T18:13:45+00:00
|
2015-03-21T16:55:12+00:00
|
[
"quant-ph",
"cond-mat.quant-gas",
"physics.atom-ph"
] |
Coherent Excitation Transfer in a Spin Chain of Three Rydberg Atoms
|
We study coherent excitation hopping in a spin chain realized using highly
excited individually addressable Rydberg atoms. The dynamics are fully
described in terms of an XY spin Hamiltonian with a long range resonant
dipole-dipole coupling that scales as the inverse third power of the lattice
spacing, $C_3/R^3$. The experimental data demonstrate the importance of next
neighbor interactions which are manifest as revivals in the excitation
dynamics. The results suggest that arrays of Rydberg atoms are ideally suited
to large scale, high-fidelity quantum simulation of spin dynamics.
|
1408.1057v1
|
## THE C ∗ -ALGEBRA GENERATED BY IRREDUCIBLE TOEPLITZ AND COMPOSITION OPERATORS
MASOUD SALEHI SARVESTANI, MASSOUD AMINI
Abstract. We describe the C ∗ -algebra generated by an irreducible Toeplitz operator T ψ , with continuous symbol ψ on the unit circle T , and finitely many composition operators on the Hardy space H 2 induced by certain linearfractional self-maps of the unit disc, modulo the ideal of compact operators K H ( 2 ). For composition operators with automorphism symbols, we show that this algebra is not isomorphic to the one generated by the shift and composition operators.
## 1. Introduction
The Hardy space H 2 = H 2 ( D ) is the collection of all analytic functions f on the open unit disk D satisfying the norm condition
$$\| f \| ^ { 2 } \coloneqq \lim _ { r \to 1 } \frac { 1 } { 2 \pi } \int _ { 0 } ^ { 2 \pi } | f ( r e ^ { i \theta } ) | ^ { 2 } d \theta < \infty.$$
For any analytic self-map ϕ of the open unit disk D , a bounded composition operator on H 2 is defined by
$$C _ { \varphi } \colon H ^ { 2 } \rightarrow H ^ { 2 }, \, C _ { \varphi } ( f ) = f \circ \varphi.$$
If f ∈ H 2 , then the radial limit f ( e iθ ) := lim r → 1 f ( re iθ ) exists almost everywhere on the unit circle T . Hence we can consider H 2 as a subspace of L 2 ( T ). Let φ is a bounded measurable function on T and P H 2 be the orthogonal projection of L 2 ( T ) (associated with normalized arc-length measure on T ) onto H 2 . The Toeplitz operator T φ is defined on H 2 by T f φ = P H 2 ( φf ) for all f ∈ H 2 . Coburn in [4, 5] shows that the unital C ∗ -algebra C ∗ ( T z ) generated by the unilateral shift operator T z contains compact operators on H 2 as an ideal and every element a ∈ C ∗ ( T z ) has a unique representation a = T φ + k for some φ ∈ C ( T ) and k ∈ K := K H ( 2 ). He shows that C ∗ ( T z ) / K is ∗ -isomorphic to C ( T ), and determines essential spectrum of Toeplitz operators with continuous symbol.
Recently the unital C ∗ -algebra generated by the shift operator T z and the composition operator C ϕ for a linear-fractional self-map ϕ of D is studied. For a linearfractional self-map ϕ on D , if ‖ ϕ ‖ ∞ < 1 then C ϕ is a compact operator on H 2 [15]. Therefore one should consider those linear-fractional self-maps ϕ which satisfy ‖ ϕ ‖ ∞ = 1. If moreover ϕ is an automorphism of D , then C ∗ ( T , C z ϕ ) / K is ∗ -isomorphic to the crossed product C ( T glyph[multicloseright] ) ϕ Z [7, 8]. When ϕ is not an automorphism there are three deferent cases:
2000 Mathematics Subject Classification. 47B33, 47B32.
Key words and phrases. C ∗ -algebras, Shift operator, irreducible Toeplitz operator, composition operator, linear-fractional map, automorphism of the unit disk.
- (i) ϕ has only one fixed point γ which is on the unit circle T (i.e. ϕ is a parabolic map). In this case C ∗ ( T , C z ϕ ) / K is a commutative C ∗ -algebra isomorphic to the minimal unitization of C γ ( T ) ⊕ C 0 ([0 , 1]), where C γ ( T ) is the set of functions in C ( T ) vanishing at γ ∈ T and C 0 ([0 , 1]) is the set of all f ∈ C ([0 , 1]) vanishing at zero [14].
glyph[negationslash]
- (ii) ϕ has a fixed point γ ∈ T and fixes another point in C ∪{∞} (equivalently ϕ has a fixed point γ ∈ T and ϕ ′ ( γ ) = 1 ). In this case C ∗ ( T , C z ϕ ) / K is ∗ -isomorphic to the minimal unitization of C γ ( T ) ⊕ ( C 0 ([0 , 1]) glyph[multicloseright] Z ) [14].
- (iii) ϕ fixes no point of T but there exist distinct points γ, η ∈ T with ϕ γ ( ) = η . In this case C ∗ ( T , C z ϕ ) / K is the C ∗ -subalgebra D of C ( T ) ⊕ M C 2 ( ([0 , 1])) defined by
$$\mathcal { D } = \left \{ ( f, V ) \in C ( \mathbb { T } ) \oplus M _ { 2 } ( C ( [ 0, 1 ] ) ) \, \colon V ( 0 ) = \left [ \begin{array} { c c } f ( \gamma ) & 0 \\ 0 & f ( \eta ) \end{array} \right ] \right \}$$
[9].
This paper generalizes the above results. The generalization is two fold. We replace the shift operator T z by an irreducible Toeplitz operator T ψ with continuous symbol ψ on T , and a single composition operator with finitely many composition operators on the Hardy space H 2 induced by certain linear-fractional self-maps of D . The paper is organized as follows. In section 2 we review basic facts and known results which are used later in the paper. In section 3 we find the C ∗ -algebra C ∗ ( T , C z ϕ 1 , ..., C ϕ n ) / K , where ϕ , ..., ϕ 1 n are as in the case (ii). In section 4 we replace the shift operator T z by an irreducible Toeplitz operator T ψ with symbol ψ ∈ C ( T ) and obtain more general results in the above cases. When ϕ is an automorphism, the composition operator C ϕ often generates the unilateral shift operator T z . We investigate this case in section 5.
## 2. Preliminaries
In this section we review some of the known results which are used in the next sections. Here we use C 0 ([0 , 1]) to denote the set of functions in C ([0 , 1]) vanishing at zero and [ T ] to denote the coset of operator T ∈ B H ( 2 ) in the Calkin algebra B H ( 2 ) /K H ( 2 ).
↦
↦
glyph[negationslash]
A linear-fractional self-map ρ with fixed pint γ ∈ T is parabolic if and only if ρ ′ ( γ ) = 1. In this case, ρ is conjugate to a translation on the right half plan Ω := { z ∈ C : Rez > 0 } via the conformal map α : z → ( γ + ) ( z / γ -z ) of D onto Ω. Therefor α ◦ ρ ◦ α -1 is the translation map z → z + a for some a ∈ C with non-negative real part. We denote the map ρ by ρ γ,a . This is an automorphism of D if and only if Rea = 0. For γ ∈ T , the set P γ := { C ρ γ,a : Rea > 0 } consists of all composition operators induced by parabolic non-automorphism self-maps of D fixing γ . If ϕ is a parabolic non-automorphism self-map of D , then C ϕ is irreducible [11], and C C ∗ ϕ ϕ -C C ϕ ∗ ϕ is a non-zero compact operator ( C ϕ is essentially normal) [1]. Therefore the unital C ∗ -algebra C ∗ ( P γ ) is irreducible and C ∗ ( P γ ) ∩ K = { 0 } . By Theorem 2.4.9 in [12], C ∗ ( P γ ) contains all compact operators on H 2 . Since the elements of P γ satisfy ρ γ,a ◦ ρ γ,b = ρ γ,a + b , C ∗ ( P γ ) / K is a unital commutative C ∗ -algebra. The following theorem completely describes this C ∗ -algebra.
Theorem 2.1. [10, Theorem 3 and Corollary 2] There is a unique ∗ -isomorphism Σ : C ([0 , 1]) → C ∗ ( P γ ) / K such that Σ( x a ) = [ C ρ γ,a ] for Rea > 0 . Moreover if ρ is a parabolic non-automorphism self-map of D fixing γ , then C ∗ ( C ρ ) = C ∗ ( P γ ) .
For a linear-fractional self-map ϕ of D ,
$$U _ { \varphi } = C _ { \varphi } ( C _ { \varphi } ^ { * } C _ { \varphi } ) ^ { - 1 / 2 }$$
is the partial isometry in the polar decomposition of C ϕ . Since both C ϕ and C ∗ ϕ are injective, U ϕ is a unitary operator. The following results are useful in finding the C ∗ -algebra generated by Toeplitz operators and composition operators induced by linear-fractional self-maps of D .
Theorem 2.2. [9, section 4] Suppose that ϕ is a linear-fractional non-automorphism self-map of D sending γ ∈ T to η ∈ T . For every f ∈ C ( T ) there exists compact operators k and k ′ such that
$$T _ { f } C _ { \varphi } = f ( \gamma ) C _ { \varphi } + k, \ C _ { \varphi } T _ { f } = f ( \eta ) C _ { \varphi } + k ^ { ^ { \prime } }.$$
Theorem 2.3. [7] Let ϕ , ϕ 1 2 are automorphisms of D and f ∈ C ( T ) . If U ϕ 1 , U ϕ 2 are unitary parts of the polar decomposition of C ϕ 1 , C ϕ 2 , respectively, then the operators U ϕ 1 U ϕ 2 -U ϕ oϕ 2 1 and U ϕ 1 T U f ∗ ϕ 1 -T foϕ 1 are compact.
For γ ∈ T and non-negative real number t , following [14], we consider the automorphism Ψ γ,t of D defined by
$$\Psi _ { \gamma, t } ( z ) = \frac { ( t + 1 ) z + ( 1 - t ) } { ( 1 - t ) \bar { \gamma } z + ( 1 + t ) },$$
which fixes γ , and satisfies Ψ ′ γ,t ( z ) = t . The set { Ψ γ,t : t > 0 } is an abelian group as Ψ γ,t 1 ◦ Ψ γ,t 2 = Ψ γ,t 1 t 2 for all t 1 , t 2 > 0.
If ϕ, ϕ 1 , · · · , ϕ n are linear-fractional non-automorphism self-maps of D fixing γ ∈ T , by equation (4.6) in [14]
$$C ^ { * } ( C _ { \varphi _ { 1 } }, \dots, C _ { \varphi _ { 1 } }, \mathfrak { K } ) =$$
$$\int _ { \mu } C ^ { * } ( \{ C _ { \rho _ { \gamma, a } } U _ { \Psi _ { \gamma, \varphi _ { 1 } ^ { ^ { \prime } } ( \gamma ) ^ { m _ { 1 } } \dots \varphi _ { n } ^ { ^ { \prime } } ( \gamma ) ^ { m _ { n } } } } \colon R e a > 0, ( m _ { 1 }, \dots, m _ { n } ) \in \mathbb { Z } ^ { n } \} ).$$
$$\cdot 1 \right ) & & C ^ { * } ( C _ { \varphi _ { 1 } }, \dots, C _ { \varphi _ { 1 } }, \mathfrak { K } ) = \\ & C ^ { * } ( \{ C _ { \alpha _ { 2 } } \cdot U _ { \Psi } \, \quad \cdot \quad, \quad \colon R e a > 0, ( m _ { 1 }, \dots, m _ { n } ) \in \mathbb { Z } ^ { n } \} ).$$
In particular
$$C ^ { * } ( C _ { \varphi }, \mathfrak { K } ) = C ^ { * } ( \{ C _ { \rho _ { \gamma, a } } U _ { \Psi _ { \gamma, \varphi ^ { ^ { \prime } } } ( \gamma ) ^ { n } } \colon R e a > 0, n \in \mathbb { Z } \} ).$$
glyph[negationslash]
Theorem 2.4. [14, Theorem 4.4] Let G be a collection of automorphisms of D that fix γ ∈ T . If G is an abelian group and η ′ ( γ ) = 1 for all η ∈ G \ { id } then C ∗ ( { [ C ρ γ,a U η ] : Rea > 0 , η ∈ G } ) is ∗ -isomorphic to the minimal unitization of C 0 ([0 , 1]) glyph[multicloseright] α G d where the action α : G d → Aut C ( 0 ([0 , 1])) is defined by α η ( f )( x ) = f ( x η ′ ( γ ) ) for η ∈ G f , ∈ C 0 ([0 , 1]) , and x ∈ [0 , 1] .
By (2.1), (2.2) and Theorem 2.4, the C ∗ -algebras C ∗ ( C ϕ 1 , ..., C ϕ n , K ) / K and C ∗ ( C , ϕ K ) / K are determined as follows.
glyph[negationslash]
Corollary 2.5. [14, Theorems 4.6 and 4.7] Let ϕ, ϕ 1 , · · · , ϕ n are linear-fractional non-automorphism self-maps of D that fix γ ∈ T , ϕ ′ ( γ ) = 1 and ln ϕ , ..., ′ 1 ln ϕ ′ n are linearly independent over Z . Define the actions α : Z → Aut C ( 0 ([0 , 1])) and α ′ : Z n → Aut C ( 0 ([0 , 1])) by α n ( f )( x ) = f ( x ϕ ′ ( γ ) n ) and α ′ ( m ,...,m 1 n ) ( f )( x ) = f ( x ϕ ′ ( γ ) m 1 ...ϕ ′ ( γ ) mn ) , respectively, for f ∈ C 0 ([0 , 1]) , n ∈ Z , ( m ,..., m 1 n ) ∈ Z n and x ∈ [0 , 1] . Then C ∗ ( C , ϕ K ) / K and C ∗ ( C ϕ 1 , ..., C ϕ n , K ) / K are ∗ -isomorphic to the minimal unitizations of C 0 ([0 , 1]) glyph[multicloseright] α Z and C 0 ([0 , 1]) glyph[multicloseright] α ′ Z n , respectively.
## 3. linear-fractional non-automorphism self-maps
Quertermous in [14] shows that if ϕ is a linear-fractional non-automorphism self-map of D fixing γ ∈ T then C ∗ ( T , C z ϕ ) / K is ∗ -isomorphic to the minimal unitization of C γ ( T ) ⊕ C 0 ([0 , 1]) where C γ ( T ) is the set of all f ∈ C ( T ) vanishing at γ . We extend this result to finitely many composition operators induced by linear-fractional non-automorphism self-maps of D with a common fixed point on the unit circle. Our approach is similar to that of Quertermous in [14], but there are some complications. As in previous section we use the notation [ T ] for the coset of T in the Calkin algebra. Let t 1 , .., t n be nonzero positive real numbers, γ ∈ T and Σ be the map defined in Theorem 2.1. Consider
$$\mathcal { N } _ { \gamma, t _ { 1 }, \dots, t _ { n } } = \{ \Sigma ( g ) [ U _ { \Psi _ { \gamma, t _ { 1 } ^ { m _ { 1 } } \dots t _ { n } ^ { m _ { n } } } } ] \, \colon g \in C _ { 0 } ( [ 0, 1 ] ), ( m _ { 1 }, \dots, m _ { n } ) \in \mathbb { Z } ^ { n } \}$$
and let A γ,t 1 ,...,t n be the non-unital C ∗ -algebra generated by N γ,t 1 ,...,t n . By Theorem 2.1 and the fact that Ψ γ, 1 is the identity map of D , A γ, 1 ,..., 1 ∼ = C 0 ([0 , 1]).
Proposition 3.1. If ϕ , ..., ϕ 1 n are linear-fractional non-automorphism self-maps of D fixing γ ∈ T , then
$$C ^ { * } ( T _ { z }, C _ { \varphi _ { 1 } }, \dots, C _ { \varphi _ { n } } ) / \mathfrak { K } = C ^ { * } ( \{ [ T _ { f } ] \colon f \in C ( \mathbb { T } ) \} \cup \mathcal { A } _ { \gamma, \varphi _ { 1 } ^ { ^ { \prime } } ( \gamma ), \dots, \varphi _ { n } ^ { ^ { \prime } } ( \gamma ) } ).$$
Moreover if ln ϕ ′ 1 ( γ , ..., ) ln ϕ ′ n ( γ ) are linearly independent over Z then
$$\mathcal { A } _ { \gamma, \varphi _ { 1 } ^ { ^ { \prime } } ( \gamma ), \dots, \varphi _ { n } ^ { ^ { \prime } } ( \gamma ) } \cong C _ { 0 } ( [ 0, 1 ] ) \rtimes _ { \alpha ^ { ^ { \prime } } } \mathbb { Z } ^ { n }.$$
Proof. First we note that if ϕ is a linear-fractional non-automorphism self-map of D fixing γ ∈ D , then ϕ ′ ( γ ) > 0 ([15]). By Theorem 2.1, Σ( x a ) = [ C ρ γ,a ] for Rea > 0. Since the closed linear span of { x a : Rea > 0 is dense in C 0 ([0 , 1]), A γ,ϕ ′ 1 ( γ ,...,ϕ ) ′ n ( γ ) is the same as
$$C ^ { * } ( \{ \Sigma ( g ) & [ U _ { \Psi _ { \gamma, \varphi _ { 1 } ^ { ^ { \prime } } ( \gamma ) ^ { m _ { 1 } } \dots \varphi _ { n } ^ { ^ { \prime } } ( \gamma ) ^ { m _ { n } } } } ] \, \colon g \in C _ { 0 } ( [ 0, 1 ] ), ( m _ { 1 }, \dots, m _ { n } ) \in \mathbb { Z } ^ { n } \} ) \\ & = C ^ { * } ( \{ [ C _ { \rho _ { \gamma, a } } U _ { \Psi _ { \gamma, \varphi _ { 1 } ^ { ^ { \prime } } ( \gamma ) ^ { m _ { 1 } } \dots \varphi _ { n } ^ { ^ { \prime } } ( \gamma ) ^ { m _ { n } } } } ] \, \colon R e a > 0, ( m _ { 1 }, \dots, m _ { n } ) \in \mathbb { Z } ^ { n } \} ).$$
Therefore by (2.1)
$$C ^ { * } ( T _ { z }, C _ { \varphi _ { 1 } }, \dots, C _ { \varphi _ { n } } ) / \mathfrak { K } = C ^ { * } ( \{ [ T _ { f } ] \colon f \in C ( \mathbb { T } ) \} \cup \mathcal { A } _ { \gamma, \varphi _ { 1 } ^ { ^ { \prime } } ( \gamma ), \dots, \varphi _ { n } ^ { ^ { \prime } } ( \gamma ) } ).$$
The last statement is a consequence of Corollary 2.5.
We set
$$\mathcal { C } _ { \gamma, t _ { 1 }, \dots, t _ { n } } = \{ [ T _ { f } ] + A \, \colon f \in C ^ { * } ( \psi ), A \in \mathcal { A } _ { \gamma, t _ { 1 }, \dots, t _ { n } } \}.$$
By Proposition 3.1, C ∗ ( T , C z ϕ 1 , ..., C ϕ n ) / K = C ∗ ( C γ,ϕ ′ 1 ( γ ,...,ϕ ) ′ n ( γ ) ). We show that C γ,ϕ ′ 1 ( γ ,...,ϕ ) ′ n ( γ ) is indeed a C ∗ -algebra and describe it. We need the following Lemma.
$$L e m m a \ 3. 2. \ [ 1 4, L e m m a \ 6. 3 ] \ I f \, \gamma \in \mathbb { T }, \, f \in C ( \mathbb { T } ) \ a n d \ A \in \mathcal { A } _ { \gamma, t _ { 1 }, \dots, t _ { n } }, \, t h e n \\ [ T _ { f } ] A = f ( \gamma ) A = A [ T _ { f } ].$$
Moreover if [ T f ] + A = [0] then f ≡ 0 and A = 0 .
The first part of above lemma says that C γ,ϕ ′ 1 ( γ ,...,ϕ ) ′ n ( γ ) is closed under multiplication and it is a dense ∗ -subalgebra of C ∗ ( T , C z ϕ 1 , ..., C ϕ n ) / K .
glyph[square]
Theorem 3.3. If ϕ , ..., ϕ 1 n are linear-fractional non-automorphism self-maps of D fixing γ ∈ T , then C ∗ ( T , C z ϕ 1 , ..., C ϕ 1 ) / K is ∗ -isomorphic to the minimal unitization of C γ ( T ) ⊕A γ,ϕ ′ 1 ( γ ,...,ϕ ) ′ n ( γ ) .
Proof. Let B is the minimal unitization of C γ ( T ) ⊕A γ,ϕ ′ 1 ( γ ,...,ϕ ) ′ n ( γ ) . Since C γ ( T ) = { f -f ( γ ) : f ∈ C ( T ) } we may define the map ∆ : B → C γ,ϕ ′ 1 ( γ ,...,ϕ ) ′ n ( γ ) by
$$\Delta ( ( f - f ( \gamma ), A ) + f ( \gamma ) I ) = [ T _ { f } ] + A,$$
for f ∈ C T ( ) and A ∈ A γ,ϕ ′ 1 ( γ ,...,ϕ ) ′ n ( γ ) . By part two of Lemma 3.2, ∆ is injective. Let α = ( f 1 -f 1 ( γ , A ) 1 ) + f 1 ( γ I ) and β = ( f 2 -f 2 ( γ , A ) 2 ) + f 2 ( γ I ) , for some f , f 1 2 ∈ C T ( ) and A , A 1 2 ∈ A . Then
∆( αβ ) = ∆( f f 1 2 -f 1 ( γ f ) 2 -f 2 ( γ f ) 1 + f 1 ( γ f ) 2 ( γ , A ) 1 A 2 )
- + f 1 ( γ )( f 2 -f 2 ( γ , A ) 2 ) + f 2 ( γ )( f 1 -f 1 ( γ , A ) 1 ) + f 1 ( γ f ) 2 ( γ )
- = ∆( f f 1 2 -f 1 ( γ f ) 2 ( γ , A ) 1 A 2 + f 1 ( γ A ) 1 + f 2 ( γ A ) 2 ) + f 1 ( γ f ) 2 ( γ I )
- = [ T f 1 f 2 ] + A A 1 2 + f 1 ( γ A ) 2 + f 2 ( γ A . ) 1
Therefore, again by Lemma 3.2,
∆( αβ ) = [ T f 1 f 2 ] + A A 1 2 +[ T f 1 ] A 2 +[ T f 2 ] A 1 = ([ T f 1 ] + A 1 )([ T f 2 ] + A 2 ) = ∆( α )∆( β . )
Hence ∆ is an injective ∗ -homomorphism, and thus an isometry with closed range C γ,ϕ ′ 1 ( γ ,...,ϕ ) ′ n ( γ ) . Therefore the C ∗ -algebra
$$C ^ { * } ( T _ { z }, C _ { \varphi _ { 1 } }, \dots, C _ { \varphi _ { 1 } } ) / \mathfrak { K } = \mathcal { C } _ { \gamma, \varphi _ { 1 } ^ { ^ { \prime } } ( \gamma ), \dots, \varphi _ { n } ^ { ^ { \prime } } ( \gamma ) }$$
is ∗ -isomorphic to the minimal unitization of C γ ( T ) ⊕A γ,ϕ ′ 1 ( γ ,...,ϕ ) ′ n ( γ ) . glyph[square]
The following result immediately follows from the above theorem and Proposition 3.1.
Corollary 3.4. If ϕ , ..., ϕ 1 n are linear-fractional non-automorphism self-maps of D fixing γ ∈ T and ln ϕ ′ 1 ( γ , ..., ) ln ϕ ′ n ( γ ) are linearly independent over Z , then C ∗ ( T , C z ϕ 1 , ..., C ϕ n ) / K ∗ -isomorphic to the minimal unitization of the direct sum C γ ( T ) ⊕ ( C 0 ([0 , 1]) glyph[multicloseright] α ′ Z n ) .
## 4. Irreducible Toeplitz operators
↦
Let X be a compact Hausdorff space and A be a C ∗ -subalgebra of C X ( ) containing the constants. For x, y ∈ X , put x ∼ y if and only if f ( x ) = f ( y ) for all f in A . Then ∼ is an equivalence relation on X . Let [ x ] denote the equivalence class of x and [ X ] be the quotient space and equip [ X ] with the weak topology induced by A . Let X/ ∼ be the quotient space equipped with the quotient topology. Then A is ∗ -isomorphic to C ([ X ]) and a C ∗ -subalgebra of C X/ ( ∼ ) via f → ˜ where f ˜ ([ f x ]) := f ( x ) for x ∈ X . Note that [ X ] is always Hausdorff and it is homeomorphic to X/ ∼ when the latter is Hausdorff [5].
If D is an irreducible C ∗ -subalgebra of C ∗ ( T z ), then it contains a nonzero compact operator [5]. Hence D contains all of compact operators on H 2 [12]. We set
$$D _ { 0 } = \{ f \in C ( \mathbb { T } ) \, \colon T _ { f } \in D \}.$$
Theorem 4.1. [5, Theorem 2 and 5] If D is an irreducible C ∗ -subalgebra of C ∗ ( T z ) , then D 0 is a C ∗ -subalgebra of C ( T ) and D/ K is ∗ -isomorphic to C ([ T ]) , where [ T ] is the quotient with respect to the equivalence relation induced by D 0 .
Let T ψ be an irreducible Toeplitz operator (i.e. the only closed vector subspaces of H 2 reducing for T ψ are 0 and H 2 ) with continuous symbol ψ (see [2, 4, 5, 13]). Then D = C ∗ ( T ψ ) is irreducible and by Theorem 6 in [5], D 0 is the C ∗ -algebra of C ( T ) generated by ψ . Hence by Theorem 4.1, D/ K is ∗ -isomorphic to C ([ T ]) where [ T ] is the quotient with respect to the equivalence relation induced by ψ (that is x ∼ y if and only if ψ x ( ) = ψ y ( )).
Note that T z is irreducible and there are other irreducible Toeplitz operators (for example, see Example 1 and 2 in [13]). If D = C ∗ ( T ψ ) = C ∗ ( T z ) For some continuous function ψ , then D 0 = C ( T ) is generated by ψ and by the StoneWeierstrass theorem, ψ must be one-to-one on the unit circle. Therefore we are interested in the case that ψ is not one-to-one on T .
Let T ψ be an irreducible Toeplitz operator and [ T ] be the quotient space with respect to the equivalence relation induced by ψ . Put
$$C _ { [ \gamma ] } ( [ \mathbb { T } ] ) \coloneqq \{ f \in C ( [ \mathbb { T } ] ) \, \colon f ( [ \gamma ] ) = 0 \}$$
and let B γ,t be the minimal unitization of C [ γ ] ([ T ]) ⊕A γ,t .
Theorem 4.2. If T ψ is an irreducible Toeplitz operator on Hardy space H 2 with symbol ψ ∈ C ( T ) and ϕ is a linear-fractional non-automorphism self-map of T fixing γ ∈ T , then C ∗ ( T ψ , C ϕ ) / K is ∗ -isomorphic to B γ,ϕ ′ ( γ ) .
Proof. Similar to the proof of Proposition 3.1,
$$C ^ { * } ( T _ { \psi }, C _ { \varphi } ) / \mathfrak { K } = C ^ { * } ( \{ [ T _ { \phi } ] \, \colon \varphi \in C ^ { * } ( \psi ) \} \cup \mathcal { A } _ { \gamma, \varphi ^ { ^ { \prime } } ( \gamma ) } ).$$
For t > 0, set
$$\mathcal { C } _ { \gamma, t, \psi } \coloneqq \{ [ T _ { \phi } ] + A \, \colon \phi \in C ^ { * } ( \psi ), A \in \mathcal { A } _ { \gamma, t } \}.$$
By (4.1), C ∗ ( T ψ , C ϕ ) / K = C ∗ ( C γ,t,ψ ). We show that C γ,t,ψ is a C ∗ -algebra and describe it. It is clear that C γ,t,ψ is closed under taking linear combination and adjoint. On the other hand, by Lemma 3.2 and the fact that for φ , φ 1 2 ∈ C ∗ ( ψ ), [ T φ 1 ][ T φ 2 ] = [ T φ φ 1 2 ] (since T φ φ 1 2 -T φ 1 T φ 2 is a compact operator), C γ,t,ψ is also closed under multiplication. Hence C γ,t,ψ is a dense ∗ -subalgebra of C ∗ ( T ψ , C ϕ ) / K . Similar to the proof of Theorem 3.3, we define the map F : B γ,ϕ ′ ( γ ) → C γ,ϕ ′ ( γ ,ψ ) , by
$$\mathcal { F } ( ( \tilde { f } - \tilde { f } ( [ \gamma ] ), A ) + \tilde { f } ( [ \gamma ] ) I = [ T _ { f } ] + A,$$
for f ∈ C ∗ ( ψ ) and A ∈ A γ,ϕ ′ ( γ ) . By Lemma 3.2, F is an injective ∗ -homomorphism and its image is C γ,ϕ ′ ( γ ,ψ ) = C ∗ ( T ψ , C ϕ ) / K . glyph[square]
The following results are straightforward consequences of Theorems 2.1, 4.2 and Corollary 2.5.
Corollary 4.3. If T ψ is irreducible with symbol ψ in C ( T ) and ρ is a parabolic non-automorphism self-map of D fixing γ ∈ T then, C ∗ ( T ψ , C ϕ ) / K is ∗ -isomorphic to the minimal unitization of C [ γ ] ([ T ]) ⊕ C 0 ([0 , 1]) .
glyph[negationslash]
Corollary 4.4. If T ψ is irreducible with symbol ψ in C ( T ) and ϕ is linear-fractional non-automorphism self-map of D fixing γ ∈ T such that ϕ ′ ( γ ) = 1 , then C ∗ ( T ψ , C ϕ ) / K is ∗ -isomorphic to the minimal unitization of C [ γ ] ([ T ]) ⊕ ( C 0 ([0 , 1]) glyph[multicloseright] α Z ) , where the action α is defined as in Corollary 2.5.
Corollary 4.5. If T ψ is irreducible with symbol ψ in C ( T ) and ϕ , ..., ϕ 1 n are linearfractional non-automorphism self-maps of D fixing γ ∈ T such that ln ϕ ′ 1 ( γ , ..., ) ln ϕ ′ n ( γ ) are linearly independent over Z , then C ∗ ( T ψ , C ϕ 1 , ..., C ϕ n ) / K is ∗ -isomorphic to the minimal unitization of C [ γ ] ([ T ]) ⊕ ( C 0 ([0 , 1]) glyph[multicloseright] α ′ Z n ) .
glyph[negationslash]
Now consider the case that ϕ is a linear-fractional non-automorphism self-map of D such that ϕ γ ( ) = η for some ϕ = η ∈ T . Kriete, MacCluer and Moorhouse investigated this case in [9]. We summarize their results as follows.
glyph[negationslash]
Theorem 4.6. [9] Let ϕ be a linear-fractional non-automorphism self-map of D with ϕ γ ( ) = η for some ϕ = η ∈ T . Then for every a ∈ A = C ∗ ( T , C z ϕ ) / K there is a unique ω ∈ C ( T ) and unique functions f, g, h and k in C 0 ([0 , 1]) such that
$$a = [ T _ { \omega } ] + f ( [ C _ { \varphi } ^ { * } C _ { \varphi } ] ) + g ( [ C _ { \varphi } C _ { \varphi } ^ { * } ] ) + [ U _ { \varphi } ] h ( [ C _ { \varphi } ^ { * } C _ { \varphi } ] ) + [ U _ { \varphi } ^ { * } ] k ( [ C _ { \varphi } C _ { \varphi } ^ { * } ] ).$$
Moreover the map Φ : A → C ( T ) ⊕ M 2 ( C ([0 , 1])) defined by
$$\Phi ( a ) = \left ( \omega, \left [ \begin{array} { c c } \omega ( \gamma ) + g & h \\ k & \omega ( \eta ) + f \end{array} \right ] \right )$$
is a ∗ -isomorphism of A onto the following C ∗ -subalgebra of C ( T ) ⊕ M 2 ( C ([0 , 1]))
$$\mathcal { D } = \left \{ ( \omega, V ) \in C ( \mathbb { T } ) \oplus \mathbb { M } _ { 2 } ( C ( [ 0, 1 ] ) ) \, \colon V ( 0 ) = \left [ \begin{array} { c c } \omega ( \gamma ) & 0 \\ 0 & \omega ( \eta ) \end{array} \right ] \right \}.$$
We replace the shift operator by an arbitrary irreducible Toeplitz operator T ψ with continuous symbol.
Theorem 4.7. Let ϕ be a linear-fractional non-automorphism self-map of D such that ϕ γ ( ) = η for distinct points γ, η ∈ T and T ψ be irreducible with continuous symbol ψ on T . Then every element b in B = C ∗ ( T ψ , C ϕ ) / K has a unique representation of the form
$$b = [ T _ { \omega } ] + f ( [ C _ { \varphi } ^ { * } C _ { \varphi } ] ) + g ( [ C _ { \varphi } C _ { \varphi } ^ { * } ] ) + [ U _ { \varphi } ] h ( [ C _ { \varphi } ^ { * } C _ { \varphi } ] ) + [ U _ { \varphi } ^ { * } ] k ( [ C _ { \varphi } C _ { \varphi } ^ { * } ] )$$
where ω ∈ C ∗ ( ψ ) and f, g, h and k are in C 0 ([0 , 1]) . Moreover B is ∗ -isomorphic to the C ∗ -subalgebra D of C ([ T ]) ⊕ M C 2 ( ([0 , 1])) defined by
$$\mathcal { D } = \left \{ ( f, S ) \in C ( [ \mathbb { T } ] ) \oplus \mathbb { M } _ { 2 } ( C ( [ 0, 1 ] ) ) \, \colon S ( 0 ) = \left [ \begin{array} { c c } f ( [ \gamma ] ) & 0 \\ 0 & f ( [ \eta ] ) \end{array} \right ] \right \}.$$
Proof. Since B is a C ∗ -subalgebra of C ∗ ( T , C z ϕ ) / K , by Theorem 4.6, for every element b ∈ B there is a unique ω ∈ C ( T ) and unique functions f, g, h and k in C 0 ([0 , 1]) such that
$$b = [ T _ { \omega } ] + f ( [ C _ { \varphi } ^ { * } C _ { \varphi } ] ) + g ( [ C _ { \varphi } C _ { \varphi } ^ { * } ] ) + [ U _ { \varphi } ] h ( [ C _ { \varphi } ^ { * } C _ { \varphi } ] ) + [ U _ { \varphi } ^ { * } ] k ( [ C _ { \varphi } C _ { \varphi } ^ { * } ] ).$$
We show that ω ∈ C ∗ ( ψ ). By Theorem 2.2, for each ε > 0, there is an element b ε ∈ B such that ‖ b ε - ‖ b < ε and b ε = p ([ T ψ ] , [ T ∗ ψ ]) + q ([ C ϕ ] , [ C ∗ ϕ ]), for some polynomials p and q . It is straightforward to show that p ([ T ψ ] , [ T ∗ ψ ]) = [ T p ψ,ψ ( ¯ ) ]. Hence by Theorem 4.6, there are unique functions f , g ε ε , h ε and k ε in C 0 ([0 , 1]) such that
$$b _ { \varepsilon } - b = [ T _ { \omega - p ( \psi, \bar { \psi } ) } ] + f _ { \varepsilon } & ( [ C _ { \varphi } ^ { * } C _ { \varphi } ] ) + g _ { \varepsilon } ( [ C _ { \varphi } C _ { \varphi } ^ { * } ] ) + [ U _ { \varphi } ] h _ { \varepsilon } ( [ C _ { \varphi } ^ { * } C _ { \varphi } ] ) \\ + & [ U _ { \varphi } ^ { * } ] k _ { \varepsilon } ( [ C _ { \varphi } C _ { \varphi } ^ { * } ] )$$
and
$$\left \| \left ( \omega - p ( \psi, \bar { \psi } ), \left [ \begin{array} { c c } \omega ( \gamma ) + g _ { \varepsilon } & h _ { \varepsilon } \\ k _ { \varepsilon } & \omega ( \eta ) + f _ { \varepsilon } \end{array} \right ] \right ) \right \| = \| b _ { \varepsilon } - b \| < \varepsilon.$$
Thus ‖ ω -p ψ, ψ ( ¯ ) ‖ < ε and ω ∈ C ∗ ( ψ ). By Theorem 4.6, B is ∗ -isomorphic to the following C ∗ -subalgebra of C ∗ ( ψ ) ⊕ M C 2 ( ([0 , 1]))
$$\mathcal { C } = \left \{ ( \omega, S ) \in C ^ { * } ( \psi ) \oplus \mathbb { M } _ { 2 } ( C ( [ 0, 1 ] ) ) \, \colon S ( 0 ) = \left [ \begin{array} { c c } \omega ( \gamma ) & 0 \\ 0 & \omega ( \eta ) \end{array} \right ] \right \}.$$
Hence by Theorem 4.1, B is ∗ -isomorphic to
$$\stackrel { \kappa } { \mathcal { D } } = \left \{ ( f, S ) \in \stackrel { \kappa } { C } ( [ \mathbb { T } ] ) \oplus \mathbb { M } _ { 2 } ( C ( [ 0, 1 ] ) ) \, \colon S ( 0 ) = \left [ \begin{array} { c } f ( [ \gamma ] ) & 0 \\ 0 & f ( [ \eta ] ) \end{array} \right ] \right \}.$$
## 5. composition operators with automorphism symbols
A self-map ϕ of the unit disk D is an automorphism if ϕ is a one-to-one holomorphic map of D onto D . We denote the class of automorphisms of D by Aut ( D ). A well-know consequence of Schwarz Lemma shows that every element ϕ ∈ Aut ( D ) has the form
$$\frac { s - z } { 1 - \tilde { e } \gamma },$$
$$\varphi ( z ) = \omega \frac { s - z } { 1 - \bar { s } z },$$
for some ω ∈ T , where s = ϕ -1 (0) ∈ D .
A Fuchsian group Γ is a discrete group of automorphisms of D . Fix a point z 0 ∈ D . The limit set of Γ is the set of limit points of the orbit { ϕ z ( 0 ) : ϕ ∈ Γ } in D . This is a closed subset of the unit circle and does not depend on the choice of z 0 . The limit set of a Fuchsian group has either 0,1,2, or infinitely many elements. When the limit set is infinite, it is perfect and nowhere dense (hence uncountable). A Fuchsian group Γ is called non-elementary if its limit set is infinite.
Jury in [7] describes the C ∗ -algebra C ∗ ( { C ϕ : ϕ ∈ Γ ) } / K when Γ is a nonelementary Fuchsian group. A basic point in the proof is that the non-elementary condition on Γ guarantees that the C ∗ -algebra C ∗ ( { C ϕ : ϕ ∈ Γ ) contains the } unilateral shift T z . In the next proposition we weaken this condition on Γ and still get the shift operator. We need the following lemma.
Lemma 5.1. [7, Lemma 3.2] If ϕ is an automorphism of D with a = ϕ -1 (0) and
$$f ( z ) = \frac { 1 - \bar { a } z } { ( 1 - | a | ^ { 2 } ) ^ { 1 / 2 } },$$
then C C ϕ ∗ ϕ = T T z ∗ z .
Proposition 5.2. Let Γ be a group of automorphisms on D . If the orbit { ϕ (0) : ϕ ∈ Γ } consists of at least two linear independent points over R , then C ∗ ( { C ϕ : ϕ ∈ Γ ) } contains the unilateral shift T z .
Proof. Let ϕ , ϕ 1 2 are two distinct elements of Γ. If a 1 = ϕ 1 (0) and a 2 = ϕ 2 (0), then by Lemma 5.1,
$$( 1 - | a _ { i } | ^ { 2 } ) C _ { \varphi _ { i } ^ { - 1 } } C _ { \varphi _ { i } ^ { - 1 } } ^ { * } = ( I - \bar { a } _ { i } T _ { z } ) ( I - \bar { a } _ { i } T _ { z } ) ^ { * },$$
for i = 1 2. Therefore the unital , C -algebra D := C ( { C ϕ : ϕ ∈ Γ ) contains }
∗ ∗ s i = ¯ a T i z + a T i ∗ z -| ¯ a i | 2 T T z ∗ z , ( i = 1 2) , .
If a 1 = α + iβ, a 2 = α ′ + iβ ′ then a simple computation shows that
$$d e t \left ( \begin{array} { c c } a _ { 1 } & a _ { 2 } \\ \bar { a _ { 1 } } & \bar { a _ { 2 } } \end{array} \right ) = - 2 i \, d e t \left ( \begin{array} { c c } \alpha & \alpha ^ { ^ { \prime } } \\ \beta & \beta ^ { ^ { \prime } } \end{array} \right ).$$
Hence if a , a 1 2 are linearly independent over R then the linear system
$$\left ( \begin{array} { c c } a _ { 1 } & a _ { 2 } \\ \bar { a _ { 1 } } & \bar { a _ { 2 } } \end{array} \right ) \left ( \begin{array} { c } x \\ y \end{array} \right ) = \left ( \begin{array} { c } 1 \\ 0 \end{array} \right )$$
has a unique solution. Therefore a linear combination of s , s 1 2 is of the form T ∗ z + tT T z ∗ z , for some scaler t ∈ C . If t = 0, then D contains T z , otherwise, since
$$I + t T _ { z } & = ( T _ { z } ^ { * } + t T _ { z } T _ { z } ^ { * } ) ( T _ { z } ^ { * } + t T _ { z } T _ { z } ^ { * } ) ^ { * } - \bar { t } ( T _ { z } ^ { * } + t T _ { z } T _ { z } ^ { * } ) \in \mathcal { D }, \\ \text{again } T _ { z } & \in \mathcal { D }.$$
Corollary 5.3. Let Γ be a Fuchsion group. If the limit set of Γ contains at least two linearly independent points, then C ∗ ( { C ϕ : ϕ ∈ Γ ) } contains the unilateral shift operator.
Proof. If z , z 1 2 are linearly independent points of the limit set, then one can choose two elements of the orbit { ϕ (0) : ϕ ∈ Γ } near to z , z 1 2 that are linearly independent. Hence by Proposition 5.2, the above C ∗ -algebra contains the unilateral shift operator. glyph[square]
As a consequence if Γ be a non-elementary Fuchsion group, then the above corollary shows that C ∗ ( { C ϕ : ϕ ∈ Γ ) contains } T z . Since the action of a Fuchsian group is amenable and topologically free, we have the following result, which could be proved by Proposition 5.2, similar to Theorem 3.1 in [7].
Corollary 5.4. Let Γ be a Fucsian group and the orbit { ϕ (0) : ϕ ∈ Γ } consists of at least two linear independent points over R . Then there is an exact sequence
$$0 \to \mathfrak { K } \to C ^ { * } ( \{ C _ { \varphi } \, \colon \varphi \in \Gamma \} ) \to C ( \mathbb { T } ) \rtimes \Gamma \to 0.$$
As an example of this situation, let Γ be a Fuchsian group and contains a hyperbolic map (a linear-fractional map with two distinct fixed points on the unit circle) whose fixed points are not diagonal (endpoints of one diameter). Then the C ∗ -algebra C ∗ ( { C ϕ : ϕ ∈ Γ ) contains } T z . This follows from the fact that fixed points of hyperbolic and parabolic maps are contained in the limit set.
Corollary 5.5. Let ϕ is an automorphism of the unit disk. If ϕ n (0) and ϕ m (0) are linearly independent over R , for some n, m ∈ Z , then the unital C ∗ -algebra C ∗ ( C ϕ ) contains the shift operator.
Proof. Since C ϕ -1 is the inverse of C ϕ , the C ∗ -algebra C ∗ ( C ϕ ) contains C ϕ n , for all n ∈ Z . Now the result follows from Proposition 5.2. glyph[square]
glyph[negationslash]
glyph[negationslash]
$$\varphi ( 0 ) = \omega s, \ \varphi ^ { 2 } ( 0 ) = \omega s \frac { 1 - \omega } { 1 - | s | ^ { 2 } \omega }$$
Let ϕ ∈ Aut ( D ) be of the form ϕ z ( ) = ω s -z 1 -¯ sz for some ω ∈ T and s ∈ D . If ω is not real ( ω = ± 1) and s = 0 then T z ∈ C ∗ ( C ϕ ). Indeed, a simple calculation shows that glyph[negationslash]
and if ω = x + iy ( y = 0), then
$$\frac { 1 - \omega } { 1 - | s | ^ { 2 } \omega } = \frac { 1 + | s | ^ { 2 } ( 1 - x ) + i ( | s | ^ { 2 } y - y ) } { | 1 - | s | ^ { 2 } \omega | ^ { 2 } }.$$
glyph[negationslash]
Therefore 1 -ω 1 -| | s 2 ω is not real and ϕ (0) and ϕ 2 (0) are linearly independent over R . Hence Corollary 5.5 shows that T z ∈ C ∗ ( C ϕ ). If ω is real (1 or -1) or s = 0, then all ϕ n (0)'s are dependent.
Jury in [8] finds the C ∗ -algebra C ∗ ( T , C z ϕ ) / K , for ϕ ∈ Aut ( D ), as a crossed product C ∗ -algebra. We do the same when the shift operator is replaced by a general irreducible Toeplitz operator T ψ . The above example shows that if ϕ ∈ Aut ( D ) be of the form (5.1) for some non-real ω ∈ T and non-zero s ∈ D , then the structure of C ∗ ( T , C z ϕ ) / K = C ∗ ( C ϕ ) / K does not change, if one replaces T z with T ψ . Here we check the case s = 0.
Theorem 5.6. Let T ψ is irreducible and ϕ ∈ Aut ( D ) . If the composition operator C ϕ is unitary and ϕ ψ ( ( T )) = ψ ( T ) , then there is an exact sequence of C ∗ -algebras
$$0 \to \mathfrak { K } \to C ^ { * } ( T _ { \psi }, C _ { \varphi } ) \to C ( \psi ( \mathbb { T } ) ) ) \rtimes _ { \varphi } \mathbb { Z } \to 0.$$
Proof. The image X = ψ ( T ) of ψ is a compact Hausdorff space and ϕ n ( X ) = X , for all n ∈ Z . Now Z acts on X by
$$\beta \colon \mathbb { Z } \rightarrow H o m e ( X )$$
↦
$$n \mapsto \beta _ { n }, \ \beta _ { n } ( x ) = \varphi ^ { n } ( x ),$$
for n ∈ Z and x ∈ X . This induces an action of Z on C X ( ) given by
$$\alpha \, \colon \mathbb { Z } \to A u t ( C ( X ) )$$
$$\alpha _ { n } ( f ) ( x ) = f ( \varphi ^ { - n } ( x ) ).$$
The C ∗ -algebra C ∗ ( T ψ , C ϕ ) / K is generated by C ∗ ( T ψ ) / K ∼ = C X ( ) and unitaries [ C ϕ n ]. On the other hand, Theorem 2.3 shows that the unitary representation n → [ C ϕ -n ] satisfies the covariance relation [ C ϕ ] f C [ ∗ ϕ ] = α n ( f ). Hence there is a surjective ∗ -homomorphism from the full crossed product C X ( ) glyph[multicloseright] ϕ Z to C ∗ ( T ψ , C ϕ ) / K . But the action of the amenable group Z on compact Hausdorff space X is amenable and topologically free (i.e. for each n ∈ Z , the set of points that are fixed by ϕ n has empty interior) thus similar to the proof of Theorem 2.1 in [8], the above ∗ -homomorphism is also injective and hence an isometry. glyph[square]
Corollary 5.7. Let ϕ ∈ Aut ( D ) be of the form (5.1) with s = 0 . If T ψ is irreducible and ϕ ψ ( ( T )) = ψ ( T ) , then there is an exact sequence of C ∗ -algebras
$$0 \to \mathfrak { K } \to C ^ { * } ( T _ { \psi }, C _ { \varphi } ) \to C ( \psi ( \mathbb { T } ) ) ) \rtimes _ { \varphi } \mathbb { Z } \to 0.$$
Proof. In this case ϕ z ( ) = ωz , for some ω ∈ T . It is easy to check that C ϕ is an isometry from H 2 onto H 2 . Hence by the above theorem, C ∗ ( T ψ , C ϕ ) / K is isomorphic to the crossed product C ψ ( ( T ))) glyph[multicloseright] ϕ Z . glyph[square]
One may wonder if for ϕ ∈ Aut ( D ) of the form (5.1) with s = 0, there is a function ψ that satisfies in the hypothesis of the above corollary. For example, let ϕ z ( ) = ze i 2 π 3 . Theorem 1 in [13] gives a sufficient condition for irreducibility of Toeplitz operator T ψ : if the restriction of a function ψ ∈ H 2 to a Borel subset S ⊆ T , with positive normalized Haar measure on the unit circle, is one-to-one and the sets ψ S ( ) and ψ ( T \ S ) are disjoint, then T ψ is irreducible. By the Riemann mapping theorem (for example, see [6]) there exists a biholomorphic (bijective and holomorphic) map ψ from the unit disk D onto the simply connected set
$$& \text{anonymous, } \text{map } \text{map } \text{$\nu$ on one one one one one} \\ & \mathbb { D } - ( [ 1 / 2, 1 ) \cup [ - 1 / 4 + \sqrt { 3 } / 4 i, - 1 / 2 + \sqrt { 3 } / 2 i ) \cup [ - 1 / 4 - \sqrt { 3 } / 4 i, - 1 / 2 - \sqrt { 3 } / 2 i ) )$$
in the complex plane. A result of Carath´odory in [3] states that e ψ extends continuously to the closure of the unit disk. Moreover one can choose ψ (see Figure 1) such that
$$\psi ( A ) = A ^ { ^ { \prime } }, \psi ( B ) = B ^ { ^ { \prime } }, \psi ( C ) = C ^ { ^ { \prime } }, \psi ( K ) = K, \psi ( K ^ { ^ { \prime } } ) = K ^ { \prime }, \psi ( K ^ { ^ { \prime \prime } } ) = K ^ { ^ { \prime \prime } },$$
and
$$\psi ( D ) = \psi ( D ^ { ^ { \prime } } ) = A, \psi ( E ) = \psi ( E ^ { ^ { \prime } } ) = B, \psi ( F ) = \psi ( F ^ { ^ { \prime } } ) = C.$$
where capital letters show points in the closure of the unit disk. The restriction of ψ to T , also denotes by ψ , is an element of H 2 , one-to-one on
$$S = \{ e ^ { i \theta } \, \colon \theta \in ( \frac { 2 \pi } { 1 5 }, \frac { 8 \pi } { 1 5 } ) \cup ( \frac { 4 \pi } { 5 }, \frac { 6 \pi } { 5 } ) \cup ( \frac { 2 2 \pi } { 1 5 }, \frac { 2 8 \pi } { 1 5 } ) \},$$
glyph[negationslash]
ψ S ( ) and ψ ( T \ S ) are disjoint and the image of ψ is invariant under ϕ , that is, ϕ ψ ( ( T )) = ψ ( T )). Moreover since ψ is not one-to-one on T , C ∗ ( T ψ ) = C ∗ ( T z ).
Figure 1.
Figure 2.
More generally, if the automorphism ϕ is of the form ϕ z ( ) = ze i 2 p q π where p and q are relatively prime integers with q positive, then by a similar construction,
there is a function ψ that satisfies the conditions of the above corollary and is not one-to-one on the unit circle and
$$\psi ( \mathbb { T } ) = \mathbb { T } \cup ( \bigcup _ { n = 0 } ^ { n = q - 1 } \varphi ^ { n } ( [ 1 / 2, 1 ) ),$$
(see Figure 2). We show that the C ∗ -algebras C ψ ( ( T )) glyph[multicloseright] ϕ Z and C ( T glyph[multicloseright] ) ϕ Z are non isomorphic. This is of course easy when p = 0, that is, ϕ z ( ) = z on T and the action of Z on T and ψ ( T ) is trivial. In this case,
$$C ( \psi ( \mathbb { T } ) ) \rtimes _ { i d } \mathbb { Z } \cong C ( \psi ( \mathbb { T } ) ) \otimes C ^ { * } ( \mathbb { Z } ) \cong C ( \psi ( \mathbb { T } ) ) \otimes C ( \mathbb { T } ) \cong C ( \psi ( \mathbb { T } ) \times \mathbb { T } )$$
glyph[negationslash]
and C ( T glyph[multicloseright] ) id Z ∼ = C ( T × T ) . The spectrum of these C ∗ -algebras are ψ ( T ) × T = ( T ∪ [1 / 2 1)) , × T and T 2 , which are not homeomorphic: if, on the contrary, Φ is a homeomorphism from ( T ∪ [1 / 2 1)) , × T onto T 2 and S is the half of T on the left side of the imaginary axis, then Φ( S × T ) = X × Y is connected, and so are X and Y . If X = T and Y = T then T \ X and T \ Y are connected subsets of T . Hence (( T \ X ) × T ) ∪ ( T × ( T \ Y )) ⊆ T 2 is connected, and so is Φ -1 ((( T \ X ) × T ) ∪ ( T × ( T \ Y ))) = ( ψ ( T ) \ S ) × T , a contradiction, as ψ ( T ) \ S is not connected. Similarly X = T or Y = T lead to a contradiction.
glyph[negationslash]
For the more general example, we need some preparation to show that the crossed products are non isomorphic (we refer the reader to [16] for more details). Let Y is a topological space. The T 0 -ization of Y is the quotient space ( Y ) ∼ = Y/ ∼ where ∼ is the equivalence relation on Y defined by x ∼ y if { x } = { y } . The space ( Y ) ∼ , equipped with the quotient topology, is a T 0 topological space (see Lemma 6.10 in [16]).
Let G be a topological group acting on a topological space X from left. The orbit of x ∈ X is the set G · x = { s · x : s ∈ G } . The stability group at x is G x := { s ∈ G : s · x = x } . The set of orbits is denoted by G X \ and is called the orbit space. It is equipped with the weakest topology making the natural quotient map p : X → G X \ continuous.
By Lemma 3.35 in [16], if X is second countable or locally compact, then so is G X \ . When G is a locally compact abelian group and ̂ G is the character group of G , one could define an equivalence relation on X × ̂ G by ( x, τ ) ∼ ( y, σ ) if G x · = G y · and τσ ¯ ∈ G ⊥ x . The space X × ̂ G/ ∼ is given the quotient topology. Note that if X is a T 0 -space, then X × ̂ G/ ∼ = X × ̂ G/G ⊥ x . By Remark 8.42 in [16], X × ̂ G/ ∼ is homeomorphic to ( G X \ ) ∼ × ̂ G/ ∼ .
Proposition 5.8. Let ψ be as above and ϕ z ( ) = ze i 2 p q π , with p and q relatively prime integers and q > 0 . Then the spectrum of the C ∗ -algebra C ψ ( ( T )) glyph[multicloseright] ϕ Z is homeomorphic to ( T ∪ [1 / 2 1)) , × T .
Proof. Consider the map
$$f \, \colon \mathbb { Z } \langle \psi ( \mathbb { T } ) \to \mathbb { T } \cup [ 1 / 2, 1 )$$
↦
$$\mathbb { Z } \cdot z \mapsto \begin{cases} \ z ^ { q } & \text{if $z\in \mathbb{T}$,} \\ z ^ { q } / | z | ^ { q - 1 } & \text{otherwise.} \end{cases}$$
↦
$$g \colon \psi ( \mathbb { T } ) \to \mathbb { T } \cup [ 1 / 2, 1 ) \\ z \mapsto \begin{cases} z ^ { q } & \text{if $z\in \mathbb{T}$,} \\ z ^ { q } / | z | ^ { q - 1 } & \text{otherwise} \end{cases}$$
$$z \mapsto \begin{cases} \begin{smallmatrix} \cdot \cdot \\ z ^ { q } / | z | ^ { q - 1 } & \text{otherwise} \end{smallmatrix} \end{cases}$$
Since the map
is continuous, g -1 ( U ) is open in ψ ( T ) for each open set U in T ∪ [1 / 2 1). , Since the quotient map p : ψ ( T ) → Z \ ψ ( T ) is open (Lemma 3.25 in [16]), f -1 ( U ) = p g ( -1 ( U )) is open in Z \ ψ ( T ) and so f is continuous. On the other hand, Z \ ψ ( T ) is compact, therefore f is a homeomorphism. Also Z \ ψ ( T ) is T 0 and Z and ψ ( T ) are second countable, hence by Theorem 8.39 and Remark 8.42 in [16],
$$P r i m ( C ( \psi ( \mathbb { T } ) ) \rtimes _ { \varphi } \mathbb { Z } ) \cong ( \mathbb { Z } \langle \varphi ( \mathbb { T } ) ) ^ { \sim } \times \widehat { \mathbb { Z } } / \sim \cong \mathbb { Z } \langle \varphi ( \mathbb { T } ) \times \widehat { \mathbb { Z } } / ( q \mathbb { Z } ) ^ { \perp }$$
$$\cong ( \mathbb { T } \cup [ 1 / 2, 1 ) ) \times \widehat { q \mathbb { Z } } \cong ( \mathbb { T } \cup [ 1 / 2, 1 ) ) \times \mathbb { T }.$$
But all Z -orbits Z · z are finite (and so closed in ψ ( T )), therefore by Theorem 8.44 in [16], C ψ ( ( T )) glyph[multicloseright] ϕ Z is a liminal C ∗ -algebra and by Theorem 5.6.4 in [12], Prim C ψ ( ( ( T )) glyph[multicloseright] ϕ Z ) is homeomorphic to the spectrum ( C ψ ( ( T )) glyph[multicloseright] ϕ Z ) ∧ , as required. glyph[square]
Now we could finish the argument for the general example. The spectrums of the C ∗ -algebras C ψ ( ( T )) glyph[multicloseright] ϕ Z and C ( T glyph[multicloseright] ) ϕ Z are ψ ( T ) × T and T 2 , which are not homeomorphic. Hence these C ∗ -algebras are not isomorphic.
## References
- [1] P. Bourdon, D. Levi, S. Narayan, J. Shapiro, Which linear fractional composition operators are essentially normal?, J. Math. Anal. Appl. 280 (2003), 30-53.
- [2] A. Brown and P. R. Halmos, Algebraic properties of Toeplitz operators, J. Reine Angew. Math. 213 (1964), 89-102.
- [3] C. Carath´ eodory, Uber die gegenseitige Beziehung der R¨nder bei der konformen Abbildung ¨ a des Inneren einer Jordanschen Kurve auf einen Kreis, Mathematische Annalen. 73(2) (1913), 305-320.
- [4] L. Coburn, The C ∗ -algebra generated by an isometry I, Bull. Amer. Math. Soc. 73 (1967), 722-726.
- [5] L. Coburn, The C ∗ -algebra generated by an isometry II, Trans. Amer. Math. Soc. 137 (1969), 211-217.
- [6] J. Gray, On the history of thr Riemann mapping theorem, Rendiconti Circolo Matematico di Palermo, serie II. supplemento N. 34 (1994), 47-94.
- [7] M. T. Jury, C ∗ -algebras generated by groups of composition operators, Indiana Univ. Math. J. 56 (2007), 3171-3192.
- [8] M. T. Jury, The Fredholm index for elements of Toeplitz-composition C ∗ -algebras, Integral Equations Operator Theory 58 (2007), 341-362.
- [9] T. Kriete, B. MacCluer, J. Moorhouse, Toeplitz-composition C ∗ -algebras, J. Operator Theory 58 (2007), 135-156.
- [10] T. Kriete, B. MacCluer, J. Moorhouse, Composition operators within singly generated composition C ∗ -algebras, Israel J. Math. 179 (2010), 449-477.
- [11] A. Montes-Rodr´ ıguez, M. Ponce-Escudero and S. Shkarin, Invariant subspaces of parabolic self-maps in the Hardy space, Math. Res. Lett. 17 (2010), no.1, 99107.
- [12] G. Murphy, C ∗ -Algebras and Operator Theory, Academic Press, New York, 1990.
- [13] E. A. Nordgren, Reducing subspaces of analytic Toeplitz operators, Duke Math. J. 34 (1967), 175-182.
- [14] K. S. Quertermous, Fixed point composition and Toeplitz-composition C ∗ -algebras, J. Funct. Anal. 265 (2013), 743-764.
- [15] J. H. Shapiro, Composition Operators and Classical Function Theory, Springer Verlag, Berlin, 1993.
- [16] D. P. Williams, Crossed Products of C*-Algebras, Amer. Math. Soc., Providence, RI, 2007.
( Masoud Salehi Sarvestani ) Department of Pure Mathematics, Faculty of Mathematical Sciences, University of Tarbiat Modares, Tehran, Iran
( Massoud Amini ) Department of Pure Mathematics, Faculty of Mathematical Sciences, University of Tarbiat Modares, Tehran, Iran
E-mail address :
[email protected]
| null |
[
"Masoud Salehi Sarvestani",
"Massoud Amini"
] |
2014-08-05T18:30:02+00:00
|
2014-08-05T18:30:02+00:00
|
[
"math.OA"
] |
The C*-algebra generated by irreducible Toeplitz and composition operators
|
We describe the C*-algebra generated by an irreducible Toeplitz operator
$T_{\psi}$, with continuous symbol $\psi $ on the unit circle $\mathbb{T}$, and
finitely many composition operators on the Hardy space $H^2$ induced by certain
linear-fractional self-maps of the unit disc, modulo the ideal of compact
operators $K(H^2)$. For composition operators with automorphism symbols, we
show that this algebra is not isomorphic to the one generated by the shift and
composition operators.
|
1408.1058v4
|
## Enumeration of monochromatic three term arithmetic progressions in two-colorings of cyclic groups
Erik Sj¨land o
November 11, 2014
## Abstract
One of the toughest problems in Ramsey theory is to determine the existence of monochromatic arithmetic progressions in groups whose elements have been colored. We study the harder problem to not only determine the existence of monochromatic arithmetic progressions, but to also count them. We reformulate the enumeration in real algebraic geometry and then use state of the art computational methods in semidefinite programming and representation theory to derive sharp, or an explicit constant from sharp, lower bounds for the cyclic group of any order.
## Contents
| 1 | Introduction | 2 |
|-----|-----------------------------------------------------------------------------|-----|
| 2 | Results | 2 |
| 3 | Polynomial optimization | 3 |
| 4 | Counting monochromatic arithmetic progressions using semidef- inite program | 5 |
| 5 | Proofs of Theorems 2.1 and Corollary 2.2 | 7 |
## 1 Introduction
In Ramsey theory there are two closely related concepts of forcing structures, by colors or density: If we color the integers with a finite number of colors, then there are monochromatic arithmetic progressions of any length. If we choose any subset of the integers with positive density, then there are also arithmetic progressions of any length.
Getting finite, one can ask how large does n need to be for us to find a monochromatic progression in 1 2 , , . . . , n , given a certain number of colors and a desired length of progressions? And more difficult than concluding the existence of monochromatic arithmetic progressions, could we count them? That is the goal of our research project.
To count the monochromatic arithmetic progressions, we reformulate it as a problem of real algebraic geometry that can be attacked by state of the art optimization theory and provide human readable proofs. This is not the ordinary relaxation right off from binary combinatorial problems to linear or semidefinite programming, but one that allows us to use representation theory for modding out symmetries.
In the next section we present our results, and after that we present our methods and proofs.
## 2 Results
The case n mod 24 ∈ { 1 5 7 11 13 17 19 23 , , , , , , , } in the following theorem was previously known [1]. To prove the theorem they used orthogonal arrays, and we do not see how their methods could be generalized to the other cases. Our method is based on completely different methods, and to our knowledge all other cases are new.
Theorem 2.1. Let n be a positive integer and let R (3 , Z n , 2) denote the minimal number of monochromatic 3 -term arithmetic progressions in any two-coloring of Z n . n / 2 8 -c 1 n + c 2 ≤ R (3 , Z n , 2) ≤ n / 2 8 -c n 1 + c 3 for all values of n , where the constants depends on the modular arithmetic and are tabulated in the following table.
| n mod 24 | c 1 | c 2 | c 3 |
|------------------------------------|-------|-------|--------|
| 1 , 5 , 7 , 11 , 13 , 17 , 19 , 23 | 1 / 2 | 3 / 8 | 3 / 8 |
| 8 , 16 | 1 | 0 | 0 |
| 2 , 10 | 1 | 3 / 2 | 3 / 2 |
| 4 , 20 | 1 | 0 | 2 |
| 14 , 22 | 1 | 3 / 2 | 3 / 2 |
| 3 , 9 , 15 , 21 | 7 / 6 | 3 / 8 | 27 / 8 |
| 0 | 5 / 3 | 0 | 0 |
| 12 | 5 / 3 | 0 | 18 |
| 6 , 18 | 5 / 3 | 1 / 2 | 27 / 2 |
It requires little work to see that the following result follows:
Corollary 2.2. Let n be a positive integer. Let R (3 , Z n , 2) and R (3 , D 2 n , 2) denote the minimal number of monochromatic 3 -term arithmetic progressions in any two-coloring of Z n and D 2 n respectively. The following equality holds
$$R ( 3, D _ { 2 n }, 2 ) = 2 R ( 3, \mathbb { Z } _ { n }, 2 ).$$
In particular n / 2 4 -2 c 1 n + 2 c 2 ≤ R D ( 2 n ; 3) ≤ n / 2 4 -2 c 1 n + 2 c 3 where the constants can be found in the table of Theorem 2.1.
## 3 Polynomial optimization
The established results can be understood without understanding any polynomial optimization, but finding a sum of squares certificate by hand would be very difficult. In order to find the algebraic certificates in Theorem 2.1 results from real algebraic geometry had to be used to rewrite the problem of counting arithmetic progressions as a semidefinite program. Patterns were then found in the numerical solutions of the semidefinite programs, and through careful analysis of the numerical patterns we found the algebraic certificates that proves our main result. In this article we will only give some elementary definitions and examples for the unfamiliar reader that are sufficient for our approach. For the interested reader we refer to the extensive survey by Laurent [4]. We refer to [6] for a survey on how the methods were used and implemented for the particular problem.
A polynomial optimization problem is a problem on the form
$$\rho _ { * } = \inf \quad f ( x ) \\ \text{subject to } \quad g _ { 1 } ( x ) \geq 0, \dots, g _ { m } ( x ) \geq 0, \\ x \in \mathbb { R } ^ { n },$$
where f ( x , g ) 1 ( x , . . . , g ) m ( x ) are given polynomials. A strategy to solve a polynomial optimization problem is to introduce a new variable λ , and study the related problem:
$$\begin{array} { c c } \rho ^ { * } = \sup & \lambda \\ \text{subject to } & f ( x ) - \lambda \geq 0, g _ { 1 } ( x ) \geq 0, \dots, g _ { m } ( x ) \geq 0 \\ & \lambda \in \mathbb { R }, x \in \mathbb { R } ^ { n } \end{array}$$
where f ( x , g ) 1 ( x , . . . , g ) m ( x ) are the same polynomials. For a discussion on the relationship between these problems we refer to the book by Lasserre [3], containing among other things the proof of strong duality; that ρ ∗ = ρ ∗ .
In real algebraic geometry one tries to find various relationships between nonnegative polynomials and sums of squares, which are known as Positivstellens¨tze. a Let X be a formal indeterminate, and define
$$\sigma ^ { * } = \sup \quad & \lambda \\ \text{subject to} \quad & f ( X ) - \lambda = \sigma _ { 0 } + \sum _ { i = 1 } ^ { m } \sigma _ { i } g _ { i } \\ & \sigma _ { i } \text{ is a sum of squares in } X.$$
X.
It is easy to see that σ ∗ ≤ ρ ∗ , and it can be proven that σ ∗ = ρ ∗ under some technical conditions (Archimedean). This is known as Putinar's Positivstellensatz. The problem can be relaxed to σ ∗ d , in which there are no monomials of degree larger than d . It is easy to see that σ ∗ d 1 ≤ σ ∗ d 2 if d 1 < d 2 , and Lasserre [2] has proven that if the Archimedean condition hold, then
$$\lim _ { d \to \infty } \sigma _ { d } ^ { * } = \sigma ^ { * } = \rho ^ { * } = \rho _ { * }.$$
The reason we rewrote the positivity condition as a sum of squares condition is that the latter is equivalent to a semidefinite condition: f ( x ) is a sum of squares of degree 2 d if and only if it is possible to write f ( x ) = v T d Qv d where v d is the vector of all monomials up to degree d and Q is some positive semidefinite matrix. This makes it possible to find σ ∗ d and a sum of squares based certificate for the lower bound of our original polynomial, f ( x ) = σ ∗ d + σ 0 + ∑ m i =1 σ g i i ≥ σ ∗ d , using semidefinite programming.
## Example 3.1. To find a lower bound to
$$\rho _ { * } = \inf \quad x _ { 1 } x _ { 2 }$$
we write
$$\rho ^ { * } = \sup \ \lambda \\ \text{subject to} \ \$$
and look at the degree 3 relaxation, where v 1 = [1 , X 1 , X 2 ] T :
$$\begin{array} {$$
Note that this is a semidefinite program with variables λ, Q , Q 0 + 1 , Q + 2 , Q -1 , Q -2 , which we can solve using any software for semidefinite programming. A lower bound for x x 1 2 on the set -1 ≤ x , x 1 2 ≤ 1 is obtained by solving the relaxed optimization problem. We find a sum of squares based certificate:
$$\text{optimization prootem. } w e \, \jmath n a \ a \$$
Since this is a simple example, it is easy to verify that this is the sharpest possible lower bounds using other methods. Examples with many variables and high degrees cannot be solved exact using other methods, and obtaining a lower bound using semidefinite programming is the current state-of-the art for many polynomial optimization problems.
## 4 Counting monochromatic arithmetic progressions using semidefinite program
An arithmetic progression in Z n of length k is a k -set ( k distinct element) { a, a + b, . . . , a +( k -1) b } where a ∈ Z n and b ∈ Z n /integerdivide { 0 } . When summing over all arithmetic progression we note that for example { 1 2 3 , , } , { 1 3 2 , , } , { 2 1 3 , , } , { 2 3 1 , , } , { 3 1 2 , , } , { 3 2 1 , , } all denote the same set, and hence we only use one representative to avoid double counting.
Let χ : Z n →{-1 1 , } be a 2-coloring of the group Z n , and let x g = χ g ( ) for all g ∈ Z n . Let also x = [ x , . . . , x 0 n -1 ] denote the vector of all variables x g . For a, b, c ∈ Z n , let us introduce the polynomial
$$p ( x _ { a }, x _ { b }, x _ { c } ) \ &$$
which has the property that
$$p ( x _ { a }, x _ { b }, x _ { c } ) = \$$
In other words, if { a, b, c } is an arithmetic progression, then the polynomial p is one if { a, b, c } is a monochromatic arithmetic progression and zero otherwise. It follows that the minimum number of monochromatic arithmetic progression of length 3 in a 2-coloring of Z n is
$$R ( 3, \mathbb { Z } _ { n }, 2 ) = \min _ {$$
To find a lower bound for R (3 , Z n , 2) we relax the integer quadratic optimization problem to a quadratic optimization problem on the hypercube. Since it is a relaxation, i.e. any solution of the integer program is also a solution to the hypercube problem, we have
$$\text{$wq$} \, \Psi \, \text{$wq$$$
where
$$p _ { n } = \sum _ { \{ a, b, c \} \text{$$
A priori we do not know how much we loose by doing the relaxation from the integer hypercube to the continuous hypercube. As it turns out for this problem
we loose very little as the certificates in the main theorem are at most a constant from the solution we would get from the integer program.
Finding a lower bound for the homogeneous degree 2 polynomial p n immediately gives us a lower bound for R (3 , Z n , 2). The monomial x x a b occurs in p n as many times as the pair ( a, b ) occurs in a 3-arithmetic progression, which depends on n . To find a lower bound to min x ∈ -[ 1 1] , n p n it is suitable to use the state-of-the-art methods surveyed in Section 3.
In particular, let us use the degree 3 relaxation of Putinar's Positivstellensatz, and let the maximal lower bound using this relaxation be denoted λ ∗ . Let v 1 = [ X ,. . . , X 0 n -1 , 1] T be the vector of all monomials of degree less or equal to one. We get
$$\begin{array} {$$
which is a semidefinite program that can be solved numerically for fixed n .
The coefficients in the sum of squares certificate we get using these methods are numerical, and hence to get an algebraic positivity certificate one has to analyze the solutions further. There is no general way of doing this, and in most cases it is difficult to find good lower bounds using this procedure. We need algebraic certificates to provide solutions to all cyclic groups, not only for the cases we can solve numerically.
It was enough to solve the semidefinite program above for n ≤ 20 before we found the patterns that lead to Theorem 2.1. All the necessary computations could easily be carried out on a laptop within minutes to find numerical solutions. To extend the numerical solutions for n ≤ 20 into algebraic certificates for all n was more cumbersome and required more work.
To go from numerical solutions to algebraic certificates is for most problems very difficult. For a specific group, all the information required to find a lower bound is contained in an eigenvalue decomposition of the involved matrices, but there is no general way of finding the optimal algebraic lower bound when the eigenvalues and eigenvectors have decimal expansions that cannot trivially be translated into algebraic numbers. If one is interested in a rational approximation to the lower bound one can use methods by Parrilo and Peyrl [5]. These methods gives a nice certificate for a specific problem but provide little help when one want to find certificates for an infinite family of problems.
To find an algebraic certificate for all cyclic groups one of the tricks we used was to restrict the SDP above further without changing the optimal value. We required some entires to equal one another and forced some entries to be zero. There are also many other ways to restrict the SDP further that works for other problems. Another trick one can try is to change the objective function slightly to try to force the SDP to have only one optimal solution instead of infinitely many.
## 5 Proofs of Theorems 2.1 and Corollary 2.2
To make the computations in the proofs that follow readable, let us introduce additional notation:
$$\sigma ( a ; b _ { 0 }, b _ { 1 }, \dots, b _ { n - 1 } ) = a + \sum _ { i, j \in \mathbb { Z } _ { n } } b _ { j - i } X _ { i } X _ { j }.$$
By elementary calculations we have the following equalities, which we need in the proofs:
$$\text{in the proofs} \colon \\ I _ { 1 } = \sum _ { i \in \mathbb { Z } _ { n } } ( 1 - X _ { i } ^ { 2 } ) = \sigma ( n ; - 1, 0, \dots, 0 ), \\ I _ { 2 } = \left ( \sum _ { i \in \mathbb { Z } _ { n } } X _ { i } \right ) ^ { 2 } = \sigma \left ( 0 ; 1, 2, \dots, 2 \right ), \\ I _ { 3 } = \left ( \sum _ { i \in \mathbb { Z } _ { n } } ( - 1 ) ^ { i } X _ { i } \right ) ^ { 2 } = \sigma \left ( 0 ; 1, - 2, 2, - 2, \dots, 2, - 2 \right ), \\ I _ { 4 } = \sum _ { i \in \mathbb { Z } _ { n / 2 - 1 } } ( X _ { i } - X _ { i + n / 2 } ) ^ { 2 } = \sigma \left ( 0 ; 1, 0, \dots, 0, - 2, 0, \dots, 0 \right ), \\ I _ { 5 } \ = \ \sum _ { \substack { ( X _ { i } - X _ { i + n / 3 } ) ^ { 2 } + ( X _ { i } - X _ { i + 2 n / 3 } ) ^ { 2 } + ( X _ { i + n / 3 } - X _ { i + 2 n / 3 } ) ^ { 2 } } + ( X _ { i + n / 3 } - X _ { i + 2 n / 3 } ) ^ { 2 }$$
$$I _ { 5 } \ = & \sum _ { i \in \mathbb { Z } _ { n / 3 - 1 } } ( X _ { i } - X _ { i + n / 3 } ) ^ { 2 } + ( X _ { i } - X _ { i + 2 n / 3 } ) ^ { 2 } + ( X _ { i + n / 3 } - X _ { i + 2 n / 3 } ) ^ { 2 } \\ = & \sigma \left ( 0 ; 2, 0, \dots, 0, - 2, 0, \dots, 0, - 2, 0, \dots, 0 \right ). \\ \text{The first one is non-negative due to the boundary conditions -} 1 \leq x _ { i } \leq 1$$
The first one is non-negative due to the boundary conditions -1 ≤ x i ≤ 1 and the other ones are non-negative since they are sums of squares.
To prove Theorem 2.1 we need a small Lemma.
Lemma 5.1. Let us consider Z n for any positive integer n such that 4 divides n +2 . The number of monochromatic 3 -arithmetic progressions is even for all 2 -colorings of Z n .
Proof. It is easy to see that the total number of arithmetic progressions is even. Let R ⊆ { 0 , . . . , n - } 1 be the red elements and B = { 0 1 , , . . . , n - } 1 /integerdivide R be the blue elements in a coloring of Z n . Let G V,E ( ) be the graph with vertices labeled by { 1 , . . . , n - } 1 and edges e ij between the vertices i and j whenever { 0 , i, j } forms an arithmetic progression.
Pick a vertex a in G and denote the vertices adjacent to a by { n , . . . , n 1 d } . If a = n/ 2 it is easy to see that d = 0, if a is another odd labeled vertex we have d = 2 since the integer a is in the arithmetic progressions { n -a, 0 , a } and { 0 , a, 2 a } and if a is in an even labeled vertex then d = 4 since the integer a is in the arithmetic progressions {-a, 0 , a } { , 0 , a, 2 a , } { 0 , a/ 2 , a } { , 0 , a/ 2+ n/ , a 2 } . In particular we note that d is always even.
| E | is even because there are an even number arithmetic progressions { n -a, 0 , a } and for every arithmetic progression { 0 , a, 2 a } there is an arithmetic progression { 0 , n -a, n -2 a } .
Let again a ∈ G , and let N b be the number of blue adjacent vertices, N r the number of red adjacent vertices, d = N b + N r and let N mixed denote the number of edges from a red vertex to a blue vertex in G . Suppose WLOG that a is blue. Switching the color of a from blue to red would imply that the number of edges from a red vertex to a blue vertex in G changes to N ′ mixed = N mixed + N b -N r . N b -N r is even (possibly negative) since d is even, hence since N mixed is 0 if all vertices are blue N ′ mixed must be even.
Let N mono be the number of monochromatic arithmetic progressions in Z n . Since there are an even number of edges from a red vertex to a blue vertex in G and since | E | is even the number of edges with same-colored endpoints is even. Thus switching the color of 0 would imply that there are N ′ mono = N mono + N new monochromatic arithmetic progressions in Z n where N new is even (possibly negative). The same argument holds if we change 0 to any other vertex, and thus any change of N mono is by an even number. Since N mono is even if all of Z n have the same color it follows by induction that N mono is even.
Proof of Theorem 2.1. Recall that for x ∈ {-1 1 , } n we have
$$R ( 3, \mathbb { Z } _ { n }, 2 ) = \sum _ { \{ a, b, c \} \text{ is an A.P. in } \mathbb { Z } _ { n } } \frac { x _ { a } x _ { b } + x _ { a } x _ { c } + x _ { b } x _ { c } + 1 } { 4 }.$$
To find a lower bound R (3 , Z n , 2) we need to find a lower bound for
$$p _ { n } = \sum _ { \{ a, b, c \} \text{ is an A.P. in } \mathbb { Z } _ { n } } x _ { a } x _ { b } + x _ { a } x _ { c } + x _ { b } x _ { c }.$$
To express p n ( X ) in the σ -notation is the same as to count how many arithmetic progressions a pair ( a, b ) ∈ { Z n × Z n : a < b } is in for all pairs ( a, b ), which depends on the modular arithmetic of n .
/negationslash n mod 2 = 1, n mod 3 = 0: It is easy to see that if 2 and 3 does not divide n , then ( a, b ) is in exactly three different arithmetic progressions of Z n ,
$$( a, b, c _ { 1 } ), ( a, c _ { 2 }, b ), ( c _ { 3 }, a, b ),$$
and hence
$$p _ { n } ( X ) = \sigma ( 0 ; 0, 3, \dots, 3 ).$$
Since p n ( X ) = 3 2 I 2 + 3 2 I 1 -3 n 2 ≥ -3 n 2 , and since there are ( n 2 ) distinct arithmetic progressions in Z n we get
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \geq \frac { \binom { n } { 2 } - \frac { 3 n } { 2 } } { 4 } = \frac { n ^ { 2 } - 4 n } { 8 } = \frac { ( n - 1 ) ( n - 3 ) - 3 } { 8 }.$$
Since n is odd 8 divides ( n -1)( n -3), and thus
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \geq \left \lceil \frac { ( n - 1 ) ( n - 3 ) - 3 } { 8 } \right \rceil = \frac { ( n - 1 ) ( n - 3 ) } { 8 }$$
since we know that the number of monochromatic arithmetic progressions is an integer.
The next step is to show that the lower bound is sharp by finding a coloring with ( n -1)( n -3) 8 monochromatic arithmetic progressions. To achieve this we color 0 , 1 2 , , . . . , ( n -1) / 2 red and ( n +1) 2 / , . . . , n -2 , n -1 blue. All monochromatic progressions are given by their end points, and the end points should be of the same parity of the coloring. There are two cases, n ≡ 1 and n ≡ 3 modulo 4, to treat separately. The different types of elements for the different cases are tabulated below:
| | red even | red odd | blue even | blue odd |
|----------|-------------|--------------|--------------|--------------|
| n ≡ 1(4) | ( n +3) / 4 | ( n - 1) / 4 | ( n - 1) / 4 | ( n - 1) / 4 |
| n ≡ 3(4) | ( n +1) / 4 | ( n +1) / 4 | ( n +1) / 4 | ( n - 3) / 4 |
In both cases we get
$$\dots \cdots \circ \cdots \\ R ( 3, \mathbb { Z } _ { n }, 2 ) & \quad \leq \binom { ( n + 3 ) / 4 } { 2 } + 3 \binom { ( n - 1 ) / 4 } { 2 } \\ & \quad = 3 \binom { ( n + 1 ) / 4 } { 2 } + \binom { ( n - 3 ) / 4 } { 2 } \\ & \quad = \frac { ( n - 1 ) ( n - 3 ) } { 8 } \\ \intertext { w e t h t h o h m d i e h a r n }$$
which shows that the bound is sharp.
## n mod 8 = 0, n mod 3 = 0:
/negationslash
/negationslash
If 2 divides n and b = a + n/ 2 it follows that ( a, b ) is in the 4 pairs ( a, b, c 1 ), ( a, c 2 , b ), ( a, c 3 , b ), ( c 4 , a, b ) if a -b is even and the 2 pairs ( a, b, c 1 ), ( c 2 , a, b ) if a -b is odd. When b = a + n/ 2 then ( a, a + n/ , a 2 + n ) = ( a, b, a ) and ( a -n/ , a, a 2 + n/ 2) = ( b, a, b ) are degenerate arithmetic progressions which we do not count, hence
$$p _ { n } ( X ) = \sigma ( 0 ; 0, 3, \dots, 3 ) + \sigma ( 0 ; 0, - 1, 1, - 1, \dots, 1, - 1 ) + \sigma ( 0 ; 0, \dots, 0, - 2, 0, \dots, 0 ).$$
Since p n ( X ) = 3 2 I 2 + 1 2 I 3 + I 4 +3 I 1 -3 n ≥ -3 n , and since there are ( n 2 ) -n 2 different (the term -n 2 comes from the fact that ( a, a + n 2 , a ) is not an arithmetic progression for any a ) arithmetic progressions in Z n we get
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \geq \frac { \binom { n } { 2 } - \frac { n } { 2 } - 3 n } { 4 } = \frac { n ^ { 2 } - 8 n } { 8 }.$$
To get an upper bound for R (3 , Z n , 2) we present a coloring with as few monochromatic arithmetic progressions as possible. Partition Z n into disjoint parts G 1 = { 0 , . . . , n/ 4 - } 1 , G 2 = { n/ , . . . , n/ 4 2 - } 1 , G 3 = { n/ , . . . , 2 3 n/ 4 - } 1 and G 4 = { 3 n/ , . . . , n 4 -1 } . Color G 1 and G 3 red, and G 2 and G 4 blue. Inside G 1 there will be 2 i monochromatic arithmetic progressions of the form ( a, a + n/ 8 -i, a + n/ 4 -2 ) for 1 i ≤ i ≤ n/ 8 -1, and hence a total number of ∑ n/ 8 -1 i =1 2 i = 2 n 8 ( n 8 -1) 2 = n 2 64 -n 8 arithmetic progressions. For any arithmetic progression { a, b, c } in G 1 , there is an arithmetic progression { a, b + n/ , c 2 } with
a, c in G 1 and b + n/ 2 in G 3 . Since all elements of G 1 and G 3 have the same color, all the mentioned arithmetic progressions are monochromatic. By symmetry we get that the total number of monochromatic arithmetic progressions in the coloring is 8 · ( n 2 64 -n 8 ), hence
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \leq \frac { n ^ { 2 } - 8 n } { 8 }.$$
/negationslash n mod 8 = 2, n mod 3 = 0: The arguments in the case n mod 8 = 0, n mod 3 = 0 to show that R (3 , Z n , 2) ≥ n 2 -8 n 8 requires only that n is even, which holds also in this case. By Lemma 5.1 we can sharper the lower bound by rounding up to an even number;
/negationslash
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \geq \frac { ( n - 2 ) ( n - 6 ) } { 8 }.$$
Let us partition Z n into disjoint parts G 1 = { 0 , . . . , ( n + 2) / 4 -1 } , G 2 = { ( n + 2) / 4 , . . . , n/ 2 -1 } , G 3 = { n/ , . . . , 2 (3 n + 2) / 4 -1 } and G 4 = { (3 n + 2) / 4 , . . . , n - } 1 . Color G 1 and G 3 red, and G 2 and G 4 blue. Note that | G 1 | = | G 3 | = 1 + | G 2 | = 1 + | G 4 | .
Inside G 1 there will be 2 i monochromatic arithmetic progressions of the form ( a, a +( n +2) 8 / -i, a +( n +2) 4 / -2 ) for 1 i ≤ i ≤ ( n +2) 8 / -1, and hence a total number of ∑ ( n +2) 8 / -1 i =1 2 i = 2 n +2 8 ( n +2 8 -1) 2 = ( n +2) 2 64 -n +2 8 arithmetic progressions.
Similarly, inside G 1 there will be 1 + 2 i monochromatic arithmetic progressions of the form ( a, a +( n +2) 8 / -i, a +( n +2) 4 / -2 ) for 0 i ≤ i ≤ ( n +2) 8 / -2, and hence a total number of ∑ ( n +2) 8 / -2 i =0 (2 i +1) = n +2 8 -1 + 2 ( n +2 8 -1)( n +2 8 -2) 2 arithmetic progressions.
To count all monochromatic arithmetic progressions in { 1 , . . . , n } is equivalent to count 4 #( A.P. in · G 1 )+4 #( A.P. in · G 2 ). Carrying out the elementary calculations we get
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \leq \frac { ( n - 2 ) ( n - 6 ) } { 8 }.$$
/negationslash n mod 8 = 4, n mod 3 = 0: By the same arguments as in the case with n mod 8 = 0, n mod 3 = 0 we get
/negationslash
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \geq \frac { n ^ { 2 } - 8 n } { 8 }.$$
Let us partition Z n into disjoint parts G 1 = { 0 1 , , . . . , n/ 4 -1 } , G 2 = { n/ , . . . , n/ 4 2 -1 } , G 3 = { n/ , . . . , 2 3 n/ 4 -1 } and G 4 = { 3 n/ , . . . , n 4 -1 } . Color G 1 and G 3 red, and G 2 and G 4 blue.
Inside G 1 there will be 2 i -1 monochromatic arithmetic progressions of the form ( a, a + ( n + 4) / 8 -i, a + ( n + 4) / 4 -2 ) for 1 i ≤ i ≤ ( n + 4) / 8 -1, and hence a total number of ∑ ( n +4) 8 / -1 i =1 (2 i -1) = -( n +4 8 -1) + 2 n +4 8 ( n +4 8 -1) 2 = -n 8 + 4 8 + n +4 8 n -4 8 = n 2 64 -n 8 + 2 8 arithmetic progressions.
To count all monochromatic arithmetic progressions in Z n is equivalent to count 8 · #( A.P. in G 1 ), hence
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \leq \frac { n ^ { 2 } - 8 n } { 8 } + 2.$$
/negationslash n mod 8 = 6, n mod 3 = 0: The arguments in the case n mod 8 = 0, n mod 3 = 0 to show that R (3 , Z n , 2) ≥ n 2 -8 n 8 holds also in this case. By Lemma 5.1 we can sharper the lower bound to an even number;
/negationslash
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \geq \frac { ( n - 2 ) ( n - 6 ) } { 8 }.$$
Let us partition Z n into disjoint parts G 1 = { 0 1 , , . . . , ( n +2) 4 / - } 1 , G 2 = { ( n + 2) / 4 , . . . , n/ 2 -1 } , G 3 = { n/ , . . . , 2 (3 n + 2) / 4 -1 } and G 4 = { (3 n + 2) / 4 , . . . , n - } 1 . Color G 1 and G 3 red, and G 2 and G 4 blue.
Inside G 1 there will be 2 i -1 monochromatic arithmetic progressions of the form ( a, a + ( n + 6) / 8 -i, a + ( n + 6) / 4 -2 ) for 1 i ≤ i ≤ ( n + 6) / 8 -1, and hence a total number of ∑ ( n +6) 8 / -1 i =1 (2 i -1) = -( n +6 8 -1) + 2 n +6 8 ( n +6 8 -1) 2 = -n 8 + 2 8 + n +6 8 n -2 8 = n 2 64 -4 n 64 + 4 64 arithmetic progressions.
To count all monochromatic arithmetic progressions in Z n is equivalent to count 4 · #( A.P. in G 1 ) + 4 · #( A.P. in G 2 ). Carrying out the elementary calculations we get
Similarly, inside G 2 there will be 2 i monochromatic arithmetic progressions of the form ( a, a +( n +6) 8 / -i, a +( n +6) 4 / -2 ) for 1 i ≤ i ≤ ( n +6) 8 / -2, and hence a total number of ∑ ( n +6) 8 / -2 i =1 (2 ) = 2 i ( n +6 8 -2)( n +6 8 -1) 2 = n -10 8 n -2 8 = n 2 64 -12 n 64 + 20 64 arithmetic progressions.
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \leq \frac { ( n - 2 ) ( n - 6 ) } { 8 }.$$
n mod 2 = 1, n mod 3 = 0: If 3 divides n one can easily see that the only difference from when 3 and n are coprime is that the triples ( a, a + n/ , a 3 +2 n/ 3), ( a +2 n/ , a, a 3 + n/ 3) and ( a, a +2 n/ , a 3 + n/ 3) corresponds to the same arithmetic progression. In this case
$$p _ { n } ( X ) = \sigma ( 0 ; 0, 3, \dots, 3 ) + \sigma ( 0 ; 0, \dots, 0, - 2, 0, \dots, 0, - 2, 0, \dots, 0 ).$$
Since p n ( X ) = 3 2 I 2 + I 5 + 7 2 I 1 -7 2 n ≥ -7 2 n , and since there are ( n 2 ) -2 n 3 different (the term -2 n 3 comes from the fact that we do not want to count the triples ( a, a + n 3 , a + 2 n 3 ) more than once) arithmetic progressions we get
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \geq \frac { \binom { n } { 2 } - \frac { 2 n } { 3 } - \frac { 7 n } { 2 } } { 4 } = \frac { n ^ { 2 } } { 8 } - \frac { 7 n } { 6 }.$$
To find a good coloring we split Z n into disjoint parts G , . . . , G 1 6 with | G 1 | = | G 3 | = | G 5 | = | G 2 | + 1 = | G 4 | +1 = | G 6 | +1 = n +3 6 sorted such that if i < j and we pick a ∈ G i and b ∈ G j , then a < b . G G 1 , 3 and G 5 are colored red,
and G G 2 , 4 and G 6 are colored blue. All monochromatic progressions are given by their end points, and the end points should be of the same parity of the coloring. The way we partitioned the elements makes sure that every pair of monochromatic end points has a middle point in the same color. We start by considering pairs of end points in G 1 and G 2 . We multiply this number by 9 to get all pairs with i < j , i th element in G a , j th element in G b for all a -b ≡ 0(2), and then we add n 3 to get all arithmeric progressions of the type ( a, a + n/ , a 3 + 2 n/ 3). There are two cases, n +3 6 ≡ 1 and n +3 6 ≡ 0 modulo 2, to treat separately. The different types of elements for the different cases are tabulated below:
$$\begin{array} { l l l } \frac { \left | \text{even in $G_{1}$ odd in $G_{1}$ even in $G_{2}$ odd in $G_{2}$} } { ( n + 3 ) / 6 \equiv 1 ( 2 ) } & ( n + 9 ) / 1 2 & ( n - 3 ) / 1 2 & ( n - 3 ) / 1 2 & ( n - 3 ) / 1 2 \\ ( n + 3 ) / 6 \equiv 0 ( 2 ) & ( n + 3 ) / 1 2 & ( n + 3 ) / 1 2 & ( n + 3 ) / 1 2 & ( n - 9 ) / 1 2 \end{array}$$
In both cases we get
$$\dots \dots \dots \dots \dots \dots \dots \\ R ( 3, \mathbb { Z } _ { n }, 2 ) \quad & \leq 9 ( \binom { ( n + 9 ) / 1 2 } { 2 } + 3 \binom { ( n - 3 ) / 1 2 } { 2 } ) + \frac { n } { 3 } \\ & = 9 ( 3 \binom { ( n + 3 ) / 1 2 } { 2 } + \binom { ( n - 9 ) / 1 2 } { 2 } ) + \frac { n } { 3 } \\ & = \frac { n ^ { 2 } } { 8 } - \frac { 7 n } { 6 } + \frac { 2 7 } { 8 }. \\ \intertext { two the ie an into r o n d e in o w o l o n d t h $ R ( 3 $ \mathbb { P } $ } $ \rangle $ ie an into r o n d e $ }$$
Since this is an integer, and since we know that R (3 , Z n , 2) is an integer, we can improve the lower bound slightly:
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \geq \frac { n ^ { 2 } } { 8 } - \frac { 7 n } { 6 } + \frac { 3 } { 8 }.$$
n mod 8 = 0, n mod 3 = 0: By combining the arguments for when n mod 3 = 0 and when n mod 2 = 0 we find
$$p _ { n } ( X ) = \sigma ( 0 ; 0, 3, \dots, 3 ) + \sigma ( 0 ; 0, - 1, 1, - 1, \dots, 1, - 1 ) + \\ + \sigma ( 0 ; 0, \dots, 0, - 2, 0, \dots, 0 ) + \sigma ( 0 ; 0, \dots, 0, - 2, 0, \dots, 0, - 2, 0, \dots, 0 ).$$
Since p n ( X ) = 3 2 I 2 + 1 2 I 3 + I 4 + I 5 +5 I 1 -5 n ≥ -5 n , and since there are ( n 2 ) -2 n 3 -n 2 different arithmetic progressions we get
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \geq \frac { \binom { n } { 2 } - \frac { 2 n } { 3 } - \frac { n } { 2 } - 5 n } { 4 } = \frac { n ^ { 2 } } { 8 } - \frac { 5 n } { 3 }.$$
To find a good coloring we split Z n into disjoint parts G , . . . , G 1 12 with | G 1 | = · · · = | G 12 | = n 12 sorted such that if i < j and we pick a ∈ G i and b ∈ G j , then a < b . Color G , G , G , G , G 1 3 5 7 9 , G 11 red and G , G , G , G 2 4 6 8 , G 10 , G 12 blue. By symmetry there will be equally many blue as red arithmetic progressions, so we restrict ourselves to counting red arithmetic progressions. Since n is even, any arithmetic progressions in Z n needs to have end points of the same parity. Let e , e 2 2 denote two elements of the same parity. If e 1 ∈ G 1 and
e 2 ∈ G 1 ∪ G 5 ∪ G 9 then there are 2 red arithmetic progressions ( e , a, e 1 2 ) and ( e , a 1 + n/ , e 2 2 ), whereas if e 1 ∈ G 1 and e 2 ∈ Z n /integerdivide G 1 ∪ G 5 ∪ G 9 there are no red monochromatic arithmetic progressions in Z n with e 1 and e 2 as endpoints. Counting all possibilities one finds that if there are m red arithmetic progressions in G 1 then there are 72 m + n 3 + n monochromatic arithmetic progressions in Z n . The term 72 m is easily found using symmetries to find all arithmetic progressions 'similar' to an arithmetic progression fully contained in G 1 , n 3 comes from that all triples ( a, a + n/ , a 3 +2 n/ 3) are monochromatic, and the term n comes from the arithmetic progressions of the type ( a, a + n/ , a 6 + n/ 3).
To finish the calculation of the number of arithmetic progressions in the specified coloring we need to find the number of arithmetic progressions in G 1 . We have:
$$\frac { \left | \text{ even in } G _ { 1 } \text{ odd in } G _ { 1 } } { n / 1 2 \equiv 0 ( 2 ) }$$
and thus we get
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \ & \leq 2 \cdot 7 2 \binom { n / 2 4 } { 2 } + \frac { n } { 3 } + n \\ & = \frac { n ^ { 2 } } { 8 } - \frac { 5 n } { 3 }. \\. \quad \cdot \, \begin{array} { c c c } R ( 3, \cdot ) \\ \end{array} \, \begin{array} { c } R ( 3, \cdot ) \\ \end{array} \, \begin{array} { c } R ( 3, \cdot ) \\ \end{array} \, \begin{array} { c } R ( 3, \cdot ) \\ \end{array} \, \begin{array} { c } R ( 3, \cdot ) \\ \end{array} \, \begin{array} { c } R ( 3, \cdot ) \\ \end{array} \, \begin{array} { c } R ( 3, \cdot ) \\ \end{array} \, \begin{array} { c } R ( 3, \cdot ) \\ \end{array} \, \begin{array} { c } R ( 3, \cdot ) \\ \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c c c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, 0 } \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3 3, \dots } \\ \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, { \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \ dots } \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 2, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, 0 } \end{c } R ( 3, \dots } \end{array} \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{ array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} { c } R ( 3, \dots } \, \begin{array} \\ \end{array} \, \begin{array} { c } R ( 3, \dots } \end{array} \, \begin{array} { c } R ( 3, \dots } \end{array} \end{array} \end{array} \end{array} \end{array} \end{}$$
n mod 8 = 4, n mod 3 = 0: As in the case when n mod 8 = 0, n mod 3 = 0 we get
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \geq \frac { \binom { n } { 2 } - \frac { 2 n } { 3 } - \frac { n } { 2 } - 5 n } { 4 } = \frac { n ^ { 2 } } { 8 } - \frac { 5 n } { 3 }. \\ \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \end{array}.$$
An equivalent analysis as for the case n mod 8 = 0, n mod 3 = 0 also shows that if there are m red arithmetic progressions in G 1 then there are 72 m + n 3 + n monochromatic arithmetic progressions in Z n .
The calculation for the number of arithmetic progressions in G 1 is different than the previous case:
$$\frac { \left | \text{ even in } G _ { 1 } \quad \text{odd in } G _ { 1 } } { n / 1 2 \equiv 1 ( 2 ) } \left | \text{ even in } G _ { 1 } \quad \text{odd in } G _ { 1 } \right |$$
and thus we get
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \ & \leq 7 2 \binom { ( n + 1 2 ) / 2 4 } { 2 } + 7 2 \binom { ( n - 1 2 ) / 2 4 } { 2 } + \frac { n } { 3 } + n \\ & = \frac { n ^ { 2 } } { 8 } - \frac { 5 n } { 3 } + 1 8. \\ \cdot \cdot \ \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \ } \hat { \. }$$
n mod 4 = 2, n mod 3 = 0: As in the case when n mod 8 = 0, n mod 3 = 0 we get
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \geq \frac { \binom { n } { 2 } - \frac { 2 n } { 3 } - \frac { n } { 2 } - 5 n } { 4 } = \frac { n ^ { 2 } } { 8 } - \frac { 5 n } { 3 }. \\. \quad. \ \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots$$
Since n is not divisible by 12 we cannot pursue the problem identically to the case when n mod 8 = 0, n mod 3 = 0. We have to make a small adjustment
and let | G 1 | = | G 3 | = | G 5 | = | G 7 | = | G 9 | = | G 11 | = | G 2 | + 1 = | G 4 | + 1 = | G 6 | +1 = | G 8 | +1 = | G 10 | +1 = | G 12 | +1 = n +6 12 . When the same analysis as in that case we see that if there are m 1 red arithmetic progressions in G 1 and m 2 blue arithmetic progressions in G 2 , then there are 36 m 1 + 36 m 2 + n 3 + n monochromatic arithmetic progressions in Z n .
To count the total number of arithmetic progressions note that we have the following two cases:
$$\frac { ( n + 6 ) / 1 2 \equiv 0 ( 2 ) } { ( n + 6 ) / 1 2 \equiv 1 ( 2 ) } \left | \begin{array} { c c c } \text{even in $G_{1}$} \quad \text{odd in $G_{1}$} \quad \text{even in $G_{2}$} \quad \text{odd in $G_{2}$} \\ ( n + 6 ) / 2 4 \quad ( n + 6 ) / 2 4 \quad ( n + 6 ) / 2 4 \quad ( n + 6 ) / 2 4 \quad ( n - 1 8 ) / 2 4 \\ ( n + 1 8 ) / 2 4 \quad ( n - 6 ) / 2 4 \quad ( n - 6 ) / 2 4 \quad ( n - 6 ) / 2 4 \end{array} \right |$$
Thus it follows that
$$R ( 3, \mathbb { Z } _ { n }, 2 ) & \leq 3 6 \cdot 3 \binom { ( n + 6 ) / 2 4 } { 2 } + 3 6 \binom { ( n - 1 8 ) / 2 4 } { 2 } + \frac { n } { 3 } + n \\ & \leq 3 6 \cdot 3 \binom { ( n - 6 ) / 2 4 } { 2 } + 3 6 \binom { ( n + 1 8 ) / 2 4 } { 2 } + \frac { n } { 3 } + n \\ & = \frac { n ^ { 2 } } { 8 } - \frac { 5 n } { 3 } + \frac { 2 7 } { 2 }. \\ \text{Since this is an intoror and since } R ( 3 \ \mathbb { V } _ { n } \ \frac { 3 } { 3 } \text{ has to be an intoror it follows that}$$
Since this is an integer and since R (3 , Z n , 2) has to be an integer it follows that we can improve the lower bound slightly:
$$R ( 3, \mathbb { Z } _ { n }, 2 ) \geq \frac { n ^ { 2 } } { 8 } - \frac { 5 n } { 3 } + \frac { 1 } { 2 }.$$
$$\stackrel { > } { - } \frac { n } { 8 } - \frac { \upsilon n } { 3 } + \frac { 1 } { 2 }.$$
Proof of Corollary 2.2. Let us denote the elements of the group D 2 n by 1, r , . . . , r n -1 , s , sr , . . . , sr n -1 . Since the sets of elements { 1 , r, . . . , r n -1 } and { s, sr, . . . , sr n -1 } both contain the same number of arithmetic progressions as Z n we get that R D ( 2 n ; 3) ≥ 2 R ( Z n ; 3)+ m , where m is the number of arithmetic progressions that does not only require rotation action. In D 2 n it holds that rs = sr -1 and s 2 = 1, hence ( r , r i i sr , r sr j i j sr j ) = ( r , r i i sr , r ssr j i -j r j ) = ( r , r i i sr , r j i ) is a degenerate arithmetic progression for any choice of i and j . The same happens for arithmetic progressions ( sr , sr sr i i j , sr i sr sr j j ). Since these are all possible arithmetic progressions of D 2 n that are not rotations it follows that m = 0. Since all arithmetic progressions containing reflections are degenerate we can color the sets { 1 , r, . . . , r n -1 } and { s, sr, . . . , sr n -1 } independently, and thus R D ( 2 n ; 3) = 2 R ( Z n ; 3) as desired.
## Acknowledgements
I would like to thank Alexander Engstr¨m for introducing me to this problem o and for his advice when I didn't know how to proceed. I also want to thank Markus Schweighofer and Cynthia Vinzant for their valuable feedback and corrections.
## References
- [1] Peter Cameron, Javier Cilleruelo and Oriol Serra. On monochromatic solutions of equations in groups. Rev. Mat. Iberoam. 23 (2007), no. 1, 385-395.
- [2] Jean Bernard Lasserre. Global optimization with polynomials and the problem of moments. SIAM J. Optim. 11 (2001), no. 3, 796-817.
- [3] Jean Bernard Lasserre. Moments, positive polynomials and their applications. Imperial College Press Optimization Series, 1. Imperial College Press, London, 2010. 361 pp.
- [4] Monique Laurent. Sums of squares, moment matrices and optimization over polynomials. Chapter of 'Emerging applications of algebraic geometry'. Springer, New York, 2009. 157-270.
- [5] Pablo A. Parrilo and Helfried Peyrl. Computing sum of squares decompositions with rational coefficients. Theoret. Comput. Sci. 409 (2008), no. 2, 269-281.
- [6] Erik Sj¨land. Using real algebraic geometry to solve combinatorial problems o with symmetries. Preprint available at http://arxiv.org/abs/1408.1065.
| null |
[
"Erik Sjöland"
] |
2014-08-05T18:31:31+00:00
|
2014-11-10T18:52:04+00:00
|
[
"math.CO"
] |
Enumeration of monochromatic three term arithmetic progressions in two-colorings of cyclic groups
|
One of the toughest problems in Ramsey theory is to determine the existence
of monochromatic arithmetic progressions in groups whose elements have been
colored. We study the harder problem to not only determine the existence of
monochromatic arithmetic progressions, but to also count them. We reformulate
the enumeration in real algebraic geometry and then use state of the art
computational methods in semidefinite programming and representation theory to
derive sharp, or an explicit constant from sharp, lower bounds for the cyclic
group of any order.
|
1408.1059v2
|
## Azimuthal velocity profiles in Rayleigh-stable Taylor-Couette flow and implied axial angular momentum transport
Freja Nordsiek , 1 Sander G. Huisman , 2 Roeland C. A. van der Veen , 2 Chao Sun 2 † , Detlef Lohse 2 ‡ , and Daniel P. Lathrop 1 ¶
1 Department of Physics and Institute for Research in Electronics and Applied Physics, University of Maryland, College Park, MD 20742, USA
2 Physics of Fluids Group, Faculty of Science and Technology, MESA+ Institute, and Burgers Center for Fluid Dynamics, University of Twente, 7500AE Enschede, The Netherlands
(Received ?; revised ?; accepted ?. - To be entered by editorial office)
We present azimuthal velocity profiles measured in a Taylor-Couette apparatus, which has been used as a model of stellar and planetary accretion disks. The apparatus has a cylinder radius ratio of η = 0 716, an aspect-ratio of Γ = 11 74, and the plates closing the . . cylinders in the axial direction are attached to the outer cylinder. We investigate angular momentum transport and Ekman pumping in the Rayleigh-stable regime. The regime is linearly stable and is characterized by radially increasing specific angular momentum. We present several Rayleigh-stable profiles for shear Reynolds numbers Re S ∼ O ( 10 5 ) , both for Ω i > Ω o > 0 (quasi-Keplerian regime) and Ω o > Ω i > 0 (sub-rotating regime) where Ω i,o is the inner/outer cylinder rotation rate. None of the velocity profiles matches the non-vortical laminar Taylor-Couette profile. The deviation from that profile increased as solid-body rotation is approached at fixed Re S . Flow super-rotation, an angular velocity greater than that of both cylinders, is observed in the sub-rotating regime. The velocity profiles give lower bounds for the torques required to rotate the inner cylinder that were larger than the torques for the case of laminar Taylor-Couette flow. The quasi-Keplerian profiles are composed of a well mixed inner region, having approximately constant angular momentum, connected to an outer region in solid-body rotation with the outer cylinder and attached axial boundaries. These regions suggest that the angular momentum is transported axially to the axial boundaries. Therefore, Taylor-Couette flow with closing plates attached to the outer cylinder is an imperfect model for accretion disk flows, especially with regard to their stability.
## Key words:
## 1. Introduction
Rotating shear flows are common in nature. Geophysical and astrophysical examples include the interiors of planets and stars, planetary atmospheres, and stellar and planetary accretion disks. Since direct observations and measurements are hard to perform for many of these flows, laboratory models that incorporate the essential features of these flows can be useful. A common simple rotating shear flow that can be implemented in the
†
Email address for correspondence: [email protected]
‡
¶
Email address for correspondence: [email protected]
Email address for correspondence: [email protected]
laboratory is Taylor-Couette (TC) flow, which is the flow in the fluid-filled gap between two coaxial rotating cylinders. Taylor-Couette flow has found particular applicability as a model for astrophysical accretion disks in determining their stability properties and the outward angular momentum flux which is necessary in order for material to be transported inward onto the central body (Zeldovich 1981; Richard & Zahn 1999; Richard 2001; Dubrulle et al. 2005 a ; Ji & Balbus 2013). Taylor-Couette experiments have produced contradictory answers to these questions, causing great debate centered on the effects of the no-slip axial boundaries found in Taylor-Couette experiments which do not match the open stratified boundaries of accretion disks Balbus (2011); Avila (2012); Schartman et al. (2012); Ji & Balbus (2013); Edlund & Ji (2014).
We can define a Reynolds number for the inner (outer) cylinder using the radius r i ( r o ), the rotation rate Ω i (Ω ), and the fluid's kinematic viscosity o ν , giving
$$R e _ { i } = \frac { \Omega _ { i } \, r _ { i } \, ( r _ { o } - r _ { i } )$$
Rather than using Re i and Re o , we use the shear Reynolds number Re S and the so so-called q parameter, detailed below, to compare different parts of the parameter space. They have a more intuitive relation to the shear and the global rotation. The shear Reynolds number Re S ∝ | Ω i -Ω o | , which quantifies shear, is defined as
$$R e _ { S } = \frac { 2 } { 1 + \eta } \left | R e _ { i } - \eta R e _ { o } \right |$$
where η = r /r i o is the radius ratio (Dubrulle et al. 2005 a ). Next to η , another important geometric quantity is the aspect-ratio Γ = L/ r ( o -r i ), which is the ratio of the height of the cylinders L to the gap width. To quantify the global rotation, we use the q parameter (Ji et al. 2006; Schartman et al. 2012) defined through the relation
$$\frac { \Omega _ { i } } { \Omega _ { o } } = \eta ^ { - q } \,. \quad \quad \quad$$
The q parameter is real for co-rotating cylinders, the case exclusively dealt with in this paper. Hence, we will define both Ω i and Ω o to be both positive throughout this paper. Solid-body rotation Ω i = Ω o corresponds to q = 0, Ω i > Ω o gives q > 0, Ω i < Ω o gives q < 0, and pure inner and pure outer rotation correspond to q = + ∞ and q = -∞ respectively.
Different dimensionless parameters other than Re S and q have been used, which are presented here for ease of comparison. Rather than using Re S to quantify the shear, previous work on our apparatus (van Gils et al. 2011 b , 2012) has used a Taylor number Ta = ( σ Re S ) 2 where σ = (1 + η ) 4 / ( 2 √ η ) 4 is a geometric Prandtl number (Eckhardt, Grossmann & Lohse 2007), which equals 1 057 . for our η = 0 716. . With a nearly identical geometry, Paoletti & Lathrop (2011) used a different definition of the Reynolds number, namely Re = √ σ Re S = √ Ta/σ . Another parameter quantifying global rotation is the rotation parameter, R Ω (Dubrulle et al. 2005 a ) defined as R Ω = (1 -η ) ( Re i + Re o ) / ( ηRe o -Re i ).
At low Re S , before the formation of Taylor-vortices, and in the absence of Ekman pumping from axial boundaries (e.g. periodic or free-slip axial boundary conditions); the azimuthal velocity profile is
$$u _ { \theta, \text{lam} } ( r ) = A r + \frac { B } { r }, \ \ A = \frac { \Omega$$
Figure 1: Taylor-Couette parameter space considered here (red arrow). The different regions are shown as well as three important lines of constant q : the Rayleigh line ( q = 2), the Keplerian line ( q = 3 2), and solid-body rotation ( / q = 0). The whole Rayleigh-stable region ( q < 2) is shaded with different colors for the quasi-Keplerian regime (light blue) and the sub-rotating regime (dark blue).
We will refer to this as laminar Taylor-Couette flow.
Some rotating flows have radially increasing specific angular momentum (sign ( ∂glyph[lscript]/∂ r ) = sign glyph[lscript] ), where glyph[lscript] = r 2 ω is the specific angular momentum and ω = u /r θ is the fluid angular velocity. Such flows (see figure 1); as long as they are purely hydrodynamic, barotropic, and stably stratified as we consider here; are stable to infinitesimal perturbations (i.e. linearly stable) according to the Rayleigh criterion (Rayleigh 1917). For Taylor-Couette flow, this corresponds to q < 2. Flows for which q > 2 are linearly unstable at sufficiently high Reynolds numbers (Taylor 1923), which is often called the centrifugal instability. Hence q = 2 is referred to as the Rayleigh line. The Rayleigh-stable region includes subrotation (Ω i < Ω ), solid-body rotation (Ω o i = Ω ), and super-rotation (Ω o i > Ω ). The o flow in the super-rotating region is often referred to as quasi-Keplerian, since it includes cylinder rotation rates ( q = 3 2) obeying Kepler's 3rd law relating orbital radius and / period. This regime is of particular relevance to astrophysical systems such as accretion disks since they are Rayleigh-stable with azimuthal flow profiles in the plane of the disk that are expected to not deviate significantly from Kepler's 3rd law when ignoring the disk's self-gravitation and relativistic effects (Richard & Zahn 1999; Richard 2001; Dubrulle et al. 2005 a ; Ji & Balbus 2013).
Accretion disks are Rayleigh-stable but are known to have accretion rates requiring radial fluxes of angular momentum far greater than the flux provided by viscous diffusion in laminar Taylor-Couette flow-like disks, indicating that they are in fact unstable (Richard & Zahn 1999; Richard 2001; Dubrulle et al. 2005 a ; Ji & Balbus 2013). There has been a search for the instabilities at play in these flows. Disks sufficiently ionized to be electrically conductive are known to be unstable via the Magneto Rotational In-
stability (MRI) (Ji & Balbus 2013, and description therein). For weakly ionized disks or parts of disks, investigation has focused on stability in the presence of stratification (Dubrulle et al. 2005 b ; Le Bars & Le Gal 2007; Le Diz`s & Riedinger 2010) and stability e to finite amplitude perturbations (non-linear stability) which has been the subject of several Taylor-Couette experiments including our own in this paper (Richard 2001; Ji et al. 2006; Paoletti & Lathrop 2011; Paoletti et al. 2012; Schartman et al. 2012; Edlund & Ji 2014).
For an incompressible fluid in the Rayleigh-stable region of Taylor-Couette flow and compressible accretion disk flow, the possibility of a non-linear instability has not yet been ruled out for all Re S . Plane Couette flow and pipe flow are both examples of linearly stable flows that have non-linear instabilities at sufficient Re (Grossmann 2000; Avila et al. 2011; Shi et al. 2013, and references therein). Maretzke et al. (2014) found transient growth, a necessary prerequisite for a non-linear instability, in Rayleigh-stable Taylor-Couette flow. Accretion disks have very high Reynolds numbers with Re S possibly as high as 10 14 (Paoletti et al. 2012; Ji & Balbus 2013). Therefore, it is reasonable to ask whether Rayleigh-stable Taylor-Couette flow is non-linearly stable or unstable.
In prior experimental work; visualization via Kalliroscope particles, angular momentum transport measurements, and velocimetry measurements were done; yielding contradictory results on the presence of a non-linear instability, especially for quasi-Keplerian flow (Wendt 1933; Taylor 1936 a b , ; Coles 1965; Richard 2001; Ji et al. 2006; Borrero-Echeverry et al. 2010; Paoletti & Lathrop 2011; Burin & Czarnocki 2012; Schartman et al. 2012; Paoletti et al. 2012; Edlund & Ji 2014). These experiments have, to varying degree, Ekman pumping driven by the no-slip boundary conditions on the axial boundaries. The Ekman pumping could destabilize the flow depending on the axial end configuration in a way that would not be found in astrophysical accretion disks (Balbus 2011; Avila 2012; Schartman et al. 2012; Ji & Balbus 2013; Edlund & Ji 2014), which have open stratified axial boundaries. Axial boundaries that rotate with the outer cylinder, such as those on the apparatus presented in this paper, were found to have Ekman pumping effects that spanned the whole flow volume (Avila 2012; Schartman et al. 2012; Edlund & Ji 2014), which might explain the large, and likely turbulent, angular momentum transport found by the Maryland experiment (Paoletti & Lathrop 2011) in contrast to the low angular momentum transport steady laminar flow found in the Princeton MRI and HTX experiments which reduced the Ekman pumping by splitting the axial boundaries into rings rotated at speeds intermediate that of the two cylinders (Ji et al. 2006; Schartman et al. 2012; Edlund & Ji 2014).
The effect of the Ekman pumping in wide-gap ( η < 0 34) low aspect-ratio (Γ . < 3) Rayleigh-stable experiments, such as the Princeton MRI and HTX experiments, on the flow state and angular momentum transport has been the subject of several investigations. When the axial boundaries are attached to the outer cylinder as opposed to rotating at intermediate speeds, there are large fluctuations and mixing near the inner cylinder (Dunst 1972; Edlund & Ji 2014) and quiescent flow rotating close to Ω o near the outer cylinder (Dunst 1972; Kageyama et al. 2004; Schartman et al. 2012; Edlund & Ji 2014). Speeding up the part of the axial boundaries near the inner cylinder causes the fluctuations near the inner cylinder to decrease and the azimuthal velocities to more closely match laminar Taylor-Couette flow (Edlund & Ji 2014). In the reduced Ekman pumping configuration, perturbations by jets from the inner cylinder were found to decay for Re S glyph[lessorequalslant] 10 6 (Edlund & Ji 2014).
The effects of the Ekman pumping in medium-gap ( η ≈ 0 7) . larger aspect-ratio (Γ ∼ 10) experiments, such as the Maryland and our experiments, has not received as much attention, though it has been expected to be similar, which would resolve the
Figure 2: Sketch of the T 3 C apparatus used for the measurements presented in this paper. The inner cylinder is split into three sections, having a 2 5 mm gap between each section . marked as positions 1 and 2.
contradictory results. At low Re S < 10 , axial boundaries attached to the outer cylinder 4 were found to destabilize the flow (Avila 2012). For Re S glyph[lessorequalslant] 10 5 Direct Numerical Simulations (DNS) with periodic axial boundaries in the quasi-Keplerian regime, Ostilla-M´nico o et al. (2014) found that initial turbulent states always decayed to laminar Taylor-Couette flow. In this paper, we present azimuthal velocimetry profiles in both the quasi-Keplerian and the sub-rotating regimes in a geometry similar to the Maryland experiment. We compare them to the profiles in laminar Taylor-Couette flow and discuss their structure to better elucidate the effects of Ekman pumping on the flow and the large angular momentum transport associated with axial boundaries fixed to the outer cylinder for our geometry (Paoletti & Lathrop 2011).
The paper is organized as follows: section 2 describes the experiment and the parameter space explored, section 3 presents the azimuthal velocity and specific angular momentum profiles, section 4 presents further analysis and discussion of the azimuthal profiles including the primarily axial transport of angular momentum, and section 5 summarizes the results and presents conclusions.
## 2. Experiment and explored parameter space
The apparatus is described in detail in van Gils et al. (2011 a ) and is shown schematically in figure 2. A brief summary is given here. The inner cylinder has an outer radius of r i = 20 00 cm and the transparent . outer cylinder has an inner radius of r o = 27 94 cm, which gives . η = 0 716. They can independently rotate up to a maximum . of | Ω i / π 2 | = 20 Hz and | Ω o / 2 π | = 10 Hz, respectively. The total height is L = 93 2 cm, . which gives Γ = 11 74. The axial boundaries are attached to the outer cylinder. The inner . cylinder is split into three sections. The height of the middle section is L mid = 53 6 cm, .
Table 1: The q values for which velocity profiles were measured and their corresponding rotation rate ratios (Ω i / Ω ) and rotation parameter, o R Ω . We also give the region of the Taylor-Couette parameter space the measurement is in, and for what Re S measurements were taken.
| | | | | Re S | Re S | Re S |
|-----------|-----------|------------|-------------------|---------------|---------------|---------------|
| q | Ω i / Ω o | R Ω | Region | 2 . 07 × 10 4 | 1 . 04 × 10 5 | 7 . 81 × 10 5 |
| 2 . 100 | 2 . 018 | - 0 . 9533 | Rayleigh unstable | | glyph[check] | glyph[check] |
| 1 . 909 | 1 . 893 | - 1 . 047 | quasi-Keplerian | | glyph[check] | glyph[check] |
| 1 . 500 | 1 . 651 | - 1 . 333 | quasi-Keplerian | glyph[check] | glyph[check] | glyph[check] |
| 1 . 258 | 1 . 523 | - 1 . 587 | quasi-Keplerian | | glyph[check] | |
| 0 . 692 | 1 . 260 | - 2 . 900 | quasi-Keplerian | | glyph[check] | |
| 0 . 333 | 1 . 118 | - 6 . 062 | quasi-Keplerian | | glyph[check] | |
| - 0 . 500 | 0 . 8461 | 4 . 141 | sub-rotating | | glyph[check] | |
| - 1 . 000 | 0 . 7158 | 2 . 113 | sub-rotating | | glyph[check] | |
| - 2 . 000 | 0 . 5124 | 1 . 113 | sub-rotating | | glyph[check] | |
and the end sections have equal heights of L e = 19 35 cm. There is a 2 5 mm gap between . . each section, labeled 1 and 2 in figure 2. The system was filled with water and operated at room temperature with cooling applied at the axial boundaries. This geometry is similar to the apparatus used by Paoletti & Lathrop (2011), who have a η = 0 7245 and . Γ = 11 47 geometry. .
The azimuthal velocity profiles were obtained using Laser Doppler Anemometry (LDA). The LDA configuration used backscatter from seed particles in a measurement volume of approximately 0 07 mm . × 0 07 mm . × 0 3 mm. Dantec PSP-5 particles with a 5 . µ m diameter and 1 03 g . / cm 3 density were used. The optical effect of the outer cylinder curvature on the LDA measurements was corrected by using the calculations of Huisman et al. (2012 b ). The velocimetry was calibrated using radial and axial profiles of solidbody rotation at different rotation rates. The error in the mean velocity profiles from the calibration, which was the dominant source in the mean profiles, was smaller than 0 1%. For all LDA measurements a statistical convergence of 1% was achieved, which . translates to between 1% and 6% of | Ω i -Ω o | , which prevents investigation into fluctuations and deviations from axisymmetry. When measuring close to the inner cylinder, reflections from the metal inner cylinder were found to be problematic. Hence, the radial profiles presented in this paper were done at the axial height of the 2 5 mm gap between . the bottom and middle inner cylinder sections, which corresponds to an axial height z/L = 0 209 off the bottom, so that the LDA laser would be absorbed in the gap as . opposed to being reflected off the cylinder surface. The axial dependence of the angular velocity was found to be less than 2% of | Ω i -Ω o | from axial profiles at midgap from midheight to 1 5 cm off the bottom, and between radial profiles over the outer half of the . gap at five heights z = 0 195 { . , 0 223 . , 0 414 . , 0 464 . , 0 927 . } m off the bottom, which are at z/L = { 0 209 . , 0 238 . , 0 444 . , 0 497 . , 0 995 . } . The last one, z/L = 0 995 is 5 mm from the . top axial boundary. Thus, a radial profile at z/L = 0 209 is representative, other than . possibly for radial positions closer than 2 5 mm to the inner cylinder. The boundary . layers on the axial boundaries are confined to within 5 mm of the boundaries.
Velocimetry was performed for five quasi-Keplerian q values including Keplerian ( q = 1 500), three sub-rotating values of . q , and one unstable but very close to the Rayleigh line q value ( q = 2 100); . which are all listed in table 1. The value q = 1 909 was .
chosen to match the simulations of Avila (2012) on a nearly identical geometry and the Princeton experimental work at q = 1 9 (Ji . et al. 2006; Schartman et al. 2012). Also, q = { 1 909 . , 1 500 . , 1 258 . , 0 692 . } were chosen to match ongoing torque measurements on the Maryland experiment. Measurements for all values of q were taken at Re S = 1 04 . × 10 , the three values of 5 q glyph[greaterorequalslant] 1 500 at . Re S = 7 81 . × 10 , and 5 q = 1 500 at . Re S = 2 07 . × 10 . All of the azimuthal velocity profiles, radial profiles at all 5 heights 4 and the axial profile at midgap, are available in the supplementary material. Each pair of q and Re S was reached by starting with both cylinders at rest, linearly increasing Ω i and Ω o to their final values over 120 s while maintaining constant q , and then waiting at least 600 s for transients to decay before doing measurements.
## 3. Results on the azimuthal profiles
It is convenient to look at the velocity profiles in terms of the normalized radial position and the normalized angular velocity given by
$$\widetilde { r } = \frac { r - r _ { i } } { d } \,,$$
$$\widetilde { \omega } = \frac { \omega - \Omega _ { o } } { \Omega _ { i } - \Omega _ { o } }$$
where d = r o -r i is the width of the gap. The expression for ˜ r gives ˜ r = 0 at the inner cylinder and ˜ r = 1 at the outer cylinder. Regardless of which cylinder has the larger angular velocity, the expression for ˜ ω gives ˜ ω = 0 whenever ω = Ω o and ˜ ω = 1 whenever ω = Ω . For the quasi-Keplerian regime, i ˜ ω > 1 indicates a super-rotating flow with ω > Ω i > Ω o > 0, and ˜ ω < 0 indicates a sub-rotating flow with Ω i > Ω o > ω . Due to the sign change in the denominator for the sub-rotating regime, ˜ ω > 1 indicates a sub-rotating flow with Ω o > Ω i > ω and ˜ ω < 0 implies a super-rotating flow with ω > Ω o > Ω i > 0. The laminar Taylor-Couette profile, which in these normalized variables is independent of Ω i and Ω , is o
$$\widetilde { \omega } _ { \text{lam} } & = \frac { \eta ^ { 2 } \left ( 1 - \widetilde { r } \right ) \left ( \widetilde { r } \left ( 1 - \eta \right ) + 1 + \eta \right ) } { \left ( 1 + \eta \right ) \left ( \widetilde { r } \left ( 1 - \eta \right ) + \eta \right ) ^ { 2 } } \. \quad \quad \quad$$
The ˜ ω profiles for all values of q at Re S = 1 04 . × 10 5 are compared to each other and to the laminar Taylor-Couette profile in figure 3. None of the profiles matched the laminar Taylor-Couette profile. Approaching solid-body rotation ( q → 0) at fixed Re S in both regimes, deviation from the laminar Taylor-Couette profile increased and the part of the profile near the inner cylinder steepened. For the quasi-Keplerian regime, as we approach solid-body rotation, the rest of the profile flattens towards ˜ ω = 0. For the sub-rotating regime, ˜ ω < 0 away from the inner cylinder. This indicates that the fluid is super-rotating in terms of angular velocity compared to both cylinders ( ω > Ω o > Ω i > 0) with the degree of super-rotation, as a fraction of | Ω i -Ω o | , increasing as we approach solid-body rotation. This flow super-rotation will be further discussed in section 4.1.
The resulting profiles of the specific angular momentum glyph[lscript] = r 2 ω at Re S = 1 04 . × 10 5 are shown in figure 4. For the quasi-Keplerian regime, the specific angular momentum profiles all follow the same pattern of having an inner flat region connected to an outer region rotating at Ω o , which will be discussed further in section 4.2. The flat region in glyph[lscript] indicates that the flow was well mixed in that region.
Keplerian ( q = +1 500) profiles for three different . Re S are compared in figure 5. They all have a similar shape; but as Re S is increased, ˜ ω decreases towards solid-body rotation
Figure 3: Comparison of the normalized angular velocity ˜ ω = ( ω -Ω ) o / (Ω i -Ω ) proo files across the gap for different values of the q parameter at Re S = 1 04 . × 10 . (a) shows 5 the full scale of ˜ ω (error bars are smaller than the symbols) and (b) shows an expansion around ˜ ω = 0, using the same symbols to emphasize the parts of the profiles close to rotation at Ω . Connecting lines are drawn to guide the eye. The profile for laminar o Taylor-Couette flow is drawn for comparison.
at Ω , especially in the outer parts of the gap. In terms of the specific angular momentum, o increasing Re S leads to a sharper transition between the flat region and the rotation at Ω o region.
## 4. Further analysis and discussion
4.1. Super-rotating flow for the sub-rotating regime
As seen in figure 3 for all three sub-rotating profiles, ˜ ω < 0 except near the inner cylinder indicating flow super-rotation (figure 3b). The flow super-rotation can be quantified by taking the minimum ˜ ω in the profile to be the strength of the super-rotation, and finding its radial position along with where the linear interpolation of where the profile crosses ω = Ω o ( ˜ ω = 0) to super-rotation. The strength of the super-rotation is shown in figure 6a, and the radial locations of the maximum super-rotation and of ˜ ω = 0 are shown in figure 6b. Approaching solid-body rotation ( q → 0) at fixed non-zero Re S , the strength of super-rotation increases, and the radial positions of the super-rotation maximum and of ˜ ω = 0 both move towards the inner cylinder. This is a singular limit, which is very different from the limit q → 0 in which case one would get ω r ( ) = Ω i = Ω . The distance o between these radial positions was approximately the same for all three q , namely a value of 0 2 gap-widths. The flow super-rotation was seen at all five heights for which radial .
Figure 4: The specific angular momentum ( glyph[lscript] = r 2 ω ) profiles across the gap for the different q values at Re S ≈ 1 04 . × 10 . 5 The red circles ( · ) are the specific angular momentum profiles of the flow, with connecting lines to guide the eye. Error bars are smaller than the symbols. The solid black line ( -) and dashed blue line ( --) are the specific angular momentum profiles for ω r ( ˜ ) = Ω i and ω r ( ˜ ) = Ω , respectively. o The vertical axes have the same units and the horizontal axes are the same for all plots. The Keplerian configuration is shown at the top-right.
Rotation at
Rotation at
Ω
Ω
ReS
= 2 07
.
i
o
×
ReS
ReS
= 1 03
.
= 7 79
.
0.4
˜
r
(b)
Figure 5: Comparison of the normalized angular velocity and specific angular momentum profiles across the gap for the Keplerian cylinder rotation ratio ( q = +1 500) at three . different Re S . (a) the normalized angular velocity ˜ ω with the laminar Taylor-Couette profile drawn for comparison and (b) the specific angular momentum glyph[lscript] normalized by r 2 i Ω o with lines for solid-body rotation at the inner and outer cylinder rotation rates ( r 2 Ω and i r 2 Ω , respectively). The error bars are smaller than the symbol heights. o
profiles of the velocity were taken. They vary from each other by ∆ ˜ ω < 0 01 axially over . the outer half of the gap. The 0 2 gap-width separation was seen at the other heights for . q = -0 503, but could not be resolved for . q = {-1 001 . , -1 994 . } since the point where ˜ ω = 0 lies in the inner half of the gap.
The specific angular momentum profiles in figure 4 were slightly greater than for solidbody rotation with the outer cylinder, except close to the inner cylinder, which is another way of saying there is flow super-rotation. The Navier-Stokes equation does not constrain angular velocities to be bound by Ω i and Ω o due to its non-linear term, unlike the temperature field in Rayleigh-B´nard flow, which is contrained between the two plate temperae tures as the temperature advection equation is linear. Even with the super-rotation, we still have ∂glyph[lscript]/∂ r > 0 over the parts of the gap that are resolved; and glyph[lscript] is bound between the specific angular momenta of the outer cylinder and the axial boundaries at ˜ r = 0, which are the locations of the largest and smallest glyph[lscript] on the axial boundaries, respectively. Angular momentum is transported to the inner cylinder in this regime since the torque on the inner cylinder is negative (Paoletti & Lathrop 2011). With inward advection of angular momentum across the gap (there is also the possibility of axial transport), the outer cylinder and axial boundaries must be the source of angular momentum to sustain the flow super-rotation against spin down to ω = Ω . This also allows one to estimate the o maximum flow super-rotation that could be seen. If fluid from the outer cylinder having specific angular momentum glyph[lscript] = r 2 o Ω o is transported to the inner cylinder while conserv-
×
×
10 4
10 5
10 5
0.6
)
o
Ω
2
r
i
(
/lscript /
3.5
3
2.5
2
1.5
1
0
0.2
0.8
1
q
Figure 6: Super-rotating flow strength and locations in the sub-rotating regime at Re S = 1 04 . × 10 . (a) Flow super-rotation ( 5 ω > Ω o > Ω i > 0) strength ω -Ω o at each q as a percentage of | Ω i -Ω o | , Ω , and Ω . (b) The radial positions where the profile crosses i o ˜ ω = 0 ( ω = Ω ) to be super-rotating, and the radial position where the super-rotation o is at its maximum (minimum ˜ ω ). Connecting lines are drawn to guide the eye in both plots.
ing glyph[lscript] , it will have an angular velocity ω s = Ω o /η 2 . Normalizing the flow super-rotation ω s -Ω o respectively by | Ω i -Ω o | , Ω , and Ω , we get i o
$$\frac { \omega _ { s } - \Omega _ { o } } { | \Omega _ { i } - \Omega _ { o } | } \leqslant \left ( \frac { 1 - \eta ^ { 2 } } { \eta ^ { 2 } } \right ) \left | \frac { 1 } { \eta ^ { - q } - 1 } \right | \,,$$
$$\frac { \cdot } { \Omega _ { i } } \leqslant \left ( \frac { 1 - \eta ^ { 2 } } { \eta ^ { 2 } _ { j } } \right ) \eta ^ { q } \,,$$
$$\frac { \omega _ { s } - \Omega _ { o } } { \Omega _ { o } } \leqslant \frac { 1 - \eta ^ { 2 } } { \eta ^ { 2 } }$$
as estimates of the super-rotation upper bound. For our η = 0 716, . ( 1 -η 2 ) /η 2 = 0 95. . The flow super-rotations we see in figure 6a are one to two orders of magnitude smaller than the estimated bounds. As equation (4.1) diverges as q → 0, an open question is whether the magnitude of the flow super-rotation normalized by | Ω i -Ω o | diverges as q → 0 at fixed non-zero Re S .
## 4.2. Quasi-Keplerian angular momentum profile and transport
For all the quasi-Keplerian profiles in figure 4, there is a pattern in the profiles. Namely, they are split into three regions: an inner region whose angular momentum profile is
nearly flat with a slight positive slope, an outer region where the flow is nearly in solidbody rotation at Ω , and a middle transition region in which the angular momentum o profile curves upward from being flat to solid-body rotation at Ω . At o q = 1 908, the . inner region extends over nearly the whole gap. As q decreases for fixed Re S , the inner region shrinks until for q = 0 333 it is nearly absent, with the outer region having grown . to be almost the whole gap.
As seen in figure 5b, as Re S is increased, the inner and outer regions appear to grow while the middle region shrinks. The same pattern is seen going from Re S = 1 03 . × 10 5 to Re S = 7 82 . × 10 5 for q = 1 909, which is not shown here but can be seen in the data . in the supplementary material. The pattern suggests that in the limit Re S → ∞ , the middle region might disappear entirely. If we approximate the inner region as a completely flat angular momentum profile, approximate the outer region as rotating at exactly Ω , o ignore any boundary layer on the inner cylinder, and assume that the pattern holds for the rest of the quasi-Keplerian regime and that no flow state transitions at higher Re S break it; then the angular velocity profile for the quasi-Keplerian in the asymptotic limit Re S →∞ regime in our geometry would be
$$\omega ( r ) = \begin{cases} \begin{array} { c c } ( r _ { i } / r ) ^ { 2 } \, \Omega _ { i } \quad & \text{for $r<r_{c}$} \\ \Omega _ { o } \quad & \text{for $r\geqslant r_{c}$} \end{array} \end{cases}$$
with
$$r _ { c } = r _ { i } \sqrt { \frac { \Omega _ { i } } { \Omega _ { o } } } = r _ { i } \, \eta ^ { - q / 2 }$$
$$\widetilde { r } _ { c } = \frac { \stackrel {. } { \eta } } { 1 - \eta } \left ( \eta ^ { - q / 2 } - 1 \right )$$
where r c is the transition radius between the flat angular momentum profile and solidbody rotation at Ω . For large but finite o Re S , equation (4.4) can serve as an approximate profile. This approximate profile was derived independently by Dunst (1972) by assuming that the inner region had a flat angular momentum profile, based on his observation of a well-mixed inner region in his Taylor-Couette experiment.
For the approximately constant specific angular momentum glyph[lscript] = r 2 ω = rv θ inner region, ∂v /∂r < θ 0. Then for the outer regions rotating at approximately Ω , i ∂v /∂r > θ 0. Hence, we can quantify the radial position of the transition region by finding the radial positions for which the azimuthal velocity profiles v θ ( r ) are at their minimum. They are shown in figure 7. In the quasi-Keplerian regime, we find that the position of the minimum velocity corresponds very well with ˜ r c in equation (4.6), giving merit to the approximate profiles of equation (4.4). Outside of the quasi-Keplerian regime, the position of the minimum is located at the inner cylinder for q < 0, and at the outer cylinder for q > 2.
The approximately flat angular momentum profile in the inner region, when away from the Rayleigh line where the laminar Taylor-Couette profile is flat, indicates that the angular momentum is well mixed with advection-dominated transport in the radial direction. In contrast, there is likely little radial angular momentum transport by advection or diffusion in the outer region as the profile is close to solid-body at Ω o . A large amount of angular momentum is transported radially from the inner cylinder based on the torque measurements with the similar Maryland experiment (Paoletti & Lathrop 2011) and on the upcoming analysis of section 4.3. The large amount of angular momentum transported off the inner cylinder and mixed in the inner region has to go somewhere, but the
Figure 7: Radial positions of the minimum in the azimuthal velocity v θ as a function of q (x-axis) and Re S (different symbols). In the quasi-Keplerian regime, ˜ r c from equation (4.6) is shown for comparison (solid line).
outer region, if present, is likely not transporting much angular momentum. Then, when an outer region is present such as when q < 2 far from the Rayleigh line, most of the angular momentum must be transported axially to the axial boundaries in the inner and possibly middle regions, as shown schematically in figure 8. As q increases at fixed Re S towards the Rayleigh line, the outer region disappears and an increasing fraction of the angular momentum can be transported to the outer cylinder through the middle region instead of being transported to the axial boundaries. For q glyph[greaterorequalslant] 2, there is no middle region and a boundary layer forms close to the outer cylinder that steepens with increasing q (van Gils et al. 2012), indicating that an increasing fraction of the angular momentum is transported to the outer cylinder instead of to the axial boundaries. Finally, nearly all of the angular momentum is transported to the outer cylinder.
These features are also seen in wide-gap low aspect-ratio experiments. Using dye injection from the inner cylinder, Dunst (1972) found a well mixed inner region and a quiescent outer region with poor mixing. In figure 9, the angular velocity and the specific angular momentum profiles for q ≈ 1 9 from Schartman . et al. (2012), Edlund & Ji (2014), and Kageyama et al. (2004) are compared to each other and to the results from our apparatus. They all deviate from the laminar Taylor-Couette profile and show the same three regions with a relatively flat glyph[lscript] close to the inner cylinder and rotate close to Ω o close to the outer cylinder. However, the relatively flat glyph[lscript] inner region is offset downward from the specific angular momentum on the inner cylinder, indicating the presence of a boundary layer on the inner cylinder more significant than in our experiment. The experiments of Kageyama et al. (2004) and possibly Edlund & Ji (2014) also exhibit flow sub-rotation ( ω < Ω o < Ω ) in the middle and outer regions. The axial transport of i angular momentum and the presence of three regions in the quasi-Keplerian azimuthal velocity profiles appear to be more general than just occuring in our specific apparatus
Figure 8: Schematic drawing of the angular momentum transport in the quasi-Keplerian regime when the axial boundaries are attached to the outer cylinder. Red dashed lines denote the boundaries between the inner, middle, and outer flow regions. Black arrows denote the transport of angular momentum. The radius and aspect-ratios ( η and Γ) have been changed for visual clarity. Angular momentum is transported radially off the inner cylinder and then transported axially to the axial boundaries in the inner and middle regions.
with its geometry and ranges of Re S and q , although the strength of the boundary layer on the inner cylinder appears to depend on η and/or Γ.
## 4.3. Torque on the inner cylinder
The velocity gradients near the inner cylinder were larger than in laminar Taylor-Couette as the ˜ ω values at the point closest to the inner cylinder in figures 3 and 5a are below that of the laminar Taylor-Couette profile. This steepness means that the torque on the inner cylinder must be larger than in laminar Taylor-Couette flow. If boundary layers were present, the profiles would be even steeper at the inner cylinder, and thus the torques even larger.
The azimuthal shear stress, when averaged azimuthally, is τ = -ρνr ( ∂ω/∂r ) where ρ is the fluid density (see page 48, Landau & Lifshitz 1987). The torque T on a cylinder of radius r from just the shear stress is T ν = 2 πr Lτ 2 , which in terms of the angular velocity is
$$T _ { \nu } = - 2 \pi \rho \nu L r ^ { 3 } \frac { \partial \omega } { \partial r } \. \quad \ \ \ \ \ \ \ ( 4. 7 )$$
As laminar Taylor-Couette flow has no Reynolds stresses and ω is uniform over a cylinder of radius r from equation (1.4), the total laminar Taylor-Couette torque is
$$T _ { \text{lam} } = \frac { 2 \pi \rho \nu ^ { 2 } L \eta } { \left ( 1 - \eta \right ) ^ { 2 } } R e _ { S } \, \text{sign} \left ( R e _ { i } - \eta R e _ { o } \right ).$$
(c)
Figure 9: Comparison of the azimuthal velocimetry for q ≈ 1 9 . between different experiments. This includes our experiment (T C) 3 with η = 0 716 . at q = 1 909 . and Re S = 7 82 . × 10 , 5 Schartman et al. (2012) with η = 0 348 . at q = 1 908 . and Re S = 5 05 . × 10 , Edlund & Ji (2014) with 5 η = 0 340 at . q = 1 803 and . Re S = 4 34 . × 10 , 5 and Kageyama et al. (2004) with η = 0 255 at . q = 1 896 and . Re S = 1 30 . × 10 . The 6 profiles for Schartman et al. (2012) and Edlund & Ji (2014) were constructed by extracting velocities from their figures (6 and 2, respectively). Kageyama et al. (2004) did velocimetry at five different axial heights. Profiles for Kageyama et al. (2004) were constructed by splitting the range ˜ r ∈ [0 , 1] into bins of width 0 02 and averaging the . ω within each bin. Normalized angular velocities ˜ ω profiles are compared (a) at full scale and (b) expanded around ˜ ω = 0 using the same symbols to emphasize the parts of the profiles close to rotation at Ω o . Dashed lines are the laminar Taylor-Couette profiles for each experiment. (c) The specific angular momentum ( glyph[lscript] = r 2 ω ) profiles for each experiment side by side with the same horizontal axes and with vertical axes in the same units. The solid black line ( -) and dashed blue line ( --) are the specific angular momentum profiles for ω r ( ˜ ) = Ω i and ω r ( ˜ ) = Ω , respectively. o
1
Table 2: The ratios of the lower bounds of the torque on the inner cylinder T ν to the laminar Taylor-Couette torque T lam for each set of measurements, ordered by Re S and then by q .
| q | Re S | T ν /T lam |
|-----------|---------------|-----------------|
| 1 . 493 | 2 . 07 × 10 4 | 2 . 32 ± 0 . 05 |
| 2 . 098 | 1 . 03 × 10 5 | 2 . 99 ± 0 . 04 |
| 1 . 908 | 1 . 03 × 10 5 | 2 . 75 ± 0 . 04 |
| 1 . 501 | 1 . 03 × 10 5 | 2 . 56 ± 0 . 05 |
| 1 . 255 | 1 . 03 × 10 5 | 2 . 49 ± 0 . 06 |
| 0 . 693 | 1 . 04 × 10 5 | 3 . 02 ± 0 . 09 |
| 0 . 333 | 1 . 05 × 10 5 | 5 . 12 ± 0 . 18 |
| - 0 . 503 | 1 . 05 × 10 5 | 9 . 79 ± 0 . 12 |
| - 1 . 001 | 1 . 04 × 10 5 | 8 . 55 ± 0 . 06 |
| - 1 . 994 | 1 . 03 × 10 5 | 7 . 45 ± 0 . 04 |
| 2 . 102 | 7 . 83 × 10 5 | 3 . 66 ± 0 . 04 |
| 1 . 909 | 7 . 82 × 10 5 | 3 . 50 ± 0 . 04 |
| 1 . 500 | 7 . 79 × 10 5 | 3 . 31 ± 0 . 05 |
Assuming a turbulent boundary layer, the thickness y 0 of the viscous sublayer on the inner cylinder is y 0 = ( ν/u ∗ ) y + 0 where u ∗ = √ | τ | /ρ is the friction velocity, ρ is the fluid density, and y + 0 is the sublayer thickness in dimensionless units (Schlichting 1979). From measurements in our apparatus for pure inner cylinder rotation at comparable Re S , y + 0 is in the range of 5-10 (Huisman et al. 2013). Then for Re S = 10 , we get 5 y 0 glyph[lessorequalslant] 2 mm since T glyph[greaterorequalslant] T lam and y + 0 glyph[lessorequalslant] 10. Since r -r i = 2 6 mm was the point closest . to the inner cylinder where the flow velocity was resolved, our azimuthal velocimetry did not extend into the viscous sublayer. Due to not resolving the viscous sublayer, the torque in our apparatus cannot be obtained from the velocity profiles; meaning direct comparisons cannot be done to the torque measurements of Paoletti & Lathrop (2011) on the Maryland experiment with near identical geometry. However, lower bounds on the torque can be obtained because the azimuthal profiles can give the shear stress, instead of both the shear and Reynolds stresses.
To get the lower bound for the torque on the inner cylinder, ∂ω/∂r was obtained from the difference between ω at the point closest to the inner cylinder ( r -r i = 2 6 mm . which is ˜ r = 0 033) and Ω . i at the inner cylinder. It must be noted that the velocity profile was taken at the axial height of one of the small separations in the inner cylinder, which is 2 5 mm thick, and therefore the gradients in . ω we calculate might be perturbed compared to other axial heights due to the vicinity to the separation.
The torque lower bounds are listed in table 2. The lower bounds were all larger than the laminar Taylor-Couette torque, which supports the | T/T lam | glyph[greatermuch] 1 result of Paoletti & Lathrop (2011) on the similar Maryland experiment in both regimes for Re S > 3 5 . × 10 . 5 The measurements in this paper extend this result of Paoletti & Lathrop (2011) towards solid-body rotation in both regimes.
For the quasi-Keplerian regime, we can use the approximate flatness of the specific angular momentum profile in the inner region to make an analytical approximate torque lower bound. Treating the inner region as having a flat specific angular momentum profile
Figure 10: Comparison of inner cylinder torque lower bounds calculated at z/L = 0 209 . to the torques measured by Paoletti et al. (2012). Torques are normalized by the torque for pure inner rotation at the same Re S . The scaling of Paoletti et al. (2012) for Re S > 3 5 . × 10 5 is the thick black solid line. The torque ratio lower bounds obtained from the velocity profiles are the symbols, coded by Re S . The lower bound torques from the quasiKeplerian flat inner angular momentum profile approximation in equation (4.9) for each Re S are the thin solid lines with the same colors as the symbols, which increase with decreasing Re S . Error bars are smaller than the symbols.
from the inner cylinder with no boundary layer as in equation (4.4), the ratio of the torque lower bound T ν, flat to the laminar Taylor-Couette torque is
$$\frac { T _ { \nu, \text{flat} } } { T _ { \text{lam} } } = \frac { 1 - \eta ^ { 2 } } { 1 - \eta ^ { q } } \quad \text{for } \ 0 < q \leqslant 2 \.$$
The ratio is always larger than one, approaching one at q = 2. It diverges as q → 0, which is due to the width of the inner region shrinking towards zero since r c → r i in equation (4.5). The decrease in r c means that ω changes from Ω i to Ω o over an ever smaller radial distance, giving a sharper gradient of ω in the inner region, which becomes infinite as q → 0. However, if the inner region of a flat angular momentum profile disappears entirely as q → 0 at a given Re S , then this lower bound may no longer hold. For Re S = 1 04 . × 10 , the inner region might be close to disappearing by 5 q = 0 333 . based on the angular momentum profiles in figure 4. As the middle region shrinks with increasing Re S (figure 5b), the q at which the inner region might disappear decreases with increasing Re S .
The torque lower bounds can be compared to the torque scaling that Paoletti et al. (2012) fit to the Maryland torque measurements in the Rayleigh-stable and unstable regimes (Paoletti & Lathrop 2011) and the torque measurements on the apparatus presented in this paper in the unstable regime (van Gils et al. 2011 b ). The scaling was for
the ratio of the torque on the inner cylinder to the torque T + ∞ for pure inner rotation ( q = + ∞ ) at the same Re S , which in this paper was obtained from torque measurements in the very similar Maryland experiment (equation (9) in Lathrop et al. 1992). The lower bounds are compared to the torque scaling Re S > 3 5 . × 10 5 (equation (12) in Paoletti et al. 2012) in figure 10.
As the torque ratios must be positive, the torque scaling must start curving upwards on the sub-rotating regime side when approaching solid-body rotation at some q > -1 5 . to avoid crossing zero. The three sub-rotating regime torque lower bounds for Re S = 1 04 . × 10 5 give | T /T | + ∞ values that are larger than those for Re S > . 3 5 × 10 5 (Paoletti et al. 2012). Thus, the | T /T | + ∞ scaling of Paoletti et al. (2012) must increase if extended to Re S = 1 04 . × 10 . 5
On the quasi-Keplerian side, comparisons can be made between our measurements at Re S = 7 81 . × 10 5 to those of Paoletti et al. (2012) for Re S > . 3 5 × 10 . Our 5 | T /T | + ∞ lower bounds from both the measured velocity profiles and the flat inner region approximation from equation (4.9), are considerably smaller than those of Paoletti et al. (2012). As our lower bounds only considered shear stress (diffusion), the difference in torques on the inner cylinder must be due to Reynolds stresses (advection) in the region of ˜ r glyph[lessorequalslant] 0 033. . The divergence of the torque lower bound for the flat inner region angular momentum approximation as q → 0 suggests that the flat quasi-Keplerian | T /T | + ∞ scaling of Paoletti et al. (2012) will deviate from being flat if extended to q < 1, unless the inner region disappears or is distorted close to solid-body rotation.
## 5. Summary and conclusions
In summary, azimuthal velocity profiles were obtained for several Rayleigh-stable (and one unstable) cylinder rotation rate ratios for the ranges 2 098 . glyph[greaterorequalslant] q glyph[greaterorequalslant] 0 333 and . -1 994 . glyph[lessorequalslant] q glyph[lessorequalslant] -0 503. They were all done for . Re S = 1 04 . × 10 , a few configurations at 5 Re S = 7 81 . × 10 , and just the Keplerian configuration also at 5 Re S = 2 07 . × 10 . For all values of 4 q , the profiles deviate from the laminar Taylor-Couette profile. The deviation increases as solid-body rotation is approached ( q → 0) at fixed non-zero Re S . The deviation consists of a steepening of the normalized angular velocity ˜ ω profile close to the inner cylinder for all q , and the flow in the outer parts of the gap approaching solid-body rotation with the outer cylinder and attached axial boundaries for the quasi-Keplerian regime.
For the sub-rotating regime, the flow exhibits super-rotation compared to both cylinders ( ω > Ω o > Ω ), except close to the inner cylinder. As solid-body rotation is api proached at fixed Re S = 1 04 . × 10 , the strength of the super-rotation increases, reaching 5 6% of Ω o -Ω for i q = -0 503, and the radial positions of the maximum of super-rotation . and where the flow switched from Ω i < ω < Ω o to super-rotation moves closer to the inner cylinder. The flow super-rotation must be sustained by inward angular momentum transport from the outer cylinder or axial boundaries. To the best of our knowledge, flow super-rotation for q < 0 has not been previously observed in the literature. This includes pure outer-rotation ( q = -∞ ) in our apparatus (van Gils et al. 2011 a ) and in those of Taylor (1936 b ), Wendt (1933), and Burin & Czarnocki (2012).
For the quasi-Keplerian regime, the specific angular momentum profiles show that the flow can be split into three regions across the gap: an inner region where the angular momentum profile is approximately flat, an outer region where the flow is close to solidbody rotation at Ω , and a middle transition region between the two. Starting near the o Rayleigh line, the middle and outer regions are almost non-existent; and then as solidbody rotation is approached at fixed non-zero Re S , the inner region shrinks while the outer region grows till the inner region is almost non-existent at q = 0 333. As . Re S
is increased, the middle region shrinks. We speculate that as Re S → ∞ , the middle region will disappear and the profile will converge towards equation (4.4) (independently derived from dye injection observations by Dunst 1972). This model profile is a good approximation by Re S = 7 81 . × 10 . The outer region, if present, likely transports little 5 angular momentum, meaning that almost all of the angular momentum is transported to the axial boundaries. Work is still needed to check how the mixing is achieved in the inner region and whether the inner region is turbulent. One must also determine the exact nature of the flow in the middle and outer regions, and see whether the region pattern is found for other η and Γ. Measuring the separate torques on the inner cylinder, outer cylinder, and axial boundaries or doing very high resolution Particle Image Velocimetry as Huisman et al. (2012 a ) did for pure inner rotation on the same apparatus would be good ways to determine what fraction of the angular momentum goes to the axial boundaries versus the outer cylinder and elucidate the axial transport mechanism. We do not see flow sub-rotation in our experiment except possibly for q = 0 693 and 0 333, . . but ˜ ω ≈ 0 within the measurement precision at those q . However, Kageyama et al. (2004) and possibly Edlund & Ji (2014) found sub-rotating flow for q > 0.
The slope of the angular velocity profile at the inner cylinder is steeper than in laminar Taylor-Couette flow for both the sub-rotating and the quasi-Keplerian regimes. Therefore, the torque required to rotate the inner cylinder must be larger than the laminar Taylor-Couette value, which supports the super-laminar torque measurements on the geometrically similar Maryland experiment (Paoletti & Lathrop 2011). Due to not resolving the viscous sublayer on the inner cylinder, only lower bounds for the torque on the inner cylinder could be obtained via a viscous stress calculation. The | T /T | + ∞ lower bounds from the velocity measurements and the approximate asymptotic profile were compared to the | T /T | + ∞ scaling found by Paoletti et al. (2012) for Re S > 3 5 . × 10 . 5 In the sub-rotating regime, their | T /T | + ∞ scaling needs to be increased if extended to Re S = 10 5 or towards solid-body rotation. In the quasi-Keplerian regime, the comparison also shows that the bulk of the transport of angular momentum off the inner cylinder is by Reynolds stresses (advection) and the | T /T | + ∞ scaling of Paoletti et al. (2012) may require modification as q → 0.
Our velocity profiles provide experimental confirmation of the expectation that the Ekman pumping from the axial boundaries was what destabilized the flow in the Maryland experiment, which has a nearly identical geometry as our apparatus, in the Rayleighstable regime causing large super-laminar torques on the inner cylinder (Balbus 2011; Avila 2012; Ji & Balbus 2013; Edlund & Ji 2014). This work, combined with the work of Avila (2012), Schartman et al. (2012), and Edlund & Ji (2014), resolves the apparent discrepancy between the approximately laminar Taylor-Couette angular momentum transport in the wide-gap, low aspect-ratio experiments with axial boundaries split into rings rotating at speeds intermediate that of the cylinders such as the Princeton MRI and HTX experiments (Ji et al. 2006; Schartman et al. 2012; Edlund & Ji 2014) and large super-laminar angular momentum transport in the medium-gap higher aspect-Maryland experiment with axial boundaries attached to the outer cylinder (Paoletti & Lathrop 2011). Moreover, we found that the Ekman pumping from the axial boundaries does more than just destabilize the flow in the Rayleigh-stable regime when the axial boundaries are attached to the outer cylinder. In the quasi-Keplerian regime, it causes the flow to be split radially into three regions and nearly all of the angular momentum to be transported to the axial boundaries instead of the outer cylinder when an outer region is present. The Ekman pumping essentially causes the axial boundaries to become the primary sink of angular momentum. In the sub-rotating regime, we discovered flow super-rotation, which is also likely due to the Ekman pumping.
Astrophysical accretion disks have open axial boundaries, which do not cause Ekman pumping, and are thought or assumed to have primarily radial transport of angular momentum (Zeldovich 1981; Richard & Zahn 1999; Richard 2001; Dubrulle et al. 2005 a ; Ji & Balbus 2013; Ostilla-M´nico o et al. 2014). Due to the strong Ekman pumping effects, including the primarily axial transport of angular momentum, Taylor-Couette flow with an aspect ratio up to Γ ∼ 10 with no-slip axial boundaries attached to the outer cylinder is an imperfect model of accretion disks, especially with regard to stability. Ideally, one would like to have axial boundaries that are free-slip or rotate at different rates along their radius such that they match the mean rotation rate of what the flow would be in the absence of axial boundaries, which may not be the laminar Taylor-Couette profile.
There are practical options available to experimental Taylor-Couette flow to mitigate the Ekman pumping and make a better model of accretion disks. One practical way is to make an experiment where the aspect ratio is great enough that the axial tranport mechanism of the angular momentum saturates and Ekman pumping can no longer directly affect the flow near midheight. However, tall experiments are difficult to handle and expensive to make, and work would be needed to ascertain whether indirect effects would still be a problem. Another way, which has been followed by the Princeton group (Ji et al. 2006; Schartman et al. 2012; Edlund & Ji 2014), is to split the axial boundaries into rings that are rotated at speeds intermediate to those of the cylinders. This reduces the strength of the Ekman pumping as well as better confining it to the axial boundaries. Implementing the independently rotating rings is difficult and there is still Ekman pumping due to having only a finite number of independently rotating rings. If the working fluid is a liquid, the top boundary can be made into an open boundary by having gas above it, reducing the Ekman pumping at the top by three orders of magnitude in the case of water and air, though it does introduce the problem of gravity waves on the top surface. For the velocities that are used in the present experiments, air could be entrained by these waves. Similarly, density-mismatched fluids such as mercury and water or stratification (e.g. salt solutions) can be used on the bottom boundary to confine the Ekman circulation near the bottom by reducing axial circulation, although this also introduces the problem of gravity waves and mixing which would destroy the stratification. In order to accurately represent an accretion disk, one probably has to combine more than one of these methods.
## Acknowledgements
Wewould like to acknowledge helpful discussions and advice from Dennis P.M. van Gils, Siegfried Grossmann, Rodolfo Ostilla-M´nico, Daniel S. Zimmerman, Eric M. Edlund, o and Hantao Ji. We thank Hantao Ji for providing us the velocimetry data from Kageyama et al. (2004). We also acknowledge work and advice from the technicians Gert-Wim Bruggert, Martin Bos, and Bas Benschop; and financial support from the Technology Foundation STW of The Netherlands from an ERC Advanced Grant and the National Science Foundation of the USA (Grant No. NSF-DMR 0906109).
## REFERENCES
Avila, K., Moxey, D., de Lozar, A., Avila, M., Barkley, D. & Hof, B. 2011 The Onset of Turbulence in Pipe Flow. Science 333 , 192-6.
Avila, M. 2012 Stability and Angular-Momentum Transport of Fluid Flows between Corotating Cylinders. Physical Review Letters 108 (12), 124501.
Balbus, S. A. 2011 Fluid dynamics: A turbulent matter. Nature 470 , 475-476.
Borrero-Echeverry, D., Schatz, M. F. & Tagg, R. 2010 Transient turbulence in TaylorCouette flow. Physical Review E 81 (2), 025301.
Burin, M. J. & Czarnocki, C. J. 2012 Subcritical transition and spiral turbulence in circular Couette flow. Journal of Fluid Mechanics 709 , 106-122.
Coles, D. 1965 Transition in circular Couette flow. Journal of Fluid Mechanics 21 , 385-425. Dubrulle, B., Dauchot, O., Daviaud, F., Longaretti, P.-Y., Richard, D. & Zahn, J.-P. 2005 a Stability and turbulent transport in Taylor-Couette flow from analysis of experimental data. Physics of Fluids 17 (9), 095103.
Dubrulle, B., Mari´, L., Normand, C., Richard, D., Hersant, F. & Zahn, J.-P. e 2005 b An hydrodynamic shear instability in stratified disks. A&A 429 , 1-13.
Dunst, M. 1972 An experimental and analytical investigation of angular momentum exchange in a rotating fluid. Journal of Fluid Mechanics 55 , 301-310.
Eckhardt, B., Grossmann, S. & Lohse, D. 2007 Torque scaling in turbulent Taylor Couette flow between independently rotating cylinders. Journal of Fluid Mechanics 581 , 221.
Edlund, E. M. & Ji, H. 2014 Nonlinear stability of laboratory quasi-Keplerian flows. Physics Review E 89 (2), 021004.
Grossmann, S. 2000 The onset of shear flow turbulence. Reviews of Modern Physics 72 , 603618.
Huisman, S. G., Scharnowski, S., Cierpka, C., K¨hler, C. J., Lohse, D. & Sun, C. a 2013 Logarithmic Boundary Layers in Strong Taylor-Couette Turbulence. Physical Review Letters 110 (26), 264501.
Huisman, S. G., van Gils, D. P. M., Grossmann, S., Sun, C. & Lohse, D. 2012 a Ultimate Turbulent Taylor-Couette Flow. Physical Review Letters 108 (2), 024501.
Huisman, S. G., van Gils, D. P. M. & Sun, C. 2012 b Applying laser Doppler anemometry inside a Taylor-Couette geometry using a ray-tracer to correct for curvature effects. European Journal of Mechanics B Fluids 36 , 115-119.
Ji, Hantao & Balbus, Steven 2013 Angular momentum transport in astrophysics and in the lab. Physics Today 66 (8), 27.
Ji, H., Burin, M., Schartman, E. & Goodman, J. 2006 Hydrodynamic turbulence cannot transport angular momentum effectively in astrophysical disks. Nature 444 , 343-346.
Kageyama, A., Ji, H., Goodman, J., Chen, F. & Shoshan, E. 2004 Numerical and Experimental Investigation of Circulation in Short Cylinders. Journal of the Physical Society of Japan 73 , 2424.
Landau, L.D. & Lifshitz, E.M. 1987 Fluid Mechanics , 2nd edn., Course of Theoretical Physics , vol. 6. Pergamon Press.
Lathrop, D. P., Fineberg, J. & Swinney, H. L. 1992 Transition to shear-driven turbulence in Couette-Taylor flow. Physical Review A 46 , 6390-6405.
Le Bars, M. & Le Gal, P. 2007 Experimental Analysis of the Stratorotational Instability in a Cylindrical Couette Flow. Physical Review Letters 99 (6), 064502.
Le Diz`s, S. & Riedinger, X. e 2010 The strato-rotational instability of Taylor-Couette and Keplerian flows. Journal of Fluid Mechanics 660 , 147-161.
Maretzke, Simon, Hof, Bj¨rn & Avila, Marc o 2014 Transient growth in linearly stable Taylor-Couette flows. Journal of Fluid Mechanics 742 , 254-290.
Ostilla-M´nico, R., Verzicco, R., Grossmann, S. & Lohse, D. o 2014 Turbulence decay towards the linearly stable regime of Taylor-Couette flow. Journal of Fluid Mechanics 748 , 3.
Paoletti, M. S. & Lathrop, D. P. 2011 Angular Momentum Transport in Turbulent Flow between Independently Rotating Cylinders. Physical Review Letters 106 (2), 024501.
Paoletti, M. S., van Gils, D. P. M., Dubrulle, B., Sun, C., Lohse, D. & Lathrop, D. P.
2012 Angular momentum transport and turbulence in laboratory models of Keplerian flows.
## A&A 547 , A64.
Rayleigh, Lord 1917 On the Dynamics of Revolving Fluids. Proceedings of the Royal Society of London Series A 93 (648), 148-154.
Richard, Denis 2001 Instabilit´s e hydrodynamiques dans les ´ ecoulements en rotation diff´rentielle. Ph.d. thesis, Universit´ Paris-Diderot - Paris VII. e e
Richard, D. & Zahn, J.-P. 1999 Turbulence in differentially rotating flows. What can be learned from the Couette-Taylor experiment. A&A 347 , 734-738.
Schartman, E., Ji, H., Burin, M. J. & Goodman, J. 2012 Stability of quasi-Keplerian shear flow in a laboratory experiment. A&A 543 , A94.
Schlichting, Hermann T. 1979 Boundary Layer Theory . McGraw-Hill Science/Engineering/Math.
Shi, L., Avila, M. & Hof, B. 2013 Scale Invariance at the Onset of Turbulence in Couette Flow. Physical Review Letters 110 (20), 204502.
Taylor, G. I. 1923 Stability of a Viscous Liquid Contained between Two Rotating Cylinders. Royal Society of London Philosophical Transactions Series A 223 , 289-343.
Taylor, G. I. 1936 a Fluid Friction between Rotating Cylinders. I. Torque Measurements. Proceedings of the Royal Society of London Series A 157 (892), 546-564.
Taylor, G. I. 1936 b Fluid Friction between Rotating Cylinders. II. Distribution of Velocity between Concentric Cylinders when Outer One Is Rotating and Inner One Is at Rest. Proceedings of the Royal Society of London Series A 157 , 565-578.
van Gils, D. P. M., Bruggert, G.-W., Lathrop, D. P., Sun, C. & Lohse, D. 2011 a The Twente turbulent Taylor-Couette (T3C) facility: Strongly turbulent (multiphase) flow between two independently rotating cylinders. Review of Scientific Instruments 82 (2), 025105.
van Gils, D. P. M., Huisman, S. G., Bruggert, G.-W., Sun, C. & Lohse, D. 2011 b Torque Scaling in Turbulent Taylor-Couette Flow with Co- and Counterrotating Cylinders. Physical Review Letters 106 (2), 024502.
van Gils, D. P. M., Huisman, S. G., Grossmann, S., Sun, C. & Lohse, D. 2012 Optimal Taylor-Couette turbulence. Journal of Fluid Mechanics 706 , 118-149.
Wendt, Fritz 1933 Turbulente Str¨mungen zwischen zwei rotierenden konaxialen Zylindern. o Ingenieurs et architectes suisses 4 , 577-595.
Zeldovich, Y. B. 1981 On the Friction of Fluids Between Rotating Cylinders. Proceedings of the Royal Society of London Series A 374 , 299-312.
|
10.1017/jfm.2015.275
|
[
"Freja Nordsiek",
"Sander G. Huisman",
"Roeland C. A. van der Veen",
"Chao Sun",
"Detlef Lohse",
"Daniel P. Lathrop"
] |
2014-08-05T18:32:22+00:00
|
2015-02-03T22:39:31+00:00
|
[
"physics.flu-dyn"
] |
Azimuthal velocity profiles in Rayleigh-stable Taylor-Couette flow and implied axial angular momentum transport
|
We present azimuthal velocity profiles measured in a Taylor-Couette
apparatus, which has been used as a model of stellar and planetary accretion
disks. The apparatus has a cylinder radius ratio of $\eta = 0.716$, an
aspect-ratio of $\Gamma = 11.74$, and the plates closing the cylinders in the
axial direction are attached to the outer cylinder. We investigate angular
momentum transport and Ekman pumping in the Rayleigh-stable regime. The regime
is linearly stable and is characterized by radially increasing specific angular
momentum. We present several Rayleigh-stable profiles for shear Reynolds
numbers $Re_S \sim O(10^5) \,$, both for $\Omega_i > \Omega_o > 0$
(quasi-Keplerian regime) and $\Omega_o > \Omega_i > 0$ (sub-rotating regime)
where $\Omega_{i,o}$ is the inner/outer cylinder rotation rate. None of the
velocity profiles matches the non-vortical laminar Taylor-Couette profile. The
deviation from that profile increased as solid-body rotation is approached at
fixed $Re_S$. Flow super-rotation, an angular velocity greater than that of
both cylinders, is observed in the sub-rotating regime. The velocity profiles
give lower bounds for the torques required to rotate the inner cylinder that
were larger than the torques for the case of laminar Taylor-Couette flow. The
quasi-Keplerian profiles are composed of a well mixed inner region, having
approximately constant angular momentum, connected to an outer region in
solid-body rotation with the outer cylinder and attached axial boundaries.
These regions suggest that the angular momentum is transported axially to the
axial boundaries. Therefore, Taylor-Couette flow with closing plates attached
to the outer cylinder is an imperfect model for accretion disk flows,
especially with regard to their stability.
|
1408.1060v1
|
## A hybrid RANS/LES framework to investigate spatially developing turbulent boundary layers
Sunil K. Arolla a,1, ∗
a Sibley School of Mechanical and Aerospace Engineering, Cornell University, Ithaca, NY 14853, USA
## Abstract
A hybrid RANS/LES framework is developed based on a recently proposed Improved Delayed Detached Eddy Simulation (IDDES) model combined with a variant of recycling and rescaling method of generating inflow turbulence. This framework was applied to investigate spatially developing flat plate turbulent boundary layer up to momentum thickness Reynolds number, R θ = 31000 and the results are compared with the available experimental data. Good agreement was obtained for the global quantities such as mean velocity and skin friction at all momentum thickness Reynolds numbers considered. The trends obtained for the Reynolds stress components are in the right direction. At high R θ , the shear stress distribution shows significant differences close to the wall indicating scope for further improving the near-wall modeling in such methods.
Keywords: IDDES; turbulent boundary layer; hybrid RANS/LES; recycling and rescaling
## 1. Introduction
In the industrial CFD, there is an increasing trend for using eddy-resolving methods, where at least a part of the turbulence spectrum is resolved in at least a part of the flow configuration. For high Reynolds number wallbounded flows such as those over aircrafts' wings, ships' hulls, and in the
∗ Corresponding author
Email address:
[email protected]
1 Postdoctoral research associate
(Sunil K. Arolla )
August 6, 2014
atmospheric turbulent boundary layer, the use of large eddy simulation with near-wall resolution (LES-NWR) is infeasible owing to the expensive grid resolution requirements. So, hybrid RANS/LES and wall-modeled LES approaches have become models of choice for such applications. The essential idea is to employ RANS model in the near-wall attached boundary layer region (or a wall model) and LES away from the surface. This reduces the computational cost significantly. There has been intensive research in this area and several variants of such models have been proposed in the recent years (Froehlich and von Terzi, 2008; Shur et al., 2008; Girimaji, 2006; Wang and Moin, 2002; Marusic et al., 2010; Gritskevich et al., 2012). However, to establish confidence in using such methods for practical design, rigorous testing over a wide range of flows need to performed. Spatially developing turbulent boundary layer is one such benchmark problem to evaluate and improve the models.
The particular type of hybrid RANS/LES formulation chosen in this work belongs to the Detached Eddy Simulation category. In this method, switch from RANS to LES takes place depending on the local grid spacing and turbulence length scale. This approach relies on inherent flow instability for a quick generation of turbulent content in the LES mode. Due to the explicit dependence on the grid resolution, DES could result in grid induced separation if the grid is not chosen carefully. To address this issue, a shielding function is introduced to prevent the model from switching to LES within the near-wall attached boundary layer region (Menter and Kunz, 2002; Spalart et al., 2006a). This approach is called Delayed Detached Eddy Simulation. The original intent of DES is to use it for massively separated flow regions, but it has an ability to function as a type of wall-modeled LES. Initial attempts made to use this formulation for channel flow resulted in an another issue called log-layer mismatch (Nikitin et al., 2000; Piomelli et al., 2003). To address this problem, Shur et al. (2008) proposed a modification of the length scale used in the LES region that gives Improved Delayed Detached Eddy Simulation (IDDES) formulation. Gritskevich et al. (2012) fine-tuned these approaches for k -ω SST as a background RANS model. This IDDES model can function similar to a wall-modeled LES approach for the attached boundary layers. Prior to the current work, IDDES has not been tested on high Reynolds number flat plate turbulent boundary layers rigorously.
While the experiments on turbulent boundary layers reached up to R θ = 31000 (DeGraaff and Eaton, 2000), direct numerical simulations of the spatially developing, smooth-wall turbulent boundary layer has hitherto been
limited to R θ = 3 -5 × 10 . 3 Wall-resolved LES has been shown to work well at moderately high Reynolds numbers (Schlatter and Orl¨, 2010), but the ¨ u cost scales approximately as Reynolds number to the power 1.8 (Pope, 2000) and hence is not feasible for using at large R θ . Recently, Inoue and Pullin (2011) performed wall-modeled LES simulations of zero-pressure-gradient flat plate turbulent boundary layer simulations up to R θ = O (10 12 ). They concluded that wall-modeled LES can be used to predict useful properties such as skin friction and mean velocity profiles without having to use fine grid resolution. Their simulations displayed some near-wall issues which they attribute to the wall modeling and the choice of grid resolution. Their work forms the motivation for investigating the accuracy of a hybrid RANS/LES type of approach for flat plate turbulent boundary layer at high momentum thickness Reynolds numbers.
In this work, a variant of recycling and rescaling method of generating inflow turbulence (Arolla and Durbin, 2014) is used in combination with k -ω SST based IDDES approach. This framework is applied to investigate spatially developing flat plate turbulent boundary layers up to R θ = 31000. The objective is to understand the applicability and scope of the IDDES approach for using at high Reynolds number attached boundary layers. The hybrid RANS/LES framework presented in this work can also be used for investigating spatially developing turbulent boundary layers in the presence of pressure gradients.
## 2. Hybrid RANS/LES framework
The computational framework used in this research is that of OpenFOAM finite volume based incompressible flow solver. The governing equations solved are:
$$\partial _ { i } \hat { u } _ { i } \ = \ 0$$
$$\partial _ { t } \hat { u } _ { i } + \partial _ { j } \hat { u } _ { j } \hat { u } _ { i } \ = \ - \frac { 1 } { \rho } \partial _ { i } \hat { p } + \nu \nabla ^ { 2 } \hat { u } _ { i } - \partial _ { i } \tau _ { i j } \quad \ \ \ \ ( 2 )$$
where ˆ u i is the mean velocity field. The unclosed term that arises due to averaging is closed by invoking an IDDES model. As an example for hybrid RANS/LES approaches, a recently proposed simplified version of IDDES for k -ω SST model (Gritskevich et al., 2012) has been implemented within
OpenFOAM framework. Stated briefly, the transport equations for turbulent kinetic energy ( k ) and specific dissipation rate ( ω ) are of the following form:
$$\frac { \partial k } { \partial t } + u _ { j } \frac { \partial k } { \underset { \lambda } { \underset { \lambda } { \underset { \lambda } { \sum } } } } = P _ { k } - \sqrt { k ^ { 3 } } / l _ { I D D E S } + \frac { \partial } { \underset { \lambda } { \underset { \lambda } { \sum } } } \left [ \left ( \nu + \frac { \nu _ { T } } { \sigma _ { k } } \right ) \frac { \partial k } { \underset { \lambda } { \underset { \lambda } { \sum } } } \right ]$$
$$\frac { \partial \omega } { \partial t } + u _ { j } \frac { \dot { \partial } \omega } { \partial x _ { j } } = P _ { \omega } - D _ { \omega } + C D _ { \omega } + \frac { \dot { \partial } } { \partial x _ { j } } \, \bar { \left [ \left ( \nu + \frac { \nu _ { T } } { \sigma _ { \omega } } \right ) \frac { \dot { \partial } \bar { \omega } } { \partial x _ { j } } \right ] }$$
where P k is the production of turbulent kinetic energy, P ω is the production of specific dissipation rate, D ω is the dissipation of ω , and CD ω is the crossdiffusion term in ω . The term l IDDES is the length scale that operates the switch between RANS and LES. The eddy viscosity is of the form ν T = k/ω with a limiter for separated flows. For detailed description of these terms and empirical constants, see Gritskevich et al. (2012). The governing equations solved are similar to that of LES, but sub-grid scale stress term is replaced by a modeled Reynolds stress.
For the numerical simulations presented in this article, Pressure Implicit with Splitting of Operator (PISO) algorithm is employed. A second order accurate backward implicit scheme for time discretization and a second order central scheme (with filtering for high-frequency ringing) for spatial discretization is used.
## 2.1. A variant of recycling and rescaling method for inflow turbulence generation
The essential idea presented in this paper is based on the work by Spalart et al. (2006b) to simplify the inflow generation algorithm using the following physical arguments:
- · The near-wall turbulence regenerates itself much faster than the outer region turbulence → Apply outer layer scaling throughout.
- · When the recycling station is located quite close to the inflow, which is desirable in terms of computing cost, the conflict between inner and outer region scaling essentially vanishes → Short recycling distance
- · Corrections to the wall normal velocity component v have very little effect → Omitted
Figure 1: A schematic of the computational domain used for flat plate turbulent boundary layer simulation. The recycling plane is located at x r = 5 δ 0 from the inflow boundary.
In the current work, momentum thickness based scaling is used in place of 99% boundary layer thickness. This avoids the need of locating the edge of the boundary layer. Moreover, using integral quantities like momentum thickness (or displacement thickness) is numerically robust. Most experiments report the momentum thickness Reynolds number at the inflow and hence back-to-back simulations can be set-up easily. A spanwise mirroring method (Jewkes et al., 2011) is adopted as it was found to be adequate in the current work for disorganizing unphysical spanwise durable structures.
The steps involved in the inflow generation algorithm are:
- 1. Extract the velocity field, u x , y, z, t ( r ), at the recycling plane located at x r and project on to the inflow boundary (see figure 1).
- 2. Perform spanwise averaging to get U y, t ( ) = 〈 u x, y, t ( ) 〉 z . A simple indexing algorithm is used for the averaging. It involves looping over all the faces and index faces with the same wall-normal coordinate. Since the recycling plane is fixed, this indexing can be stored at the preprocessing step itself and reused at each timestep.
- 3. Find the freestream velocity U ∞ = U y ( max , t ).
- 4. Integrate the velocity profile to compute the momentum thickness:
$$\theta _ { r } = \int _ { 0 } ^ { y _ { \max } } \frac { U ( y, t ) } { U _ { \infty } } \left ( 1 - \frac { U ( y, t ) } { U _ { \infty } } \right ) d y$$
- 5. Compute the rescaling factor, γ = θ /θ r in , where θ in is the desired momentum thickness at the inflow.
- 6. Rescale only the x-component of the velocity field:
$$u ( x _ { i n }, y, z, t ) = u ( x _ { r$$
For the Improved Delayed Detached Eddy Simulation (IDDES) model considered in this work, the rescaling operation on the underlying SST variant of k -ω model requires the following:
$$k ( x _ { i n }, y, z, t ) = k ( x _ { r$$
$$\omega ( x _ { i n }, y, z, t ) = \omega ($$
where t -∆ means the velocity from the previous time step is used t for convenience. A linear interpolation is used to compute velocity at the rescaled y-coordinate.
- 7. Apply spanwise mirroring to disorganize unphysical structures which would otherwise be recycled and take much time to get dampened by the spanwise diffusion.
$$u ( x _ { i n }, y, z, t ) = u ( x _ { i$$
where ∆ z is considered to be equal to the spanwise period. Note that w has to be negative to ensure spatial coherence once mirrored Jewkes et al. (2011).
- 8. Check for constant mass flow rate at the inflow by verifying the bulk velocity.
This framework can be used for investigating spatially developing turbulent boundary layers. The current focus is on zero-pressure gradient flat plate problem.
## 3. Flow configuration
The flow configuration considered is that of a flat plate turbulent boundary layer with inflow momentum thickness Reynolds number in the range, R θ = 1520 -31000. The computational domain has dimensions 12 δ 0 × 3 δ 0 × 3 δ 0
in the streamwise, wall-normal, and spanwise directions, respectively where δ 0 is the 99% boundary layer thickness at the recycling plane. The mesh resolution is chosen based on the wall-modeled LES simulations of Inoue and Pullin (2011), as given in table 1. Uniform mesh is used in the streamwise and spanwise directions while a hyperbolic tangent stretching is used in the wall-normal direction to cluster points close to the wall. The recycling station was located at 5 δ 0 downstream of the inlet and the simulation provides its own inflow. The bottom wall is treated as a no-slip wall, top boundary is a slip wall, and at the outflow an advective boundary condition is used. Initial conditions are provided using a mean velocity profile given by Spalding law with random fluctuations with a maximum amplitude of 10% of the freestream value superimposed on the mean value. The time step used is approximately two viscous time units (∆ t ≈ 2 ν/u 2 τ ). The simulation was run for 1000 inertial timescales ( δ /U 0 ∞ ) to eliminate transients and the statistics are collected over another 1000 timescales.
Table 1: Simulation parameters
| Case | R θ | N x | N y | N z | ∆ + x | ∆ + y max | ∆ + z |
|--------|-------|-------|-------|-------|---------|-------------|---------|
| A | 1520 | 140 | 96 | 116 | 60 | 20 | 15 |
| B | 5200 | 240 | 128 | 128 | 104 | 42 | 42 |
| C | 13000 | 240 | 128 | 128 | 245 | 98 | 98 |
| D | 31000 | 240 | 128 | 128 | 546 | 219 | 219 |
## 4. Results and discussion
Mean velocity profiles plotted in inner-inner coordinates at different R θ are presented in figure 2. Comparison with the experimental data of DeGraaff and Eaton (2000) shows that good agreement is obtained for all the cases considered. The mean velocity at the edge of the boundary layer normalized by the friction velocity, U + e = U ∞ /u τ , increases with the increase in R θ . The velocity profiles scale well when plotted in the inner coordinates, which is in agreement with the experimental observations.
The trends of Reynolds normal stress components are plotted in figure 3. It is generally assumed that if the mean flow near the wall scales on inner coordinates, then so too must the Reynolds stresses, which rely on the mean gradients for their energy Tennekes and Lumley (1972). Thus, the Reynolds
stresses are almost always normalized by inner scales, u τ and ν/u τ . Experimental observations of DeGraaff and Eaton (2000), however, found that the streamwise component of the Reynolds stress scales on mixed coordinates (inner-outer), and the wall-normal component on the inner coordinates. When a hybrid RANS/LES is used for studying turbulent boundary layers, the near-wall behavior of Reynolds stress depends on the underlying RANS model being used. Here, we use SST variant of the k -ω as RANS model which gives accurate predictions for the mean velocity, but the near-wall peak of the turbulent kinetic energy is underpredicted due to its formulation (Durbin and Pettersson Reif, 2011). So, it is not expected to get accurate results for the Reynolds stress components with hybrid RANS/LES which is why no direct comparisons with the experimental data are made here. The trends plotted show that the near-wall peak of < u ′ 2 > is not constant when R θ is varied and the location of the peak shifts slightly to the right. This is also true for the wall-normal component of the Reynolds stress. These trends are in good agreement with the observations of DeGraaff and Eaton (2000), but the inner scaling observed on < v ′ 2 > component is not predicted by hybrid RANS/LES. This could be an artifact of the particular choice of the RANS model itself.
Figure 4 presents shear stress distribution in comparison with the experimental data. It has two branches for each R θ . Branch 1 gives the viscous shear ( τ viscous ), and Branch 2 gives the turbulence shear ( τ turbulence ). Note that turbulence shear includes both the resolved and modeled parts as given in the following equation:
$$\begin{array} {$$
where µ t is the eddy viscosity calculated from the underlying SST k -ω model (Menter, 1994). Close to the wall, viscous shear is dominant. Away from the wall (about y + > 100), turbulence shear becomes dominant. These two branches cross at y + ≈ 10 and τ/τ wall ≈ 0 5 except at very high mo-. mentum thickness Reynolds number. The agreement with the experimental data is good for the cases with R θ of 1520 and 5200. At higher Reynolds numbers, there is significant error close to the wall which could be attributed to the near-wall modeling. The 'bi-modal' shape of the turbulence shear is, perhaps, due to the switch between RANS and LES. The resolved field is
showing larger peak in the turbulence shear than in the experiments. These discrepancies indicate that there is scope for further optimizing the parameters used in the length scales of the IDDES model for high R θ .
The Reynolds stress production calculated from the resolved field is plotted in figure 5. It is interesting to see that the peak production decreases with increase in R θ . This means that RANS model is active up to a large y + at higher R θ . The maximum of the production from the resolved field occurs where
$$- \, \rho < u ^ { \prime } v ^ { \prime } > = \mu ( \frac { d U } { d y } )$$
This corresponds to a location, y + ≈ 15, where Branch 1 and Branch 2 meet in the shear stress distribution plotted in figure 4. Assuming that the total shear stress is approximately constant and equal to τ w in the log region implies that the production should lie along the line
$$- < u ^ { \prime } v ^ { \prime } > ^ { + } \left ( \frac { d U ^ { + } } { d y ^ { + } } \right ) = \frac { 1 } { \kappa y ^ { + } }.$$
This line is also plotted in figure 5 with κ = 0 41. . In fact, this is the assumption used in deriving the model constants for the RANS modeling. At higher momentum thickness Reynolds number, log region corresponds to only a small range of y + in our simulations. It is, perhaps, possible to vary the grid resolution to increase the extent of the LES region, but that has not been tried out in this work.
Skin friction is the most important global quantity of interest to the designers. It is plotted in figure 6 along with the experimental data and a data correlation from the 1/7th power law. IDDES shows good agreement with the data at all Reynolds numbers considered. Shape factor, H = δ /θ ∗ is plotted in figure 7 in comparison with the DNS data of Schlatter and Orl¨ ¨ u (2010). The trends are predicted well for this quantity. Vortical structures as identified with the Q-criterion are plotted at the lowest and the highest momentum thickness Reynolds numbers in figures 8-9. As expected, vortical structures resolved at R θ = 31000 are much finer scales compared with that at R θ = 1520.
## 5. Conclusions
A hybrid RANS/LES framework based on k -ω SST based IDDES model and a variant of recycling and rescaling method of generating inflow tur-
5
5
(c) Case C:
800000
10"
(a) Case A:
y
R
θ
= 1520
0000000000
y
10
(b) Case B:
R
θ
= 5200
R
θ
= 13000
Figure 2: Mean velocity profiles in comparison with the experimental data. Solid lines: IDDES; Symbols: DeGraaff and Eaton (2000)
5
(a) Streamwise component
(b) Wall-normal component
Figure 3: Streamwise and wall-normal components of Reynolds stress: Variation at different R θ . The arrow indicates direction of increasing momentum thickness Reynolds number.
bulence is presented. Using this framework, numerical simulations of spatially developing flat plate turbulent boundary layers were performed up to R θ = 31000. The following are the principal conclusions drawn from this work:
- · IDDES predicted the global quantities such as mean velocity and skin friction accurately in comparison with the experimental data.
- · The detailed shear stress distribution indicate scope for improving the near-wall modeling at high R θ . It is necessary to further optimize the parameters used in the model for a wide range of Reynolds numbers.
- · Vortical structures show expected behavior at large R θ .
This work demonstrates the applicability of IDDES to predict global quantities at high Reynolds numbers without having to use very fine grid resolution. Future work will involve investigating the spatially developing boundary layers under the influence of pressure gradients. The framework presented in this paper will be useful in investigating such benchmark problems in more detail and possibly improving the closure models.
(a) Case A:
R
θ
= 1520
(b) Case B:
R
θ
= 5200
Figure 4: Shear stress distribution in comparison with the experimental data. Solid lines: IDDES, Symbols: DeGraaff and Eaton (2000). Branch 1: Viscous shear; Branch 2: Turbulence shear (resolved plus modeled).
Figure 5: Reynolds stress production variation with R θ (calculated from the resolved velocity field).
Figure 6: Skin friction variation with R θ . The dashed line is plotted using a correlation based on the 1 7 th power law.
Figure 7: Shape factor, H, as a function of R θ . Filled symbols are IDDES, Open symbols are from DNS data of Schlatter and Orl¨ (2010). ¨ u
Figure 8: Vortical structures identified using isosurfaces of Q-criterion at R θ = 1520.
Figure 9: Vortical structures identified using isosurfaces of Q-criterion at R θ = 31000.
## References
| Arolla, S. K., Durbin, P. A., 2014. Generating inflow turbulence for eddy simulation of turbomachinery flows. In: 52nd AIAA Aerospace Sciences Meeting, National Harbor, Maryland. |
|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| DeGraaff, D. B., Eaton, J. K., 2000. Reynolds-number scaling of the flat-plate turbulent boundary layer. J. Fluid Mech. 422, 319-346. |
| Durbin, P. A., Pettersson Reif, B. A., 2011. Statistical theory and mathe- matical modeling for turbulent flows. John Wiley & Sons. |
| Froehlich, J., von Terzi, D., 2008. Hybrid RANS/LES methods for simulation of turbulent flows. Progress in Aerospace Sciences 44 (5), 349-377. |
| Girimaji, S. S., 2006. Partially-averaged Navier-Stokes model for turbulence: A Reynolds-averaged Navier-Stokes to direct numerical simulation bridg- ing method. J. Appl. Mech. 73 (3), 413-421. |
| Gritskevich, M. S., Garbaruk, A. V., Schuetze, J., Menter, F. R., 2012. Development of DDES and IDDES Formulations for the k - ω Shear Stress Transport Model. Flow Turbulence Combust 88, 431-449. |
| Inoue, M., Pullin, D. I., 2011. Large-eddy simulation of the zero-pressure- gradient turbulent boundary layer up to Re θ = o (10 12 ). J. Fluid Mech. 686, 507-533. |
| Jewkes, J., Chung, Y., Carpenter, P., 2011. Modification to a turbulent inflow generation method for boundary-layer flows. AIAA Journal 49 (1), 247- 250. |
| Marusic, I., Mathis, R., Hutchins, N., 2010. Predictive wall-model for wall- bounded turbulent flow. Science 329 (5988), 193-196. |
| Menter, F. R., 1994. Two-equation eddy-viscosity turbulence models for en- gineering applications. AIAA Journal 32 (8), 269-289. |
| Menter, F. R., Kunz, M., 2002. Adaptation of Eddy-Viscosity Turbu- lence Models to Unsteady Separated Flow Behind Vehicles. R. McCallen, Browand, F., J, Ross (Eds.), Symposium on the Aerodynamics of Heavy Vehicles: Trucks, Buses, and Trains (2-6). |
| Nikitin, N. V., Nicoud, F., Wasistho, B., Squires, K. D., Spalart, P., July 2000. An approach to wall modeling in large-eddy simulations. Phys. Fluids 12 (7), 1629-1632. |
|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Piomelli, U., Balaras, E., Pasinato, H., Squires, K. D., Spalart, P., 2003. The inner-outer layer interface in large-eddy simulations with wall-layer models. Int. J. Heat Fluid Flow 24, 538-550. |
| Pope, S. B., 2000. Turbulent flows. Cambridge University Press. |
| Schlatter, P., ¨ Orl¨, u R., 2010. Assessment of direct numerical simulation data of turbulent boundary layers. J. Fluid Mech. 659, 116-126. |
| Shur, M. L., Spalart, P., Strelets, M. K., Travin, A. K., 2008. Ahybrid RANS- LES approach with delayed-DES and wall-modelled LES capabilities. Int. J. Heat Fluid Flow 299, 1638-1649. |
| Spalart, P., Deck, S., Shur, M. L., Squires, K. D., Strelets, M. K., Travin, A. K., 2006a. A new version of detached-eddy simulation, resistant to ambiguous grid densities. Theor. Comput. Fluid Dyn. 20, 181-195. |
| Spalart, P., Strelets, M., Travin, A., 2006b. Direct numerical simulation of large-eddy-break-up devices in a boundary layer. Int. J. Heat Fluid Flow 27, 902-910. |
| Tennekes, H., Lumley, J. L., 1972. A first course in turbulence. The MIT Press. |
| Wang, M., Moin, P., 2002. Dynamic wall modeling for large-eddy simulation of complex turbulent flows. Phys. Fluids 14 (7), 2043-2051. |
| null |
[
"Sunil K. Arolla"
] |
2014-08-04T04:06:02+00:00
|
2014-08-04T04:06:02+00:00
|
[
"physics.flu-dyn"
] |
A hybrid RANS/LES framework to investigate spatially developing turbulent boundary layers
|
A hybrid RANS/LES framework is developed based on a recently proposed
Improved Delayed Detached Eddy Simulation (IDDES) model combined with a variant
of recycling and rescaling method of generating inflow turbulence. This
framework was applied to investigate spatially developing flat plate turbulent
boundary layer up to momentum thickness Reynolds number, $R_{\theta} = 31000$
and the results are compared with the available experimental data. Good
agreement was obtained for the global quantities such as mean velocity and skin
friction at all momentum thickness Reynolds numbers considered. The trends
obtained for the Reynolds stress components are in the right direction. At high
$R_{\theta$}, the shear stress distribution shows significant differences close
to the wall indicating scope for further improving the near-wall modeling in
such methods.
|
1408.1061v3
|
## Consistent perturbative treatment of the subohmic spin-boson model yielding arbitrarily small T /T 2 1 decoherence time ratios
Kyung-Joo Noh and Uwe R. Fischer
Seoul National University, Department of Physics and Astronomy Center for Theoretical Physics, Seoul 151-747, Korea
(Dated: December 10, 2014)
We present a perturbative treatment of the subohmic spin-boson model which remedies a crucial flaw in previous treatments. The problem is traced back to the incorrect application of a Markov type approximation to specific terms in the temporal evolution of the reduced density matrix. The modified solution is consistent both with numerical simulations and the exact solution obtained when the bath-coupling spin-space direction is parallel to the qubit energy-basis spin. We therefore demonstrate that the subohmic spin-boson model is capable of describing arbitrarily small ratios of the T 2 and T 1 decoherence times, associated to the decay of the off-diagonal and diagonal reduced density-matrix elements, respectively. An analytical formula for T /T 2 1 at the absolute zero of temperature is provided in the limit of a subohmic bath with vanishing spectral power law exponent. Small ratios closely mimic the experimental results for solid state (flux) qubits, which are subject predominantly to low-frequency electromagnetic noise, and we suggest a reanalysis of the corresponding experimental data in terms of a nonanalytic decay of off-diagonal coherence.
Since the advent of quantum information technology, the problem of two-state systems aka qubits, immersed in environmental (bath) degrees of freedom has increasingly gained importance, cf., e.g., 1-3 . The spin-boson model, succinctly describing such a physical situation, namely an open two-state quantum system interacting with a bath, accounts for two-state energy relaxation and dephasing, and therefore has been studied extensively 4,5 . It was, for example, used to describe physical contexts as diverse as flux qubits implemented using superconducting quantum interference devices (SQUIDs) 6-10 , electron transfer in biomolecules 11 , and phonon coupling in atomic tunneling 12 .
be much larger than the system frequency spacing ω s . The strength of coupling is parametrized by the dimensionless η . Baths having n < 1 , n = 1, and n > 1 are called subohmic, ohmic, and superohmic, respectively. In a subohmic bath, for decreasing n , low-frequency oscillators play an increasingly important role.
The spin-boson model introduces a continuum of independent simple harmonic oscillators with a given spectral weight distribution as the environment, assuming them to couple with the qubit linearly. The Hamiltonian of the entire (closed) system therefore reads (setting glyph[planckover2pi1] = 1)
$$\stackrel { \Xi } { \stackrel { \ast } { \stackrel { \ast } { \stackrel { \ast } { \stackrel { \ast } { \stackrel { \ast } { \ast } } } } } } \\ \stackrel { \stackrel { \ast } { \stackrel { \ast } { \stackrel { \ast } { \stackrel { \ast } { \ast } } } } } } \\ \stackrel { \stackrel { \ast } { \stackrel { \ast } { \stackrel { \ast } { \ast } } } } } \stackrel { \ast } { \stackrel { \ast } { \stackrel { \ast } { \ast } } } } \stackrel { \ast } { \stackrel { \ast } { \stackrel { \ast } { \ast } } } } \stackrel { \ast } { \stackrel { \ast } { \stackrel { \ast } { \ast } } } } \stackrel { \text{ular c} } { \stackrel { \ast } { \stackrel { \ast } { \ast } } } } \\ \stackrel { \stackrel { \ast } { \stackrel { \ast } { \ast } } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } \, \\ \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackPLE { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { < \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } } \stackrel { \ast } { \stackrel { \ast } { \ast } } }$$
Here, -→ σ = ( σ , σ x y , σ z ) are the usual Pauli matrices and ˆ b † k , ˆ b k are creation and annihilation operators of harmonic oscillators in the (infinitely extended) bath, labelled by the quantum number(s) k , which can stand, e.g., for the momentum of the bath excitations. The coupling direction -→ n is parametrized by two angles θ ϕ , as -→ n = (sin θ cos ϕ, sin θ sin ϕ, cos θ ).
The bath is conventionally characterized by the quantity J ω ( ) ≡ ∑ k λ δ ω 2 k ( -ω k ), which determines the dynamics of the spin-boson model. This spectral density of the bath is effectively a density of states summation weighed by the coupling strength squared λ 2 k , hence the name. Usually it is assumed that J ω ( ) is of a power law form up to a cutoff, J ω ( ) = ηω 1 -n s ω e n -ω ωc , where ω c is the (physical) cutoff frequency of the bath, assumed to
The Rabi model, which is a ( θ = π/ 2) variant of the spin-boson model with a single-mode 'bath' (e.g. a twolevel atom in a cavity), is integrable 13 . In principle, an exact solution of the spin-boson model in the same perpendicular coupling case θ = π/ 2 was also obtained 14 , which however does not lend itself to a description of the effective qubit dynamics because tracing out the bath is still highly nontrivial. Furthermore, in the simple case that the coupling direction and the energy basis direction of the system are parallel ( θ = 0), the spin-boson model is also exactly solvable 15,16 .
Away from the limiting cases of parallel and perpendicular coupling, θ = 0 , π/ 2, perturbative approaches have been employed 6,7,9,16 . However, as pointed out in 16 , the previous perturbative approaches show a pathological behavior for the subohmic heat bath by predicting the instantaneous loss of phase coherence for any finite coupling η . In addition, in the low temperature limit, the perturbative solution does not agree even qualitatively with the exact solution for θ = 0. Previous perturbative approaches thus cannot be self-consistent. We aim at remedying this situation, by developing a consistent perturbative approach to the subohmic spin-boson model which both agrees with the exactly solvable case θ = 0 and is free of the aforementioned pathological behavior.
The time evolution of the density matrix of the closed system in the interaction picture, ˜ ( ), is determined by ρ I t the following von Neumann equation
$$\frac { d \tilde { \rho } _ { I } ( t ) } { d t } = - i [ V _ { I } ( t ), \tilde { \rho } _ { I } ( t ) ],$$
where Schr¨dinger and interaction picture operators are o related by A I ( ) t ≡ U † H 0 ( ) t AU H 0 ( ). t We define the re-
duced density matrix by tracing out the bath, ρ I ( ) t ≡ Tr B I ˜ ( ), and employ the following three conditions ρ t 17,18
- { } 1 The Born approximation which treats Eq.(2) perturbatively to second order accuracy in V I ( ). t
- { } 2 The initial product state assumption ˜(0) = ρ ρ (0) ⊗ ρ B (0) where ρ B (0) is the canonical density matrix at temperature T (or ˜ (0) = ρ I ρ I (0) ⊗ ρ B (0)).
- { } 3 Tr B [ V I ( ) t , ρ B (0)] = 0, due to V I having odd and ρ B (0) having even spatial parity.
The equation (2) can then be transformed into the perturbative expression (see Eqs. 8.1-8.15 in Ref. 17 )
$$\frac { d \rho _ { I } ( t ) } { d t } = - \int _ { 0 } ^ { t } d s \text{Tr} _ { B } [ V _ { I } ( t ), [ V _ { I } ( s ), \rho _ { I } ( t ) \otimes \rho _ { B } ( 0 ) ] ]. \ ( 3 ) \quad \text{where} \quad \text{phasis} \quad \text{$\sim}.$$
For the purpose of our derivation to follow, we note that in the literature, one encounters rather widely differing notions of what is called 'Born' (i.e., weak coupling) and 'Markov' (i.e., short-time memory) approximations, which are not necessarily equivalent. For example, Ref. 19 derives Eq.(3) by assuming a product state for arbitrary times ˜ ( ) = ρ I t ρ I ( ) t ⊗ ρ B (0) (instead of only initially as in { 2 } ), and terms this 'Born' approximation. Then the 'Markov' approximation is imposed that ρ I ( s ) in the integrand is replaced by ρ I ( ) t to arrive at (3). On the other hand, 17,18 derives Eq.(3) by using the above stated requirements { 1 } { -3 } , utilizing the Born approximation in the form of { 1 } , which can be justified rigorously 17 . No (variety of) Markov approximation is employed in deriving (3). In addition, the Markov type approximation { 4 } which we will use and explain in detail below is distinct from that used in, e.g., 19,20 .
Employing the spin-boson Hamiltonian (1) in the master equation (3), and adopting the variable change s → t -s , one obtains
$$\frac { d \rho _ { I } ( t ) } { d t } = & - \int _ { 0 } ^ { t } d s \int _ { 0 } ^ { \infty } d \omega J ( \omega ) \quad & \rho \\ \times & \left \{ \coth \left [ \frac { \beta \omega } { 2 } \right ] \cos { ( \omega s ) } [ \overrightarrow { n } \cdot \overrightarrow { \sigma } ( t ), [ \overrightarrow { n } \cdot \overrightarrow { \sigma } ( t - s ), \rho _ { I } ( t ) ) ] \right ] \\ - i \sin { ( \omega s ) } [ \overrightarrow { n } \cdot \overrightarrow { \sigma } ( t ), \{ \overrightarrow { n } \cdot \overrightarrow { \sigma } ( t - s ), \rho _ { I } ( t ) \} ] \right \}. \quad ( 4 ) \quad \text{To see} \\ \cdot \quad \text{$d$} \quad.$$
Here, the t dependence of the operators - · -→ n → σ ( ) t is determined by the spin part of the Hamiltonian, i.e., by H ( s ) 0 = 1 2 ω σ s z . Previous studies on the spin-boson model 6,7,9,16 deduced a solution for the evolution of the reduced density matrix ρ I ( ) of Eq.(4), which, when writt ten in the presently employed notation, reads
$$\rho _ { I, i i } ( t ) & = \rho _ { I, i i } ^ { \text{eq} } ( 1 - \exp \left [ - \frac { t } { T _ { 1 } } \right ] ) + \rho _ { I, i i } ( 0 ) \exp \left [ - \frac { t } { T _ { 1 } } \right ], \quad \text{$\text{$\text{are}$} \\ i & = \{ 1, 2 \}, \quad \text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text$$\text{$\text{$\text{$\text{$\text{$\text$$\text$$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$ \text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\tt$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$
$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$
\rho _ { I, i i i } } } ( t ) & = \rho _ { I, i i i } } ( 1 - \exp \left [ - \frac { t } { T _ { 1 } } } \right ] ) + \rho _ { I, i i i } ( 0 ) \exp \left [ - \frac { t } { T _ { 1 } } } \right ], \quad \text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$$\text{#}} } } } } } } } } \end{ \text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{#}} } } } } } } } } } } )$$
$$\iota & = \iota _ { 1 }, \, \tilde { \jmath }, \, \\ \rho _ { I, 1 2 } ( t ) & = \rho _ { I, 1 2 } ( 0 ) \exp \left [ - \frac { t } { T _ { 2 } } \right ] \exp \left [ - i \frac { t } { T _ { 3 } } \right ], \, \quad \text{
< \text{
\rho _ { I, 2 1 } ( t ) & = \rho _ { I, 2 1 } ( 0 ) \exp \left [ - \frac { t } { T _ { 2 } } \right ] \exp \left [ i \frac { t } { T _ { 3 } } \right ], \, \quad \text{$$
where the indices refer to qubit space, and ρ eq I, 11 = (1 + exp[ βω s ]) -1 and ρ eq I, 22 = (1 + exp[ -βω s ]) -1 are canonical equilibrium distributions at the inverse temperature β . Finally, the various rates in (5) satisfy
$$\underset { \text{matrix} } { \rho ( 0 ) \otimes } \quad \frac { 1 } { T _ { 1 } } = 2 \pi \sin ^ { 2 } \theta J ( \omega _ { s } ) \coth \left [ \frac { \beta \omega _ { s } } { 2 } \right ], \quad \quad \quad ( 6 )$$
$$\int _ { \substack { \text{max} \\ \text{ld and} } } ^ { \text{max} } \frac { 1 } { T _ { 2 } } = \frac { 1 } { 2 T _ { 1 } } + \frac { 1 } { T _ { \phi } }, \quad \frac { 1 } { T _ { \phi } } = \bar { 4 \pi k _ { B } } ^ { \bar { \ } } T \cos ^ { 2 } \theta \, \left ( \frac { J ( x ) } { x } \right ) _ { x \to 0 } \quad ( 7 )$$
$$\text{Id and } \begin{array} { c } \tau _ { 2 } & \tau _ { 1 } & \tau _ { \phi } & \tau _ { \phi } & \vee \tau _ { x \to 0 } \\ \text{ie per-} \\ \text{$\frac{1}{T_{3} } = 2\sin^{2}\theta$} \int _ { 0 } ^ { \infty } \int _ { 0 } ^ { \infty } d s d \omega J ( \omega ) \coth \left [ \frac { \beta \omega } { 2 } \right ] \cos ( \omega s ) \sin ( \omega _ { s } s ), \\ \text{$\frac{1}{2} \theta$} \end{array}$$
where T 1 and T 2 are called (energy) relaxation and dephasing time, respectively. The coupling angle ϕ does not occur in the above: The two expressions containing ϕ , ( - · -→ → n σ ) 12 = sin θe -iϕ and ( - · -→ → n σ ) 21 = sin θe iϕ enter as absolute squares.
We observe that 1 /T φ in Eq.(7) diverges for the subohmic case n < 1 at any finite T as ( x n -1 ) x → 0 ; also note that at absolute zero, T φ cannot be determined from the formula at all. Accordingly, if Eq.(7) is valid for the subohmic regime, this indicates that the dephasing time T 2 should always vanish. This appears to be unphysical, because it implies that phase coherence is destroyed instantaneously as soon as the system comes into contact with a subohmic heat bath, at any value of the coupling strength η . We therefore suspect that Eqs.(5)-(8) do not represent a proper perturbative solution of (4).
Furthermore, we now argue that Eqs.(5)-(8) are in addition inconsistent with the exact solution for the commuting case θ = 0, i.e, when - · -→ → n σ = σ z . Using a unitary transformation 15 , the exact full density matrix can be found from (2), only using the initial product state assumption { 2 } . The solution for the reduced density matrix, at any coupling strength η , is then glyph[negationslash]
$$\dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \delta \rho _ { I, 1 1 } ( t ) & = \rho _ { I, 1 1 } ( 0 ), \quad \rho _ { I, 2 2 } ( t ) = \rho _ { I, 2 2 } ( 0 ), \\ \rho _ { I, i j } ( t ) & = \rho _ { I, i j } ( 0 ) \exp [ - Q ( t ) ], \quad i \neq j, \quad \left ( 9 \right ) \\ Q ( t ) & = 4 \int _ { 0 } ^ { \infty } d \omega J ( \omega ) \coth \left [ \frac { \beta \omega } { 2 } \right ] \frac { 1 - \cos ( \omega t ) } { \omega ^ { 2 } }. \\ \cdots \quad T \cap \cos t h a \text{ in connection} \, \text{ from } \mathcal { F } \cos T \, \mathcal { S } \, \mathcal { W } \, \mathcal { W } \, \text{ with } a - \mathcal { N } \, \mathcal { W }$$
To see the inconsistency between Eqs.(5)-(8) (with θ = 0) and Eq.(9) in a manifest manner, we make Eqs.(5)-(8) concrete for an ohmic heat bath ( n = 1):
glyph[negationslash]
$$& \rho _ { I, 1 1 } ( t ) = \rho _ { I, 1 1 } ( 0 ), \quad \rho _ { I, 2 2 } ( t ) = \rho _ { I, 2 2 } ( 0 ) \\ & \rho _ { I, i j } ( t ) = \rho _ { I, i j } ( 0 ) \exp [ - 4 \pi \eta k _ { B } T t ], \quad i \neq j \quad ( 1 0 )$$
In the high temperature limit ( k B T glyph[greatermuch] ω c ), using the approximation coth[ βω 2 ] glyph[similarequal] 2 k B T ω for ω below the cutoff energy, Q t ( ) is approximated as 4 πηk B Tt ; hence Eq.(5)(8) and Eq.(9) for large T give the same results and are consistent with each other. However, at the absolute zero of temperature, Eq. (9) yields the following solution which obviously does not agree with Eqs.(5)-(8), i.e., by setting coth[ βω 2 ] = 1 ∀ ω , we get glyph[negationslash]
$$& \rho _ { I, 1 1 } ( t ) = \rho _ { I, 1 1 } ( 0 ), \quad \rho _ { I, 2 2 } ( t ) = \rho _ { I, 2 2 } ( 0 ), \\ & \rho _ { I, i j } ( t ) = \rho _ { I, i j } ( 0 ) \left ( 1 + ( \omega _ { c } t ) ^ { 2 } \right ) ^ { - 2 \eta } \quad i \neq j. \quad ( 1 1 )$$
Eq.(10) exhibits an exponential decay of the off-diagonal elements of the reduced density matrix, and the decay rate vanishes at T = 0. On the other hand, according to the exact solution in Eq. (11), we can still observe a polynomial decay in the off-diagonal elements even at T = 0. The inconsistency between Eqs. (5)-(8) and Eq. (9) under the approximations employed so far is therefore apparent. We next provide a perturbative solution which avoids the above consistency problems.
We first observe that the derivation of Eqs. (5)-(8) from Eq. (4) requires the following two additional conditions to be implemented apart from { 1 } -{ 3 } :
- { } 4 The Markov approximation (short-time memory approximation), in the sense of replacing a finite time integral, ∫ t 0 , by one extending to positive infinity, ∫ ∞ 0 in Eq. (4).
- { } 5 ω -1 s glyph[lessmuch] T i ∀ i ; that is, averaging out rapidly oscillating terms in Eq. (4).
We proceed to demonstrate that the approximation { 4 } forms the reason for the pathological behavior displayed by Eqs.(5)-(8). In particular, the problem becomes manifest when T φ defined in (7) is involved in the solution process (or T 2 when T 1 ( θ, ω s ) is considered fixed) 21 .
To initiate our discussion, we first observe that one encounters the following four integrals I i when writing down the right-hand side of Eq. (4):
$$\text{down the right-hand size of $eq_{1}(4)$} \colon & \quad \rho _ { I, 2 1 } ( i ) \\ I _ { 1 } + i I _ { 2 } & = \int _ { 0 } ^ { t } d s \left [ \int _ { 0 } ^ { \infty } d \omega J ( \omega ) \coth \left [ \frac { \beta \omega } { 2 } \right ] \cos ( \omega s ) \right ] e ^ { i \omega _ { s } s }, \\ I _ { 3 } & = \int _ { 0 } ^ { t } d s \left [ \int _ { 0 } ^ { \infty } d \omega J ( \omega ) \sin ( \omega s ) \right ] \sin ( \omega _ { s } s ), \\ I _ { 4 } & = \int _ { 0 } ^ { t } d s \left [ \int _ { 0 } ^ { \infty } d \omega J ( \omega ) \coth \left [ \frac { \beta \omega } { 2 } \right ] \cos ( \omega s ) \right ], \, ( 1 2 ) \quad \text{defini} \quad \text{places} \\ \text{where we used the Euler formula} \exp [ i \phi ] = \cos \phi + i \sin \phi$$
where we used the Euler formula exp[ iφ ] = cos φ + sin i φ to combine the first two integrals into one expression. For a spectral density J ω ( ) of power law form, the energy integrals over ω in square brackets yield polynomially decreasing functions of s at T = 0. The Markov type approximation of replacing ∫ t 0 with ∫ ∞ 0 can then be legitimately applied to I 1 , I 2 , I 3 , because these integrals converge sufficiently fast due to rapidly oscillating functions in the integrand such as cos( ω s s ) and sin( ω s s ), assuming that condition { 5 } holds. In addition, even though I 1 , I 2 , I 3 are functions of both ω s and t , the ω s dependence dominates the t dependence, because the Markovian t →∞ limit of these integrals is determined by ω s . On the other hand, due to the absence of such oscillating functions in the integrand, the integral I 4 = I 4 ( ) t converges much more slowly. Hence it follows that the detailed t dependence of I 4 should be treated carefully, and straightforwardly applying the Markov approximation { 4 } to I 4 is not warranted. As a result, because only I 4 is related to the evaluation of T φ , one can obtain a more adequate solution for the reduced density matrix by discarding the Markov approximation { 4 } for the corresponding 'critical' terms involving I 4 .
Applying the Markov type approximation { 4 } to I 1 , I 2 , I 3 , using I 1 ( ∞ , ω s ) = π 2 J ω ( s ) coth( βω s 2 ), I 3 ( ∞ , ω s ) = π 2 J ω ( s ) and noting the definitions for T , T 1 2 , T 3 in Eqs. (6)-(8), we derive the following differential equations for the reduced density matrix from Eq.(4),
$$= 0. \quad & \text{tial equations for the reduced density matrix from Eq.(4),} \\ \text{nd} \\ \text{rent}. \quad & \text{$\partial_{t} \rho_{I,11}(t)= -\pi \sin ^ { 2 } \theta J(\omega _ { s } ) \left ( \coth \left [ \frac { \beta \omega _ { s } } { 2 } \right ] + 1 \right ) \rho_{I,11}(t)$} \\ \text{from} \\ \text{as to} \\ \text{$\partial_{t} \rho_{I,12}(t)= -\left ( \frac { 1 } { 2 T _ { 1 } } + \frac { i } { T _ { 3 } } + 4 \cos ^ { 2 } \theta I _ { 4 }(t) \right ) \rho_{I,12}(t),} \\ \text{nory} \\ \text{inite} \\ \text{e in} \\ \text{$e$ in} \\ \text{$\secil}$} \\ \text{$\text{$\text{$the first and last line can be rewritten in a compactified}$} }$$
The first and last line can be rewritten in a compactified notation, using that ρ I, 11 ( ) + t ρ I, 22 ( ) t = 1, ∂ ρ t I,ii ( ) t = -1 T 1 ( ρ I,ii ( ) t -ρ eq I,ii ) , i = { 1 2 , } . The linear first-order differential equations (13) are then readily solved by
$$\Pi ^ { ( 1 + J ) } _ { 1 } & \text{differential equations (13)} \, \text{are then readily solved by} \\ \text{laved} \\ \text{lun} \\ \text{rt} \text{ in} \\ \text{rt} \text{ing} \\ \text{t} \text{ one} \\ \text{ritting} \\ \text{e} ^ { i \omega _ { s } s }, \\ e ^ { i \omega _ { s } s }, \\ e ^ { i \omega _ { s } s }, \\ \end{matrix} \, \begin{array} { c } \rho _ { I, 1 1 } ( t ) = \rho _ { I, 1 1 } ^ { e q } \left ( 1 - \exp \left [ - \frac { t } { T _ { 1 } } \right ] \right ) + \rho _ { I, 1 1 } ( 0 ) \exp \left [ - \frac { t } { T _ { 1 } } \right ] \, \text{exp} \left [ - K ( t ) \right ], \\ \rho _ { I, 2 1 } ( t ) = \rho _ { I, 2 1 } ( 0 ) \exp \left [ - \frac { t } { 2 T _ { 1 } } \right ] \, \text{exp} \left [ i \frac { t } { T _ { 1 } } \right ] \, \text{exp} \left [ - K ( t ) \right ], \\ \end{array} \\ e ^ { i \omega _ { s } s },$$
defining a function K t ( ) = 4 cos 2 θ ∫ t 0 dsI 4 ( s ), which replaces t/T φ in (5),
$$\inf _ { \begin{array} { c } \phi \\ \text{on.} \end{array} } K ( t ) = 4 \cos ^ { 2 } \theta \int _ { 0 } ^ { \infty } d \omega J ( \omega ) \coth \left ( \frac { \beta \omega } { 2 } \right ) \frac { 1 - \cos ( \omega t ) } { \omega ^ { 2 } }. \\ \text{on.} \end{array} \left ( \frac { 1 5 } { \pi } \right ) \right ) \frac { 1 - \cos ( \omega t ) } { \omega ^ { 2 } }.$$
Eq. (14) represents an improved perturbative solution for Eq. (4) in that it does not display any of the unphysical pathology exhibited in Eq. (7), and it is reduced to the exact solution displayed in Eq. (9) if we set θ = 0. It thus demonstrates perfect consistency with the exact solution for the commuting case. Note that with θ = 0, K t ( ) becomes Q t ( ) in (9), and the rates 1 T 1 , 1 T 3 vanish.
At the absolute zero of temperature, by setting coth( βω 2 ) = 1 for all ω , K t ( ) can be calculated for our assumed spectral density of the form J ω ( ) = ηω 1 -n s ω e n -ω ωc . For the subohmic case (0 ≤ n < 1), K t ( ) does not diverge even if there is no cutoff energy ω c . In this regime, we may thus let ω c → ∞ . Then, K t ( ) is given by K t ( ) = 2 πη cos 2 θ ω t s for n = 0 and K t ( ) = 4 η cos 2 θ sin( nπ 2 ) Γ( n ) 1 -n ( ω t s ) 1 -n for 0 < n < 1. For the ohmic and superohmic cases ( n ≥ 1), on the other hand, K t ( ) converges only if there is a finite cutoff energy ω c . For example, K t ( ) = 2 η cos 2 θ ln(1+( ω t c ) 2 ) for n = 1 and K t ( ) = 4 η cos 2 θ ω c ω s ( ω t c ) 2 1+( ω t c ) 2 for n = 2. The effect of
FIG. 1. The decay factor exp[ -K t ( )] in the ohmic and subohmic regime, 0 ≤ n ≤ 1; η = 5 × 10 -3 and θ = 60deg; the ultraviolet cutoff ω c = 100 ω s . For comparison, the dominant decay factor exp[ -t/ 2 T 1 ] is shown by the dotted line.
K t ( ) is generally strongly cutoff frequency dependent in the superohmic case ( n > 1, not shown in Fig. 1), and increasingly so with spectral power n . The superohmic case thus needs a more elaborate treatment, which we do not pursue here.
The functions K t ( ) corresponding to different powers n can be numerically compared with each other, see Fig. 1 for the ohmic and subohmic cases. In particular, when we approach the zero of temperature, for n = 0, we obtain an off-diagonal decay which is exactly exponential because then K(t) is linear in t. The influence of K t ( ) then becomes comparable to exp[ -t/ 2 T 1 ] (which dominates for the ohmic regime)-cf. Eq. (14) and Fig. 1.
As a result of exponential decay, we can define in this limit of n → 0 the dephasing time T 2 properly 22 . Combining Eq. (14) and K t ( ) = 2 πη cos 2 θ ω t s ( n = 0 and T = 0), the dephasing rate 1 T 2 = 1 2 T 1 +2 πηω s cos 2 θ. Combining this with Eq. (6) the ratio T 2 T 1 can be obtained analytically at the absolute zero of temperature for n → 0,
$$\frac { T _ { 2 } } { T _ { 1 } } = \frac { 2 \sin ^ { 2 } \theta } { 2 - \sin ^ { 2 } \theta }. \quad \quad \ \ ( 1 6 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
Hence, contrary to the ohmic case where the ratio T /T 2 1 = 2 regardless of the coupling angle θ , the decoherence time ratio can have any value between 0 and 2 depending on the coupling angle θ . This is of particular relevance for the interpretation of studies on flux qubits created using SQUIDs, see for example 23-29 , which are typically dominated by low-frequency electromagnetic noise. The experiments report a wide range of decoherence time ratios T 2 T 1 ∼ 0 01 . · · · 2 rather than this ratio being always identical to two. Thus we conclude from (16) that the perturbatively treated subohmic heat bath furnishes a realistic physical description for these experimentally realized systems.
We have validated Eq. (16) by checking its consistency with the numerical solution of (4) for the extremely subohmic case with η = 5 × 10 -3 . As can be seen in Fig. 2,
FIG. 2. Decoherence time ratio T 2 T 1 for the extremely subohmic case ( n = 0) as a function of coupling angle θ . The analytical result (Eq. (16)) is shown by the red line and the numerical results by solving Eq. (4) are indicated by dots. The parameters η and ω c are the same as in Fig.1.
Eq. (16) agrees very well with the numerical results (for a smaller coupling η = 1 × 10 -3 , numerics and analytics become essentially indistinguishable on the scale of the figure). On the other hand, with Eqs. (5)-(8), it is readily concluded that it is impossible to get a similarly consistent formula for T /T 2 1 in the subohmic regime.
To conclude, we have reinvestigated the perturbative solution for the reduced density matrix dynamics of the subohmic spin-boson model. It was shown that Eqs. (14)-(15) represent a solution both consistent with the exact solution for zero coupling angle, θ = 0, and agrees with the numerical solution of Eq. (4) for any coupling direction θ . As a result, our solution is capable of describing the physically realistic situation that the dephasing effect of the off-diagonal reduced density matrix elements is much larger than the relaxation effect embodied in the diagonal terms, T 2 glyph[lessmuch] T 1 . In addition, the nonlinearity of K t ( ) we obtained in the subohmic regime 0 < n < 1 sheds new light on the interpretation of experiments measuring the off-diagonal decay rates, for example, Refs. 26,27,29 . Specifically, we suggest a (re)analysis of the experimental data in terms of a nonanalytic behavior of the off-diagonal decay, involving K t ( ) ∝ t 1 -n . We note in this connection that the corresponding formula for K t ( ), under the conditions explained in the above, in fact also holds for -1 < n < 0.
glyph[negationslash]
For the parallel case ( θ = 0), in which there is no energy exchange between spin and heat bath, the perturbative approach reproduces the exact solution Eq. (9), and hence is valid in the long time limit as well. On the other hand, when there is energy exchange between the spin and the heat bath ( θ = 0), nonperturbative studies of the subohmic spin-boson model using the renormalization group flow equation method have pointed out that the perturbative approach for the subohmic case will eventually fail in the long time limit. This is due to the development of a localized, so-called 'trapped', state for the pseudospin (in a double-well analogy). The
phase transition point to the trapped state, η c = η c ( n ), depends on the power law of the spectral density 30-33 . Despite the potential occurence of a phase transition, the perturbative approach can be applied for the shorttime dynamics 31,32 , which in view of the limited coherence times of, in particular, flux qubits will generally suffice. Moreover, a self-consistent perturbative treat-
- 1 L. Amico, R. Fazio, A. Osterloh, and V. Vedral, Rev. Mod. Phys. 80 , 517 (2008).
- 2 D.P. DiVincenzo and D. Loss, Phys. Rev. B 71 , 035318 (2005).
- 3 C. H. van der Wal, F. K. Wilhelm, C. P. J. M. Harmans, and J. E. Mooij, Eur. Phys. J. B 31 , 111 (2003).
- 4 U. Weiss, Quantum Dissipative Systems , 4th edition, World Scientific (2012).
- 5 A.J. Leggett, S. Chakravarty, A.T. Dorsey, M.P.A. Fisher, A. Garg, and W. Zwerger, Rev. Mod. Phys. 59 , 1 (1987).
- 6 Y. Makhlin, G. Sch¨n, and A. Shnirman, Rev. Mod. Phys. o 73 , 357 (2001).
- 7 M. Governale, M. Grifoni, and G. Sch¨n, Chem. Phys. o 268 , 273 (2001).
- 8 L. Tian, S. Lloyd, and T.P. Orlando, Phys. Rev. B 65 , 144516 (2002).
- 9 G. Burkard, R. H. Koch, and D. P. DiVincenzo, Phys. Rev. B 69 , 064503 (2004).
- 10 T. Kato, A. A. Golubov, and Y. Nakamura, Phys. Rev. B 76 , 172502 (2007).
- 11 A. Garg, J. N. Onuchic, and V. Ambegaokar, J. Chem. Phys. 83 , 4491 (1985).
- 12 J. P. Sethna, Phys. Rev. B 24 , 698 (1981).
- 13 D. Braak, Phys. Rev. Lett. 107 , 100401 (2011).
- 14 B. Gardas, J. Phys. A: Math. Theor. 44 , 195301 (2011).
- 15 L.-M. Kuang, H.-S. Zeng, and Z.-Y. Tong, Phys. Rev. A 60 , 3815 (1999).
- 16 A. Shirman and G. Sch¨n, 'Dephasing and Renormalizao tion in Quantum Two Level Systems,' in Quantum Noise in Mesoscopic Physics , Y. Nazarov (Ed.), Springer (2003).
- 17 R.R. Puri, Mathematical Methods of Quantum Optics , Springer (2001).
- 18 http://www.quantiki.org/wiki/Master equation contains a concise summary of the derivation in 17 .
- 19 H.-P. Breuer and F. Petruccione, The Theory of Open
ment as presented here provides an explicit analytic expression for the reduced density matrix dynamics of the spin-boson model (which the renormalization group approaches did not yield), from which the experimentally observed decoherence times can be defined.
This research was supported by the NRF of Korea, Grant No. 2014R1A2A2A01006535.
- Quantum Systems , Oxford University Press (2002).
- 20 More specifically, the variety of Markov approximation which is used to derive Eq.3.117 in 19 [and which is equivalent to Eq. (3)], from Eq.3.116 therein.
- 21 Non-Markovianity in the perturbative spin-boson model was previously also considered in M. Thoss, H. Wang, and W.H. Miller, J. Chem. Phys. 115 , 2991 (2001).
- 22 When 0 < n < 1, the decay is no longer exponential, due to K t ( ) ∝ t 1 -n . We have performed an analysis of this somewhat more cumbersome case as well, and found the same excellent agreement of numerics and analytics.
- 23 D. Vion, A. Aassime, A. Cottet, P. Joyez, H. Pothier, C. Urbina, D. Esteve, and M. H. Devoret, Science 296 , 886 (2002).
- 24 E. Collin, G. Ithier, A. Aassime, P. Joyez, D. Vion, and D. Esteve, Phys. Rev. Lett. 93 , 157005 (2004).
- 25 I. Chiorescu, Y. Nakamura, C. J. P. M. Harmans, and J. E. Mooij, Science 299 , 1869 (2003).
- 26 F. Yoshihara, K. Harrabi, A. O. Niskanen, Y. Nakamura, and J. S. Tsai, Phys. Rev. Lett. 97 , 167001 (2006).
- 27 K. Kakuyanagi, T. Meno, S. Saito, H. Nakano, K. Semba, H. Takayanagi, F. Deppe, and A. Shnirman, Phys. Rev. Lett. 98 , 047004 (2007).
- 28 J. Bylander, S. Gustavsson, F. Yan, F. Yoshihara, K. Harrabi, G. Fitch, D. G. Cory, Y. Nakamura, J.-S. Tsai, and W.D. Oliver, Nature Physics 7 , 565 (2011).
- 29 M. Stern, G. Catelani, Y. Kubo, C. Grezes, A. Bienfait, D. Vion, D. Esteve, and P. Bertet, Phys. Rev. Lett. 113 , 123601 (2014).
- 30 S. K. Kehrein and A. Mielke, Phys. Lett. A 219 , 313 (1996).
- 31 F. B. Anders, R. Bulla, and M. Vojta, Phys. Rev. Lett. 98 , 210402 (2007).
- 32 P. Nalbach and M. Thorwart, Phys. Rev. B 81 , 054308 (2010).
- 33 S. Kirchner, K. Ingersent, and Q. Si, Phys. Rev. B 85 , 075113 (2012).
|
10.1103/PhysRevB.90.220302
|
[
"Kyung-Joo Noh",
"Uwe R. Fischer"
] |
2014-08-05T18:37:40+00:00
|
2014-12-09T11:57:19+00:00
|
[
"cond-mat.mes-hall",
"quant-ph"
] |
Consistent perturbative treatment of the subohmic spin-boson model yielding arbitrarily small $T_2/T_1$ decoherence time ratios
|
We present a perturbative treatment of the subohmic spin-boson model which
remedies a crucial flaw in previous treatments. The problem is traced back to
the incorrect application of a Markov type approximation to specific terms in
the temporal evolution of the reduced density matrix. The modified solution is
consistent both with numerical simulations and the exact solution obtained when
the bath-coupling spin-space direction is parallel to the qubit energy-basis
spin. We therefore demonstrate that %in distinction to previous findings the
subohmic spin-boson model is capable of %consistently describing arbitrarily
small %(less than two) ratios of the $T_2$ and $T_1$ decoherence times,
associated to the decay of the off-diagonal and diagonal reduced density-matrix
elements, respectively. An analytical formula for $T_2/T_1$ at the absolute
zero of temperature is provided in the limit of a subohmic bath with vanishing
spectral power law exponent. Small ratios closely mimic the experimental
results for solid state (flux) qubits, which are subject predominantly to
low-frequency electromagnetic noise, and we suggest a reanalysis of the
corresponding experimental data in terms of a $nonanalytic$ decay of
off-diagonal coherence.
|
1408.1062v2
|
## Identifying signatures of photothermal current in a double-gated semiconducting nanotube
G. Buchs 1 , ∗ , S. Bagiante , G. A. Steele 2 3 , ∗
- 1 Centre Suisse d'Electronique et de Microtechnique (CSEM), Jaquet-Droz 1, 2002 Neuchˆtel, Switzerland. a
- 2 Laboratory for Micro- and Nanotechnology, Paul Scherrer Institut, 5232 Villigen PSI, Switzerland.
- 3 Kavli Institute of NanoScience, Delft University of Technology, PO Box 5046, 2600 GA, Delft, The
## Netherlands.
∗ Correspondence and requests for materials should be address to G.A.S. and/or G.B. (email: [email protected]; [email protected]).
The remarkable electrical and optical properties of single-walled carbon nanotubes (SWNT) have allowed for engineering device prototypes showing great potential for applications such as photodectors and solar cells. However, any path towards industrial maturity requires a detailed understanding of the fundamental mechanisms governing the process of photocurrent generation. Here, we present scanning photocurrent microscopy measurements on a double-gated suspended semiconducting SWNT and show that both photovoltaic and photothermal mechanisms are relevant for the interpretation of the photocurrent. We find that the dominant or non-dominant character of one or the other processes depends on the doping profile, and that the magnitude of each contribution is strongly influenced by the series resistance from the band alignment with the metal contacts. These results provide new insight into the interpretation of features in scanning photocurrent microscopy and lay the foundation for the understanding of optoelectronic devices made from SWNTs.
## Introduction
A better understanding of charge transport mechanisms in nanoscale devices, especially the role of electrical contact interfaces and band bending, has been made possible in the last decade with the development of scanning photocurrent microscopy (SPCM), a dedicated local probe technique exploiting local electrical currents generation from light absorption. This technique has been used for characterizing various systems like e.g. Si nanowires [1, 2, 3], colloidal quantum dots [4], VO 2 nanobeams [5], carbon nanotubes [6, 7, 8, 9, 10, 11, 12, 13, 15] as well as 2D materials like graphene [16, 17, 18, 19, 20, 21, 22, 23] and MoS 2 [24, 25]. In such nanoscale systems, two mechanisms have been identified for the generation of photocurrent: i) photovoltaic processes where photo-excited carriers are separated by built-in electrical fields and ii) photothermal processes where thermoelectric forces drive carriers through light-induced thermal gradients.
Single-walled carbon nanotubes (SWNT) are long, one dimensional conductors with a band structure ranging from quasimetallic (bandgap E g ∼ 30 meV) to semiconducting character ( E g typically in the range of 0.1 to 1 eV) depending on chirality and diameter. These properties make SWNTs an ideal platform for exploring photocurrent generation with SPCM. In early single-walled nanotubes SPCM work the interpretation of photocurrent was mostly based on photovoltaic mechanisms [6, 7, 8, 9, 11]. The importance of photothermal effects has been suggested in the context of measurements of bulk SWNT films [12, 13, 14] and very recently, SPCM work on graphene and individual metallic SWNTs has emphasized the importance of photothermal mechanisms in materials with no or small bandgaps [22, 26]. The question of the role of photothermal mechanisms in larger bandgap semiconducting nanotubes has been studied very recently in double-gated [26] and single-gated [27] suspended carbon nanotube devices. These two studies report contradictory results, leaving the understanding of fundamental mechanisms underlying photocurrent generation in semiconducting nanotubes unclear.
Here we report on the study of a suspended semiconducting nanotube device where we show that both photovoltaic and photothermal mechanisms compete in the generation of photocurrent. In particular, we find that the dominant or non-dominant character of one or the other processes is a function of the doping profile and that the magnitude of each contribution is strongly influenced by the band alignment with the metal contacts through the resulting contact resistance.
## Results
Description of the device. The device consists of a suspended nanotube grown between platinum electrodes over a predefined 3 µ m wide and 1 µ m deep trench with four gates defined at the bottom (see Methods). A schematic and a scanning electron microscopy image of a typical device used in this study are shown in Figs. 1(a) and 1(b), respectively. In our SPCM experiments, gate pairs G1-G2 and G3-G4 are connected to independent tunable voltage sources V G1 -G2 and V G3 -G4 , respectively and source-drain voltage V SD is set to 0 V. All measurements are performed at atmospheric pressure and room temperature. A transistor characteristic curve of the device recorded by sweeping all gate voltages simultaneously with V SD = 1 mV is displayed in Fig. 1(c). A large conductance for holes ( R ≈ 300kΩ) and a very low conductance for electrons ( R > GΩ) are observed. From the largely suppressed conductance when the device is pinched-off, we conclude that the nanotube is semiconducting with a relatively large bandgap of the order of a few to several hundreds of meV [15], consistent with the diameter distribution expected from the growth recipe used [28].
SPCM Measurements. Photocurrent (PC) images and corresponding qualitative band diagrams for p-n, n-n, and p-p doping configurations are shown in Fig. 2. For n-n and p-n (n-p) configurations, strong PC signatures observed at locations corresponding to local depletion regions show signs which are consistent with photovoltaic processes [15, 7]. The PC image for the p-p configuration shows a more complex pattern with alternating sign changes along the nanotube axis. The sign of PC spots at the drain and source electrode edges is consistent with the photovoltaic mechanism. However, the sign of the PC patterns observed inside the trench cannot be explained even qualitatively by photovoltaics effects, since it indicates that electrons are travelling 'uphill' along the electrostatic potential profile. Similar observations have recently been reported for single gated devices and have been attributed to photothermoelectric effects [27].
In order to explore the possibility of the presence of such photothermoelectric effects in our device, we studied the PC generation mechanisms by inducing non-uniform charge density profiles in the suspended nanotube channel using the two separate gate pairs. For this, we recorded 2D maps of the PC as a function of the gate pairs voltages ( V G1 -G2 and V G2 -G3 ) for a fixed location of the laser spot along the nanotube axis. The resulting 2D map for the laser spot located at the center of the suspended portion of the SWNT is shown in Fig. 3(a). Four doping regions, clockwise p-n, n-n, n-p and p-p, are identified
across the gates voltages space and are delimited by two dashed black lines. A first clear feature of the image is a strongly suppressed PC in the n-n region. This can be easily understood from the very high contact resistance ( R > GΩ) measured in this regime (Fig. 1(c)). The p-n and n-p regions show PC signatures with a sign that is consistent with the photovoltaic mechanism (Note that the slight clockwise tilt angle of the strip shaped pattern is most probably due to the fact that the laser spot is not exactly centred on the n-p (or p-n) depletion zone but slightly shifted towards the drain contact. This has no incidence on the main conclusions of our present work). However, the p-p region shows PC features which are not consistent with a purely photovoltaic mechanism. This can be verified if one considers the line trace in Fig. 3(b) corresponding to a transition from p-n to p-p + doping configuration along the vertical dashed line in the PC map, which reveals two sign reversals of the PC signal. From a purely photovoltaic mechanism, one would expect a monotonic increase of the PC signal from the p-n minimum to a positive value in the p-p + region with a zero crossing at the symmetric p-p doping profile. This would show only one single PC sign reversal corresponding to a sign reversal of the electric field at the laser spot location. The exact same analysis applies for the PC signal recorded along the horizontal dashed line in the 2D map (Fig. 3(c)).
Model for photothermal currents. From the above analysis, it is clear that the photocurrent in the p-p region of Fig. 3 cannot be described by a photovoltaic mechanism, even at a qualitative level. In Fig. 4, we show how the observations in Fig. 3 can be explained with a simple model where the suspended nanotube is separated in two regions of different carriers density on the left and right side of the laser spot. The model also includes the effect of laser heating at the junction between the two regions. The left (drain contact) and right (source contact) portions have average Seebeck coefficients S 1 and S 2 , respectively. The laser light induces a temperature increase ∆ T with respect to the contacts which can be regarded as heat sinks. We consider linear temperature profiles from the S 1 -S 2 interface to the contacts [27, 12]. Photothermal currents are induced by the local electromotive field E PT = - ∇ S T where ∇ T is the temperature gradient along the nanotube axis and S is the Seebeck coefficient. Thus the generated photothermal current can be expressed by I PT = -R -1 · ∫ S x ( ) ∇ Tdx where R is the overall contact resistance. Applied to our thermocouple model, we find:
$$I _ { \text{PT} } = R ^ { - 1 } \cdot | \Delta T | \cdot ( S _ { 2 } - S _ { 1 } )$$
The sign of the above expression (conventional flow notation) is then given by the Seebeck coefficients
- , whose absolute values and signs depend on the doping levels. A qualitative profile of the Seebeck coefficient as a function of the Fermi level for a semiconducting SWNT is shown in Fig. 4(a) [29, 27].
Analysis and interpretation of the experimental results. Band diagrams corresponding to doping configurations along the vertical line in Fig. 3(a), i.e. p-n, p-p -and p-p + are drawn in Figs. 4(c)-(e), respectively. The photothermal and photovoltaic currents are indicated with arrows oriented with respect to the electronic flow notation (colors correspond to the current sign with respect to the conventional flow notation, i.e. red for positive and blue for negative). From this, we find that the photovoltaic and photothermal currents in the case of the p-n (and n-p) configuration flow in the same direction. However, we find that photovoltaic and photothermal currents are flowing in opposite directions for p-p -and p-p + (p --p and p + -p). To understand these opposite flowing currents, it is important to contrast the mechanisms driving the two. For photovoltaic currents, electron flow is driven by differences in electrostatic voltage: electrons move to the point of lowest energy. For photothermal currents, electrons are driven by differences in chemical potential and in some cases can cause electrons to flow 'uphill'. The photothermal current signs given by our model for gates voltages configurations corresponding to a p-doped nanotube channel (Fig. 4(d)-(e)) are consistent with the experimental results (p-p region in Fig. 3(a)), indicating that photothermoelectric effects are dominant in our device in this regime. From equation 1 and the maximum PC in the p-p region of about 20 pA corresponding to a Seebeck coefficient difference S 2 -S 1 ≈ 140 µ V/K [29], we estimate the laser-induced temperature increase ∆ T to be in the order of ∼ 50 mK. This is smaller than what was reported in a previous work [26]. This could be a result of the approximation of an abrupt junction in our model, from different net laser power reaching the sample, or from a different ratio between excitation energy and optical resonances of the nanotube.
From the data in Fig. 3, it is also quite clear that the largest photocurrent occurs in the p-p region, suggesting that photothermal mechanisms are the most dominant in our device for all gate voltages. However, we note that one has to be careful in comparing photocurrent levels from different doping configurations of the device since the photocurrent is also sensitive to series resistances to the contacts. A more relevant comparison figure for different doping levels is the photo-induced voltage: for our device, in the p-n configuration, the photovoltage should be quite large, on the order of the band gap ( ∼ 100 meV). In contrast, in the p-p regions where the photocurrent is the largest, the photothermal voltage estimated from the device resistance ( R ∼ 300 kΩ) and observed current is on the order of ≈ 10 µ
eV. Although the photovoltage in the p-n and n-p regions should be significantly larger than in the p-p region, the photocurrent is comparable or smaller due to the very large n-type contact series resistance ( R > GΩ).
In previous studies with graphene [22, 30] and metallic nanotubes [26], 2D gate maps such as the one in Fig. 3(a) showed a very characteristic 6-fold rotational symmetry pattern that was used as a fingerprint for identifying a dominating photothermal effect. In contrast, our semiconducting suspended nanotube shows instead a more 4-fold-like pattern. This lack of symmetry finds its origin in the fact that we are working with a semiconducting material with asymmetric n- and p-type contact resistances. In particular, the suppressed PC signal in the n-n region is due to the very large n-type contact resistance ( R > G Ω). For graphene (a semi-metal) and metallic SWNTs, the lack of an energy bandgap allows for similarly low p-type and n-type contact resistances, resulting in equal intensity PC signatures both in p-p and n-n regions of the 2D maps, thus forming a 6-fold pattern.
Furthermore, we note that our conclusion for dominant photothermal PC in the p-type regime is in contradiction with a recent work [26] in which it was suggested that the photovoltaic mechanism dominates photocurrent generation in suspended semiconducting nanotubes for all doping configurations. We address these apparent contradictory results by noting that compared to our device, the devices in Ref. [26] had a much lower contact resistance for n-type doping ( ∼ 100 MΩ) and a much higher resistance for p-type doping ( ∼ 10 MΩ). This suggests a more symmetric effective work function alignment in their devices, with the Fermi level of the leads pinned near the mid-gap level (this difference in work function alignment is likely due to the vacuum environment and the vacuum thermal annealing procedure reported, which likely removed an adsorbed water layer on the surface of the silicon oxide in the trench). With the larger Schottky barriers for p-type doping in Ref. [26] the photothermal currents are likely suppressed to a level close to or below the noise floor of the measurement by the large series resistance to the contacts.
## Discussion
In summary, we have shown that for the interest of interpreting the origin of photocurrent features in SPCM with semiconducting nanotubes, photothermal effects can absolutely not be a-priori excluded. Similar to systems like graphene and metallic SWNTs, suspended semiconducting SWNTs can show strong photothermoelectric effects through measurement of photothermal currents induced by inhomo-
geneous doping profiles along thermal gradients. However, the dominant or non-dominant character of photothermal currents compared with photovoltaic currents strongly depends on the doping profile of the device and on the n- and p-type contact resistances.
## Methods
Sample fabrication. The fabrication of our device began with a p++ silicon wafer used as a backgate covered by 285 nm of thermal silicon oxide. On top of this, gate electrodes made of 5/25 nm W/Pt were defined using electron-beam lithography, followed by the deposition of a 1100 nm thick SiO 2 layer. A 1000 nm deep trench was dry etched, leaving a thin oxide layer on top of the gates. A 5/25 nm W/Pt layer was then deposited to serve as source and drain contacts, and a single-walled carbon nanotube was grown at the last fabrication step at a temperature of 900 ◦ C from patterned Mo/Fe catalysts [31, 28].
SPCM setup. Our SPCM setup [15] consists of a confocal microscope system with the objective ( NA = 0.8) illuminated by λ = 532 nm continuous wave collimated laser light. The diffraction limited spot with a measured diameter of ∼ 330 nm is scanned across the sample surface by means of two orthogonal galvo-mirrors (x,y) combined to a telecentric lens system while the PC signal as well as the reflected light intensity are recorded simultaneously in order to determine the absolute position of the detected PC features.
Measurements. All measurements presented in this work are performed at atmospheric pressure and room temperature. For SPCM measurement, the source-drain voltage is maintained to 0 V and typical light intensities of 5 kW/cm 2 are used.
## References
- [1] Ahn, Y., Dunning, J. & Park, J. Scanning photocurrent imaging and electronic band studies in silicon nanowire field effect transistors. Nano Lett. 5 , 1367-1370 (2005).
- [2] Koren, E., Rosenwaks, Y., Allen, J. E., Hemesath, E. R. & Lauhon, L. J. Nonuniform doping distribution along silicon nanowires measured by kelvin probe force microscopy and scanning photocurrent microscopy. Appl. Phys. Lett. 95 , 092105 (2009).
- [3] Allen, J. E., Hemesath, E. R. & Lauhon, L. J. Scanning photocurrent microscopy analysis of si nanowire field-effect transistors fabricated by surface etching of the channel. Nano Lett. 9 , 19031908 (2009).
- [4] Prins, F. et al. Fast and efficient photodetection in nanoscale quantum-dot junctions. Nano Lett. 12 , 5740-5743 (2012).
- [5] Kasirga, T. S. et al. Photoresponse of a highly correlated material determined by scanning photocurrent microscopy. Nature Nanotech 7 , 723-727 (2012).
- [6] Balasubramanian, K., Burghard, M., Kern, K., Scolari, M. & Mews, A. Photocurrent imaging of charge transport barriers in carbon nanotube devices. Nano Lett. 5 , 507-510 (2005).
- [7] Ahn, Y. H., Tsen, A. W., Kim, B., Park, Y. W. & Park, J. Photocurrent imaging of p-n junctions in ambipolar carbon nanotube transistors. Nano Lett. 7 , 3320-3323 (2007).
- [8] Lee, E. J. et al. Electronic-band-structure mapping of nanotube transistors by scanning photocurrent microscopy. Small 3 , 2038-2042 (2007).
- [9] Freitag, M. et al. Imaging of the schottky barriers and charge depletion in carbon nanotube transistors. Nano Lett. 7 , 2037-2042 (2007).
- [10] Tsen, A. W., Donev, L. A. K., H.Kurt, Herman, L. H. & Park, J. Imaging the electrical conductance of individual carbon nanotubes with photothermal current microscopy. Nature Nanotech. 4 , 108-113 (2008).
- [11] Gabor, N. M., Zhong, Z., Bosnick, K., Park, J. & McEuen, P. L. Extremely efficient multiple electron-hole pair generation in carbon nanotube photodiodes. Science 325 , 1367-1371 (2009).
- [12] St-Antoine, B. C., Menard, D. & Martel, R. Position sensitive photothermoelectric effect in suspended single-walled carbon nanotube films. Nano Lett. 9 , 3503-3508 (2009).
- [13] St-Antoine, B. C., Menard, D. & Martel, R. Photothermoelectric effects in single-walled carbon nanotube films: Reinterpreting scanning photocurrent experiments. Nano Res 5 , 73-81 (2012).
- [14] Nanot, S., Cummings, A. W., Pint, C. L., Ikeuchi, A., Akiho, T., Sueoka, K., Hauge, R. H., L´onard, e F., & Kono, J. Broadband, Polarization-Sensitive Photodetector Based on Optically-Thick Films of Macroscopically Long, Dense, and Aligned Carbon Nanotubes. Sci. Rep. 3 , 1335 (2013).
- [15] Buchs, G., Barkelid, M., Bagiante, S., Steele, G. A. & Zwiller, V. Imaging the formation of a p-n junction in a suspended carbon nanotube with scanning photocurrent microscopy. J. Appl. Phys. 110 , 074308 (2011).
- [16] Allen, J. E., Hemesath, E. R. & Lauhon, L. J. Role of contacts in graphene transistors: A scanning photocurrent study. Phys. Rev. B 79 , 245430 (2009).
- [17] Lee, E. J. H., Balasubramanian, K., Weitz, R. T., Burghard, M. & Kern, K. Contact and edge effects in graphene devices. Nature Nanotech. 3 , 486-490 (2008).
- [18] Lemme, M. C. et al. Gate-activated photoresponse in a graphene pn junction. Nano Lett. 11 , 4134-4137 (2011).
- [19] Peters, E. C., Lee, E. J. H., Burghard, M. & Kern, K. Gate dependent photocurrents at a graphene p-n junction. Appl. Phys. Lett. 96 , 193102 (2010).
- [20] Xia, F. et al. Photocurrent imaging and efficient photon detection in a graphene transistor. Nano Lett. 9 , 1039-1044 (2009).
- [21] Park, J., Ahn, Y. H. & Ruiz-Vargas, C. Imaging of photocurrent generation and collection in singlelayer graphene. Nano Lett. 9 , 1742-1746 (2009).
- [22] Gabor, N. M. et al. Hot carrier-assisted intrinsic photoresponse in graphene. Science 334 , 648-652 (2011).
- [23] Herring, P. K. et al. Photoresponse of an electrically tunable ambipolar graphene infrared thermocouple. Nano Lett. 14 , 901-907 (2014).
- [24] Wu, C. et al. Elucidating the photoresponse of ultrathin MoS 2 field-effect transistors by scanning photocurrent microscopy. J. Phys. Chem. Lett. 4 , 2508-2513 (2013).
- [25] Buscema, M. et al. Large and tunable photothermoelectric effect in single-layer MoS 2 . Nano Lett. 13 , 358-363 (2013).
- [26] Barkelid, M. & Zwiller, V. Photocurrent generation in semiconducting and metallic carbon nanotubes. Nature Photon. 8 , 47-51 (2013).
- [27] DeBorde, T., Aspitarte, L., Sharf, T., Kevek, J. W. & Minot, E. Photothermoelectric effect in suspended semiconducting carbon nanotubes. ACS Nano 8 , 216-221 (2014).
- [28] Kong, J., Soh, H. T., Cassell, A. M., Quate, C. F. & Dai, H. Synthesis of individual single-walled carbon nanotubes on patterned silicon wafers. Nature 395 , 878-881 (1998).
- [29] Small, J. P., Perez, K. M. & Kim, P. Modulation of thermoelectric power of individual carbon nanotubes. Phys. Rev. Lett. 25 , 256801 (2003).
- [30] Song, J. C. W., Rudner, M. S., Marcus, C. M. & Levitov, L. S. Hot carrier transport and photocurrent response in graphene. Nano Lett. 11 , 4688-4692 (2011).
- [31] Steele, G. A., Gotz, G. & Kouwenhoven, L. P. Tunable few-electron double quantum dots and Klein tunnelling in ultra-clean carbon nanotubes. Nature Nanotech. 3 , 363-367 (2009).
## Acknowledgments
The authors acknowledge Val Zwiller for experimental support. This research was supported by a Marie Curie Intra European Fellowship within the 7th European Community Framework Programme (G.B.) and the Netherlands Organization for Scientific Research, NWO (G.A.S.) and the University of Catania (S.B.).
## Author contributions
G.B. and S.B. performed the experiments; S.B. fabricated the sample; G.B. and G.A.S wrote the manuscript; all authors discussed the results and contributed to the manuscript.
The authors declare that they have no competing financial interests.
Figure 1 : Description and characterization of the device. a ) Schematic of the device. Gates G1G2, G3-G4 are connected to tunable voltage sources V G1 -G2 and V G2 -G3 , respectively. The diffractionlimited laser spot ( λ =532 nm) is accurately positioned along the suspended nanotube axis and induced photocurrent is recorded at the source (S) contact. ( b ) Scanning electron microscope image of a typical device. Scale bar: 5 µ m. ( c ) Transistor curve of the device recorded by sweeping gate voltages V G1 -G2 -G3 -G4 simultaneously with V SD = 1 mV. Onsets: qualitative band diagrams in p- and n-doping regimes, corresponding to 'ON' and 'OFF' states, respectively.
Figure 2 : Photocurrent images for different doping configurations. ( a )-( c ) Qualitative band diagrams (left) and photocurrent images (right) measured for p-n, n-n, and p-p doping configurations, respectively. Location of trench edges (vertical black dashed lines) and gates with applied voltages are indicated. Source ( S ) and drain ( D ) contacts with corresponding electrical connections are highlighted in (a) and a yellow dashed line shows the position of the nanotube axis. The location of photocurrent extrema are indicated by dashed circles. The sign of the photocurrent corresponds to the conventional flow notation (red: positive, blue: negative). The arrows on the band diagrams are oriented according to the electron flow notation. Panel (c) illustrates the dominant photothermal mechanism observed in the p-p configuration where electrons are travelling 'uphill' (dashed arrows in the band diagram). Patterns around the maximum intensity PC spots in the PC images are due to diffraction effects at the trench gates [15].
b
(pA)
PC
I
c
(pA)
PC
I
10
0
-10
-20
10
0
-10
-20
8
8
4
0
V
G3-G4
p -
(V)
-p
0
-4
p +
-4
-8
-8
V
G1-G2
(V)
Figure 3 : Gate dependence of the photocurrent demonstrating a clear dominating photothermal response. ( a ) 2D map of the measured photocurrent ( I PC ) versus V G1 -G2 and V G3 -G4 recorded with laser focussed near the middle section of the suspended nanotube, close to the p-n photocurrent minimum shown in Figure 2(a). Black dashed lines delimit four p-n, n-n, n-p and p-p doping regions. ( b ) and ( c ) I PC line traces recorded along the horizontal, respectively vertical white dashed lines in panel (a). n-p, p --p and p + -p (p-n, p-p -and p-p + ) doping regions are highlighted.
p-n
4
n-p p-p +
p-p -
-p
Figure 4 : A simple model for photothermal currents travelling against an electric field. ( a )
Qualitative behavior of the Seebeck coefficient of a semiconducting carbon nanotube versus Fermi level. ( b ) Thermocouple model with two separately gated nanotube regions with corresponding Seebeck coefficients S 1 and S 2 . Laser excitation region located at the two regions interface gives rise to a temperature increase ∆ T . ( c ) -( e ) Band diagrams for p-n, p-p -and p-p + doping configurations with corresponding photovoltaic and photothermal current arrows oriented with respect to the electronic flow notation. The color corresponds to the current sign with respect to the conventional flow notation.
|
10.1038/ncomms5987
|
[
"Gilles Buchs",
"Salvatore Bagiante",
"Gary A. Steele"
] |
2014-08-05T18:40:22+00:00
|
2014-10-28T11:32:06+00:00
|
[
"cond-mat.mes-hall"
] |
Identifying signatures of photothermal current in a double-gated semiconducting nanotube
|
The remarkable electrical and optical properties or single-walled carbon
nanotubes (SWNT) allowed for engineering device prototypes showing great
potential for applications such as photodectors and solar cells. However, any
path towards industrial maturity requires a detailed understanding of the
fundamental mechanisms governing the process of photocurrent generation. Here,
we present scanning photocurrent microscopy measurements on a double-gated
suspended semiconducting SWNT and show that both photovoltaic and photothermal
mechanisms are relevant for the interpretation of the photocurrent. We find
that the dominant or non-dominant character of one or the other processes
depends on the doping profile, and that the magnitude of each contribution is
strongly influenced by the series resistance from the band alignment with the
metal contacts. These results provide new insight into the interpretation of
features in scanning photocurrent microscopy and lay the foundation for the
understanding of optoelectronic devices made from SWNTs.
|
1408.1063v3
|
## Enumeration of three term arithmetic progressions in fixed density sets
## Erik Sj¨land o
## November 11, 2014
## Abstract
Additive combinatorics is built around the famous theorem by Szemer´ edi which asserts existence of arithmetic progressions of any length among the integers. There exist several different proofs of the theorem based on very different techniques. Szemer´ edi's theorem is an existence statement, whereas the ultimate goal in combinatorics is always to make enumeration statements. In this article we develop new methods based on real algebraic geometry to obtain several quantitative statements on the number of arithmetic progressions in fixed density sets. We further discuss the possibility of a generalization of Szemer´ edi's theorem using methods from real algebraic geometry.
## Contents
| 1 | Introduction | Introduction | 3 |
|-----|--------------------------------------------------------|--------------------------------------------------------|-----|
| 2 | Results | Results | 3 |
| 3 | Background | Background | 6 |
| | 3.1 | Existence of arithmetic progressions . . . . . . . . | 6 |
| | 3.2 | From existence to counting . . . . . . . . . . . . . | 11 |
| 4 | Methods from real algebraic geometry | Methods from real algebraic geometry | 14 |
| | 4.1 | Polynomial optimization . . . . . . . . . . . . . . . | 15 |
| | 4.2 | Exploiting symmetries in semidefinite programming | 17 |
| 5 | Gray code, bubble language and fixed-density necklaces | Gray code, bubble language and fixed-density necklaces | 19 |
| 6 | Problem formulated as a semidefinite program | Problem formulated as a semidefinite program | 24 |
| 7 | Proofs of Theorems 2.1, 2.2, 2.3 and Corollary 2.4 | Proofs of Theorems 2.1, 2.2, 2.3 and Corollary 2.4 | 28 |
| 8 | Additional numerical results | Additional numerical results | 40 |
| 9 | Discussion | Discussion | 44 |
|-----|--------------|------------------------------------------------------------------------------------------------------------------------------------------|------|
| | 9.1 | How Szemer´di's e theorem could perhaps be generalized for arith- metic progressions of length 3 . . . . . . . . . . . . . . . . . . . . | 44 |
| | 9.2 | Longer arithmetic progressions . . . . . . . . . . . . . . . . . . . | 47 |
| 10 | Appendix | Appendix | 56 |
## 1 Introduction
Finding an optimal lower bound for the number of arithmetic progressions of length 3 in any subset of fixed cardinality in a set is extremely difficult, and finding this for the cyclic group would imply close to optimal bounds to Szemer´ edi's theorem. In this paper we develop new methods that theoretically will achieve the optimal lower bound by Putinar's Positivstellensatz, and provide a relaxation scheme based on the Lasserre hierarchy providing lower bounds to the number of arithmetic progressions. The methods are general and can be applied to any fixed density set. We apply the methods to find the first known lower bound for the number of arithmetic progressions in any subset S of the cyclic group Z p with a fixed density | S | p for any prime p (Theorem 2.1), and investigate how much this can be improved using a similar certificate (Theorem 2.3 and Corollary 2.4). We discuss a possible way to improve the bounds further, which in theory could ultimately generalize Szemer´di's Theorem, in e Section 9. We provide sharper lower bounds with algebraic certificate for all primes p ≤ 17 (Theorem 2.2), and in Section 8 we discuss lower bounds found using a numerical certificate for all primes p ≤ 613 and the correct results up to p ≤ 32 found by searching through all different possibilities. We have provided tables of all these exact lower bounds in the appendix.
## 2 Results
The easiest is always to consider small examples. Let W k, n , D/n ( [ ] ) denote the optimal lower bound for the number of arithmetic progressions of length k in any subset of [ n ] of cardinality D . Let W k, ( Z n , D/n ) denote the optimal lower bound for the number of arithmetic progressions of length k in any subset of Z n of cardinality D . Using the gray code for fixed density necklaces that is introduced in Section 5 we can find the exact value for W k, ( Z n , D/n ) when n ≤ 32 and D ∈ { 0 1 , , . . . , n } for any k ∈ { 1 , . . . , n } . We have done this for k = 3 4 5, and the results can be found in Tables 3, 4, 5, 6, 7 and 8 in the , , appendix. These results cannot be extended to all n , which is why the rest of the theorems involve lower bounds of W k, ( Z n , D/n ).
The next theorem is the first theorem quantifying how many arithmetic progressions there are in any subset S of Z p of cardinality D = | S | . The bound holds for any prime p .
Theorem 2.1. Let p be a prime number. A lower bound for the minimum number of arithmetic progressions of length 3 among all subsets of Z p of cardinality D ,
$$W ( 3, \mathbb { Z } _ { p }, D / p ) = \min \{ \sum _ { \{ i, j, k \} \ A. P. \ i n \ \mathbb { Z } _ { p } } x _ { i } x _ { j } x _ { k } \, \colon x _ { i } \in \{ 0, 1 \}, \sum _ { i = 0 } ^ { p - 1 } x _ { i } = D \},$$
$$\lambda = \frac { D ^ { 3 } - ( \frac { p + 3 } { 2 } ) D ^ { 2 } + ( \frac { p + 3 } { 2 } - 1 ) D } { p - 1 }.$$
A certificate for the lower bound is given by:
glyph[negationslash]
$$\sum _ { \{ i, j, k \} \ A. P. \ i n \ Z _ { p } } X _ { i } X _ { j } X _ { k } - \lambda = \sum _ { i = 0 } ^ { p - 1 } \sigma _ { 1, i } X _ { i } + \sum _ { i = 0 } ^ { p - 1 } \sigma _ { 2, i } X _ { i } + \sigma _ { 3 } ( D - \sum _ { i = 0 } ^ { p - 1 } X _ { i } ^ { 3 } ) \\ + \sigma _ { 4 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ),$$
where
$$\sigma _ { 1, i } \ = \frac { 1 } { p - 1 } \sum _ { 0 < j < k < ( p - 1 ) / 2 } ( X _ { j + i } - X _ { j + k + i } - X _ { n - j - k + i } + X _ { n - j + i } ) ^ { 2 }$$
$$\sigma _ { 2, i } \ = \frac { 1 } { p - 1 } ( D X _ { i } - \sum _ { j = 0 } ^ { p - 1 } X _ { j } ) ^ { 2 }$$
$$\sigma _ { 3 } \ = \frac { ( D - 1 ) ^ { 2 } } { p - 1 }$$
$$\sigma _ { 4 } \ = \frac { 4 D - p + 3 } { 2 ( p - 1 ) }.$$
$$\cdot$$
One can hope to find a sharper bound than the one provided in Theorem 2.1 that holds for all primes, but how to find this is not trivial. Next we provide sharper bounds for some low primes. The reason we can find nice algebraic bounds for p ≤ 17 has to do with that the trigonometric functions in Theorem 2.3 are on a nice form. We could find results for slightly larger p , but it would require more work and the bounds would contain messy combinations of trigonometric functions.
Theorem 2.2. Let p be prime and let W (3 , Z p , D/p ) denote the minimum number of arithmetic progressions of length 3 among all subsets of Z p of cardinality D . There are algebraic certificates with polynomials up to degree 3 giving the following bounds
$$\Lambda _ { 0 } ^ { 3 } W ( 3, \mathbb { Z } _ { 5 }, D / 5 ) \geq \frac { D ^ { 3 } - 3 D ^ { 2 } + 2 D } { 6 },$$
$$W ( 3, \mathbb { Z } _ { 7 }, D / 7 ) \geq \frac { D ^ { 3 } - 4 D ^ { 2 } + 3 D } { 8 },$$
$$W ( 3, \mathbb { Z } _ { 1 3 }, D / 1 3 ) \geq \frac { 2 1 - 2 \sqrt { 3 } } { 2 8 6 } D ^ { 3 } + \frac { 2 8 \sqrt { 3 } - 1 5 1 } { 2 8 6 } D ^ { 2 } + \frac { 5 - \sqrt { 3 } } { 1 1 } D$$
$$W ( 3, \mathbb { Z } _ { 1 1 }, D / 1 1 ) \geq \frac { \sqrt { 5 } D ^ { 3 } + ( 1 5 - 1 2 \sqrt { 5 } ) D ^ { 2 } + ( - 1 5 + 1 1 \sqrt { 5 } ) D } { 3 0 },$$
and
$$W ( 3, \mathbb { Z } _ { 1 7 }, D / 1 7 ) \geq \frac { 1 } { 2 4 } D ^ { 3 } - \frac { 1 } { 4 } D ^ { 2 } + \frac { 5 } { 2 4 } D.$$
In fact, finding bounds sharper than those in Theorem 2.1 can be done very efficiently for relatively small primes using a simple degree 3 relaxation. As shown in the following theorem we can reformulate a certain relaxation as a linear program:
Theorem 2.3. Let r be a primitive root of the prime p . Let further
$$V _ { i j } = \left | \left \{ \{ 0, 1, r ^ { i } \} \, \colon \{ 0, 1, r ^ { i } \} = \{ 0, r ^ { t }, r ^ { j + t } \} \, f o r \, t = 0, \dots, p - 2 \right \} \right |$$
for all i, j ∈ { 0 , . . . , p - } 1 ,
$$C _ { i j } = c o s ( \frac { 2 \pi ( i - 1 ) ( j - 1 ) } { p - 1 } )$$
for all i, j ∈ { 0 , . . . , p - } 1
$$u = [ u _ { 0 }, u _ { 1 }, \dots, u _ { \frac { p - 3 } { 2 } }, u _ { \frac { p - 1 } { 2 } }, u _ { \frac { p - 3 } { 2 } }, \dots, u _ { 1 } ] ^ { T },$$
$$u _ { + } = 1 ^ { T } u = u _ { 0 } + 2 u _ { 1 } + \cdots + 2 u _ { ( p - 3 ) / 2 } + u _ { ( p - 1 ) / 2 }$$
and
$$v _ { i } = \begin{cases} \ 1 & \text{if $r^{i}=2$} \\ \ 0 & \text{otherwise.} \end{cases}$$
for i ∈ { 0 , . . . , p - } 1 .
The following optimization problems attain the same optimal value:
(a)
$$\max \{ \lambda \, \colon \sum _ { \{ i, j, k \} \ A. P. \ i n \, \mathbb { Z } _ { p } } X _ { i } X _ { j } X _ { k } - \lambda = S \}$$
where
$$S = \sum _ { \substack { i = 0 \\ i = 0 } } ^ { p - 1 } \sum _ { j = 0 } ^ { p - 1 } ( \sum _ { k = 1 } ^ { p - 1 } a _ { i j k } X _ { k } ) ^ { 2 } x _ { i } + b \sum _ { i = 0 } ^ { p - 1 } ( D X _ { i } - \sum _ { j = 0 } ^ { p - 1 } X _ { j } ) ^ { 2 } X _ { i } \\ + c ( \sum _ { i = 0 } ^ { p - 1 } X _ { i } ^ { 3 } - D ) + d ( \sum _ { i, j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) \\.$$
for a ijk , c, d ∈ R and b ≥ 0 .
$$\max \{ \frac { u _ { + } } { p - 1 } ( D - 1 - \frac { u _ { 0 } } { u _ { + } } ( p - 1 ) ) D ( D - 1 ) \colon C u \geq 0, V u = v \}.$$
The following corollary, which follows from Theorem 2.1 and Theorem 2.3 shows that there is room to improve the bounds further.
Corollary 2.4. Let p be prime and denote the optimal value to problem (a) in Theorem 2.3 by λ p ( D ) :
$$\lambda _ { p } ( D ) = \max \{ \lambda \, \colon \sum _ { \{ i, j, k \} \ A. P. \ i n \ \mathbb { Z } _ { p } } X _ { i } X _ { j } X _ { k } - \lambda = S \}$$
where
$$S = \sum _ { \substack { i = 0 \\ j = 0 } } ^ { p - 1 } \sum _ { j = 0 } ^ { p - 1 } ( \sum _ { k = 1 } ^ { p - 1 } a _ { i j k } X _ { k } ) ^ { 2 } X _ { i } + b \sum _ { i = 0 } ^ { p - 1 } ( D X _ { i } - \sum _ { j = 0 } ^ { p - 1 } X _ { j } ) ^ { 2 } X _ { i } \\ + c ( \sum _ { i = 0 } ^ { p - 1 } X _ { i } ^ { 3 } - D ) + d ( \sum _ { i, j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) )$$
for a ijk , c, d ∈ R and b ≥ 0 . For all p
$$\left \lceil \frac { p + 3 } { 4 } \right \rceil \leq \min \{ D \in \mathbb { Z } _ { + } \, \colon \lambda _ { p } ( D ) > 0 \} \leq \frac { p + 3 } { 2 }.$$
## 3 Background
## 3.1 Existence of arithmetic progressions
This section provides a literature review on existence type theorems about arithmetic progressions. We start with some basic definitions and discuss many of the major contributions to the field.
Definition 3.1. For S ⊆ N , let us define the (upper) density
$$\rho ( S ) = \lim _ { n \to \infty } \frac { | S \cap \{ 1, 2, \dots, n \} | } { n }.$$
As an example, the density of the even positive integers is ρ (2 N ) = 1 2. / Formally we define arithmetic progressions for the integers and for finite groups in the following way:
Definition 3.2. An arithmetic progression of length k ∈ Z + in the positive integers is given by a, a + b, a +2 b, . . . , a +( k -1) b , where a, b ∈ Z + .
Definition 3.3. An arithmetic progression of length k ∈ Z + in a finite group G is a set of k distinct elements { a, b · a, b · b · a, . . . , b k -1 · a } , where a ∈ G and b ∈ G glyph[integerdivide] { 0 } .
To clarify, for example { 1 2 3 , , } , { 1 3 2 , , } , { 2 1 3 , , } , { 2 3 1 , , } , { 3 1 2 , , } and { 3 2 1 , , } should be considered as the same arithmetic progression as it is the same set of 3 distinct elements, hence when we write
$$\sum _ { \{ i, j, k \} } \, \text{is an A.P. in } G$$
only one representative for every arithmetic progression is used.
It is interesting to study arithmetic progressions since it measures how structured a set is. A set is usually either very random in nature or has a structure similar to that of an arithmetic progression. In additive combinatorics one tries to answer questions about the length of the longest arithmetic progressions contained in very unstructured sets, how structured sets are contained in generalizations of arithmetic progressions and many similar questions. In the last century a lot of results in additive combinatorics have been proven using probabilistic methods, Fourier-analytic methods, sum set estimates, graph theory, ergodic theory and algebraic methods. To answer questions about existence of arithmetic progression in subsets of the integers is one of the most difficult questions in additive combinatorics, and the question that has gotten the most attention. We refer to the book by Tao and Vu [79] for an accessible introduction to the subject.
Van der Waerden proved in 1927 [81] that for any given positive integers r and k there is a number N such that if the integers { 1 , . . . , N } are colored in r colors, then there exists a monochromatic arithmetic progression of length k . This is one of the key results of Ramsey theory. The smallest such N will be denoted W k r ( ) and is called the van der Waerden number . Already for r = 2 we have a huge gap between the lower bound
$$W _ { 2 } ( k ) \geq p 2 ^ { p }$$
for primes by Berlekamp [6] and the upper bound
$$W _ { 2 } ( k ) \leq 2 ^ { 2 ^ { 2 ^ { 2 ^ { k + 9 } } } }$$
by Gowers [18]. Finding bounds for van der Waerden-like numbers is still an active area of research, recent contributions include [34, 21, 1, 2].
Results about arithmetic progressions in fixed-density sets implies results about monochromatic arithmetic progressions in colorings and are thus more general. The first non-trivial result about existence of arithmetic progressions in sets is due to Roth in 1953. The proof is based on the Hardy-Littlewood circle method, and because of this result Roth was awarded the fields medal in 1958.
Theorem 3.4 (Roth's Theorem [57]) . If S ⊆ N and ρ S ( ) > 0 , then it follows that we can find arithmetic progressions in S of length 3 .
In 1969 Szemer´ edi strengthened Roth result to existence of arithmetic progressions of length 4 [70] and in 1975 he generalized it to arithmetic progressions of arbitrary length [71]. In order to prove the result Szemer´di developed the e famous regularity lemma, which states that every large enough graph can be partitioned into subgraphs of about the same size so that the edges between the different subgraphs behave fairly randomly. For these and other combinatorial results he was awarded the Abel prize in 2012.
Theorem 3.5 (Szemer´ edi's Theorem [71]) . If S ⊆ N and ρ S ( ) > 0 , then it follows that we can find arbitrarily long arithmetic progressions in S .
Definition 3.6. We denote by N k,δ ( ) the smallest positive integer such that for any M ≥ N k,δ ( ) it holds that any subset of { 1 , . . . , M } of cardinality δM contains an arithmetic progression of length k for 0 < δ ≤ 1 . Similarly we define δ N, k ( ) to be the smallest density such that any subset of { 1 , . . . , M } of cardinality δ N, k M ( ) contains an arithmetic progression of length k for any M ≥ N .
Szemer´ edi's theorem implies the existence of an N k,δ ( ) for any k and δ but fails to provide any good bounds for this number. Another proof of Szemer´di's e theorem using ergodic methods was mainly due to Furstenberg in 1982 [14]. The ergodic methods makes the proof easier, and they have implied several other results in additive combinatorics including a multi-dimensional version and a polynomial version of Szemer´di's theorem, and the related density Halese Jewett theorem. Tao gave a quantitative version based on ergodic theory in 2006 [75], but no effective bound for N k,δ ( ) has been found using ergodic methods.
Even though Roth's theorem is old, there has been more recent progress on improving the quantitative bounds N (3 , δ ). Most results are formulated bounding δ N, ( 3) instead of N (3 , δ ), and so we follow this notation. Roth's original proof showed that there is a constant C such that δ N, ( 3) ≤ C log log N . This was improved by Heath-Brown and Szemer´ edi [27, 72] showing that there exist constants C 1 and C 2 such that δ N, ( 3) ≤ C 1 (log N ) C 2 and later by Bourgain
[8] showing that δ N, ( 3) ≤ C ( log log N log N ) 1 / 2 for some constant C . More recently Bourgain [9] improved the bound to δ N, ( 3) ≤ C (log log N ) 2 (log N ) 2 / 3 , then Sanders [61] improved it to δ N, ( 3) ≤ C (log N ) 3 / 4 -o (1) and then Sanders [60] improved it further to δ N, ( 3) ≤ C (log N ) 1+ (1) o for some large enough constant C .
Versions of Roth's theorem for other groups have also been of interest lately. The result was generalized to hold for arbitrary abelian groups of odd order by Meshulam in 1995 [42]. A way to generalize Roth's theorem is through generalizing the (triangle) removal lemma, which is related to Szemer´di's regulaity e lemma. Green generalized the removal lemma to abelian groups [24], and Kr´l, a Serra and Vela generalized it to hold for all finite groups, with a version of Roth's theorem for all finite groups as a consequence [33]:
Theorem 3.7. Let G be a finite group of odd order N and S a subset of its elements. If the number of solutions of the equation x · z = y 2 with x, y, z ∈ S is o N ( 2 ) , then the size of S is o N ( ) .
To get the result to also hold for non-abelian groups they had to avoid using Fourier analysis. Other results in additive combinatorics for non-abelian groups are the results by Bergelson and Hindman [5], Gowers [20], Tao [76, 77, 78], Sanders [59] and Solymosi [69]. The results have been established long after the commutative counterparts, and often with much weaker bounds.
Let a ↑ b denote a b , a ↑ b ↑ c = a ↑ ( b ↑ c ) = a ( b c ) , and so on. In 1998 [17] Gowers showed that there exists an absolute constant C such that δ N, ( 4) ≤ (log log N ) -C , which was a major improvement on the previous best bound. The bound was improved by Green and Tao in 2009 [26] to δ N, ( 4) ≤ e -C √ log log N . In 2001 Gowers provided the first effective upper bound for N k,δ ( ) for k > 3 using the Fourier analytic machinery he had developed and been awarded the fields medal for in 1998.
Theorem 3.8 ([18], Theorem 18.2) . Let k be a positive integer and let 0 < δ ≤ 1 / 2 . Then
$$N ( k, \delta ) = 2 \uparrow 2 \uparrow \delta ^ { - 1 } \uparrow 2 \uparrow 2 \uparrow k + 9.$$
Before stating our next theorem, let us state what (Gower's version of) Szemer´ edi's theorem implies about W k, n , δ ( [ ] ).
Corollary 3.9. It follows by Theorem 3.8 that
$$\lim _ { n \to \infty } \min \{ \delta \in \mathbb { R } _ { + } \, \colon W ( k, [ n ], \delta ) > 0 \} = 0.$$
Conversely, if there are lower bounds λ k, n , δ ( [ ] ) ≤ W k, n , δ ( [ ] ) with the additional property that
$$\lim _ { n \to \infty } \min \{ \delta \in \mathbb { R } _ { + } \, \colon \lambda ( k, [ n ], \delta ) > 0 \} = 0$$
holds, then Szemer´di's theorem would follow. e
Proof. We want to show that for any glyph[epsilon1] > 0 and k ∈ Z + there exists N k,glyph[epsilon1] ∈ R + such that
$$\min \{ \delta \in \mathbb { R } _ { + } \, \colon W ( k, [ n ], \delta ) > 0 \} < \epsilon$$
for all n > N k,glyph[epsilon1] . By Theorem 3.8 we know that if we pick N k,glyph[epsilon1] = 2 ↑ 2 ↑ glyph[epsilon1] -1 ↑ 2 ↑ 2 ↑ k +9, then any subset S ⊂ { 1 , . . . , n } with cardinality | S | ≥ glyph[epsilon1]n , where n > N , contains an arithmetic progression of length k , hence the desired inequality holds.
On the other hand, suppose
$$\lim _ { n \to \infty } \min \{ \delta \in \mathbb { R } _ { + } \, \colon \lambda ( k, [ n ], \epsilon ) > 0 \} = 0.$$
Since λ k, n , glyph[epsilon1] ( [ ] ) ≤ W k, n , glyph[epsilon1] ( [ ] ) it follows that
$$\lim _ { n \to \infty } \min \{ \delta \in \mathbb { R } _ { + } \, \colon W ( k, [ n ], \epsilon ) > 0 \} = 0,$$
or in other words that for any glyph[epsilon1] > 0 there exists an N k,glyph[epsilon1] such that for any n > N k,glyph[epsilon1] it holds that S ⊂ { 1 , . . . , n } with cardinality | S | = nglyph[epsilon1] has an arithmetic progression of length k . This is equivalent to that there are bounds for Szmer´ edi's theorem on the form N k,glyph[epsilon1] ( ) = N k,glyph[epsilon1] .
Another proof strategy for Szemer´di's theorem relies on generalizing Szee mer´ edi's regularity lemma to hypergraphs, and was developed independently by Gowers [19] and Nagle, R¨dl, Schacht and Skokan [55, 56, 45, 53, 54]. o
All the results above on Szemer´di's theorem require a positive upper density, e but this condition might not be necessary. An open problem is how dense does a set have to be in order to contain arithmetic progressions. The following has been conjectured by Erd˝s and Tur´n. o a
Conjecture 3.10 (Erd˝ os-Tur´n conjecture) a . Let S be a subset of the positive integers. S contains arithmetic progressions of any length if ∑ x ∈ S 1 x diverges.
The primes do not have positive density, but results on arithmetic progressions among them would follow from the Erd˝s-Tur´n conjecture. Existence of o a arithmetic progressions of length 3 among any positive proportion of the prime numbers was proven by Green [23] and quantitative improvements were made by Helfgott and Roton [28]. The result for arithmetic progressions of arbitrary length was found using Szemer´di's regularity lemma and properties about the e distribution of primes by Green and Tao [25].
Theorem 3.11 (Green-Tao theorem) . There exist arbitrarily long arithmetic progressions among the prime numbers.
So far we have only discussed results on upper bounds for δ N, k ( ), one can also try to find lower bounds. The lower bounds are still very far from the upper bounds, and surprisingly not much improvements have been done on the lower bounds for over 50 years. Let C be a large enough constant. The lower bound for k = 3 due to Behrend [4]
$$\delta ( N, 3 ) \geq C \frac { 1 } { \log ^ { 1 / 4 } N } \cdot \frac { 1 } { 2 ^ { 2 \sqrt { 2 } } \sqrt { \log _ { 2 } N } }$$
was the best bound for over 60 years until Elkin [13] recently improved it to
$$\delta ( N, 3 ) \geq C \log ^ { 1 / 4 } N \cdot \frac { 1 } { 2 ^ { 2 \sqrt { 2 } } \sqrt { \log _ { 2 } N } }.$$
For larger k the we have the following bound by Rankin [50]:
$$\delta ( N, k ) \geq \exp ( - C ( \log N ) ^ { 1 / ( k - 1 ) } ).$$
Soon after Elkin improved the bound for k = 3, O'Bryant [48] improved the bounds for k = 3 and general k further: For every glyph[epsilon1] > 0 and N large enough
$$\delta ( N, 3 ) \geq \left ( \frac { 6 \cdot 2 ^ { 3 / 4 } \sqrt { 5 } } { e \pi ^ { 3 / 2 } } - \epsilon \right ) 2 ^ { - \sqrt { 8 \log _ { 2 } N } + \frac { 1 } { 4 } \log _ { 2 } \log _ { 2 } N }$$
and
$$\delta ( N, k ) \geq C _ { k } 2 ^ { - n 2 ^ { ( n - 1 ) / 2 } \, \tilde { \sqrt { \log _ { 2 } N } } + \frac { 1 } { 2 n } \log _ { 2 } \log _ { 2 } N },$$
where C k > 0 is an unspecified constant and n = log glyph[ceilingleft] k glyph[ceilingright] .
Szemer´ edi's theorem roughly states that any subset S ⊂ { 1 , . . . , N } of the integers, however random and unstructured it is, will contain very structured subsets if N and | S | are large enough. One can ask how long arithmetic progressions there are in more structured sets, such as for example S + S = { s 1 + s 2 | s , s 1 2 ∈ S } and lS = { s 1 + · · · + s l | s , . . . , s 1 l ∈ S } . In fact, the additional structure allows us to find much longer arithmetic progressions, for example we have the following result from Bourgain [7] on sums of sets:
Theorem 3.12. Let p ≥ 1 be a prime nuber and A,B additive sets in Z p with | A , | | B | ≥ δp for some C (log log p ) 3 log p < δ ≤ 1 where C > 1 is a large enough absolute constant. Then A + B contains an arithmetic progression of length at least exp( ( c δ log p ) 1 / 3 ) for some constant c > 0 .
Iterated sumsets have even more structure, and results include those of S´ ark¨zy o [62], Lev [38, 39], Nathanson and Rusza [46, 47], and Szemer´di and e Vu [73]. We state a quite general and strong theorem from [73]:
Theorem 3.13. Let d ≥ 1 . There exists constants C , D d d > 0 such that for any l ≥ 1 and A ⊂ { 1 , . . . N } of cardinality | A | ≥ C d l N d and | A | ≥ 2 , lA contains a proper arithmetic progression of length D l A d | | 1 /d
The number of arithmetic progressions in a fixed density set is closely related to the number of monochromatic arithmetic progressions in coloring of a set. All results above can easily be translated in terms of colorings.
Definition 3.14. A c -coloring of a set S is a map χ : S →{ 1 , . . . , c } splitting S into c color classes.
Proposition 3.15. If there exist arithmetic progressions in any set S ⊆ Z + with positive upper density ρ S ( ) > 0 then there exist monochromatic arithmetic progressions in any c -coloring of the positive integers.
Proof. Suppose there exist arithmetic progressions in any set S ⊆ Z + with positive upper density. Let χ ( Z + ) be a c -coloring of the positive integers. Since c is finite we have
$$\max _ { i \in \{ 1, \dots, c \} } \frac { | \{ j \in \{ 1, \dots, n \} \colon \chi ( j ) = i \} | } { n } \geq \frac { 1 } { c } > 0,$$
for any n , and thus for some i the color class S i = { j ∈ { 1 , . . . , n } : χ j ( ) = i } has positive upper density. By assumption there are arithmetic progressions in S i , and thus of color i .
## 3.2 From existence to counting
The ultimate goal in combinatorial problems is always to count the desired constellation. The existence results on arithmetic progressions provide very strong results on how many arithmetic progressions there are in certain limits,
but provides no information about the amount of arithmetic progressions far from the limits. To make the statements precise, let us define the functions of interest:
Definition 3.16. Let R k, n , c ( [ ] ) denote the minimal number of monochromatic arithmetic progressions of length k in any c -coloring of [ n ] = 1 { , . . . , n } .
Let R k, G, c ( ) denote the minimal number of monochromatic arithmetic progressions of length k in any c -coloring of a finite group G .
Let W k, n , δ ( [ ] ) denote the minimal number of arithmetic progressions of length k in any subset S ⊂ { 1 , . . . , n } of cardinality | S | = nδ .
Let W k,G,δ ( ) denote the minimal number of arithmetic progressions of length k in any subset S ⊂ G of cardinality | S | = | G δ | .
It would of course be desirable to find the exact values of these functions for any possible inputs k , n , c , δ and G , but it is not likely that they will ever be found because of how complicated the functions are. Finding the exact functions would give us sharp upper bounds for Szemer´ edi's theorem for all k , which is known to be extremely difficult already for low values of k . Since there is little hope in finding the exact functions, we aim to find as good lower bounds to the functions as possible. Good lower bounds provide a better understanding of arithmetic progressions, and if they are good enough they may additionally improve the current best existence bounds. We state all the best bounds known for various inputs that we have been able to find in the literature. Upper bounds are also provided, and are usually achieved by colorings.
R(3,[n],2) : Upper and lower bound from [49].
$$\frac { 1 6 7 5 } { 3 2 7 6 8 } n ^ { 2 } ( 1 + o ( 1 ) ) \leq R ( 3, [ n ], 2 ) \leq \frac { 1 1 7 } { 2 1 9 2 } n ^ { 2 } ( 1 + o ( 1 ) )$$
R(3, Z n ,2) : Upper and lower bound from [66] (The special case n mod 24 ∈ { 1 5 7 11 13 17 19 23 , , , , , , , } found first in [11]).
$$n ^ { 2 } / 8 - c _ { 1 } n + c _ { 2 } \leq R ( 3, \mathbb { Z } _ { n }, 2 ) \leq n ^ { 2 } / 8 - c _ { 1 } n + c _ { 3 }$$
where constants depends on n mod 24:
| n mod 24 | c 1 | c 2 | c 3 |
|------------------------------------|-------|-------|--------|
| 1 , 5 , 7 , 11 , 13 , 17 , 19 , 23 | 1 / 2 | 3 / 8 | 3 / 8 |
| 8 , 16 | 1 | 0 | 0 |
| 2 , 10 | 1 | 3 / 2 | 3 / 2 |
| 4 , 20 | 1 | 0 | 2 |
| 14 , 22 | 1 | 3 / 2 | 3 / 2 |
| 3 , 9 , 15 , 21 | 7 / 6 | 3 / 8 | 27 / 8 |
| 0 | 5 / 3 | 0 | 0 |
| 12 | 5 / 3 | 0 | 18 |
| 6 , 18 | 5 / 3 | 1 / 2 | 27 / 2 |
R(3, D 2 n ,2) : Upper and lower bound from [66].
$$n ^ { 2 } / 4 - 2 c _ { 1 } n + 2 c _ { 2 } \leq R ( 3, D _ { 2 n }, 2 ) \leq n ^ { 2 } / 4 - 2 c _ { 1 } n + 2 c _ { 3 }$$
where constants depends on n mod 24:
| n mod 24 | c 1 | c 2 | c 3 |
|------------------------------------|-------|-------|--------|
| 1 , 5 , 7 , 11 , 13 , 17 , 19 , 23 | 1 / 2 | 3 / 8 | 3 / 8 |
| 8 , 16 | 1 | 0 | 0 |
| 2 , 10 | 1 | 3 / 2 | 3 / 2 |
| 4 , 20 | 1 | 0 | 2 |
| 14 , 22 | 1 | 3 / 2 | 3 / 2 |
| 3 , 9 , 15 , 21 | 7 / 6 | 3 / 8 | 27 / 8 |
| 0 | 5 / 3 | 0 | 0 |
| 12 | 5 / 3 | 0 | 18 |
| 6 , 18 | 5 / 3 | 1 / 2 | 27 / 2 |
R(3,G,2) : Let G be any finite group. Let G k denote the set of elements of G of order k , N = | G | and N k = | G k | . Denote the Euler phi function φ k ( ) = |{ t ∈ { 1 , . . . , k } : t and k are coprime }| . Let K = { k ∈ { 5 , . . . , n } : φ k ( ) ≥ 3 k 4 } . Lower bound from [67]:
$$\sum _ { k \in K } \frac { N \cdot N _ { k } } { 8 } ( 1 - 3 \frac { k - \phi ( k ) } { \phi ( k ) } ) \leq R ( 3, G, 2 )$$
R(4, Z n ,2) : Wolf [82] contributed with several bound on R (4 , Z n , 2). The bounds were improved by Lu and Peng [40]. Note that the results in their paper differs by a factor 2, this is because they count { a, b · a, b · b · a } and { b · b · a, b · a, a } as distinct arithmetic progressions. If p is prime we have [40, Theorem 1.1]:
$$\frac { 7 } { 1 9 2 } p ^ { 2 } ( 1 + o ( 1 ) ) \leq R ( 4, \mathbb { Z } _ { p }, 2 ) \leq \frac { 1 7 } { 3 0 0 } p ^ { 2 } ( 1 + o ( 1 ) ).$$
For general n we have [40, Theorem 1.2 and Theorem 1.3]:
$$c _ { 1 } n ^ { 2 } ( 1 + o ( 1 ) ) \leq R ( 4, \mathbb { Z } _ { n }, 2 ) \leq c _ { 2 } n ^ { 2 } ( 1 + o ( 1 ) ).$$
where constants depends on n mod 4:
$$\begin{smallmatrix} n \mod 4 & c _ { 1 } & c _ { 2 } \\ 1, 3 & 7 / 1 9 2 & 1 7 / 3 0 0 \\ 0 & 2 / 6 6 & 8 5 4 3 / 1 4 5 2 0 0 0 \\ 2 & 7 / 1 9 2 & 8 5 4 3 / 1 4 5 2 0 0 0 \end{smallmatrix}$$
Furthermore we have [40, Theorem 1.5]
$$\varliminf _ { n \rightarrow \infty } R ( 4, \mathbb { Z } _ { n }, 2 ) \leq \frac { 1 } { 2 4 },$$
and the following conjecture [40, Conjecture 1.1]:
$$\inf _ { n } \{ R ( 4, \mathbb { Z } _ { n }, 2 ) \} = \frac { 1 } { 2 4 }.$$
R(4,[n],2) : Upper bound from [40, Equation (12)]:
$$R ( 4, [ n ], 2 ) \leq \frac { 1 } { 7 2 } n ^ { 2 } ( 1 + o ( 1 ) ).$$
R(5, Z n ,2) : For odd n we have [40, Theorem 1.4]:
$$R ( 5, \mathbb { Z } _ { n }, 2 ) \leq \frac { 3 6 2 9 } { 1 3 1 4 2 4 } n ^ { 2 } ( 1 + o ( 1 ) ).$$
For even n we have [40, Theorem 1.4]:
$$R ( 5, \mathbb { Z } _ { n }, 2 ) \leq \frac { 3 6 4 7 } { 1 3 1 4 2 4 } n ^ { 2 } ( 1 + o ( 1 ) ).$$
Furthermore we have [40, Theorem 1.5]
$$\varliminf _ { n \to \infty } R ( 5, \mathbb { Z } _ { n }, 2 ) \leq \frac { 1 } { 7 2 }.$$
R(5,[n],2) : Upper bound from [40, Equation (13)]:
$$R ( 5, [ n ], 2 ) \leq \frac { 1 } { 3 0 4 } n ^ { 2 } ( 1 + o ( 1 ) ).$$
W(3, Z p , δ ) Let p be prime. A lower bound is given by Theorem 2.1:
$$\frac { ( \delta p ) ^ { 3 } - \frac { p + 3 } { 2 } ( \delta p ) ^ { 2 } + \frac { p + 1 } { 2 } ( \delta p ) } { p - 1 } \leq W ( 3, \mathbb { Z } _ { p }, \delta ).$$
Since { a, b, c } is an arithmetic progression in S ∈ { 1 , . . . , n } if and only if a + c = 2 b for distinct a, b, c ∈ S there are many related problems where one tries to count the number of solutions to various linear equations. One example is Schur triples, which are triples { a, b, c } such that a + b = , have been studied c in [52, 65, 12]. Other examples include sets with no solutions to x + y = 3 z studied in [43] and monochromatic solution to equations in groups [11].
## 4 Methods from real algebraic geometry
Tools from real algebraic geometry are used to prove the main results of this paper. Even though the main result can be understood without understanding polynomial optimization, it played a vital part when finding the certificates in the main theorems. Only a few of the most important theorems will be presented. For a more extensive survey on the topic we refer to the excellent paper by Laurent [37]. For a longer review of how we applied and implemented methods from real algebraic geometry to the specific problem we refer to [68].
## 4.1 Polynomial optimization
We use polynomial optimization to count monochromatic arithmetic progressions, and so in this section we briefly review the topic.
Let f ( x , g ) 1 ( x , . . . , g ) m ( x ) be polynomials. A problem on the form
$$\begin{smallmatrix} \rho _ { * } = \inf & f ( x ) \\ \text{subject to } & g _ { 1 } ( x ) \geq 0, \dots, g _ { m } ( x ) \geq 0, \\ & x \in \mathbb { R } ^ { n }, \end{smallmatrix}$$
is a polynomial optimization problem. One way to solve a polynomial optimization problem is by studying the related problem:
$$\begin{array} { c c } \rho ^ { * } = \sup & \lambda \\ \text{subject to} & \lambda - f ( x ) \geq 0, g _ { 1 } ( x ) \geq 0, \dots, g _ { m } ( x ) \geq 0, \\ & \lambda \in \mathbb { R }, x \in \mathbb { R } ^ { n }. \end{array}$$
Definition 4.1. Let g , . . . , g 1 m ∈ R [ x , . . . , x 1 n ] be polynomials. The quadratic module generated by g , . . . , g 1 m is defined by:
$$\text{QM} ( g _ { 1 }, \dots, g _ { m } ) = \{ \sigma _ { 0 } + \sum _ { i = 1 } ^ { m } \sigma _ { i } g _ { i } | \sigma _ { 0 }, \dots, \sigma _ { m } \ a r e \ s u m s \ o f \ s q u a r e s \}.$$
Definition 4.2. A quadratic module QM( g , . . . , g 1 m ) is Archimedean if there is an N ∈ N such that
$$N - \sum _ { i = 1 } ^ { n } x _ { i } ^ { 2 } \in \text{QM} ( g _ { 1 }, \dots, g _ { m } ).$$
Theorem 4.3 (Putinar's Positivstellensatz) . Let the set K = { x ∈ R n | g 1 ( x ) ≥ 0 , . . . , g m ( x ) ≥ 0 } be compact. Assume that the associated quadratic module QM( g , . . . , g 1 m ) is Archimedean. If f is strictly positive on K , then it is possible to find sums of squares σ , . . . , σ 1 m such that f = σ 0 + ∑ m i =1 σ g i i .
In other words, if the quadratic module of the polynomial constraints in the polynomial optimization problem is Archimedean it is possible to solve the polynomial optimization problem using are sums of squares ,
$$\begin{array} { c c c c c } \cdot & \cdot & \cdot \\ & \rho ^ { * } = \sup & \lambda \\ & \text{subject to} & \lambda - f ( X ) = \sigma _ { 0 } + \sum _ { i = 1 } ^ { m } \sigma _ { i } g _ { i }, \\ & & \lambda \in \mathbb { R }, \\ & & \sigma _ { 0 }, \dots, \sigma _ { m } \text{ are sums of squares,} \end{array}$$
where λ is the actual parameter, and X is a formal indeterminate.
Let ρ j be the optimal value of the related problem in which the degrees of the sums of squares are bounded:
$$\rho _ { j } = \sup \ \lambda \\ \text{subject to} \ \lambda - f ( X ) = \sigma _ { 0 } + \sum _ { i = 1 } ^ { m } \sigma _ { i } g _ { i } \\ \sigma _ { 0 } \text{ is a sum of squares of degree at most } j, \\ \sigma _ { i } \text{ is a sums of squares of degree at most } j - \deg ( g _ { i } ).$$
It is easy to see that ρ ∗ ≥ ρ j for any j . Furthermore, the following important convergence result follows directly from Putinar's positivstellensatz (we present a simplified version of Theorem 3.4 page 805 [35], and Theorem 4.1 page 79 [36]):
Theorem 4.4. Let f, g 1 , . . . , g m ∈ R [ x ] , let K = { g 1 ≥ 0 , . . . , g m ≥ 0 } be compact and assume that QM( g , . . . , g 1 m ) is Archimedean. Assume that ρ ∗ and ρ ∗ defined as above are finite and let ρ j denote the optimal value of the j th restriction defined above. Then ρ j → ρ ∗ = ρ ∗ as j →∞ .
The family of relaxations ρ j is often referred to as the Lasserre hierarchy. Let v d denote the vector of all monomials of degree less than or equal to d . σ is a sum of squares of degree 2 d if and only if it is possible to find a positive semidefinite matrix Q such that σ = v T d Qv d . It follows that ρ j can be obtained by
$$\rho _ { j } = \sup _ { \lambda, \{ Q _ { i } \} } & \lambda \\ \text{subject to} \quad & f ( X ) - \lambda = v _ { [ \frac { j } { 2 } ] } ^ { T } Q _ { 0 } v _ { [ \frac { j } { 2 } ] } + \sum _ { i = 1 } ^ { m } v _ { [ \frac { j - \deg ( g _ { i } ) } { 2 } ] } ^ { T } Q _ { i } v _ { [ \frac { j - \deg ( g _ { i } ) } { 2 } ] } ^ { \ } g _ { i } ( X ), \\ & Q _ { i } \succeq 0 \text{ for } i = 0, 1, \dots, m.$$
We rewrite the problem in order to get it on the generic form of a semidefinite program. If we let h X ( ) = ∑ α [ h X ( )] α X α , we see that
$$\text{program. If we let } n ( \Lambda ) = \sum _ { \alpha } \left | n ( \Lambda ) \right | _ { \alpha } \Lambda \overset { \lambda } { \lambda }, \, \text{we see that} \\ \rho _ { j } = \sup _ { \lambda, \left \{ Q _ { i } \right \} } \lambda \\ \text{subject to } \left [ f ( X ) - v _ { \left \{ \frac { 1 } { 2 } \right \} } ^ { T } Q _ { 0 } v _ { \left \{ \frac { 1 } { 2 } \right \} } + \sum _ { i = 1 } ^ { m } v _ { \left \{ \frac { 1 - \deg ( g _ { i } ) } { 2 } \right \} } ^ { T } \right ] _ { \alpha } g _ { i } ( X ) \right ] _ { 0 } = \lambda, \\ \left [ f ( X ) - v _ { \left \{ \frac { 1 } { 2 } \right \} } ^ { T } Q _ { 0 } v _ { \left \{ \frac { 1 } { 2 } \right \} } + \sum _ { i = 1 } ^ { m } v _ { \left \{ \frac { 1 - \deg ( g _ { i } ) } { 2 } \right \} } ^ { T } Q _ { i } v _ { \left \{ \frac { 1 - \deg ( g _ { i } ) } { 2 } \right \} } g _ { i } ( X ) \right ] _ { \alpha } = 0 \\ \text{$ for all $\alpha$,} \\ Q _ { i } \succeq 0$ for $ i = 0, 1, \dots, m. \\ \text{Since } \left [ f ( X ) - v _ { \dots } ^ { T } Q _ { 0 } v _ { \left \{ \frac { 1 } { 2 } \right \} } + \sum _ { i = 1 } ^ { m } v _ { \left \{ \frac { 1 - \deg ( g _ { i } ) } { 2 } \right \} } ^ { T } Q _ { i } v _ { \left \{ \frac { 1 - \deg ( g _ { i } ) } { 2 } \right \} } ^ { T } \right ] _ { \alpha } g _ { i } ( X ) \right ] _ { \alpha } = \lambda.$$
Since [ f ( X ) -v T glyph[floorleft] j 2 glyph[floorright] Q v 0 glyph[floorleft] j 2 glyph[floorright] + ∑ m i =1 v T glyph[floorleft] j -deg( g i ) 2 glyph[floorright] Q v i glyph[floorleft] j -deg( g i ) 2 glyph[floorright] g i ( X )] α is a linear polynomial in the entries of the matrices Q i this is indeed a semidefinite program.
Any numerical solution to the semidefinite program includes a certificate based on positive semidefinite matrices. The certificate can be translated into a sum of squares based certificate. The sum of squares based certificate attained additionally serves as a lower bound to the original polynomial optimization problem since ρ ∗ = ρ ∗ ≥ ρ j . Sum of squares based certificates are very easy to check by hand, which is one of the major advantages of this method.
## 4.2 Exploiting symmetries in semidefinite programming
This subsection can be skipped for the reader who just wants to understand the final proofs as these methods were just necessary to find the numerical solutions that allowed us to find the final certificates. The main result we review in this section is Theorem 4.9 as it is one of the strongest results when working on semidefinite programs with symmetries. We discuss the limitations of this Theorem when applied to our specific problem in Section 6. Because of these limitations we would barely gain anything by using Theorem 4.9 instead of the easier Lemma 4.8. Although the lemma is a very basic result in representation theory, the numerical calculations done in this article would not have been possible without this lemma.
Let C and A , . . . , A 1 m be real symmetric matrices and let b 1 , . . . , b m be real numbers. In this section the objective is to reduce the order of the matrices in the semidefinite programming problem
$$\max \{ \text{tr} ( C X ) \ | \ X \text{ positive semidefinite, tr} ( A _ { i } X ) = b _ { i } \text{ for } i = 1, \dots, m \}$$
when it is invariant under the actions of a group.
As in [31] and [30], we use a ∗ -representation to reduce the dimension of the matrices. For a survey on ∗ -algebras we refer to the book by Takesaki [74]. The method we use as well as other efficient methods for invariant semidefinite programs are discussed in [3]. Other important recent contributions include [29, 16, 80, 41, 44, 51].
Definition 4.5. A matrix ∗ -algebra is a collection of matrices that is closed under addition, scalar and matrix multiplication, and transposition.
Let G be a finite group, which acts on a finite set Z and let S | Z | be the group of all permutations of Z . Let h be a homomorphism h : G → S | Z | that takes any element g ∈ G to a permutation h g = h g ( ) of Z for which h gg ′ = h h g g ′ and h g -1 = h -1 g . For every permutation h g we element-wise define the corresponding permutation matrix M g ∈ { 0 1 , } | Z |×| Z | by
$$( M _ { g } ) _ { i, j } = \begin{cases} \ 1 & \text{if $h_{g}(i)=j$,} \\ 0 & \text{otherwise} \end{cases}$$
for all i, j ∈ Z . The span of these matrices is the following matrix ∗ -algebra
$$\mathcal { A } = \left \{ \sum _ { g \in G } \lambda _ { g } M _ { g } \ | \ \lambda _ { g } \in \mathbb { R } \right \}.$$
The matrices X that commute with all permutation matrices are the invariant matrices of G . The collection of all such matrices,
$$\mathcal { A } ^ { \prime } = \{ X \in \mathbb { R } ^ { n \times n } | X M = M X \text{ for all } M \in \mathcal { A } \},$$
is called the commutant of A , and it is again a ∗ -algebra. Denote the dimension of the commutant by d = dim A ′ .
Let J be the matrix of size | Z | × | Z | of ones. The commutant has a basis of { 0 1 , } -matrices E , . . . , E 1 d such that ∑ d i =1 E i = J . We form a new normalized basis by
$$B _ { i } = \frac { 1 } { \sqrt { t r ( E _ { i } ^ { T } E _ { i } ) } } E _ { i }$$
for which tr( B B T i j ) = δ i,j where δ i,j is the Kronecker delta. The multiplication parameters λ k i,j are then defined by
$$B _ { i } B _ { j } = \sum _ { k = 1 } ^ { d } \lambda _ { i, j } ^ { k } B _ { k }$$
for i, j, k = 1 , . . . , d
.
We define d × d -matrices L , . . . , L 1 d by
$$( L _ { k } ) _ { i, j } = \lambda _ { k, j } ^ { i }$$
for i, j, k = 1 , . . . , d , which span
$$\mathcal { L } = \{ \sum _ { i = 1 } ^ { d } x _ { i } L _ { i } \colon x _ { 1 }, \dots, x _ { d } \in \mathbb { R } \}.$$
Theorem 4.6 ([31]) . The linear function φ : A → ′ R d × d defined by φ B ( i ) = L i for i = 1 , . . . , d is a bijection. The linear function additionally satisfies φ XY ( ) = φ X φ Y ( ) ( ) and φ X ( T ) = φ X ( ) T for all X,Y ∈ A ′ .
Corollary 4.7 ([31]) . ∑ d i =1 x B i i is positive semidefinite if and only if ∑ d i =1 x L i i is positive semidefinite.
The next lemma shows that we can find X ∈ A ′ . It is then possible to use Corollary 4.7 to reduce the order of the semidefinite constraint.
Lemma 4.8. There is a solution X ∈ A ′ to a G -invariant semidefinite program
$$\max \{ \text{tr} ( C X ) \ | \ X \ p o s i t i v e \ s e m i d e f i n i t e, \text{tr} ( A _ { i } X ) = b _ { i } \text{ for } i = 1, \dots, m \}.$$
Proof. Let C, A , . . . , A 1 m be | Z |×| Z | matrices commuting with M g for all g ∈ G . If X is an optimal solution to the optimization problem then the group average,
X ′ = 1 | G | ∑ g ∈ G M XM g T g , is also an optimal solution: It is feasible since
$$i \ m g { g a m g } \,, \, \infty \, \infty \, \circ \, \text{$P$\nabla$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$ } \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \ \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \tau \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \, \text{$O$} \
\quad \text{used that the well-known fact that the trace is in}$$
where we have used that the well-known fact that the trace is invariant under change of basis. By the same argument tr( CX ′ ) = tr( CX ), which implies that X ′ is optimal. It is easy to see that X ′ ∈ A ′ .
All in all we get the following theorem:
Theorem 4.9 ([31]) . The semidefinite program
$$\max \{ \text{tr} ( C X ) \ | \ X \succeq 0, \text{tr} ( A _ { i } X ) = b _ { i } \text{ for } i = 1, \dots, m \}$$
$$\}$$
has a solution X = ∑ d i =1 x B i i that can be obtained by
$$\max \{ \text{tr} ( C X ) \ | \ \sum _ { i = 1 } ^ { d } x _ { i } L _ { i } \succeq 0, \text{tr} ( A _ { i } \sum _ { j = 1 } ^ { d } B _ { j } x _ { j } ) = b _ { i } \text{ for } i = 1, \dots, m \}.$$
When d is smaller than | Z | the theorem can be used to reduce the size of the linear matrix inequality, which in turn improves the computational efficiency. This allows us to solve many problems that would otherwise not be possible.
## 5 Gray code, bubble language and fixed-density necklaces
The original Gray code was used by Frank Gray in a patent filed in 1947 and granted in 1953 [22]. For a background on Gray codes we refer to Knuth's Art of Computer Programming [32]. In this chapter we present some recent development on a Gray code for fixed-density necklaces by Ruskey, Sawada and Wiliams. These new methods are useful for counting arithmetic progressions, and the author's contribution based on these methods can be found in Section 8.
To represent all subsets S ⊂ Z n of density δ = | S /n | we use binary strings; when for example n = 5, k = 3 and δ = 4 5 all different possible / S are: 11110, 11101, 11011, 10111, 01111. All these strings are just rotations of one another. In our application it is enough to consider the lexicographically smallest 01111. To exclude rotation and generate all sets of fixed density is what is in the literature called to generate all fixed-density necklaces . There is software using gray
code for listing all fixed-density necklaces in constant amortized time implemented in C [64], based on theory developed in [58] and [63]. We only cover the material useful for fixed-density necklaces, and refer to the mentioned articles for the most general forms of the theorems.
Definition 5.1. A gray code is a generation of all combinatorial possibilities in which only a constant amount of change is required to go from any binary string to the following binary string. If there in addition only is a constant change to go from the last binary string to the first binary string, the gray code is cyclic .
To understand the concept we begin with the example of generating binary n -tuples using gray code.
Denote a binary n -tuple by a n -1 a n -2 . . . a 0 where a i ∈ { 0 1 , } , and let a ∞ = n ∑ i =1 a i + 1 mod 2 denote the parity. We can generate all possible n -tuples by only changing one number a i at the time using the following algorithm:
- 1. Initiate with a i = 0 for i = 0 , . . . , n -1 and a ∞ = 1.
- 2. Visit a a 1 2 . . . a n .
- 3. If a ∞ = 1, set j = 0, otherwise let j = min( i : a i -1 = 1).
- 4. Terminate if j = n , otherwise let a j = 1 -a j and thus also a ∞ = 1 -a ∞ and return to Step 2.
For a discussion about why this generates a cyclic gray code, and for other algorithms to generate n -tuples, we refer to the material by Knuth (subsection 7.2.1.1. pages 1-39 [32]), here we just check how the algorithm works for n = 4:
Example 5.2. We run through the given algorithm with input n = 4 to generate all binary 4 -tuples.
```
Example 5.2. We run through the given a l
all binary 4-tuples.
0000, a _ { \infty } = 1, j = 0 \Rightarrow \text{switch} a _ { 0 }, a _ { \infty }
0001, a _ { \infty } = 0, j = 1 \Rightarrow \text{switch} a _ { 1 }, a _ { \infty }
0011, a _ { \infty } = 1, j = 0 \Rightarrow \text{switch} a _ { 0 }, a _ { \infty }
0010, a _ { \infty } = 0, j = 2 \Rightarrow \text{switch} a _ { 2 }, a _ { \infty }
0110, a _ { \infty } = 1, j = 0 \Rightarrow \text{switch} a _ { 0 }, a _ { \infty }
0111, a _ { \infty } = 0, j = 1 \Rightarrow \text{switch} a _ { 1 }, a _ { \infty }
0101, a _ { \infty } = 1, j = 0 \Rightarrow \text{switch} a _ { 0 }, a _ { \infty }
0100, a _ { \infty } = 0, j = 3 \Rightarrow \text{switch} a _ { 3 }, a _ { \infty }
1100, a _ { \infty } = 1, j = 0 \Rightarrow \text{switch} a _ { 0 }, a _ { \infty }
1101, a _ { \infty } = 0, j = 1 \Rightarrow \text{switch} a _ { 1 }, a _ { \infty }
1111, a _ { \infty } = 1, j = 0 \Rightarrow \text{switch} a _ { 0 }, a _ { \infty }
1110, a _ { \infty } = 0, j = 2 \Rightarrow \text{switch} a _ { 2 }, a _ { \infty }
1010, a _ { \infty } = 1, j = 0 \Rightarrow \text{switch} a _ { 0 }, a _ { \infty }
1011, a _ { \infty } = 0, j = 1 \Rightarrow \text{switch} a _ { 1 }, a _ { \infty }
1001, a _ { \infty } = 1, j = 0 \Rightarrow \text{switch} a _ { 0 }, a _ { \infty }
1000, a _ { \infty } = 0, j = 4 \Rightarrow \text{terminate.}
```
As claimed all 16 of the binary 4 -tuples has been generated, and in every step there is only 1 bit that changes. Further note that to go from 1000 to 0000 we also need to change 1 bit, hence the code in this example is cyclic.
Definition 5.3. A bubble language is a set of binary strings with either of the following properties:
- 1. The first 01 of any string can be switched to 10 to obtain another string in the set.
- 2. The first 10 of any string can be switched to 01 to obtain another string in the set.
A wide variety of families of combinatorial objects are bubble languages, we only give a few examples and one proof here, and refer to [58] for omitted proofs and further examples. We consider fixed density necklaces as they represent all possibilities of fixed-density sets in our application. The proof for the following proposition also works for the lexicographically smallest representatives of all necklaces and for Lyndon words (restricting to the aperiodic necklaces).
Proposition 5.4. The set of the lexicographically smallest representatives of fixed-density necklaces is a bubble language.
Proof. We prove that if we switch the first 10 to a 01 in any necklace we get another necklace.
glyph[negationslash]
Consider a necklace α = a 1 . . . a n . If α = 0 1 d n -d , then suppose the first 10 appears at positions j and j +1. The string α must start with a 1 . . . a i = 0 1 s t for some s, t > 0, s + = , because otherwise there would be a lexicographically t j smaller rotation of α . Note also that if we switch the first 10 to 01 we get a string β which is lexicographically smaller than α . We need to check that every rotation of β is lexicographically larger than β itself. Rotating and switching a j with a j +1 commutes, so to keep the notation simple, let us rotate first. Let r i = a . . . a i n a 1 . . . a i -1 denote the i th rotation of α , let w i denote the string we get after swapping a j with a j +1 in r i , and let us consider 4 cases:
- 1. w , . . . , w 2 j -1 : The prefix of r i is of the form 0 u 1 v with u + v = j -i +1, and is larger than the prefix of α . Since i ≤ j -1 we have u + v ≥ 2 and the swap of a j and a j +1 thus only removes a 1 in the end of the prefix in both r i and β , which does not affect the lexicographic order of the two. We conclude that β is lexicographically smaller than w , . . . , w 1 j -1 .
- 2. w j : If s > 1 then β starts with 00 and β is thus lexicographically smaller than w j , which starts with 01. If s = 1, then since α is lexicographically smaller than r j we must have a j +1 = 0 and a j +2 . . . a 2 j = 1 . Thus we get t that β starts with 01 t -1 0 whereas w j starts with 01 t +1 . It is clear that β is lexicographically smaller than w j .
- 3. w j +1 : We note that w j +1 starts with a 1 whereas β start with a 0. β is clearly lexicographically smaller than w j +1 .
- 4. w j +2 , . . . , w n : After the rotation the swap does not occur in the prefix of r i , whereas it makes the prefix of α smaller. Thus since α is lexicographically smaller than r i we can conclude that β is lexicographically smaller than w j +2 , . . . , w n .
Definition 5.5. In co-lexicographic (co-lex) order strings are sorted by increasing value of their last symbol, i.e. lexicographic order read from right to left.
The co-lex order does not provide a gray code, but it is more intuitive to define how we can recursively generate all binary strings in co-lex order. Once we have defined everything properly it is easy to introduce the cool-lex order, which is just a small modification of co-lex order and provides the gray code we are interested in.
Any binary string can be decomposed in the form 0 1 s t g where g is the suffix consisting of all remaining zeroes and ones. We denote the empty suffix by glyph[epsilon1] . We have the following recursive formula for generating all fixed density necklaces in a co-lex order [58, Page 7]:
Proposition 5.6. Let initially s = d , t = n -d and g = glyph[epsilon1] . All necklaces of length n and density d are generated recursively in co-lexicographic order by
$$\mathcal { L } ( s, t, g ) = \begin{cases} \begin{array} { c } 0 ^ { s } 1 ^ { t } g, \mathcal { L } ( s - 1, 1, 0 1 ^ { t - 1 } g ), \dots, \mathcal { L } ( s - 1, t - j, 0 1 ^ { j } g ) & \text{if } s > 0, \\ 1 ^ { t } g \end{array} \end{cases} \\ \begin{array} { c } 0 ^ { s } 1 ^ { t } g, \mathcal { L } ( s - 1, 1, 0 1 ^ { t - 1 } g ), \dots, \mathcal { L } ( s - 1, t - j, 0 1 ^ { j } g ) & \text{if } s > 0, \\ \text{if } s = 0. \end{cases}$$
where j is the minimum value such that 0 s -1 1 t -j 01 j g is a fixed density necklace.
We can visualize the recursion with a computation tree , in which every node is labeled with a suffix g from a certain instance of the recursion. We label the root by glyph[epsilon1] , and the children of any node are labeled from left to right by 01 i with i decreasing. When the suffix tree is build we add an additional leaf to each node in the tree of the form 0 s 1 t g . The computation tree for necklaces of length 8 with density 4 can be found in Figure 1. We can write it simpler by replacing the suffix at every node with the string in the leaf next to it. This simpler tree is called the compact computation tree , and the compact computation tree for necklaces of length 8 with density 4 can be found in Figure 2. Note that by construction we get the co-lex order by a pre-order traversal of the compact computation tree.
Now we are ready to define the cool-lex order, which provides a gray code. Note that the only difference in that the term 0 s 1 t g has been changed from first to last from the co-lex order.
Definition 5.7. Let initially s = d , t = n -d and g = glyph[epsilon1] . All necklaces of length n and density d are generated recursively in cool-lex order by
$$\mathcal { L } ( s, t, g ) = \begin{cases} \mathcal { L } ( s - 1, 1, 0 1 ^ { t - 1 } g ), \dots, \mathcal { L } ( s - 1, t - j, 0 1 ^ { j } g, 0 ^ { s } 1 ^ { t } g ) & i f \, s > 0, \\ 1 ^ { t } g & i f \, s = 0. \end{cases}$$
where j is the minimum value such that 0 s -1 1 t -j 01 j g is a fixed density necklace.
Figure 1: The computation tree for necklaces of length 8 with density 4.
Figure 2: The compact computation tree for necklaces of length 8 with density 4.
Since the term 0 1 s t g is placed last in the recursion we get the cool-lex order by post-order transversal of the compact computation tree.
By either traversing the tree or using the recursive formula we can get the co-lex order and cool-lex order respectively for our example:
| co-lex | cool-lex |
|----------|------------|
| 00001111 | 00100111 |
| 00010111 | 00010111 |
| 00100111 | 00101011 |
| 00011011 | 00110011 |
| 00101011 | 00011011 |
| 00110011 | 00101101 |
| 00011101 | 01010101 |
| 00101101 | 00110101 |
| 00110101 | 00011101 |
| 01010101 | 00001111 |
All we have done up to now could be done for any bubble language, not just for fixed density necklaces. By analyzing the possible steps from one binary string to the next in the compact computation tree in post-order one gets the following result (Theorem 3.1, page 10 [58]):
Theorem 5.8. Cool-lex order provide a gray code for any bubble language.
To prove that a specific bubble language can be generated in constant amortized time requires more involved analysis, and this was done in the case for fixed-density necklaces (Theorem 2 page 11, [63]):
Theorem 5.9. Fixed-density necklaces can be generated in cool-lex Gray code order or co-lex order in constant amortized time.
The methods in this section are used to obtain the results in Tables 3, 4, 5, 6, 7 and 8 in the appendix.
## 6 Problem formulated as a semidefinite program
Recall that W k,G,δ ( ) denotes the minimal number of k -arithmetic progressions in the subset S ⊆ G with | S | = δ G | | . For this article we are primarily interested in G = Z p for prime p . In the case when k = 3 we have
$$W ( 3, G, D / | G | ) = \min \{ \sum _ { \{ i, j, k \} \text{ is an A.P. in } G } x _ { i } x _ { j } x _ { k } \colon x _ { i } \in \{ 0, 1 \}, \sum _ { i = 0 } ^ { | G | - 1 } x _ { i } = D \},$$
and using the methods introduced in Sections 4.1 and 4.2 we make the following relaxation to get an algebraic certificate for the lower bound:
$$W ( 3, G, D / | G | ) \geq \max \{ \lambda \, \colon \sum _ { \{ i, j, k \} \ A. P. \text{ in } G } X _ { i } X _ { j } X _ { k } - \lambda = S, X \in K \}$$
where K = { X ∈ [0 , 1] n : ∑ | G |-1 i =0 X i = D } and S = ∑ i S g 2 i i for S , g i i ∈ R [ X ,. . . , X 1 k ] where g i are the half-spaces defining K .
Let us use the degree 3 relaxation of Putinar's Positivstellensatz, and let the maximal lower bound using this relaxation be denoted λ ∗ . Denote the elements of G by g , . . . , g 1 | G | and let v = [1 , X g 1 , . . . , X g | G | ] T be the vector of all monomials of degree less or equal to one. We get
$$\lambda ^ { * } = \max \quad \lambda \\ \text{subject to} \colon \quad & p _ { G } - \lambda = v ^ { T } Q _ { 0 } v + \sum _ { g \in G } v ^ { T } Q _ { g } ^ { + } v X _ { g } + \sum _ { g \in G } v ^ { T } Q _ { g } ^ { - } v ( 1 - X _ { g } ) \\ & + v ^ { T } Q _ { s } ^ { + } v ( \sum _ { g \in G } X _ { g } - D ) + v ^ { T } Q _ { s } ^ { - } v ( D - \sum _ { g \in G } X _ { g } ), \\ & X _ { G } \in [ 0, 1 ] ^ { | G | }, \\ & Q _ { 0 }, Q _ { s } ^ { + }, Q _ { s } ^ { - }, Q _ { g } ^ { + }, Q _ { g } ^ { - } \succeq 0 \text{ for all } g \in G. \\ \text{This optimization problem in simulation to a problem on the form of } / 1 \text{ } \text{ that}$$
This optimization problem is equivalent to a problem on the form of (1), that is if we use the notation h x ( ) = ∑ α [ h x ( )] α x α then:
$$\text{is if we use the notation } h ( x ) = \sum _ { \alpha } [ h ( x ) ] _ { \alpha } x ^ { \alpha } \text{ then} \colon \\ \lambda ^ { * } = \max \ \lambda \\ \text{subject to } \ [ p _ { G } - v ^ { T } Q _ { 0 } v + \sum _ { g \in G } v ^ { T } Q _ { g } ^ { + } v X _ { g } + \sum _ { g \in G } v ^ { T } Q _ { g } ^ { - } v ( 1 - X _ { g } ) \\ \quad + v ^ { T } Q _ { s } ^ { + } v ( \sum _ { g \in G } x _ { g } - D ) + v ^ { T } Q _ { s } ^ { - } v ( D - \sum _ { g \in G } X _ { g } ) ] _ { 0 } = \lambda, \\ \quad \ [ p _ { G } - v ^ { T } Q _ { 0 } v + \sum _ { g \in G } v ^ { T } Q _ { g } ^ { + } v X _ { g } + \sum _ { g \in G } v ^ { T } Q _ { g } ^ { - } v ( 1 - X _ { g } ) \\ \quad + v ^ { T } Q _ { s } ^ { + } v ( \sum _ { g \in G } X _ { g } - D ) + v ^ { T } Q _ { s } ^ { - } v ( D - \sum _ { g \in G } X _ { g } ) ] _ { \alpha } = 0, \\ \quad \ X _ { G } \in [ 0, 1 ] ^ { | G | }, \\ \quad \ Q _ { 0 }, Q _ { s } ^ { + }, Q _ { s } ^ { - }, Q _ { g } ^ { + }, Q _ { g } ^ { - } \succeq 0 \text{ for all } g \in G. \\ \text{Since arithmetic progressions are invariant under affine transformations we can } \lambda \cdot \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \dots \ \delta \end{$$
Since arithmetic progressions are invariant under affine transformations we can find a solution to the problem by restricting to invariant solutions. If g 1 is the
identity element and v g = [1 , X g 1 + g , . . . , X g | G | + g ] we get
$$\text{identity element and } v _ { g } = [ 1, X _ { g _ { 1 } + g }, \dots, X _ { g _ { l G _ { 1 } + g } } ] \, \text{we get} \\ \lambda ^ { * } = \max \ \lambda \\ \text{subject to} \ \ [ p _ { G } - v ^ { T } Q _ { 0 } v + \sum _ { g \in G } v _ { g } ^ { T } Q _ { 1 } ^ { + } v _ { g } X _ { g } + \sum _ { g \in G } v _ { g } ^ { T } Q _ { 1 } ^ { - } v _ { g } ( 1 - X _ { g } ) \\ + v ^ { T } Q _ { 3 } ^ { + } v ( \sum _ { g \in G } X _ { g } - D ) + v ^ { T } Q _ { 3 } ^ { - } v ( D - \sum _ { g \in G } X _ { g } ) ] _ { 0 } = \lambda, \\ [ p _ { G } - v ^ { T } Q _ { 0 } v + \sum _ { g \in G } v _ { g } ^ { T } Q _ { 1 } ^ { + } v _ { g } ( 1 + X _ { g } ) + \sum _ { g \in G } v _ { g } ^ { T } Q _ { 1 } ^ { - } v _ { g } ( 1 - X _ { g } ) \\ + v ^ { T } Q _ { 3 } ^ { + } v ( \sum _ { g \in G } X _ { g } - D ) + v ^ { T } Q _ { 3 } ^ { - } v ( D - \sum _ { g \in G } X _ { g } ) ] _ { \alpha } = 0, \\ X _ { c } \in [ 0, 1 ] | ^ { G | } \\ Q _ { 0 } ; Q _ { 3 } ^ { + }, Q _ { 3 } ^ { - }, Q _ { 3 } ^ { + }, Q _ { 3 } ^ { - } _ { g _ { 1 } } \succeq 0, \\ Q _ { 0 } ( g _ { i }, g _ { j } ) = Q _ { 0 } ( a + b g _ { i }, a + b _ { j } ), \forall \ ( a, b ) \in G \rtimes \mathbb { Z } ^ { + }, g _ { i }, g _ { j } \in G, \\ Q _ { 0 } ( g, 1 ) = Q _ { 0 } ( 1, g ) = Q _ { 0 } ( g _ { 1 }, 1 ), \forall \ g \in G, \\ Q _ { 3 } ^ { + } ( g _ { i }, g _ { j } ) = Q _ { 3 } ^ { + } ( a + b g _ { i }, a + b g _ { j } ), \forall \ ( a, b ) \in G \rtimes \mathbb { Z } ^ { + }, g _ { i }, g _ { j } \in G, \\ Q _ { 3 } ^ { + } ( g, 1 ) = Q _ { 3 } ^ { + } ( 1, g ) = Q _ { 3 } ^ { + } ( g _ { 1 }, 1 ), \forall \ g \in G, \\ Q _ { 3 } ^ { - } ( g _ { i }, g _ { j } ) = Q _ { 3 } ^ { - } ( a + b g _ { i }, a + b g _ { j } ), \forall \ ( a, b ) \in G \rtimes \mathbb { Z } ^ { + }, g _ { i }, g _ { j } \in G, \\ Q _ { 3 } ^ { - } ( g, 1 ) = Q _ { 3 } ^ { - } ( 1, g ) = Q _ { 3 } ^ { - } ( g _ { 1 }, 1 ), \forall \ g \in G, \\ Q _ { 3 } ^ { + } _ { g _ { 1 } } ( g _ { j }, g _ { k } ) = Q _ { 3 } ^ { + } ( b g _ { j }, b g _ { k } ), \forall \ ( a, b ) \in G \rtimes \mathbb { Z } ^ { + }, g _ { i }, g _ { j } \in G, \\ Q _ { 3 } ^ { + } _ { g _ { 1 } } ( g, 1 ) = Q _ { 3 } ^ { + } ( b g, 1 ), \forall \ ( a, b ) \in G \rtimes \mathbb { Z } ^ { + }, g \in G, \\ Q _ { 3 } ^ { + } _ { g _ { 1 } } ( 1, g ) = Q _ { 3 } ^ { + } _ { g _ { 1 } } ( 1, b g ), \forall \ ( a, b ) \in G \rtimes \mathbb { Z } ^ { + }, g \in G. \\ \text{We see that this reduces the number of variables significantly and in this new} \\ \text{formulation we only need to require that five |G|+1x|G|+1-matrices are positive} \\ \text{semidefinite instead of the 2|G|+3 matrices required in the original formulation.} \\ \text{One could hope to reduce the size further using Theorem 4.9 if the commutant}$$
We see that this reduces the number of variables significantly and in this new formulation we only need to require that five | G | +1 ×| G | +1-matrices are positive semidefinite instead of the 2 | G | +3 matrices required in the original formulation. One could hope to reduce the size further using Theorem 4.9 if the commutant is small. Unfortunately though, one cannot hope to get reduce the size of the Q + s -matrices significantly since there will be one basis element of the commutant for the roughly n/ 2 pairs of 3-tuples ( g , g , g 1 i k ) and ( g , g 1 i -g k + g , g i i ), and these matrices turns out to be the most essential to be able to find λ ∗ . We have solved this semidefinite program to find lower bounds to W (3 , Z p , D/p ) for all primes 5 ≤ p ≤ 613 and D ∈ { 0 , 1 20 , 2 20 , . . . , p -1 20 , p } using the Matlab-based software CVX [15] on the Triton computer cluster, which is part of the Aalto Science-IT project. The code was parallelized and we ran it simultaneously on 100 computer nodes. All the data was generated in 5 days. The data helped us get the intuition needed to prove the theorems in this section. The theoretically most interesting piece of this data can be found in Figure 7 in Section 8. In that section the data is also discussed further.
To achieve better bounds we make a degree 5 relaxation of Putinar's Positivestellensatz. The SDP is the same as in the degree 3 relaxation above with the difference that v is the vector of all monomials of degree up to 2, and thus the Q -matrices are of size | G | 2 + | G | +1 ×| G | 2 + | G | +1 and many more equalities.
Just to get the equalities from the Q -matrices required 2000 lines of code, which was significantly more than in the degree 3 relaxation. It is not obvious to the author how the code could be simplified significantly as there are so many different cases to consider. Using the same computer cluster we found W k, ( Z p , D/p ) for all primes 5 ≤ p ≤ 19, k ∈ { 3 4 5 , , } and D ∈ { 0 , 1 20 , 2 20 , . . . , p -1 20 , p } by running parallelized code simultaneously on 100 computer nodes for 7 days. Again, the most interesting piece of data can be found in Figure 7 in Section 8.
Since
$$\sum _ { g \in G } x _ { g } = D$$
on { 0 1 , } | G | it follows that
$$\sum _ { g \in G } x _ { g } ^ { 3 } - D = 0,$$
glyph[negationslash]
$$\sum _ { g \neq h \in G } x _ { g } ^ { 2 } x _ { h } - D ( D - 1 ) = 0$$
$$\sum _ { g \neq h \neq l \in G } x _ { g } x _ { h } x _ { l } - \binom { D } { 3 } = 0$$
and glyph[negationslash]
glyph[negationslash]
on { 0 1 , } | G | , providing several new possibilities for polynomials that are nonnegative on { 0 1 , } | G | . Note that these equation holds over the discrete hypercube, but not necessarily on the continuous hypercube, so technically they have to be added before making the relaxation. Similar assertion also holds for higher order polynomials. These redundant conditions simplifies the algebraic certificates we get from Putinar's Positivstellensatz, and replacing some of the original g i s to these new constraints can in many cases simplify the computations significantly. By using these conditions and setting a lot of variables to zero we can make all polynomials homogeneous up to a constant and get lower bounds to λ ∗ . This is the main idea behind Theorem 2.3. As adding seemingly redundant conditions on this form gives more freedom in the semidefinite relaxation one gets from Putinar's Positivstellensatz we believe that degree 4 and 5 version of this theorem can be found using a similar trick. Numerical results suggests that a degree 4 and 5 certificate for a lower bound to the number of arithmetic progressions of length 3 would have many similar terms with certificates bounding arithmetic progressions of length 4 and 5 respectively.
## 7 Proofs of Theorems 2.1, 2.2, 2.3 and Corollary 2.4
To simplify notation in the proofs of this section, let glyph[negationslash]
$$\omega ( a ^ { 0 }, a ^ { 3 }, a ^ { 2 1 }, a ^ { 1 1 1 } _ { 1 }, \dots, a ^ { 1 1 1 } _ { d } ) = \ & a ^ { 0 } + a ^ { 3 } \sum _ { i = 0 } ^ { p - 1 } X _ { i } ^ { 3 } + a ^ { 2 1 } \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } + \\ & + \sum _ { i = 1 } ^ { d } a ^ { 1 1 1 } _ { i } \sum _ { ( i _ { 1 }, i _ { 2 }, i _ { 3 } ) \in o r b _ { i } } X _ { i _ { 1 } } X _ { i _ { 2 } } X _ { i _ { 3 } }$$
where d is the number of orbits orb i of the polynomials on the form X X X i 1 i 2 i 3 s under the action { i 1 , i 2 , i 3 } ↦→ { ai 1 + b, ai 2 + b, ai 3 + b } with ( b, a ) ∈ Z p × ( Z glyph[integerdivide] p { 0 } ).
Proof of Theorem 2.1. All we have to check is that the algebraic certificate glyph[negationslash]
$$\sum _ { \{ i, j, k \} \ A. P. \ \text{ in } \mathbb { Z } _ { p } } X _ { i } X _ { j } X _ { k } - \lambda = \sum _ { i = 0 } ^ { p - 1 } \sigma _ { 1, i } X _ { i } + \sum _ { i = 0 } ^ { p - 1 } \sigma _ { 2, i } X _ { i } + \sigma _ { 3 } ( D - \sum _ { i = 0 } ^ { p - 1 } X _ { i } ^ { 3 } ) \\ + \sigma _ { 4 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = S,$$
where
$$\text{where} \\ \sigma _ { 1, i } \ & = \frac { 1 } { p - 1 } \sum _ { 0 < j < k < ( p - 1 ) / 2 } \left ( X _ { j + i } - X _ { j + k + i } - X _ { n - j - k + i } + X _ { n - j + i } \right ) ^ { 2 } \\ \sigma _ { 2, i } \ & = \frac { 1 } { p - 1 } ( D X _ { i } - \sum _ { j = 0 } ^ { p - 1 } X _ { j } ) ^ { 2 } \\ \sigma _ { 3 } \ & = \frac { ( D - 1 ) ^ { 2 } } { 4 \beta - p + 3 } \\ \sigma _ { 4 } \ & = \frac { 4 \beta - p + 3 } { 2 ( p - 1 ) } \\ \text{is correct. We have:}$$
is correct. We have:
$$\text{correct. We have} \colon \\ \sum _ { i = 0 } ^ { p - 1 } \sigma _ { 1, i } X _ { i } = \omega ( 0, 0, \frac { ( p - 3 ) } { p - 1 }, \frac { p - 7 } { p - 1 }, - \frac { 6 } { p - 1 }, \dots, - \frac { 6 } { p - 1 } ), \\ \sum _ { i = 0 } ^ { p - 1 } \sigma _ { 2, i } X _ { i } = \omega ( 0, \frac { ( D - 1 ) ^ { 2 } } { p - 1 }, \frac { 3 - 2 D } { p - 1 }, \frac { 6 } { p - 1 }, \dots, \frac { 6 } { p - 1 } ), \\ \sigma _ { 3 } ( D - \sum _ { i = 0 } ^ { p - 1 } X _ { i } ^ { 3 } ) = \omega ( D \frac { ( D - 1 ) ^ { 2 } } { p - 1 }, - \frac { ( D - 1 ) ^ { 2 } } { p - 1 }, 0, \dots, 0 )$$
and glyph[negationslash]
$$\sigma _ { 4 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = \omega ( - D ( D - 1 ) \frac { 2 D - ( \frac { p + 3 } { 2 } ) } { p - 1 }, 0, \frac { 2 D - ( \frac { p + 3 } { 2 } ) } { p - 1 }, 0, \dots, 0 ).$$
Summing up we get
$$\cos _ { \substack { p } \psi } \psi & = \omega ( D \frac { ( D - 1 ) ^ { 2 } } { p - 1 } - D ( D - 1 ) \frac { 2 D - ( \frac { p + 3 } { 2 } ) } { p - 1 }, 0, 0, 1, 0, \dots, 0 ) \\ & = \omega ( - \frac { D ^ { 3 } - ( \frac { p + 3 } { 2 } ) D ^ { 2 } + ( \frac { p + 3 } { 2 } - 1 ) D } { p - 1 }, 0, 0, 1, 0, \dots, 0 ) \\ & = \sum _ { \{ i, j, k \} } X _ { i } X _ { j } X _ { k } - \lambda. \\ & \quad \sum _ { \{ i, j, k \} } \ A. P. \text{ in } Z _ { p } \\ & \quad \ \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\end{array}$$
Proof of Theorem 2.2. An algebraic certificate proving that
$$W ( 3, \mathbb { Z } _ { 5 }, 2 ) \geq \frac { D ^ { 3 } - 3 D ^ { 2 } + 2 D } { 6 }$$
is the following:
glyph[negationslash]
$$\sum _ { \{ i, j, k \} \ A. P. \text{ in } Z _ { 5 } } X _ { i } X _ { j } X _ { k } - \lambda = \sum _ { i = 0 } ^ { 4 } \sigma _ { 1, i } X _ { i } + \sigma _ { 2 } ( D - \sum _ { i = 0 } ^ { 4 } X _ { i } ^ { 3 } ) \\ + \sigma _ { 3 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = S _ { 5 }$$
where
$$\sigma _ { 1, i } = \frac { 1 } { 6 } ( D X _ { i } - \sum _ { j = 0 } ^ { 4 } X _ { j } ) ^ { 2 },$$
$$\sigma _ { 2 } = \frac { D ^ { 2 } - 2 D + 1 } { 6 },$$
$$\sigma _ { 3 } = \frac { 2 D - 3 } { 6 }$$
$$& \frac { 1 - 2 } { j = 0 } \Lambda _ { j } \rangle \\ & \frac { 1 } { 6 } \frac { 1 } { 6 }, \\ & \frac { D - 3 } { 6 }$$
and
$$\lambda = \frac { D ^ { 3 } - 3 D ^ { 2 } + 2 D } { 6 }.$$
To check that everything adds up, let us use the ω -notation:
$$\sum _ { i = 0 } ^ { 4 } \sigma _ { 1, i } X _ { i } = \omega ( 0, \frac { ( D - 1 ) ^ { 2 } } { 6 }, - \frac { 2 D - 3 } { 6 }, 1 )$$
$$\sigma _ { 2 } ( D - \sum _ { i = 0 } ^ { 4 } X _ { i } ^ { 3 } ) = \omega ( \frac { D ( D - 1 ) ^ { 2 } } { 6 }, - \frac { ( D - 1 ) ^ { 2 } } { 6 }, 0, 0 )$$
glyph[negationslash]
$$\sigma _ { 3 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = \omega ( - \frac { D ( D - 1 ) ( 2 D - 3 ) } { 6 }, 0, \frac { 2 D - 3 } { 6 }, 0 ).$$
Summing up yields
$$\dots \circ \psi _ { r } \psi _ { \nu } \dots \\ S _ { 5 } & \ = \omega ( \frac { D ( D - 1 ) ^ { 2 } } { 6 } - \frac { D ( D - 1 ) ( 2 D - 3 ) } { 6 }, 0, 0, 1 ) \\ & = \omega ( - \frac { D ^ { 3 } - 3 D ^ { 2 } + 2 D } { 6 }, 0, 0, 1 ) \\ & = \sum _ { \{ i, j, k \} } X _ { i } X _ { j } X _ { k } - \lambda. \\ \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot$$
An algebraic certificate proving that
$$W ( 3, \mathbb { Z } _ { 7 }, 2 ) \geq \frac { D ^ { 3 } - 4 D ^ { 2 } + 3 D } { 8 }$$
is the following:
glyph[negationslash]
$$\sum _ { \{ i, j, k \} \ A. P. \ \in \ \mathbb { Z } _ { 7 } } X _ { i } X _ { j } X _ { k } - \lambda = \ \sum _ { i = 0 } ^ { 6 } \sigma _ { 1, i } X _ { i } + \sigma _ { 2 } ( D - \sum _ { i = 0 } ^ { 6 } X _ { i } ^ { 3 } ) + \sum _ { i = 0 } ^ { 6 } \sigma _ { 4, i } X _ { i } \\ + \sigma _ { 3 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = S _ { 7 }$$
where
$$\sigma _ { 1, i } = \frac { 1 } { 8 } ( D X _ { i } - \sum _ { j = 0 } ^ { 6 } X _ { j } ) ^ { 2 },$$
$$\sigma _ { 2 } = \frac { D ^ { 2 } - 2 D + 1 } { 8 },$$
$$\sigma _ { 3 } = \frac { D - 2 } { 4 }$$
$$& \frac { \mu 2 D + 1 } { 8 }, \\ & \frac { D - 2 } { 4 }$$
$$\sigma _ { 4, i } = \frac { 1 } { 8 } ( X _ { i + 1 } + X _ { i + 2 } - X _ { i + 3 } + X _ { i + 4 } - X _ { i + 5 } - X _ { i + 6 } ) ^ { 2 }$$
and
$$\lambda = \frac { D ^ { 3 } - 4 D ^ { 2 } + 3 D } { 8 }.$$
To check that everything adds up, let us use the ω -notation:
$$\sum _ { i = 0 } ^ { 6 } \sigma _ { 1, i } X _ { i } = \omega ( 0, \frac { ( D - 1 ) ^ { 2 } } { 8 }, - \frac { 2 D - 3 } { 8 }, \frac { 3 } { 4 }, \frac { 3 } { 4 } )$$
$$\sum _ { i = 0 } ^ { 6 } \sigma _ { 1, i } X _ { i } = \omega ( 0, \frac { ( D - 1 ) ^ { 2 } } { 8 }, - \frac { 2 D - 3 } { 8 }, \frac { 3 } { 4 }, \frac { 3 } { 4 } ) \\ \sigma _ { 2 } ( D - \sum _ { i = 0 } ^ { 6 } X _ { i } ^ { 3 } ) = \omega ( \frac { D ( D - 1 ) ^ { 2 } } { 8 }, - \frac { ( D - 1 ) ^ { 2 } } { 8 }, 0, 0, 0 ) \\ \sigma _ { 3 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = \omega ( - \frac { D ( D - 1 ) ( D - 2 ) } { 4 }, 0, \frac { D - 2 } { 4 }, 0, 0 )$$
glyph[negationslash]
$$\sigma _ { 3 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = \omega ( - \frac { D ( D - 1 ) ( D - 2 ) } { 4 }, 0, \frac { D - 2 } { 4 }, 0, 0 )$$
$$\sum _ { i = 0 } ^ { 6 } \sigma _ { 4, i } X _ { i } = \omega ( 0, 0, \frac { 1 } { 8 }, \frac { 1 } { 4 }, - \frac { 3 } { 4 } )$$
Summing up yields
$$\begin{array} { l l } \hookrightarrow & \mathcal { k } \\ S _ { 7 } & = \omega ( \frac { D ( D - 1 ) ^ { 2 } } { 8 } - \frac { D ( D - 1 ) ( D - 2 ) } { 4 }, 0, 0, 1, 0 ) \\ & = \omega ( - \frac { D ^ { 3 } - 4 D ^ { 2 } + 3 D } { 8 }, 0, 0, 1, 0 ) \\ & = \sum _ { \{ i, j, k \} \ A. P. \text{ in } \mathbb { Z } _ { 7 } } X _ { i } X _ { j } X _ { k } - \lambda. \\. &. &. &. &. \end{array}$$
An algebraic certificate proving that
$$\lambda _ { \quad W ( 3, \mathbb { Z } _ { 1 1 }, 2 ) } \geq \frac { \cdot \cdot \cdot } { \sqrt { 5 } D ^ { 3 } + ( 1 5 - 1 2 \sqrt { 5 } ) D ^ { 2 } + ( - 1 5 + 1 1 \sqrt { 5 } ) D } { 3 0 }$$
is the following:
glyph[negationslash]
$$\sum _ { \{ i, j, k \} \ A. P. \ \text{ in } Z _ { 1 1 } } X _ { i } X _ { j } X _ { k } - \lambda = \sum _ { \begin{array} { c } i = 0 \\ + \sigma _ { 3 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = S _ { 1 1 } \end{array} } ^ { 1 0 } \sigma _ { 4, i } X _ { i }$$
where
$$\sigma _ { 1, i } = \frac { \sqrt { 5 } } { 3 0 } ( D X _ { i } - \sum _ { j = 0 } ^ { 1 0 } X _ { j } ) ^ { 2 },$$
$$\sigma _ { 2 } = \frac { \sqrt { 5 } ( D - 1 ) ^ { 2 } } { 3 0 }, \\ -$$
$$\sigma _ { 3 } = \frac { 2 \sqrt { 5 } D + 1 5 - 1 2 \sqrt { 5 } } { 3 0 }$$
$$\sigma _ { 4, i } = \frac { 9 \sqrt { 5 } - 1 5 } { 3 0 } ( ( \sum _ { j = 0 } ^ { 9 } \cos ( 4 \pi j / 1 0 ) X _ { 2 ^ { i + j } + 1 } ) ^ { 2 } + ( \sum _ { j = 0 } ^ { 9 } \sin ( 4 \pi j / 1 0 ) X _ { 2 ^ { i + j } + 1 } ) ^ { 2 } )$$
and
$$\lambda = \frac { \sqrt { 5 } D ^ { 3 } + ( 1 5 - 1 2 \sqrt { 5 } ) D ^ { 2 } + ( - 1 5 + 1 1 \sqrt { 5 } ) D } { 3 0 }.$$
To check that everything adds up, let us use the ω -notation:
$$\sum _ { i = 0 } ^ { 1 0 } \sigma _ { 1, i } X _ { i } = \omega ( 0, \sqrt { 5 } \frac { ( D - 1 ) ^ { 2 } } { 3 0 }, - \frac { \sqrt { 5 } ( 2 D - 3 ) } { 3 0 }, \frac { \sqrt { 5 } } { 5 }, \frac { \sqrt { 5 } } { 5 } )$$
$$\sigma _ { 2 } ( D - \sum _ { i = 0 } ^ { 1 0 } X _ { i } ^ { 3 } ) = \omega ( \frac { \sqrt { 5 } D ( D - 1 ) ^ { 2 } } { 3 0 }, - \frac { \sqrt { 5 } ( D - 1 ) ^ { 2 } } { 3 0 }, 0, 0, 0 )$$
glyph[negationslash]
$$\sigma _ { 3 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = \omega ( \quad - \frac { D ( D - 1 ) ( 2 \sqrt { 5 } D + 1 5 - 1 2 \sqrt { 5 } ) } { 3 0 }, 0, \\ \frac { 2 \sqrt { 5 } D + 1 5 - 1 2 \sqrt { 5 } } { 3 0 }, 0, 0 ) \\ 1 0 \quad \sim \dots \quad \sim \dots \quad \neq \dots \quad \neq \dots$$
$$\sum _ { i = 0 } ^ { 1 0 } \sigma _ { 4, i } X _ { i } = \omega ( 0, 0, \frac { 9 \sqrt { 5 } - 1 5 } { 3 0 }, 1 - \frac { \sqrt { 5 } } { 5 }, - \frac { \sqrt { 5 } } { 5 } )$$
Summing up yields
$$S _ { 1 1 } \ & = \omega ( \frac { \sqrt { 5 } D ( D - 1 ) ^ { 2 } } { 3 0 } - \frac { D ( D - 1 ) ( 2 \sqrt { 5 } D + 1 5 - 1 2 \sqrt { 5 } ) } { 3 0 }, 0, 1, 0 ) \\ & = \omega ( - \frac { \sqrt { 5 } D ^ { 3 } + ( 1 5 - 1 2 \sqrt { 5 } ) D ^ { 2 } + ( - 1 5 + 1 1 \sqrt { 5 } ) D } { 3 0 }, 0, 0, 1, 0 ) \\ & = \sum _ { \{ i, j, k \} } \ A. P. \ \text{in $Z_{1}$} } X _ { i } X _ { j } X _ { k } - \lambda. \\ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \. \ \cdot \.$$
An algebraic certificate proving that
$$\lambda _ { W ( 3, \mathbb { Z } _ { 1 3 }, 2 ) } \geq \frac { \lambda _ { \lambda } } { 2 8 6 } D ^ { 3 } + \$$
is the following:
glyph[negationslash]
$$\sum _ { \{ i, j, k \} \ A. P. \ \in \ Z _ { 1 3 } } X _ { i } X _ { j } X _ { k } - \lambda = \$$
where
$$\sigma _ { 1, i } = \frac { 2 1 - 2 \sqrt { 3 } } { 2 8 6 } ( D X _ { i } - \sum _ { j = 0 } ^ { 1 2$$
$$\sigma _ { 2 } = \frac { ( 2 1 - 2 \sqrt { 3 } ) ( D - 1 ) ^ { 2 } } { 2 8 6 }, \\ - \, \, \,$$
$$\sigma _ { 3 } = \frac { 2 ( 2 1 - 2 \sqrt { 3 } ) D + 2 8 \sqrt { 3 } - 1 5 1 } { 2 8 6 }$$
$$\sigma _ { 4, i } & = \quad \frac { 2 3 - 9 \sqrt { 3 } } { 2 8 6 } ( ( \sum _ { j = 0 } ^ { 1 1 } \cos ( 4 \pi j / 1 2 ) X _ { 2 ^ { i + j } + 1 } ) ^ { 2 } + X ( \sum _ { j = 0 } ^ { 1 1 } \sin ( 4 \pi j / 1 2 ) X _ { 2 ^ { i + j } + 1 } ) ^ { 2 } ) \\ & \quad + \frac { 5 - \sqrt { 3 } } { 2 2 } ( ( \sum _ { j = 0 } ^ { 1 1 } \cos ( 2 \pi j / 1 2 ) X _ { 2 ^ { i + j } + 1 } ) ^ { 2 } + ( \sum _ { j = 0 } ^ { 1 1 } \sin ( 2 \pi j / 1 2 ) X _ { 2 ^ { i + j } + 1 } ) ^ { 2 } ) \\ \dots$$
and
$$\lambda = \frac { 2 1 - 2 \sqrt { 3 } } { 2 8 6 } D ^ { 3 } + \frac { 2 8 \sqrt { 3 } - 1 5 1 } { 2 8 6 } D ^ { 2 } + \frac { 5 - \sqrt { 3 } } { 1 1 } D.$$
To check that everything adds up, let us use the ω -notation:
$$\sum _ { i = 0 } ^ { 1 2 } \sigma _ { 1, i } X _ { i } = \omega ( \quad 0, ( 2 1 - 2 \sqrt { 3 } ) \frac { ( D - 1 ) ^ { 2 } } { 2 8 6 }, - \frac { ( 2 1 - 2 \sqrt { 3 } ) ( 2 D - 3 ) } { 2 8 6 }, \\ 6 \frac { 2 1 - 2 \sqrt { 3 } } { 2 8 6 }, 6 \frac { 2 1 - 2 \sqrt { 3 } } { 2 8 6 } ) \\ \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
$$\sigma _ { 2 } ( D - \sum _ { i = 0 } ^ { 1 2 } X _ { i } ^ { 3 } ) = \omega ( \frac { ( 2 1 - 2 \sqrt { 3 } ) D ( D - 1 ) ^ { 2 } } { 2 8 6 }, - \frac { ( 2 1 - 2 \sqrt { 3 } ) ( D - 1 ) ^ { 2 } } { 2 8 6 }, 0, 0, 0, 0 )$$
glyph[negationslash]
$$\frac { ( 4 2 - 4 \sqrt { 3 } ) D + 2 8 \sqrt { 3 } - 1 5 1 } { 2 8 6 }, 0, 0, 0 )$$
$$\sigma _ { 3 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = \omega ( \begin{array} { c c } i = 0 \\ - \frac { D ( D - 1 ) ( ( 4 2 - 4 \sqrt { 3 } ) D + 2 8 \sqrt { 3 } - 1 5 1 ) } { 2 8 6 }, 0, \\ ( 4 2 - 4 \sqrt { 3 } ) D + 2 8 \sqrt { 3 } - 1 5 1 \\ - \frac { 1 } { 2 } \. \end{array}.$$
$$\sum _ { i = 0 } ^ { 1 2 } \sigma _ { 4, i } X _ { i } = \omega ( 0, 0, \frac { 8 8 - 2 2 \sqrt { 3 } } { 2 8 6 }, \frac { 8 0 + 6 \sqrt { 3 } } { 1 4 3 }, - 6 \frac { 2 1 - 2 \sqrt { 3 } } { 2 8 6 }, - 6 \frac { 2 1 - 2 \sqrt { 3 } } { 2 8 6 } )$$
Summing up yields
$$\text{unmimg up } y \text{two} \\ S _ { 1 3 } \ & = \omega ( - \frac { 2 1 - 2 \sqrt { 3 } } { 2 8 6 } D ^ { 3 } - \frac { 2 8 \sqrt { 3 } - 1 5 1 } { 2 8 6 } D ^ { 2 } - \frac { 5 - \sqrt { 3 } } { 1 1 } D, 0, 0, 1, 0, 0 ) \\ & = \sum _ { \{ i, j, k \} \ A. P. \text{ in } Z _ { 1 3 } } X _ { i } X _ { j } X _ { k } - \lambda.$$
An algebraic certificate proving that
$$W ( 3, \mathbb { Z } _ { 1 7 }, 2 ) \geq \frac { 1 } { 2 4 } D ^ { 3 } - \frac { 1 } { 4 } D ^ { 2 } + \frac { 5 } { 2 4 } D$$
is the following:
glyph[negationslash]
$$\sum _ { \{ i, j, k \} \ A. P. \ \text{ in } Z _ { 1 7 } } X _ { i } X _ { j } X _ { k } - \lambda = \sum _ { i = 0 } ^ { 1 6 } \sigma _ { 1, i } X _ { i } + \sigma _ { 2 } ( D - \sum _ { i = 0 } ^ { 1 6 } X _ { i } ^ { 3 } ) + \sum _ { i = 0 } ^ { 1 6 } \sigma _ { 4, i } X _ { i } \\ + \sigma _ { 3 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = S _ { 1 7 }$$
where
$$\sigma _ { 1, i } = \frac { 1 } { 2 4 } ( D X _ { i } - \sum _ { j = 0 } ^ { 1 6 } X _ { j } ) ^ { 2 }, \\ \sigma _ { 2 } = \frac { ( D - 1 ) ^ { 2 } } { 2 4 }, \\ \sigma _ { 3 } = \frac { 2 D - 6 } { 2 4 }$$
where
$$S = \sum _ { i = 0 } ^ { p - 1 } \sum _ { j = 0 } ^ { p - 1 } ( \sum _ { k = 1 } ^ { p - 1 } a _ { i j k } X _ { k } ) ^ { 2 } X _ { i } + b \sum _ { i = 0 } ^ { p - 1 } ( D X _ { i } - \sum _ { j = 0 } ^ { p - 1 } X _ { j } ) ^ { 2 } X _ { i } \\ + c ( \sum _ { i = 0 } ^ { p - 1 } X _ { i } ^ { 3 } - D ) + d ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) )$$
$$\sigma _ { 4, i } = \frac { 1 } { 8 } ( ( \sum _ { j = 0 } ^ { 1 6 } ( - 1 ) ^ { j } X _ { 3 ^ { i + j } + 1 } ) ^ { 2 }$$
and
$$\lambda = \frac { 1 } { 2 4 } D ^ { 3 } - \frac { 1 } { 4 } D ^ { 2 } + \frac { 5 } { 2 4 } D.$$
To check that everything adds up, let us use the ω -notation:
$$\sum _ { i = 0 } ^ { 1 6 } \sigma _ { 1, i } X _ { i } = \omega ( 0, \frac { ( D - 1 ) ^ { 2 } } { 2 4 }, - \frac { ( 2 D - 3 ) } { 2 4 }, \frac { 1 } { 4 }, \frac { 1 } { 4 }, \frac { 1 } { 4 } )$$
glyph[negationslash]
$$\sigma _ { 3 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = \omega ( - \frac { D ( D - 1 ) ( 2 D - 6 ) } { 2 4 }, 0, \frac { 2 D - 6 } { 2 4 }, 0, 0, 0 )$$
$$\sum _ { i = 0 } ^ { 1 6 } \sigma _ { 4, i } X _ { i } = \omega ( 0, 0, \frac { 1 } { 8 }, \frac { 3 } { 4 }, - \frac { 1 } { 4 }, - \frac { 1 } { 4 } )$$
$$\sum _ { i = 0 } ^ { 1 6 } \sigma _ { 1, i } X _ { i } = \omega ( 0, \frac { ( D - 1 ) ^ { 2 } } { 2 4 }, - \frac { ( 2 D - 3 ) } { 2 4 }, \frac { 1 } { 4 }, \frac { 1 } { 4 }, \frac { 1 } { 4 } ) \\ \sigma _ { 2 } ( D - \sum _ { i = 0 } ^ { 1 6 } X _ { i } ^ { 3 } ) = \omega ( \frac { D ( D - 1 ) ^ { 2 } } { 2 4 }, - \frac { ( D - 1 ) ^ { 2 } } { 2 4 }, 0, 0, 0 ) \\ \sigma _ { 3 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) = \omega ( - \frac { D ( D - 1 ) ( 2 D - 6 ) } { 2 4 }, 0, \frac { 2 D - 6 } { 2 4 }, 0, 0, 0 ) \\ \sum _ { i = 0 } ^ { 1 6 } \sigma _ { 4, i } X _ { i } = \omega ( 0, 0, \frac { 1 } { 8 }, \frac { 3 } { 4 }, - \frac { 1 } { 4 }, - \frac { 1 } { 4 } ) \\ \sum _ { i \neq m i \, w i d e }$$
Summing up yields
$$\begin{array} { l l } \rangle & \cdot \cdot \cdot \cdot \\ S _ { 1 7 } & = \omega ( \frac { D ( D - 1 ) ^ { 2 } } { 2 4 } - \frac { D ( D - 1 ) ( 2 D - 6 ) } { 2 4 }, 0, 0, 1, 0, 0 ) \\ & = \omega ( - \frac { 1 } { 2 4 } D ^ { 3 } + \frac { 1 } { 4 } D ^ { 2 } - \frac { 5 } { 2 4 } D, 0, 0, 1, 0, 0 ) \\ & = \sum _ { \{ i, j, k \} } \ A. P. \text{ in } \mathbb { Z } _ { 1 7 } \\ \quad \{ i, j, k \} \ A. P. \text{ in } \mathbb { Z } _ { 1 7 } \end{array}$$
To show Theorem 2.3 let us first prove two lemmas:
Lemma 7.1. The following two problems are equivalent:
(a)
$$\max \{ \lambda \, \colon \sum _ { \{ i, j, k \} \ A. P. \ i n \, \mathbb { Z } _ { p } } X _ { i } X _ { j } X _ { k } - \lambda = S \}$$
glyph[negationslash]
for a ijk ∈ R , c, d ∈ R and b ≥ 0 .
(b') Let r be a primitive root of p .
$$\max \{ b ( D - 1 - u _ { 0 } / b ) D ( D - 1 ) \, \colon U \succeq 0, V u$$
where U is given by
$$& \text{where $U$ is given by} \\ & \left [ \begin{array} {$$
V is element-wise given by
$$V _ { i j } = \left | \left \{ \{ 0, 1, r ^ { i } \} \, \colon \{ 0, 1, r$$
for all i, j ∈ { 0 , . . . , p - } 1 , u is the vector
$$[ u _ { 0 }, u _ { 1 }, \dots, u _ { \frac { p - 3 } { 2 } }, u _ { \frac { p$$
and v is element-wise given by
$$v _ { i } = \begin{cases} \ 1 & \text{if $r^{i}=2$} \\ \ 0 & \text{otherwise.} \end$$
$$f o r \, i \in \{ 0, \dots, p - 1 \}.$$
Proof. The polynomial ∑ p -1 i =0 ∑ p -1 j =0 ( ∑ p -1 k =1 a ijk X k ) 2 X i in Problem (a) can be written in matrix form. There is a unique positive semidefinite matrix U ′ i such that if ˆ X i = [ X ,. . . , X 0 i -1 , X i +1 , . . . , X p -1 ] then
$$\hat { X } _ { i } ^ { T } U _ { i } ^ { \prime } \hat { X } _ { i } = \sum _ { j = 0 } ^ { p - 1 } ( \sum _ { k = 1 } ^ { p - 1 } a _ { i j k } X _ { k } ) ^ { 2 },$$
and thus glyph[negationslash]
$$S = \sum _ { i = 0 } ^ { p - 1 } \hat { X } _ { i } ^ { T } U _ { i } ^ { \prime } \hat { X } _ { i } X _ { i } + b \sum _ { i = 0 } ^ { p - 1 } ( D X _ { i } - \sum _ { j = 0 } ^ { p - 1 } X _ { j } ) ^ { 2 } X _ { i } \\ + c ( \sum _ { i = 0 } ^ { p - 1 } X _ { i } ^ { 3 } - D ) + d ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ). \\ \dots \quad. \quad. \quad. \quad. \quad. \quad \sim \quad.$$
As arithmetic progressions are invariant under affine transformations; if ( a, b ) ∈ Zglyph[multicloseright] Z + , and { i, j, k } is an arithmetic progression, then ( a, b ) · { i, j, k } =
{ a + bi, a + bj, a + bk } is also an arithmetic progressions; ∑ p -1 i =0 ˆ X U X X T i ′ i ˆ i i can also be assumed to be invariant under affine transformations by Lemma 4.8. We get all elements of an orbit by first considering the orbit of the multiplicative action on ˆ X U X X T 0 ′ 0 ˆ 0 0 , and then translate by the additive action to get the elements in ˆ X U X X T a ′ a ˆ a a for all a ∈ { 0 , . . . , p -1 } . Since ˆ X U X X T a ′ a ˆ a a is just a permutation of ˆ X U X X T 0 ′ 0 ˆ 0 0 it follows that U ′ a glyph[followsequal] 0 if and only if U ′ 0 glyph[followsequal] 0, hence requiring that U ′ 0 is positive semidefinite is sufficient. Let now X = [ X ,X 0 r 0 , X r 1 , . . . , X r p -2 ] T where r is a primitive root of p . Since ˆ X U X X T 0 ′ 0 ˆ 0 0 is invariant under the multiplicative group it follows U ′ 0 ( r , r i j ) = U ′ 0 ( r i + k , r j + k ) = U ′ 0 ( r , r j i ) = U ′ 0 ( r , r i 2 i -j ) = U ′ 0 ( r i + k , r 2 i -j + k ) for k = 1 , . . . , p -2 and hence
We see that if we let u i = u ′ i + b we get glyph[negationslash]
$$S = X ^ { T } U X + c ( \sum _ { i = 0 } ^ { p - 1 } X _ { i } ^ { 3 } - D ) + d ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) )$$
where
$$U = \left [ \begin{array} { c c c c c c c c c } b ( 1 - D ) ^ { 2 } & b ( 1 - D ) & b ( 1 - D ) & \cdots & b ( 1 - D ) & \cdots & b ( 1 - D ) \\ b ( 1 - D ) & u _ { 0 } & u _ { 1 } & \cdots & u _ { \frac { p - 1 } { 2 } } & \cdots & u _ { 1 } \\ b ( 1 - D ) & u _ { 1 } & \ddots & \ddots & \ddots & \ddots & \vdots \\ \vdots & \vdots & \ddots & \ddots & \ddots & \ddots & u _ { \frac { p - 1 } { 2 } } \\ b ( 1 - D ) & u _ { \frac { p - 1 } { 2 } } & \ddots & \ddots & \ddots & \vdots \\ \vdots & \vdots & \ddots & \ddots & \ddots & \ddots & u _ { 1 } \\ b ( 1 - D ) & u _ { 1 } & \cdots & u _ { \frac { p - 1 } { 2 } } & \cdots & u _ { 1 } & u _ { 0 } \end{array} \right ].$$
Since
$$\sum _ { \{ i, j, k \} \ A. P. \ \text{in} \ \mathbb { Z } _ { p } } X _ { i } X _ { j } X _ { k } - \lambda = S$$
.
we conclude that b (1 -D ) 2 + c = 0, 2 (1 b -D ) + u 0 + d = 0 and that V u = v , where
$$V _ { i j } = \left | \left \{ \{ 0, 1, r ^ { i } \} \colon \{ 0, 1, r ^ { i } \} = \{ 0, r ^ { t }, r ^ { j + t } \} \text{ for } t = 0, \dots, p - 2 \right \} \right \} \right |$$
for all i, j ∈ { 0 , . . . , p - } 1 ,
$$u = [ u _ { 0 }, u _ { 1 }, \dots, u _ { \frac { p - 3 } { 2 } }, u _ { \frac { p - 1 } { 2 } }, u _ { \frac { p - 3 } { 2 } }, \dots, u _ { 1 } ] ^ { T }$$
and
$$v _ { i } = \begin{cases} \ 1 & \text{if $r^{i}=2$} \\ \ 0 & \text{otherwise.} \end{cases}$$
$$\text{for } i \in \{ 0, \dots, p - 1 \}.$$
Finally we see that
$$- D ) ^ { 2 } D + ( 2 b ( 1 - D ) + u _ { 0 } ) D ( D - 1 ),$$
$$- \lambda = - c D + - d D ( D - 1 ) = b ( 1 - D ) ^ { 2 } D + ( 2 b ( 1 - D ) + u _ { 0 } ) D ( D - 1 ),$$
and simplifying gives the desired objective function:
$$\begin{array} { r c l } \lambda & = - b ( 1 - D ) ^ { 2 } D - ( 2 b ( 1 - D ) + u _ { 0 } ) D ( D - 1 ) \\ & = ( - b ( D - 1 ) + 2 b ( D - 1 ) - u _ { 0 } ) D ( D - 1 ) \\ & = b ( D - 1 - u _ { 0 } / b ) D ( D - 1 ). \end{array}$$
Lemma 7.2. Let r be a primitive root of p and b ∈ R . Let further
$$V _ { i j } = \left | \left \{ \{ 0, 1, r ^ { i } \} \, \colon \{ 0, 1, r ^ { i } \} = \{ 0, r ^ { t }, r ^ { j + t } \} \, f o r \, t = 0, \dots, p - 2 \right \} \right |$$
for all i, j
∈ {
0
, . . . , p
- }
1
,
$$u = [ u _ { 0 }, u _ { 1 }, \dots, u _ { \frac { p - 3 } { p } }, u _ { \frac { p - 1 } { p } }, u _ { \frac { p - 3 } { p } }, \dots, u _ { 1 } ] ^ { T },$$
$$u _ { + } = 1 ^ { T } u = u _ { 0 } + 2 u _ { 1 } + \dots + 2 u _ { ( p - 3 ) / 2 } + u _ { ( p - 1 ) / 2 }$$
$$& \cdot, \, \psi _ { \frac { \mu - 2 } { 2 } }, \, \psi _ { \frac { \mu - 1 } { 2 } }, \, \psi _ { \frac { \mu - 2 } { 2 } }, \, \dots, \, \psi _ { 1 1 } \, \\ & \cdot 2 u _ { 1 } + \dots + 2 u _ { ( p - 3 ) / 2 } + u _ { ( p - 1 ) / 2 }$$
and
$$v _ { i } = \begin{cases} \ 1 & \text{if $r^{i}=2$} \\ \ 0 & \text{otherwise.} \end{cases}$$
for i ∈ { 0 , . . . , p - } 1 . The following two problems are equivalent:
(b')
$$\max \{ b ( D - 1 - u _ { 0 } / b ) D ( D - 1 ) \, \colon U \succeq 0, V u = v \}$$
where U is given by
,
| b (1 - D ) 2 | b (1 - D ) | b (1 - D ) | · · · | b (1 - D ) | · · · | b (1 - D ) |
|----------------|--------------|--------------|---------|--------------|---------|-------------------|
| b (1 - D ) | u 0 | u 1 | . . . | u p - 1 2 | . . . | u 1 |
| b (1 - D ) | u 1 | . . . | . . . | . . . | . . . | . . . |
| . . . | . . . | . . . | . . . | . . . | . . . | u p - 1 2 |
| b (1 - D ) | u p - 1 2 | . . . | . . . | . . . | . . . | . . . |
| . . . | . . . | . . . | . . . | . . . | . . . | u 1 |
| b (1 - D ) | u 1 | · · · | u p - 1 | · · · | u 1 | u 0 |
$$\max \{ \frac { u _ { + } } { p - 1 } ( D - 1 - \frac { u _ { 0 } } { u _ { + } } ( p - 1 ) ) D ( D - 1 ) \colon C u \geq 0, V u = v \} \\.$$
where
$$C _ { i j } = c o s ( \frac { 2 \pi ( i - 1 )$$
Proof. As in the previous lemma, let
$$\Lambda _ { \Lambda } ^ { \prime } = U - \left [$$
or in other words
$$U ^ { \prime } = \left [ \begin{array} { c c$$
where u ′ i = u i -b . Maximizing the objective function is equivalent to maximizing b when D -1 -u /b 0 ≥ 0 . Let U ′′ be the matrix U ′ without its first row and column. U ′′ is a circulant matrix, i.e. row i is the first row shifted i -1 steps to the right. It is easy to see that U is positive semidefinite if and only U ′′ is positive semidefinite. U ′′ is positive semidefinite if and only if all its eigenvalues are nonnegative. Let ω j = e 2 πij p -1 , where i = √ -1. The eigenvalues of a circulant matrix with the first row [ k , k 0 p -2 , k p -3 , . . . , k 1 ] are known to be on the form λ j = k 0 + k p -2 ω j + k p -3 ω 2 j + · · · + k ω 1 p -2 j for j = 0 , . . . , p -2. In our case u ′ i = u ′ p - -1 i for i = 1 , . . . , p -3 2 . By Euler's formula we get
$$\begin{array} {$$
λ j = λ p - -1 j since cos( x ) is an even function, thus it follows that all eigenvalues are nonnegative is equivalent to that
$$[ \cos ( \frac { 0 \cdot 2 \pi j } { p - 1 } )$$
for j = 1 , . . . , p -1 , which is equivalent to the condition Cu ≥
The eigenvalue corresponding to to the eigenvector [1 , . . . , 1] in U is λ 0 = ∑ p -1 i u ′ i . We get the maximal value of b when u i -u ′ i is maximized, that is when λ 0 = 0 and thus
2 0 in Problem (b). ′′
$$b = \frac { [ 1, \dots, 1 ] u } { p - 1 } = \frac { u _ { 0 } + 2 u _ { 1 } + \dots + 2 u _ { ( p - 3 ) / 2 } + u _ { ( p - 1 ) / 2 } } { p - 1 }.$$
$$\begin{smallmatrix}. \end{smallmatrix}$$
Proof of Theorem 2.3. Lemma 7.1 shows that problems (a) and (b') are equivalent and Lemma 7.2 shows that problems (b') and (b) are equivalent. It immediately follows that problems (a) and (b) are equivalent.
Proof of Corollary 2.4. Let
$$B = \min \{ D \in \mathbb { Z } _ { + } \, \colon \lambda _ { p } ( D ) > 0 \}.$$
To show that
$$B \leq \frac { p + 3 } { 2 }$$
let us use the fact that the algebraic certificate in Theorem 2.1 is on the desired form. We have
$$\i m. \ v e \max \\ B \ \ & \leq \min \{ D \in \mathbb { Z } _ { + } \, \colon \frac { D ^ { 3 } - ( \frac { p + 3 } { 2 } ) D ^ { 2 } + ( \frac { p + 3 } { 2 } - 1 ) D } { p - 1 } > 0 \} \\ & = \min \{ D \in \mathbb { Z } _ { + } \, \colon \frac { D ( D - 1 ) ( D - \frac { p + 1 } { 2 } ) } { p - 1 } > 0 \} \\ & = \frac { p + 3 } { 2 }. \\ \intertext { n t h o t h o w h o d } \, \text{$two$ how $tot$}$$
On the other hand, to show that
$$B \geq \left \lceil \frac { p + 3 } { 4 } \right \rceil$$
note that by Theorem 2.3 we have
$$B = \min \{ D \in \mathbb { Z } _ { + } \colon \frac { u _ { + } } { p - 1 } ( D - 1 - \frac { u _ { 0 } } { u _ { + } } ( p - 1 ) ) D ( D - 1 ) \}.$$
By the matrix condition V u = v we know that the sum of the u i with i > 0 which do not contribute to an arithmetic progression is zero. Furthermore, if t and s are the unique integer such that r t = 2 and r s = p -1 2 , then u t s orbit contributes to 4 arithmetic progressions and u s s orbit contributes to 2 arithmetic progressions respectively. In other words 4 u t +2 u s = 1, and hence
$$u _ { + } = u _ { 0 } + 2 u _ { t } + u _ { s } = u _ { 0 } + \frac { 1 } { 2 },$$
so
$$B = \min \{ D \in \mathbb { Z } _ { + } \colon \frac { u _ { + } } { p - 1 } ( D - 1 - \frac { u _ { 0 } } { u _ { 0 } + \frac { 1 } { 2 } } ( p - 1 ) ) D ( D - 1 ) \}.$$
To find a lower bound to B we should find a lower bound for u 0 u 0 + 1 2 . Since u 0 ≥ max p -1 i =1 u i in order for U to be positive semidefinite and since 4 u t +2 u s = 1 we have u 0 ≥ 1 6 . Hence
$$B & \geq \min \{ D \in \mathbb { Z } _ { + } \, \colon \frac { u _ { + } } { p - 1 } ( D - 1 - \frac { \frac { 1 } { 6 } } { \frac { 1 } { 6 } + \frac { 3 } { 6 } } ( p - 1 ) ) D ( D - 1 ) \}. \\ & = \lceil 1 + \frac { 1 } { 4 } ( p - 1 ) \rceil$$
## 8 Additional numerical results
We can easily generate the list of all necklaces of length n with nδ ones using the recursive formula introduced in Section 5. We can use the list to find W k, ( Z n , δ ), we merely need to count the number of arithmetic progressions in each necklace and find the minimum among these. Note that n does not have to be prime as long as we are cautious when we count arithmetic progressions to avoid double counting and degenerate arithmetic progressions.
We have found exact solutions using fixed density necklaces for W k, ( Z n , δ ) for 3 ≤ k ≤ 5, 5 ≤ n ≤ 32 and δ ∈ { 0 , 1 n , . . . , n -1 n , 1 } . To find the exact solution we used the code for finding fixed density necklaces by Sawada [64], and code implemented in matlab by the author for counting the number of arithmetic progressions in every necklace. The code was implemented on the Triton computer cluster, which is part of the Aalto Science-IT project, and we used 100 computer nodes for 5 days to carry out the calculations. The results can be found in Tables 3, 4, 5, 6, 7 and 8 in the appendix.
Using the degree 3 relaxation of the Lasserre hierarchy we have found lower bounds to W (3 , Z p , δ ) where p ≤ 300 is a prime and
$$\delta \in \{ 0, \frac { 1 } { 2 0 p }, \frac { 2 } { 2 0 p }, \dots, \frac { 2 0 p - 1 } { 2 0 p }, 1 \}.$$
For 300 ≤ p ≤ 613 we start to run into numerical problems for larger δ , which might be because of our particular choice of solver, so then we have fewer and perhaps less accurate data points. The results for δ < 1 3 seems accurate in this interval, so we feel confident using this data for analysis on the asymptotic behavior of the degree 3 relaxation.
Using the degree 5 relaxation of the Lasserre hierarchy we have found lower bounds to W k, ( Z p , δ ) for primes p ≤ 19, k = 3 4 5 and , ,
$$\delta \in \{ 0, \frac { 1 } { 2 0 p }, \frac { 2 } { 2 0 p }, \dots, \frac { 2 0 p - 1 } { 2 0 p }, 1 \}.$$
Figure 3: The minimum number of arithmetic progressions of length 3 as a function of the density in Z 17 .
In Figure 3 we show the results for k = 3 and D = δp ∈ { 0 , . . . , 17 } when p = 17, for which we have numerical data for all different relaxations we make, as well as a lower bound from Theorem 2.2. In Figure 4 we showed a zoomed in version around where W (3 , Z 17 , δ ) becomes positive. For small values of D and p we can find algebraic solutions to check when the points coincide, unfortunately this is generally not the case for the larger values of D and p which we are more interested in. The case p = 17 is a good representative for the situation for all small primes, and hence the numerical certificates as well as the algebraic certificates one can get from Theorem 2.3 work very well here. The problem is that the larger p is the worse the relaxations will get. We discuss this further in Section 9.
We also provide a list of one necklace per D that achieve the lower bound as a certificate in Tables 1 and 2. There is recent research on how to find necklaces with less arithmetic progressions than average for higher p [10, 40]. In both papers they analyze the problem of avoiding arithmetic progressions in a 2-coloring of [ n ], which is closely related to the fixed-density case, and it is likely that similar constructions are applicable here.
It is shown in the first paper that there exists colorings in which the colors
Figure 4: A zoomed in version of Figure 3.
Table 1: Necklaces that contains the least number of arithmetic progressions for each D in Z 7 .
| D | A.P.s | smallest necklace |
|-----|---------|---------------------|
| 0 | 0 | 0000000 |
| 1 | 0 | 0000001 |
| 2 | 0 | 0000011 |
| 3 | 0 | 0001011 |
| 4 | 2 | 0001111 |
| 5 | 6 | 0011111 |
| 6 | 12 | 0111111 |
| 7 | 21 | 1111111 |
Table 2: Necklaces that contains the least number of arithmetic progressions for each D in Z 31 .
| D | A.P.s | smallest necklace |
|-----|---------|---------------------------------|
| 0 | 0 | 0000000000000000000000000000000 |
| 1 | 0 | 0000000000000000000000000000001 |
| 2 | 0 | 0000000000000000000000000000011 |
| 3 | 0 | 0000000000000000000000000001011 |
| 4 | 0 | 0000000000000000000000000011011 |
| 5 | 0 | 0000000000000000000000101100011 |
| 6 | 0 | 0000000000000000000010110001011 |
| 7 | 0 | 0000000000000000001011000011011 |
| 8 | 0 | 0000000000000000011011000011011 |
| 9 | 2 | 0000000000000001011010001011011 |
| 10 | 5 | 0000000000000001101101000110111 |
| 11 | 9 | 0000000000000101011110001011011 |
| 12 | 14 | 0000000000000001111110000111111 |
| 13 | 19 | 0000000001010011110001101110011 |
| 14 | 27 | 0000000001101011110001011011011 |
| 15 | 37 | 0000000000000111111100011111111 |
| 16 | 48 | 0000000000000111111110011111111 |
| 17 | 61 | 0000000000001111111100111111111 |
| 18 | 76 | 0000000000001111111101111111111 |
| 19 | 92 | 0000000000011111111101111111111 |
| 20 | 110 | 0000000000011111111111111111111 |
| 21 | 130 | 0000000000111111111111111111111 |
| 22 | 152 | 0000000001111111111111111111111 |
| 23 | 177 | 0000000011111111111111111111111 |
| 24 | 204 | 0000000111111111111111111111111 |
| 25 | 234 | 0000001111111111111111111111111 |
| 26 | 266 | 0000011111111111111111111111111 |
| 27 | 301 | 0000111111111111111111111111111 |
| 28 | 338 | 0001111111111111111111111111111 |
| 29 | 378 | 0011111111111111111111111111111 |
| 30 | 420 | 0111111111111111111111111111111 |
| 31 | 465 | 1111111111111111111111111111111 |
come in large blocks that provide much better bounds than random colorings, and in the second paper it is shown that periodic colorings can provide even better bounds. For computational reasons we are not able to do examples for p high enough to make reasonable conjectures how fixed-density sets avoiding arithmetic progressions look like, and we have not tried to apply their methods. Note that finding a class of fixed-density necklaces does not help to imply a new proof of Szemer´di's theorem directly as we would need a certificate that all e fixed-density necklaces have a positive density. Somewhat surprisingly though we have in some cases seen correlations between good classes of colorings and coefficients in the corresponding semidefinite programs for finding a certificate. Finding a good way of coloring groups could thus prove to be useful for finding good SDP relaxations.
Another opportunity to analyze what happens as p grows is to look at the distributions of the number of arithmetic progressions for all different necklaces of fixed p , k and δ . When for example p = 7, k = 3 and δ = 4 7 there / are 3 necklaces with 2 arithmetic progressions, and 2 necklaces with 3, clearly W (7 , 3 4 , / 7) = 2 and we have the distribution in Figure 5. More interestingly, for p = 31, k = 3 and δ = 16 31 we have the situation in Figure 6; it looks like / finding the minimum number of arithmetic progressions can be approximated by analyzing the left tail of a probability distribution. The distribution looks very similar as we vary p and δ , it just gets shifted and has a different height. The end goal is to go from bounds on Z p for all prime numbers p to [ n ] for all integers n , and using a theorem similar to Corollary 3.9 to get a new proof for Szemer´ edi's theorem. Hence if it would be possible to prove that for any density δ > 0 there exists an N such that for all p ≥ N it holds that the leftmost non-zero value of the distribution of the number of arithmetic progressions in different necklaces of length p and density δ is positive it is likely possible that a proof would follow. It is unclear whether the tools for approximating the tail are strong enough to find a new proof of Szemer´di's theorem, and even if e the tools are sufficient it would most probably be highly non-trivial to find a new proof.
## 9 Discussion
## 9.1 How Szemer´ edi's theorem could perhaps be generalized for arithmetic progressions of length 3
In order to prove Szemer´di's theorem by providing bounds e
$$\lambda ( k, \mathbb { Z } _ { n }, \delta ) \leq W ( k, \mathbb { Z } _ { n }, \delta )$$
it is by Corollary 3.9, and sufficiently good tools to translate results between Z n and [ n ], necessary and sufficient to find bounds as sharp as
$$\delta _ { * } ( k ) = \lim _ { n \to \infty } \min \{ \delta \, \colon \lambda ( k, \mathbb { Z } _ { n }, \delta ) > 0 \} = 0.$$
## W(3,Z7,4/7)=2
Figure 5: The plot shows how many necklaces of length 7 and density 4 / 7 contain n 3-arithmetic progressions for n = 0 , . . . , 21.
## W(3,Z31,16/31)=48
Figure 6: The plot shows how many necklaces of length 31 and density 16 / 31 contain n 3-arithmetic progressions for n = 0 , . . . , 81.
Already finding sharp enough bounds in the case k = 3 would be a great result. Depending on how fast the convergence is, such bound might additionally improve the current best existence bounds for arithmetic progressions by Gowers [18]. It can be shown by a probabilistic argument that W k, ( Z n , δ ) with fixed k and n behaves as a degree k polynomial when δ is large enough, but it is not clear whether this should also be the case for small δ . Any additional understanding on how W k, ( Z n , δ ) looks like for small δ could possibly benefit in a better understanding on how to restrict polynomials in a relaxation of Putinar's positivstellensatz in the search for an improved lower bound λ k, ( Z n , δ ).
In Figure 7 we have in the same plot put the upper and lower bounds from Conjecture 2.4 as well as numerical upper bounds from the degree 3 and 5 relaxations of Putinar's positivstellensatz. All methods presented perform very well for small p . In the first few cases we can find algebraic solutions verifying that we find the exact solution with our approximation. It seems like using the degree 3 relaxation in its full generality rather than our simplification in Theorem 2.1 does not make a huge difference, and it seems like we need to increase the degree to make sure δ ∗ converge to 0. We have too few data points of the degree 5 relaxation to make any reasonable conjectures about the convergence of those upper bounds. We suggest as a next step to try to come up with reasonable simplification of the degree 5 relaxation that would allow us to find more data points. How to find general patterns for a simplified degree 5 relaxation based on numerical data is highly non-trivial, and it is unclear whether a degree 5 relaxation is sufficient to generalize Szemer´ edi's theorem for k = 3, or if one will find lower bounds as in Corollary 2.4. We know in theory by Putinar's positivstellensatz that if the relaxation is of a degree high enough we will find a quantitative version of Szmer´di's theorem, e but in practice we have to either limit the degree of the relaxation or add restrictions to get a simpler problem but worse lower bound. It is thus unlikely that the theoretically optimal lower bound will ever be obtained, but one cannot exclude the possibility that a lower bound strong enough to obtain a quantitative version of Szemer´di's theorem can be obtained. e As we have found algebraic bounds with polynomials of degree 3, it is likely that improved bounds with polynomials of higher degree look similar, or at least share some similarities, with the polynomials of degree 3. Another possibility would thus be to try to find a certificate based on polynomials of high degree with a lot of structure that are similar to the polynomials in Theorem 2.1.
## 9.2 Longer arithmetic progressions
Recall that W k,G,δ ( ) denotes the minimal number of arithmetic progressions in the subset S ⊆ G with | S | = δ G | | . It is easy to see that
$$W ( k, G, D / p ) = \min \{ \sum _ { \{ a _ { 1 }, \dots, a _ { k } \} \text{ is an A.P. in } G } x _ { a _ { 1 } } \cdots x _ { a _ { k } } \, \colon x _ { i } \in \{ 0, 1 \}, \sum _ { i = 0 } ^ { p - 1 } x _ { i } = D \},$$
Figure 7: Lower and upper bounds from Conjecture 2.4, numerical upper bounds from degree 3 and 5 relaxations of Putinar's Positivstellensatz and exact results found using fixed density necklaces.
and using the methods introduced in Sections 4.1 and 4.2 we make the following relaxation to get an optimization problem which if it is solved gives an algebraic certificate for the lower bound:
$$W ( 3, G, D / | G | ) \geq \max \{ \lambda \colon \sum _ { \{ a _ { 1 }, \dots, a _ { k } \} \ A. P. \text{ in } G } X _ { a _ { 1 } } \cdots X _ { a _ { k } } - \lambda = S, X \in K \}$$
where K = { X ∈ [0 , 1] n : ∑ | G |-1 i =0 X i = D } and S = ∑ i S g 2 i i for S , g i i ∈ R [ X ,. . . , X 1 k ] where g i are the half-spaces defining K .
glyph[negationslash]
Since ∑ | G |-1 i =0 x i = D on {-1 1 , } n it follows that for example ∑ | G |-1 i =0 x t i -D = 0 for all t ∈ Z + , ∑ | G |-1 i =0 x x s i t j -D D ( -1) = 0 for all s = t ∈ Z + , and so on, providing several new possibilities for polynomials that are nonnegative on {-1 1 , } n . Let us denote the set of all possible conditions of this form { R , R , . . . , 0 1 } .
These redundant conditions simplifies the algebraic certificates we get from Putinar's Positivstellensatz, and replacing some of the original g i s to these new constraints could possibly simplify the computations significantly. For any k we can write S on the form
$$S = \sum _ { i } a _ { i } S _ { i } ^ { 2 } g _ { i } + \sum _ { j } b _ { j } R _ { j } + c,$$
where a i ∈ { 0 1 , } are used to simplify the problem, b j ∈ R are constants indicating how much of the redundant condition R j should be added and c a constant. This is one option to move forward with the suggested methods as the redundant conditions add much more freedom to the problem. As a generalization of Szemer´di's theorem would be extremely difficult to accomplish we e want to stress that these are just speculations on how to improve the results in this article towards that goal. Even though there are strong theoretical results in real algebraic geometry, the current methods for applying the methods in practice have their limitations. If even possible probably many years of further research in this direction is required to find a quantitative version of Szemer´di's e theorem.
## Acknowledgements
I would like to thank Alexander Engstr¨m for introducing me to Szemer´ edi's o theorem and for suggesting how the problem could be approached. Additionally I would want to thank Markus Schweighofer and Cynthia Vinzant for their valuable feedback.
## References
- [1] Tanbir Ahmed. Some new van der Waerden numbers and some van der Waerden-type numbers. Integers 9 (2009), no. 1, 65-76.
- [2] Tanbir Ahmed. Some more van der Waerden numbers. J. Integer Seq. 16 (2013), no. 4, 13.4.4.
- [3] Christine Bachoc, Dion C. Gijswijt, Alexander Schrijver and Frank Vallentin. Invariant semidefinite programs. Chapter of 'Handbook on Semidefinite, Conic and Polynomial Optimization'. Springer, Berlin, 2012. 219269.
- [4] Felix A. Behrend, On sets of integers which contain no three terms in arithmetical progression. Proc. Natl. Acad. Sci. USA 32 (1946), no. 12, 331-332.
- [5] Vitaly Bergelson and Neil Hindman. Ramsey theory in noncommutative semigroups. Trans. Amer. Math. Soc. 330 (1992), no. 1, 433-446.
- [6] Elwyn Ralph Berlekamp. A construction for partitions which avoid long arithmetic progressions. Canad. Math. Bull. 11 (1968), no. 3, 409-414.
- [7] Jean Bourgain. On arithmetic progressions in sums of sets of integers. A tribute to Paul Erd˝os o (1990), 105-109.
- [8] Jean Bourgain. On triples in arithmetic progression. Geom. Funct. Anal. 9 (1999), no. 5, 968-984.
- [9] Jean Bourgain. Roth's theorem on progressions revisited. J. Anal. Math. 104 (2008), 155-192.
- [10] Steve Butler, Kevin P. Costello and Ron Graham. Finding patterns avoiding many monochromatic constellations. Exp. Math. 19 (2010), no. 4, 399411.
- [11] Peter Cameron, Javier Cilleruelo and Oriol Serra. On monochromatic solutions of equations in groups. Rev. Mat. Iberoam. 23 (2007), no. 1, 385-395.
- [12] Boris A. Datskovsky. On the number of monochromatic Schur triples. Adv. in Appl. Math. 31 (2003), no. 1, 193-198.
- [13] Michael Elkin. An improved construction of progression-free sets. Israel J. Math. 184 (2011), no. 1, 93-128.
- [14] Hillel Furstenberg, Yitzhak Katznelson and Donald Samuel Ornstein. The ergodic theoretical proof of Szemer´di's theorem. e Bull. Amer. Math. Soc. (N.S.) 7 (1982), no. 3, 527-552.
- [15] Michael Grant and Stephen Boyd. CVX: Matlab software for disciplined convex programming, version 1.21. cvxr.com/cvx , February 2011.
- [16] Karin Gatermann and Pablo A. Parrilo. Symmetry groups, semidefinite programs, and sums of squares. J. Pure Appl. Algebra 192 (2004), no. 1, 95-128.
| [17] | William T. Gowers. A new proof of Szemer´di's e theorem for arithmetic progressions of length four. Geom. Funct. Anal. 8 (1998), no. 3, 529-551. |
|--------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [18] | William T. Gowers. A new proof of Szemer´di's e theorem. Geom. Funct. Anal. 11 (2001), no. 3, 465-588. |
| [19] | William T. Gowers. Hypergraph regularity and the multidimensional Sze- mer´di e theorem. Ann. of Math.(2) 166 (2007), no. 3, 897-946. |
| [20] | William T. Gowers. Quasirandom groups. Combin. Probab. Comput. 17 (2008), no. 3, 363-387. |
| [21] | Ron Graham. On the growth of a van der Waerden-like function. Integers 6 (2006), A29. |
| [22] | Frank Gray. Pulse code communication. U.S. Patent No. 2,632,058. 17 Mar. 1953. |
| [23] | Ben Green. Roth's theorem in the primes. Ann. of Math.(2) 161 (2005), no. 3, 1609-1636. |
| [24] | Ben Green. A Szemer´di-type e regularity lemma in abelian groups, with applications. Geom. Funct. Anal. 15 (2005), no. 2, 340-376. |
| [25] | Ben Green and Terence Tao. The primes contain arbitrarily long arithmetic progressions. Ann. of Math.(2) 167 (2008), no. 2, 481-547. |
| [26] | Ben Green and Terence Tao. New bounds for Szemer´di's e theorem, II: Pro- gressions of length 4 in finite field geometries. Chapter of 'Analytic num- ber theory: essays in honour of Klaus Roth'. Cambridge University Press, Cambridge, 2009. 180-204. |
| [27] | David Rodney Heath-Brown. Integer sets containing no arithmetic progres- sions. J. Lond. Math. Soc. (2) 2 (1987), no. 3, 385-394. |
| [28] | Harald Andr´s e Helfgott and Anne de Roton. Improving Roth's theorem in the primes. Int. Math. Res. Not. IMRN 2011 (2011), no. 4, 767-783. |
| [29] | Yoshihiro Kanno, Makoto Ohsaki, Kazuo Murota and Naoki Katoh. Group symmetry in interior-point methods for semidefinite program. Optim. Eng. 2 (2001), no. 3, 293-320. |
| [30] | Etienne de Klerk, Cristian Dobre and Dmitrii V. Pasechnik. Numerical block diagonalization of matrix ∗ -algebras with application to semidefinite programming. Math. Program. 129 (2011), no. 1, 91-111. |
| [31] | Etienne de Klerk, Dmitrii V. Pasechnik and Alexander Schrijver. Reduc- tion of symmetric semidefinite programs using the regular ∗ -representation. |
| [32] | Donald E. Knuth. Generating All Tuples and Permutations. Art of Com- puter Programming, Volume 4, Fascicle 2, Addison-Wesley, Reading, MA, 2005. 127 pp. |
|--------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [33] | Daniel Kr´l, a Oriol Serra and Llu´s ı Vena. A combinatorial proof of the removal lemma for groups. J. Combin. Theory Ser. A 116 (2009), no. 4, 971-978. |
| [34] | Bruce Landman, Aaron Robertson and Clay Culver. Some new exact van der Waerden numbers. Integers 5 (2005), no. 2, A10. |
| [35] | Jean Bernard Lasserre. Global optimization with polynomials and the prob- lem of moments. SIAM J. Optim. 11 (2001), no. 3, 796-817. |
| [36] | Jean Bernard Lasserre. Moments, positive polynomials and their applica- tions. Imperial College Press Optimization Series, 1. Imperial College Press, London, 2010. 361 pp. |
| [37] | Monique Laurent. Sums of squares, moment matrices and optimization over polynomials. Chapter of 'Emerging applications of algebraic geometry'. Springer, New York, 2009. 157-270. |
| [38] | Vsevolod F. Lev. Optimal representations by sumsets and subset sums. J. Number Theory 62 (1997), no. 1, 127-143. |
| [39] | Vsevolod F. Lev. On small sumsets in abelian groups. Ast´risque e 258 (1999), 317-321. |
| [40] | Linyuan Lu and Xing Peng. Monochromatic 4-term arithmetic progressions in 2-colorings of Z n . J. Combin. Theory 119 (2012), Ser. A, no. 5, 1048- 1065. |
| [41] | Takanori Maehara and Kazuo Murota. A numerical algorithm for block- diagonal decomposition of matrix ∗ -algebras with general irreducible com- ponents. Jpn. J. Ind. Appl. Math. 27 (2010), no. 2, 263-293. |
| [42] | Roy Meshulam. On subsets of finite abelian groups with no 3-term arith- metic progressions. J. Combin. Theory Ser. A 71 (1995), no. 1, 168-172. |
| [43] | M´t´ a e Matolcsi and Imre Z. Ruzsa. Sets with no solutions to x+ y= 3z. European J. Combin. 34 (2013), no. 8, 1411-1414. |
| [44] | Kazuo Murota, Yoshihiro Kanno, Masakazu Kojima and Sadayoshi Ko- jima. A numerical algorithm for block-diagonal decomposition of matrix ∗ - algebras with application to semidefinite programming. Jpn. J. Ind. Appl. Math. 27 (2010), no. 1, 125-160. |
| [45] | Brendan Nagle, Vojtˇch e R¨dl o and Mathias Schacht. The counting lemma for regular k-uniform hypergraphs. Random Structures Algorithms 28 |
| [46] | Melvyn B. Nathanson. Growth of sumsets in abelian semigroups. Semigroup Forum 61 (2000), no. 1, 149-153. |
|--------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [47] | Melvyn B. Nathanson and Imre Z. Ruzsa. Polynomial growth of sumsets in abelian semigroups. J. Th´or. e Nombres Bordeaux 14 (2002), no. 2, 553- 560. |
| [48] | Kevin O'Bryant. Sets of integers that do not contain long arithmetic pro- gressions. Electron. J. Combin. 18 (2011), no. 1, P59. |
| [49] | Pablo A. Parrilo, Aaron Robertson, and Dan Saracino. On the asymp- totic minimum number of monochromatic 3-term arithmetic progressions. J. Combin. Theory 115 (2008), no. 1, Ser. A,185-192. |
| [50] | Robert A. Rankin. Sets of integers containing not more than a given number of terms in arithmetical progression. Proc. Roy. Soc. Edinburgh Sect. A 65 (1961), no. 4, 332-344. |
| [51] | Cordian Riener, Thorsten Theobald, Lina Jansson Andr´n e and Jean B. Lasserre. Exploiting symmetries in SDP-relaxations for polynomial opti- mization. Math. Oper. Res. 38 , (2013), no. 1, 122-141. |
| [52] | Aaron Robertson and Doron Zeilberger. A 2-coloring of [1 , N ] can have (1 / 22) N 2 + O ( N ) Monochromatic Schur Triples, But Not Less!. Electron. J. Combin. 5 (1998), no. 1, R19. |
| [53] | Vojtˇch e R¨dl o and Mathias Schacht. Regular partitions of hypergraphs: Regularity Lemmas Combin. Probab. Comput. 16 (2007), no. 6, 833-885. |
| [54] | Vojtˇch e R¨dl o and Mathias Schacht. Regular partitions of hypergraphs: Counting Lemmas Combin. Probab. Comput. 16 (2007), no. 6, 887-901. |
| [55] | Vojtˇch e R¨dl o and Jozef Skokan. Regularity lemma for uniform hypergraphs Random Structures Algorithms 25 (2004), no. 1, 1-42. |
| [56] | Vojtˇch e R¨dl o and Jozef Skokan. Applications of the regularity lemma for uniform hypergraphs Random Structures Algorithms 28 (2006), no. 2, 180- 194. |
| [57] | Klaus Friedrich Roth. On certain sets of integers. J. Lond. Math. Soc. 28 (1953), 104-109. |
| [58] | Frank Ruskey, Joe Sawada and Aaron Williams. Binary bubble languages and cool-lex order. J. Combin. Theory 119 (2012), no. 1, Ser. A, 155-169. |
| [59] | Tom Sanders. On a non-abelian Balog-Szemer´di-type e lemma. J. Aust. Math. Soc. 89 (2010), no. 1, 127-132. |
| [60] | Tom Sanders. On Roth's theorem on progressions. Ann. of Math. (2) 174 (2011), no. 1, 619-636. |
| [61] | Tom Sanders. On certain other sets of integers. J. Anal. Math. 116 (2012), no. 1, 53-82. |
|--------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [62] | Andr´s a S´rk¨zy. a o Finite addition theorems, I. J. Number Theory 32 (1989), no. 1, 114-130. |
| [63] | Joe Sawada and Aaron Williams. A Gray code for fixed-density necklaces and Lyndon words in constant amortized time. Theoret. Comput. Sci. 502 (2013), 46-54. |
| [64] | Joe Sawada. http://www.cis.uoguelph.ca/ ∼ sawada/prog/necklaces.c |
| [65] | Tomasz Schoen. The number of monochromatic Schur triples. European J. Combin. 20 (1999), no. 8, 855-866. |
| [66] | Erik Sj¨land. o Enumeration of monochromatic three term arithmetic progressions in two-colorings of cyclic groups. Preprint available at http://arxiv.org/abs/1408.1058. |
| [67] | Erik Sj¨land. o Enumeration of monochromatic three term arithmetic pro- gressions in two-colorings of any finite group. Preprint available at http://arxiv.org/abs/1408.1088. |
| [68] | Erik Sj¨land. o Using real algebraic geometry to solve combinatorial problems with symmetries. Preprint available at http://arxiv.org/abs/1408.1065. |
| [69] | Jozsef Solymosi. Roth-type theorems in finite groups. European J. Combin. 34 (2013), no. 8, 1454-1458. |
| [70] | Endre Szemer´di. e On sets of integers containing no four elements in arith- metic progression. Acta Math. Hungar. 20 (1969), no. 1, 89-104. |
| [71] | Endre Szemer´di. e On sets of integers containing no k elements in arithmetic progression. Acta Arith. 27 (1975), no. 1, 199-245. |
| [72] | Endre Szemer´di. e Integer sets containing no arithmetic progressions. Acta Math. Hungar. 56 (1990), no. 1, 155-158. |
| [73] | Endre Szemer´di e and Van Vu. Long arithmetic progressions in sumsets: thresholds and bounds. J. Amer. Math. Soc. 19 (2006), no. 1, 119-169. |
| [74] | Masamichi Takesaki. Theory of operator algebras I . Encyclopaedia of Math- ematical Sciences, 124. Operator Algebras and Non-commutative Geome- try, 5. Springer, Berlin, 2002. 525 pp. |
| [75] | Terence Tao. A quantitative ergodic theory proof of Szemer´di's e theorem. Electron. J. Combin. 13 (2006), no. 1, R99. |
| [76] | Terence Tao. Product set estimates for non-commutative groups. Combi- natorica 28 (2008), no. 5, 547-594. |
- [77] Terence Tao. Freiman's theorem for solvable groups. Contrib. Discrete Math. 5 (2010), no. 2, 137-184.
- [78] Terence Tao. Noncommutative sets of small doubling. European J. Combin. 34 (2013), no. 8, 1459-1465.
- [79] Terence Tao and Van Vu. Additive combinatorics. Cambridge Studies in Advanced Mathematics, 105. Cambridge University Press, Cambridge, 2006. xviii+512 pp.
- [80] Frank Vallentin. Symmetry in semidefinite programs. Linear Algebra Appl. 43 (2009), no. 1, 360-369.
- [81] Bartel Leendert van der Waerden. Beweis einer baudetschen vermutung. Nieuw Arch. Wiskd. (5) 15 (1927), no. 2, 212-216.
- [82] Julia Wolf. The minimum number of monochromatic 4-term progressions in Z p . J. Comb. 1 (2010), no. 1, 53-68.
## 10 Appendix
| n \ D | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 |
|---------|-----|-----|-----|-----|-----|-----|-----|-----|-----|-----|------|------|-------|-------|-------|-------|-------|-------|
| 5 | 0 | 0 | 0 | 1 | 4 | 10 | | | | | | | | | | | | |
| 6 | 0 | 0 | 0 | 0 | 0 | 4 | 8 | | | | | | | | | | | |
| 7 | 0 | 0 | 0 | 0 | 2 | 6 | 12 | 21 | | | | | | | | | | |
| 8 | 0 | 0 | 0 | 0 | 0 | 3 | 8 | 15 | 28 | | | | | | | | | |
| 9 | 0 | 0 | 0 | 0 | 0 | 1 | 2 | 11 | 20 | 30 | | | | | | | | |
| 10 | 0 | 0 | 0 | 0 | 0 | 2 | 4 | 10 | 16 | 28 | 45 | | | | | | | |
| 11 | 0 | 0 | 0 | 0 | 0 | 2 | 5 | 11 | 18 | 28 | 40 | 55 | | | | | | |
| 12 | 0 | 0 | 0 | 0 | 0 | 1 | 2 | 5 | 8 | 18 | 28 | 39 | 52 | | | | | |
| 13 | 0 | 0 | 0 | 0 | 0 | 1 | 4 | 8 | 14 | 22 | 32 | 45 | 60 | 78 | | | | |
| 14 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 8 | 16 | 24 | 36 | 48 | 66 | 91 | | | |
| 15 | 0 | 0 | 0 | 0 | 0 | 1 | 2 | 5 | 8 | 12 | 20 | 30 | 40 | 58 | 76 | 95 | | |
| 16 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 4 | 11 | 20 | 29 | 40 | 55 | 72 | 91 | 120 | |
| 17 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 4 | 8 | 14 | 21 | 31 | 42 | 56 | 72 | 91 | 112 | 136 |
| 18 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 8 | 12 | 16 | 34 | 52 | 70 | 88 | 110 |
| 19 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 3 | 6 | 11 | 18 | 26 | 36 | 48 | 62 | 79 | 98 | 120 |
| 20 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 5 | 11 | 19 | 28 | 39 | 51 | 65 | 80 | 104 |
| 21 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 2 | 3 | 9 | 15 | 22 | 33 | 42 | 59 | 77 | 95 |
| 22 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 8 | 14 | 20 | 32 | 44 | 58 | 72 | 92 |
| 23 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 3 | 7 | 12 | 18 | 26 | 36 | 47 | 61 | 76 | 94 |
| 24 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 2 | 6 | 10 | 20 | 30 | 39 | 48 | 70 |
| 25 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 6 | 10 | 15 | 22 | 31 | 41 | 53 | 66 | 82 |
| 26 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 4 | 10 | 16 | 24 | 32 | 44 | 56 | 72 |
| 27 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 3 | 6 | 9 | 12 | 21 | 30 | 39 | 48 | 60 |
| 28 29 | 0 0 | 0 0 | 0 | 0 0 | 0 | 0 | 0 | 0 | 0 0 | 2 | 4 | 7 | 12 16 | 20 23 | 28 32 | 37 41 | 48 52 | 63 67 |
| | | | 0 | | 0 | 0 | 0 | 0 | | 3 | 6 | 11 | | | | | | |
| 30 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 4 | 6 | 8 | 14 | 20 27 | 26 37 | 32 48 | 44 |
| 31 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 5 | 9 | 14 | 19 | | | | 61 |
| 32 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 5 | 8 | 11 | 16 | 23 | 32 | 46 |
Table 3: W (3 , Z n , D/n ) for different n and D .
| \ D | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 |
|-------------|------------|---------|------------|------------|------------|------|---------|---------|---------|---------|------|------|------|------|------|
| 5 6 7 8 9 | | | | | | | | | | | | | | | |
| 11 12 13 | 10 14 | 10 14 | 10 14 | 10 14 | 10 14 | | | | | | | | | | |
| 16 17 18 19 | 15 132 144 | 171 | 15 132 144 | 15 132 144 | 15 132 144 | | | | | | | | | | |
| 20 | 128 | 153 | 190 | | | | | | | | | | | | |
| 21 | 114 | 141 168 | 196 | | | | | | | | | | | | |
| 22 | 112 | 136 | 160 | 190 | 231 | | | | | | | | | | |
| 23 | 114 | 137 | 162 | 190 | 220 | 253 | 253 | 253 | 253 | 253 | | | | | |
| 24 | 90 | 112 | 135 | 160 | 188 | | 217 248 | 217 248 | 217 248 | 217 248 | | | | | |
| 25 | 100 | 119 | 140 | 170 | 200 | 231 | 264 | 300 | | | | | | | |
| 26 | 88 | 108 | 128 | 154 | 180 | 210 | 240 | 276 | 325 | | | | | | |
| 27 | 72 | 99 | 126 | 153 | 180 | 210 | 240 | 276 | 312 | 351 | | | | | |
| 28 | 79 | 97 | 116 | 139 | 163 | 189 | 216 | 252 | 288 | 325 | 378 | | | | |
| 29 | 83 | 101 | 120 | 142 | 166 | 193 | 222 | 254 | 288 | 325 | 364 | 406 | | | |
| 30 | 56 | 68 | 80 | 110 | 140 | 168 | 192 | 228 | 264 | 300 | 336 | 378 | 435 | | |
| 31 | 76 | 92 | 110 | 130 | 152 | 177 | 204 | 234 | 266 | 301 | 338 | 378 | 420 | 465 | |
| 32 | 60 | 78 | 95 | 115 | 135 | 157 | 180 | 211 | 244 | 277 | 312 | 351 | 392 | 435 | 496 |
Table 4: W (3 , Z n , D/n ) for different n and D .
| n \ | D | 0 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 |
|-------|-----|-------|-----|-----|-----|-----|-----|-----|-----|-----|------|------|------|------|------|------|-------|------|
| 5 | | 0 0 | 0 | 0 | 2 | 10 | | | | | | | | | | | | |
| 6 | | 0 0 | 0 | 0 | 0 | 4 | 8 | | | | | | | | | | | |
| 7 | | 0 0 | 0 | 0 | 0 | 3 | 9 | 21 | | | | | | | | | | |
| 8 | | 0 0 | 0 | 0 | 0 | 0 | 4 | 12 | 28 | | | | | | | | | |
| 9 | | 0 0 | 0 | 0 | 0 | 1 | 2 | 11 | 20 | 30 | | | | | | | | |
| 10 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 4 | 8 | 24 | 45 | | | | | | | |
| 11 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 5 | 11 | 21 | 35 | 55 | | | | | | |
| 12 | | 0 0 | 0 | 0 | 0 | 1 | 2 | 5 | 8 | 18 | 28 | 39 | 52 | | | | | |
| 13 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 2 | 6 | 13 | 22 | 36 | 54 | 78 | | | | |
| 14 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 6 | 12 | 24 | 36 | 60 | 91 | | | |
| 15 | | 0 0 | 0 | 0 | 0 | 1 | 2 | 5 | 8 | 12 | 20 | 30 | 40 | 58 | 76 | 95 | | |
| 16 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 3 | 8 | 16 | 24 | 40 | 60 | 84 | 120 | |
| 17 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 8 | 15 | 25 | 39 | 56 | 78 | 104 | 136 |
| 18 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 8 | 12 | 16 | 34 | 52 | 70 | 88 | 110 |
| 19 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 5 | 9 | 15 | 27 | 40 | 58 | 79 | 105 |
| 20 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 8 | 16 | 25 | 36 | 48 | 80 |
| 21 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 1 | 2 | 3 | 9 | 15 | 22 | 33 | 42 | 59 | 77 | 95 |
| 22 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 10 | 20 | 32 | 44 | 64 |
| 23 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 4 | 7 | 13 | 20 | 31 | 45 | 63 |
| 24 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 2 | 6 | 10 | 20 | 30 | 39 | 48 | 70 |
| 25 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 4 | 8 | 14 | 20 | 27 | 41 |
| 26 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 8 | 16 | 24 | 38 |
| 27 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 3 | 6 | 11 | 17 | 26 | 38 |
| 28 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 1 | 0 | 3 | 9 | 16 20 | 26 |
| 29 | | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | | 3 | 7 | 13 | | 28 |
| | | 0 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 8 | 17 |
| 30 31 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 5 | 10 | 15 | 23 |
| 32 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 4 | 9 | 15 |
Table 5: W (4 , Z n , D/n ) for different n and D .
| \ D | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 |
|-------------|------------|---------|------------|------------|------------|------|-------------|-------------|-------------|-------------|------|------|---------|---------|---------|
| 5 6 7 8 9 | | | | | | | | | | | | | | | |
| 11 12 13 | 10 14 | 10 14 | 10 14 | 10 14 | 10 14 | | | | | | | | | | |
| 16 17 18 19 | 15 132 135 | 171 | 15 132 135 | 15 132 135 | 15 132 135 | | | | | | | | | | |
| 20 | 112 | 144 | 190 | | | | | | | | | | | | |
| 21 | 114 | 141 168 | 196 | | | | | | | | | | | | |
| 22 | 84 | 112 140 | 180 | | 231 | | | | | | | | | | |
| 23 | 82 | 108 | 137 | 171 | 209 | 253 | 253 | 253 | 253 | 253 | | | | | |
| 24 | 90 | 112 | 135 | 160 | | | 188 217 248 | 188 217 248 | 188 217 248 | 188 217 248 | | | | | |
| 25 | 56 | 72 | 90 | 130 | 170 | 210 | 252 | 300 | | | | | | | |
| 26 | 52 | 70 | 88 | 116 | 144 | 180 | 216 | 264 | 325 | | | | | | |
| 27 | 52 | 70 | 90 | 114 | 142 | 175 | 211 | 254 | 300 | 351 | | | | | |
| 28 | 37 | 52 | 68 | 91 | 114 | 140 | 168 | 216 | 264 | 312 | 378 | | | | |
| 29 | 40 | 53 | 71 | 91 | 116 | 144 | 175 | 213 | 254 | 300 | 350 | 406 | | | |
| 30 | 27 | 38 | 50 | 66 | 83 | 101 | 120 | 168 | 216 | 264 | 312 | 365 | 435 | 435 | 435 |
| 31 | 32 | 43 | 55 | 75 | 95 | 119 | 147 | 179 | 214 | 256 | 301 | 351 | 405 465 | 405 465 | 405 465 |
| 32 | 23 | 33 | 46 | 60 | 76 | 96 | 116 | 148 | 184 | 224 | 264 | 312 | 364 | 420 | 496 |
Table 6: W (4 , Z n , D/n ) for different n and D .
Table 7: W (5 , Z n , D/n ) for different n and D .
| n \ | D | 0 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 |
|-------|-----|-------|-----|-----|-----|-----|-----|-----|-----|-----|------|------|------|------|------|------|------|------|
| 5 | | 0 0 | 0 | 0 | 0 | 10 | | | | | | | | | | | | |
| 6 | | 0 0 | 0 | 0 | 0 | 4 | 15 | | | | | | | | | | | |
| 7 | | 0 0 | 0 | 0 | 0 | 1 | 6 | 21 | | | | | | | | | | |
| 8 | | 0 0 | 0 | 0 | 0 | 0 | 2 | 10 | 28 | | | | | | | | | |
| 9 | | 0 0 | 0 | 0 | 0 | 0 | 2 | 8 | 18 | 36 | | | | | | | | |
| 10 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 20 | 45 | | | | | | | |
| 11 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 2 | 6 | 15 | 30 | 55 | | | | | | |
| 12 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 9 | 20 | 38 | 66 | | | | | |
| 13 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 6 | 14 | 28 | 48 | 78 | | | | |
| 14 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 4 | 14 | 24 | 54 | 91 | | | |
| 15 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 6 | 9 | 12 | 42 | 72 | 105 | | |
| 16 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 7 | 14 | 30 | 50 | 78 | 120 | |
| 17 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 6 | 14 | 26 | 42 | 66 | 96 | 136 |
| 18 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 3 | 8 | 18 | 29 | 47 | 72 | 106 |
| 19 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 3 | 12 | 24 | 41 | 62 | 91 |
| 20 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 8 | 12 | 16 | 56 |
| 21 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 5 | 13 | 24 | 38 | 54 |
| 22 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 6 | 14 | 24 | 42 |
| 23 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 10 | 21 | 38 |
| 24 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 5 | 12 | 28 |
| 25 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 10 |
| 26 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 | 4 | 13 |
| 27 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 4 | 13 |
| 28 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 0 | 0 0 | 0 | 0 | 2 | 6 |
| 29 | | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | | | 0 | 0 | 0 | 4 |
| 30 | | 0 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 3 |
| 31 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 2 |
| 32 | | 0 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 |
| \ D | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 |
|-----------|--------|---------|--------|--------|--------|------|-------------|-------------|-------------|-------------|------|------|-------------|-------------|-------------|
| 5 6 7 8 9 | | | | | | | | | | | | | | | |
| 11 12 13 | 10 14 | 10 14 | 10 14 | 10 14 | 10 14 | | | | | | | | | | |
| 16 17 18 | 15 153 | 15 153 | 15 153 | 15 153 | 15 153 | | | | | | | | | | |
| 20 | 96 | 136 | 190 | | | | | | | | | | | | |
| 21 | 72 | 117 162 | 210 | | | | | | | | | | | | |
| 22 | 60 | 90 120 | 170 | | 231 | | | | | | | | | | |
| 23 | 56 | 83 114 | | 153 | 198 | 253 | 253 | 253 | 253 | 253 | | | | | |
| 24 | 44 | 65 | 88 | 123 | | | 164 212 276 | 164 212 276 | 164 212 276 | 164 212 276 | | | | | |
| 25 | 20 | 30 | 40 | 90 | 140 | 190 | 240 | 300 | | | | | | | |
| 26 | 24 | 40 | 56 | 84 | 112 | 152 | 192 | 252 | 325 | | | | | | |
| 27 | 24 | 39 | 58 | 81 | 108 | 143 | 182 | 233 | 288 | 351 | | | | | |
| 28 | 10 | 23 | 36 | 55 | 74 | 96 | 120 | 180 | 240 | 300 | | 378 | 378 | 378 | 378 |
| 29 | 11 | 21 | 34 | 50 | 74 | 103 | 134 | 176 | 222 | 276 | 336 | 406 | | | |
| 30 | 8 | 12 | 16 | 24 | 32 | 40 | 48 | 108 | 168 | 228 | 288 | 352 | 435 390 465 | 435 390 465 | 435 390 465 |
| 31 | 5 | 10 | 20 | 33 | 49 | 72 | 100 | 133 | 168 | 215 | 266 | 325 | | | |
| 32 | 2 | 8 | 17 | 26 | 38 | 56 | 76 | 106 | 139 | 179 | 222 | 278 | 338 | 406 | 496 |
Table 8: W (5 , Z n , D/n ) for different n and D .
| null |
[
"Erik Sjöland"
] |
2014-08-05T18:41:01+00:00
|
2014-11-10T19:02:37+00:00
|
[
"math.CO"
] |
Enumeration of three term arithmetic progressions in fixed density sets
|
Additive combinatorics is built around the famous theorem by Szemer\'edi
which asserts existence of arithmetic progressions of any length among the
integers. There exist several different proofs of the theorem based on very
different techniques. Szemer\'edi's theorem is an existence statement, whereas
the ultimate goal in combinatorics is always to make enumeration statements. In
this article we develop new methods based on real algebraic geometry to obtain
several quantitative statements on the number of arithmetic progressions in
fixed density sets. We further discuss the possibility of a generalization of
Szemer\'edi's theorem using methods from real algebraic geometry.
|
1408.1064v2
|
## CONNECTED COMPONENTS OF PRYM EIGENFORM LOCI IN GENUS THREE
## ERWANLANNEAU AND DUC-MANH NGUYEN
ABSTRACT. This paper is devoted to the classification of connected components of Prym eigenform loci in the strata H (2 , 2) odd and H (1 , 1 2) , of the Abelian differential bundle Ω M 3 over M 3 . These loci, discovered by McMullen [Mc06], are GL (2 + , R ) -invariant submanifolds of complex dimension 3 of Ω M g that project to the locus of Riemann surfaces whose Jacobian variety has a factor admitting real multiplication by some quadratic order O D .
It turns out that these subvarieties can be classified by the discriminant D of the corresponding quadratic orders. However there algebraic varieties are not necessarily irreducible. The main result we show is that for each discriminant D the corresponding locus has one component if D ≡ 0 4 , mod 8 , two components if D ≡ 1 mod 8 , and is empty if D ≡ 5 mod 8 .
Surprisingly our result contrasts with the case of Prym eigenform loci in the strata H (1 , 1) (studied by McMullen [Mc07]) which is connected for every discriminant D .
## 1. INTRODUCTION
Since the work of McMullen [Mc03] it has been known that the properties of SL ( 2 R ) -orbit closure of translation surfaces are strongly related to the endomorphisms rings of the Jacobian of the underlying Riemann surfaces (see also [Möl06]). The algebro-geometric approach emphasized by McMullen is to detect affine homeomorphisms of the flat metric on the level of the first homology group as affine homeomorphisms induce self-adjoint endomorphisms of the Jacobian variety.
Recall that an Abelian variety A ∈ A g admits real multiplication by a totally real number field K of degree g over Q if there exists an inclusion K ↪ → End( A ) ⊗ Q such that for any k ∈ K , the action of k is self-adjoint with respect to the polarization of A . Equivalently, End( A ) contains a copy of an order O ⊂ K acting by self-adjoint endomorphism.
## 1.1. Brief facts summary in the genus 2 case. The locus
E 2 = ( { X,ω ) ∈ Ω M 2 : Jac( X ) admits real multiplication with ω as an eigenform } ,
( ) The locus E 2 is actually a (disjoint) union of subvarieties indexed by the discriminants of the orders O ⊂ End(Jac( X )) . Since orders in quadratic fields (quadratic orders) are classified by their discriminant, the unique quadratic order with discriminant D is denoted by O D . We then define plays an important role in the classification of SL(2 , R ) -orbit closures in Ω M 2 . Here A = Jac( X ) ∈ A 2 , K is a real quadratic field, and the endomorphism ring is canonically isomorphic to the ring of homomorphisms of H X, 1 ( Z ) that preserve the Hodge decomposition. The polarization comes from the intersection form J 0 0 J on the homology.
Ω E D = ( { X,ω ) ∈ E 2 : ω is as an eigenform for real multiplication by O } D .
Date : October 21, 2014.
Key words and phrases. Real multiplication, Prym locus, Translation surface.
The subvarieties Ω E D are of interest since they are GL + (2 , R ) -invariant submanifolds of Ω M 2 (see [Mc07, Mc06]). We can further stratified Ω E D by defining Ω E D ( κ ) = Ω E D ∩ H ( κ ) for κ = (2) or κ = (1 1) , . This defines complex submanifolds of dimension 2 and 3 , respectively. Hence Ω E D (2) projects to a union of algebraic curves (Teichmüller curves) in the moduli space M 2 .
- 1.2. Components of Ω E D (1 1) , and Ω E D (2) . It is well known that the set of Abelian varieties A ∈ A 2 admitting real multiplication by O D with a specified faithful representation i : O D → End( A ) is parametrized by the Hilbert modular surface X D := ( H × -H ) / SL( O D ) . In [Mc07], it has been shown that each Ω E D can be viewed as a C ∗ -bundle over a Zariski open subset of X D , and we have
$$\mathcal { E } _ { 2 } = \bigcup _ { D \geq 4, D \equiv 0, 1 \bmod 4 } \Omega E _ { D } \\ \sum \nolimits \ln \alpha \alpha \alpha \beta \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \beta \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \beta \alpha \alpha \alpha \eta \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \eta \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \eta \alpha \alpha \alpha \eta \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \beta \alpha \alpha \alpha \xi \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \tag \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \appa \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \ alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \xi \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \tag \alpha \alpha \alpha \tag \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \tag \alpha \alpha \alpha \eta \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \tag \alpha \alpha \alpha \appa \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \tag \tag \alpha \tag \tag \tag \tag \tag \tag \tag \tag \tag$$
In particular Ω E D is a connected, complex suborbifold of Ω M 2 of dimension 3 . The fact that there is only one (connected) eigenform locus for each D follows from the fact that there is only one faithful, proper, self-adjoint representation i : O D → M 4 ( Z ) up to conjugation by Sp(4 , Z ) (see [Mc07] Theorem 4.4). It follows that Ω E D (1 1) , is a Zariski open set in Ω E D . In particular Ω E D (1 1) , is connected for any quadratic discriminant D .
The classification of components of Ω E D (2) has also been obtained by McMullen [Mc05].
- 1.3. Higher genera. In [Mc06] it is shown that analogues of Ω E D exist in higher genus (up to 5 ). These subvarieties of Ω M g are called Prym eigenform loci . Surfaces in a Prym eigenform locus are pairs ( X,ω ) such that there exists a holomorphic involution τ : X → X such that g X ( ) -g Y ( ) = 2 , where Y = X/ τ 〈 〉 , τ ∗ ω = -ω , and the Prym variety Prym ( X,τ ) admits a real multiplication with ω an eigenform (see Section 2 for precise definitions). Note that the condition g X ( ) -g Y ( ) = 2 is needed for our discussion. For any genus, the set of Prym eigenforms whose Prym variety admits a multiplication by O D will be denoted by Ω E D , and the intersection of Ω E D with a stratum H ( κ ) is denoted by Ω E D ( κ ) .
The goal of this paper is to investigate the topology of the Prym eigenform loci in the strata H (2 2) , and H (1 1 2) , , of Ω M 3 . It is well known that the stratum H (2 2) , also has two components H (2 2) , odd and H (2 2) , hyp . It is not difficult to see that Prym eigenforms in H (2 2) , hyp are double covers of surfaces in Ω E D (2) (see Proposition 2.3). Thus dimΩ E D (2 2) , hyp = 2 , and Ω E D (2 2) , hyp is a (finite) union of GL (2 + , R ) closed orbits. On the other hand, we have dimΩ E D (2 2) , odd = 3 . The stratum H (1 1 2) , , is connected and we also have dimΩ E D (1 1 2) = 3 , , (see Proposition 2.6).
Our main result reveals that the situation in genus three is quite different from the one in genus. More precisely:
Theorem A. Let κ ∈ { (2 2) , odd , (1 1 2) , , } . For any discriminant D ≥ 8 , with D ≡ 0 1 , mod 4 , the loci Ω E D ( κ ) are non empty if and only if D ≡ 0 1 4 , , mod 8 , and in this case they are pairwise disjoint. Moreover the following dichotomy holds:
- (1) If D is even then Ω E D ( κ ) is connected,
- (2) If D is odd then Ω E D ( κ ) has exactly two connected components.
Remark 1.1. One of the main differences between the cases of genus two and genus three is that the polarization of the Prym variety in genus three has the form ( J 0 0 2 J ) , which is the reason why Ω E D ( κ ) is empty if D ≡ 5 mod 8 .
In genus two we have Prym ( X,τ ) = Jac( X ) and the Prym involution τ must be the hyperelliptic involution which is unique, in genus three Prym ( X,τ ) is only a factor of Jac( X ) , and there may be more than one Prym involution as we will see in Section 3. Thus it is not obvious that one can use the discriminant to distinguish different Prym eigenform loci.
It is also worth noticing that while Ω E 9 (4) and Ω E 16 (4) are empty (see [LN11] ), the loci Ω E 9 ( κ ) and Ω E 16 ( κ ) do exist for κ ∈ { (2 2) , odd , (1 1 2) , , } .
## 1.4. Triple tori. An important tool in our proof is the use of triple tori :
- (1) We say that ( X,ω ) ∈ Ω E D (2 2) , odd admits a three tori decomposition if there exists a triple of homologous saddle connections { σ , σ 0 1 , σ 2 } on X joining the two distinct zeros of ω .
If ( X,ω ) admits a three tori decomposition then it can be viewed as a connected sum of three slit tori ( X ,ω j j ) , j = 0 1 2 , , , (see Figure 1). We will always assume that X 0 is preserved by the Prym involution τ and X ,X 1 2 are exchanged by τ .
- (2) We say that ( X,ω ) ∈ Ω E D (1 1 2) , , admits a three tori decomposition if there exist two pairs of homologous saddle connections { σ , σ 0 1 } and { σ , σ ′ 0 ′ 1 } on X joining the double zero to the simple zeros such that { σ , σ ′ 0 ′ 1 } = τ ( { σ , σ 0 1 } ) .
1
FIGURE 1. Decomposition of ( X,ω ) ∈ Prym (2 2) , odd (left) and ( X,ω ) ∈ Prym (1 1 2) , , (right) into three tori.
As a corollary of our main result, we prove the following theorem, which is used in the paper [LN13b]:
/negationslash
Theorem B. For any discriminant D such that Ω E D ( κ ) = ∅ , there exist in any component of Ω E D ( κ ) surfaces which admit three-tori decompositions.
Theorem B is proved in Section 7.
- 1.5. Strategy of the proof. The important ingredients of the proof of the main theorem is the use of surgeries (see Section 5). The core of Theorem A are Theorem 6.1 on admissible saddle connection and Theorem 4.1 on non-connectedness. The proofs of Theorem 6.1 and Theorem 4.1 appear in Section 6 and 4 respectively.
- (1) An elementary way to get Prym eigenforms in H (2 2) , odd and H (1 1 2) , , is given by Lemma 2.4. Another way is the use of the surgery 'Breaking up a zero" on a Prym eigenforms in H (4) (see [KZ03]). We deduce that Ω E D (2 2) , odd and Ω E D (1 1 2) , , are non-empty whenever Ω E D (4) is non-empty.
- (2) In Section 3 we prove that the loci Ω E D ( κ ) are pairwise disjoint (Lemma 3.3 and Theorem 3.1). As we have noticed in Remark 1.1 a surface ( X,ω ) ∈ H (2 2) , odd may have more than one Prym involution. However we then show that all Prym varieties admit real multiplication by O 9 . This proves in particular that Ω E 9 (2 2) , odd is non-empty despite the fact that Ω E 9 (4) = ∅ .
- (3) To get an upper bound of the number of components of Ω E D ( κ ) our strategy is to find in each component C of Ω E D ( κ ) a surface ( X,ω ) such that we can collapse the zeros of ω along some saddle connections to get a surface in Ω E D (4) . Such saddle connections are called admissible (see Section 5). In this situation, the component C is adjacent to the locus Ω E D (4) i.e. Ω E D (4) ⊂ C . We then prove that the number of components of Ω E D ( κ ) that are adjacent to Ω E D (4) can not exceed the number of components of Ω E D (4) . Surprisingly, it turns out that there exist components that are not adjacent to Ω E D (4) . These are precisely the components of the loci Ω E 9 ( κ ) and Ω E 16 ( κ ) . This fact is proved in Theorem 6.1. This result plus the fact that Ω E D (4) has at most two connected components (by [LN11]) furnish the desired upper bound.
- (4) Finally, to get the exact count of the number of components we will show in Section 4 that if Ω E D (4) is not connected then Ω E D ( κ ) is not connected either (in contrast with the situation in genus two). This difference comes from the invariant defined in [LN11].
Acknowledgements. We would like to thank Alex Eskin, Martin Möller, and Barak Weiss for useful discussions. We also thank Alex Wright for useful discussions and comments on an earlier version of this text. We would also thank Université de Bordeaux and Institut Four ier in Grenoble for the excellent working condition during the preparation of this work. Some of the research visits which made this collaboration possible were supported by the ANR Project GeoDyM. The authors are partially supported by the ANR Project GeoDyM.
## 2. BACKGROUND
We review the necessary tools and results involved in the proof of our main result. For an introduction to translation surfaces in general, and a nice survey on this topic, see e.g. [Zor06, MaTa02].
- 2.1. Prym eigenform. Let X be a Riemann surface and τ an involution of X . We define the Prym variety of ( X,τ ) to be
$$\Pr y m ( X, \tau ) = ( \Omega ( X, \tau ) ^ { - } ) ^ { * } / H _ { 1 } ( X, \mathbb { Z } ) ^ { - }$$
where Ω( X,τ ) -= ker( τ + id) ⊂ Ω( X ) , Ω( X ) is the space of holomorphic one forms on X ), and H X, 1 ( Z ) -is the anti-invariant homology of X with respect to τ . Remark that Prym ( X,τ ) has naturally a polarization: the lattice H X, 1 ( Z ) -is equipped with the restriction of the intersection form on H X, 1 ( Z ) .
Following [Mc06] we will call a translation surface ( X,ω ) a Prym form if there exists an involution τ of X such that dim C Ω( X,τ ) -= 2 , and ω ∈ Ω( X,τ ) -. Note that the condition dim C Ω ( -X,τ ) = 2 is equivalent to g X ( ) -g Y ( ) = 2 , where Y := X/ τ 〈 〉 . In this situation, we will call τ a Prym involution of X . Note that a Riemann surface may have more than one Prym involution (see Theorem 3.1).
Recall that a (real) quadratic order is a ring isomorphic to Z [ x / x ] ( 2 + bx + ) c , the discriminant of the order is defined by D = b 2 -4 c . Orders with the same discriminants are isomorphic. Thus for any
D ∈ N , D ≡ 0 1 , mod 4 , we will write O D to designate the unique quadratic order of discriminant D . When D is not a square, O D is a finite index subring of the integer ring in the quadratic field K := Q ( √ D ) .
Let A be an Abelian variety of (complex) dimension 2 . We say that A admits a real multiplication by O D if there exists an injective ring morphism i : O D → End( A ) such that i ( O D ) is a self-adjoint proper subring of End( A ) (properness means that if f ∈ End( A ) and there exists n ∈ N ∗ such that nf ∈ i ( O D ) , then f ∈ i ( O D ) ).
Definition 2.1. We will call a translation surface ( X,ω ) a Prym eigenform , if there exists a Prym involution τ of X such that
- · Prym ( X,τ ) admits a real multiplication by some quadratic order O D ,
- · ω ∈ Ω( X,τ ) -is an eigenvector of O D .
The set of Prym eigenforms admitting real multiplication by O D will be denoted by Ω E D . In [Mc05] it is showed that Ω E D are closed, GL (2 + , R ) -invariant submanifolds of the bundle Ω M g . Up to now, these are the only known GL (2 + , R ) -invariant submanifolds of Ω M g which are neither closed orbits nor covers of Abelian differentials or quadratic differentials in lower genus. The intersection of Ω E D with a stratum H ( κ ) will be denoted by Ω E D ( κ ) . Clearly, Ω E D ( κ ) are GL (2 + , R ) -invariant submanifolds of H ( κ ) .
McMullen [Mc06] proved the existence of Prym eigenforms in genus 3 and 4 , and in particular that Ω E D (4) and Ω E D (6) are non-empty for infinitely many D . It is well known that the minimal stratum H (4) of Ω M 3 has two components H hyp (4) and H odd (4) (see [KZ03] for precise definitions) and the loci Ω E D (4) are contained in H odd (4) . In [LN11] the authors gave a complete classification of Ω E D (4) , namely:
Any translation surface in genus two is a Prym form (the Prym involution being the hyperelliptic involution). It turns out that the locus E 2 of Prym eigenforms in genus two is a disjoint union of Ω E D for D ≡ 0 1 , mod 4 and D ≥ 5 . It is a fact that Ω E D (2) is connected if D ≡ 0 4 5 , , mod 8 , and has two components otherwise ( D ≡ 1 mod 8 ). On the other hand Ω E D (1 1) , is connected for all D (see [Mc07, Mc05]).
Theorem 2.2 (Lanneau-Nguyen [LN11]) . For D ≥ 17 Ω , E D (4) is non empty if and only if D ≡ 0 1 4 , , mod 8 . All the loci Ω E D (4) are pairwise disjoint. Moreover, for the values 0 1 4 , , of discriminants, the following dichotomy holds. Either
- (1) D is odd and then Ω E D (4) has exactly two connected components, or
- (2) D is even and Ω E D (4) is connected.
In addition, each connected component of Ω E D (4) corresponds to a closed GL (2 + , R ) -orbit.
For D < 17 , only Ω E 12 (4) and Ω E 8 (4) are non-empty, each of which consists of a unique closed GL (2 + , R ) -orbit.
The first striking fact about Prym eigenforms in H (4) is that Ω E D (4) = ∅ if D ≡ 5 mod 8 , this is actually due to the signature of the polarization of the Prym variety in genus three which is different from the one in genus two. The second remarkable fact is that Ω E 9 (4) = Ω E 16 (4) = ∅ even though 9 and 16 are not 'forbidden values' of D . It is worth noticing that even though we have the same statement in the case D ≡ 1 mod 8 as McMullen's result in Ω E D (2) (namely, Ω E D (4) has two components), the reason for this disconnectedness is different. Roughly speaking, the two components of Ω E D (4) correspond to two distinct complex lines in the space Ω( X,τ ) -/similarequal H 1 ( X, R ) -, whereas in the case Ω E D (2) , the two components correspond to the same complex line (this is actually a consequence of the fact that Ω E D (1 1) , is connected), they can only be distinguished by the spin invariant (see [Mc05, Section 5]).
2.2. Prym eigenforms in H (2 2) , and H (1 1 2) , , . The stratum H (1 1 2) , , is connected while the stratum H (2 2) , has two connected components: H (2 2) , hyp and H (2 2) , odd (see [KZ03]). We will not use this classification in the sequel.
Proposition 2.3. If ( X,ω ) ∈ Ω E D (2 2) , hyp then there exists a Prym eigenform ( X , ω ′ ′ ) ∈ Ω E D ′ (2) and an unramified double cover ρ : X → X ′ such that ρ ω ∗ ′ = ω . In particular Ω E D (2 2) , hyp is a finite union of GL (2 + , R ) closed orbits, each of which is a copy of a GL (2 + , R ) -orbit in Ω E D ′ (2) .
Proof of Proposition 2.3. By definition X is a hyperelliptic Riemann surface, and the hyperelliptic involution ι exchanges the zeros of ω . Since ι commutes with all automorphisms of X , we have τ ′ := τ ◦ ι is also an involution of X (where τ is the Prym involution of X ). Set X ′ := X/ τ 〈 ′ 〉 . Note that ker( τ ′ -id) = ker( τ +id) , thus we have dimker( τ ′ -id) = 2 and X ′ is a Riemann surface of genus two. Let ρ : X → X ′ be the associated double cover. Using Riemann-Hurwitz formula, it is easy to see that ρ is unramified. Since τ ′ ∗ ω = ω , there exists a holomorphic one-form ω ′ on X ′ such that ρ ω ∗ ′ = ω . Since ( X,ω ) ∈ H (2 2) , and ρ is unramified, we conclude that ( X , ω ′ ′ ) ∈ H (2) .
Remark that ρ ∗ is an isomorphism between ker( τ ′ -id) = ker( τ ′ + id) and Ω( X ′ ) , and ρ maps H X, 1 ( Z ) -to a sublattice of index two in H X , 1 ( ′ Z ) , therefore ρ induces a two-to-one covering from Prym ( X,τ ) to Jac( X ′ ) . By assumption Prym ( X,τ ) admits a real multiplication by the order O D for which ω is an eigenvector. It follows that Jac( X ′ ) also admits a real multiplication by O D ⊗ Q for which ω ′ is an eigenvector. Thus there exists a discriminant D ′ satisfying D D ′ | such that ( X , ω ′ ′ ) ∈ Ω E D ′ (2) . This shows the first part of the proposition.
↦
But we know from [Mc07] that Ω E D ′ (2) is a union of GL (2 + , R ) closed orbits, and since the map ( X,ω ) → ( X , ω ′ ′ ) is clearly GL (2 + , R ) -equivariant, it follows that ( X,ω ) belongs to a GL (2 + , R ) -closed orbit. Since any Riemann surface X in M g admits only finitely may unramified double covers, we derive that there are only finitely many closed orbits in Ω E D (2 2) , hyp . The proposition is then proved. /square
Because of Proposition 2.3, in the rest of the paper we will focus on Ω E D (2 2) , odd and Ω E D (1 1 2) , , . Observe that if ( X,ω ) ∈ Prym (2 2) , odd then the Prym involution τ exchanges the two zeros of ω , and if ( X,ω ) ∈ Prym (1 1 2) , , then τ exchanges the simple zeros ω .
The next lemma provides us with examples of Prym eigenforms in Prym (2 2) , odd and Prym (1 1 2) , , .
Lemma 2.4 (Real multiplication by O D ) . Let ( w,h,e ) ∈ Z 3 be an integral vector satisfying
$$\begin{cases} \ w > 0, h > 0, \gcd ( w, h, e ) = 1, \\ D = e ^ { 2 } + 8 w h, \ e + \sqrt { D } > 0. \end{cases}$$
Let λ := e + √ D 2 . Note that λ 2 = eλ +2 wh . We denote by S κ ( w,h,e ) the surface defined in Figure 2 below. Then
$$S _ { \kappa } ( w, h, e ) \in \Omega E _ { D } ( \kappa ).$$
Proof of Lemma 2.4. Each surface in Figure 2 is a connected some of three slit tori, and admits an involution τ which fixes one torus and exchanges the other two (see also Section 1.4). It is not difficult to see that τ is a Prym involution, and that S κ ( w,h,e ) ∈ Prym ( κ ) . Let ( X,ω ) be one of the surfaces in Figure 2. Let X 0 be the torus invariant by τ , and X ,X 1 2 be the other two tori. By construction, there are bases ( a , b i i ) of H X , 1 ( i Z ) , i = 0 1 2 , , , such that
S (2 2) , ( w,h,e ) ∈ Ω E D (2 2) , odd
∈
FIGURE 2. Real multiplication by O D
- · τ ( a 0 ) = -a , τ 0 ( b 0 ) = -b 0 , τ ( a 1 ) = -a , τ 2 ( b 1 ) = -b 2 ,
- · For i = 1 2 , one has ω a ( i ) = w and ω b ( i ) = ıh .
- · ω a ( 0 ) = λ, ω ( b 0 ) = ıλ ,
Set a = a 1 + a , b 2 = b 1 + b 2 . Then { a , b 0 0 , a, b } is a symplectic basis of H X, 1 ( Z ) -in which the intersection form is given by the matrix ( J 0 0 2 J ) . Let T be the endomorphism of H X, 1 ( Z ) which is given in the basis ( a , b 0 0 , a, b ) by the matrix
$$T = & \left ( \begin{array} { c } e \text{Id} _ { 2 } & \left ( \begin{array} { c } 2 w & 0 \\ 0 & 2 h \end{array} \right ) \\ \left ( \begin{array} { c } h & 0 \\ 0 & w \end{array} \right ) & 0 \end{array} \right ). \\ \text{ection form on } H _ { 1 } ^ { - } ( X, \mathbb { Z } ) \text{ is given by}$$
Since the restriction of the intersection form on H -1 ( X, Z ) is given by ( J 0 0 2 J ) , it is easy to check that T is self-adjoint with respect this form. Note that in this basis ω is given by the vector ( λ, ıλ, 2 w, 2 hı ) , therefore we have T ω ∗ = λω . It follows that T ∈ End( Prym ( X,τ )) . Since T satisfies T 2 = eT + 2 wh Id 4 , T generates a self-adjoint proper subring of End( Prym ( X,τ )) isomorphic to O D for which ω is an eigenvector. Thus S κ ( w,h,w ) ∈ Ω E D ( κ ) . The lemma is proved. /square
Corollary 2.5. For any D ≥ 8 , D ≡ 0 1 4 , , mod 8 , the loci Ω E D (2 2) , odd and Ω E D (1 1 2) , , are non-empty.
$$P r o o f o f C o r o l l a r y \, 2. 5. \, \text{ Apply Lemma } 2. 4 \, \text{ for some } ( w, h, e ) \in \mathbb { Z } \, \text{ with } D = e ^ { 2 } + 8 w h. \quad \Box$$
2.3. Kernel foliation. To investigate the topology of these loci we first recall the notion of the kernel foliation. Let ( X,ω ) ∈ H ( κ ) be a translation surface. In a neighborhood of ( X,ω ) the kernel foliation leaf of ( X,ω ) consists of surfaces having the same absolute periods as ( X,ω ) and the relatives periods slightly different. This foliation has already appeared in several papers (see for example [EMZ03, Cal04, MZ08, HMSZ, MW14].
For κ ∈ { (2 2) , odd , (1 1 2) , , } , the intersection with the kernel foliation leaves gives rise to a foliation of the Prym eigenform loci Ω E D ( κ ) , the leaves of this foliation have complex dimension one. Constructions of surfaces in the intersection of the kernel foliation and Prym eigenform loci can be found in [LN13a, LN13b]. Since the leaves of this foliation has dimension one, for any ( X,ω ) ∈ Ω E D ( κ ) , we can use the notation ( X,ω )+ w , with w ∈ C and | w | small, to denote surfaces in the same leaf and close to ( X,ω ) (see [LN13b], Section 3). Moreover up to the action of GL (2 + , R ) , a neighborhood of ( X,ω ) in Ω E D ( κ ) consists of surfaces in the same kernel foliation leaf as ( X,ω ) . Namely, we have
Proposition 2.6 ([LN13a], Corollary 3.2) . Let ( X , ω ′ ′ ) ∈ Ω E D ( κ ) close enough to ( X,ω ) ∈ Ω E D ( κ ) . Then there exists a unique pair ( g, w ) , where g ∈ GL (2 + , R ) close to id , and w ∈ C with | w | small, such that ( X , ω ′ ′ ) = g · (( X,ω ) + w ) . In particular, we have dimΩ E D ( κ ) = 3 .
## 3. UNIQUENESS
/negationslash
In genus two Prym involution and hyperelliptic involution coincide, so it is unique. In higher genus, surfaces may have more than one Prym involution (see e.g. the Appendix). In [LN11], we showed that if ( X,ω ) ∈ Ω E D (4) then the Prym involution is unique. This is no longer true in H (2 2) , odd , nevertheless, if a surface in H (2 2) , odd has two Prym involutions, then both Prym varieties admit real multiplication by O 9 as we will see in Theorem 3.1. It follows in particular that if D 1 = D 2 and κ ∈ { (2 2) , odd , (1 1 2) , , } then Ω E D 1 ( κ ) ∩ Ω E D 2 ( κ ) = ∅ .
/negationslash
If ( X,ω ) ∈ H (1 1 2) , , , then there exists at most one Prym involution τ such that τ ∗ ω = -ω .
Theorem 3.1. Let ( X,ω ) ∈ H (2 2) , odd be a surface having two Prym involutions τ 1 = τ 2 such that τ ∗ 1 ω = τ ∗ 2 ω = -ω . Then there exist i 1 : O 9 → End(Prym( M,τ 1 )) and i 2 : O 9 → End(Prym( M,τ 2 )) such that i i ( O 9 ) is a self-adjoint proper subring of End( Prym ( X,τ i )) , and ω is an eigenform for both subrings i 1 ( O 9 ) and i 2 ( O 9 ) . In particular ( X,ω ) ∈ Ω E 9 (2 2) , odd .
Corollary 3.2. For κ ∈ { (2 2) , odd , (1 1 2) , , } , if D 1 = D 2 then Ω E D 1 ( κ ) ∩ Ω E D 2 ( κ ) = ∅ .
/negationslash
/negationslash
Proof of Corollary 3.2. Let ( X,ω ) ∈ Ω E D 1 ( κ ) ∩ Ω E D 2 ( κ ) . Let τ 1 and τ 2 be the corresponding Prym involutions of X . If τ 1 = τ 2 then by Theorem 3.1 one has κ = (2 2) , odd and D 1 = D 2 = 9 . If τ 1 = τ 2 then Lemma 3.3, applied to A = Prym ( X,τ 1 ) = Prym ( X,τ 2 ) , gives the uniqueness of the self-adjoint proper subring O ⊂ End( A ) , which implies D 1 = D 2 . The corollary is then proved. /square
We will need the following two elementary lemmas. The first one proves the uniqueness of the proper subring, once the Prym variety and the eigenform are given (see also [LN11, Section 5 ] for related results).
/negationslash
Lemma 3.3. Let A be an Abelian variety of dimension two. We regard A as a quotient C 2 /L , where L is a lattice isomorphic to Z 4 equipped with a non-degenerate skew-symmetric inner product 〈 , 〉 : L × L → Z which is compatible with the complex structure. Let v = 0 be a vector in C 2 . Assume that there exists a self-adjoint endomorphism ϕ of A such that ϕ v ( ) = λv , with λ ∈ R , ϕ = λ · Id . Then there exists a unique discriminant D and a unique self-adjoint proper subring O of End( A ) isomorphic to O D for which v is an eigenvector.
/negationslash
Proof. Let S = C · v be the complex line generated by v , and let S ′ denote the orthogonal complement of S with respect to 〈 , 〉 in C 2 . Note that S ′ is also a complex line in C 2 . Set w = ıv . Since ϕ is an endomorphism of A , we have ϕ w ( ) = iϕ v ( ) = λw . In other words ϕ | S = λ · id S . Since ϕ is self-adjoint, it also preserves the complex line S ′ . Thus ϕ | S ′ = λ ′ · id S ′ where λ ′ = λ .
/negationslash
Let K v be the subring of End( A ) ⊗ Q consisting of self-adjoint endomorphisms of A for which v is an eigenvector. For any f ∈ K v , the matrix of f in the decomposition C 2 = S ⊕ S ′ has the form ( λ f ( ) 0 0 λ ′ ( f ) ) .
Since the self-adjoint endomorphism ϕ preserves the lattice L its minimal polynomial χ ϕ ∈ Z [ X ] has degree 2 . By definition λ is a real root of χ ϕ . Hence λ ′ , that is a root, is also real. Moreover, since v is an eigenvector of ϕ , up to a real scalar, all the coordinates of v in a basis of L belong to K = Q ( λ ) . Remark that either K = Q , or K ⊂ R and [ K : Q ] = 2 .
↦
/negationslash
↦
We claim that K v is either isomorphic to K or to Q 2 . To see this, it suffices to notice that each element of K v is uniquely determined by its eigenvalues on S and S ′ . If λ ∈ Q then we can assume that all the coordinates of v belong to Q , hence both λ f ( ) and λ ′ ( f ) belong to Q as f is defined of over Q . Thus Λ : K v → Q 2 , f → ( λ f ( ) , λ ′ ( f )) is an isomorphism of Q vector spaces. If λ ∈ Q , then λ ∈ K = Q ( √ d ) , with d ∈ N , d is not a square. It follows that λ ′ ( f ) is the Galois conjugate of λ f ( ) in K . Consequently, Λ : K v → K, f → λ f ( ) is an isomorphism of Q -vector spaces.
/negationslash
Lemma 3.4. Let ( X,ω ) ∈ H (2 2) , /unionsq H (1 1 2) , , , and τ , τ 1 2 be two Prym involutions of X such that τ ∗ 1 ω = τ ∗ 2 ω = -ω .
Set O = K v ∩ End( A ) . By definition, O is the unique self-adjoint proper subring of End( A ) for which v is an eigenvector. It remains to show that O /similarequal Z [ X / X ] ( 2 + bX + ) c for some b, c ∈ Z , such that b 2 -4 c > 0 . We have dim Q K v = 2 . For any f ∈ O such that ( f, id) is a basis of K v as a Q -vector space, we denote by ∆( ) f the discriminant of the minimal polynomial of f . Note that we have ∆( ) f > 0 , and ∆( ) = ( f λ f ( ) -λ ′ ( f )) 2 . Set D = min ∆( { f ) : f ∈ O , ( f, id) is a basis of K v } . Let ψ be an element of O such that ∆( ψ ) = D . Let us show that O = Z ψ + Z id . Indeed, let f be an element of O , then we can write f = xψ + id y , with ( x, y ) ∈ Q 2 , and ∆( ) = f x D 2 . If x ∈ N , by replacing ψ by f -[ x ψ ] , we can find ψ ′ ∈ O such that 0 < ∆( ψ ′ ) < D , therefore we must have x ∈ N . It follows that y ∈ N . Finally, since ψ ∈ End( A ) , the minimal polynomial of ψ has the the form ψ 2 + bψ + id c , with b, c ∈ Z such that D = b 2 -4 c . The proof of the lemma is now complete. /square
- (a) If ( X,ω ) ∈ H (1 1 2) , , then τ 1 = τ 2 .
/negationslash
Proof of Lemma 3.4. Set τ = τ 1 ◦ τ 2 , one has τ ∗ ω = ω and τ fixes all the zeros of ω . We identify the neighborhood of a double zero P of ω with the unit disk ∆ ⊂ C such that P is mapped to 0 . In this local chart, ω = z 2 dz . If ω ∈ H (1 1 2) , , then P is the unique double zero of ω , therefore τ 1 ( P ) = τ 2 ( P ) = P . Since both τ , τ 1 2 are involutions, their restrictions to this neighborhood of P read τ i ( z ) = -z . Thus τ ( z ) = z , which implies that τ = id X and τ 1 = τ 2 .
- (b) If ( X,ω ) ∈ H (2 2) , odd and τ 1 = τ 2 , then there exists a branched cover p : X → Y of degree three, where Y is a torus, which is ramified only at the zeros of ω and satisfies ω = p ξ ∗ , where ξ is a holomorphic one-form on Y . Moreover, the involutions τ , τ 1 2 descend to the unique involution of Y which acts by -id on the homology and exchanges the images by p of the zeros of ω .
/negationslash
/negationslash
$$- 4 & = 2 - 2 \cdot \text{genus} ( X ) = 3 \cdot ( 2 - 2 \cdot \text{genus} ( Y ) ) - \sum _ { x \in X } ( e _ { p } ( x ) - 1 ) \leq - \sum _ { x \in X } ( e _ { p } ( x ) - 1 ) \leq - 4 \\ \text{immlices that some} ( V ) - 1 \text{ and the two zero of} \, ( i + 1 ) \, \text{are the only branch} \, \, \text{in} \, \text{of} \, n \text{ } \text{ $line} \, \tau ^ { * }, i - 1, \text{ the}$$
From now on, we will assume that ω ∈ H (2 2) , odd and τ 1 = τ 2 . The restriction of τ to the local chart of P (defined above) can be written as τ ( z ) = ζz . Since τ ∗ ω = ω , it follows that ζ 3 = 1 . Obviously ζ = 1 , otherwise τ is the identity map in a neighborhood of P and hence it is the identity on X implying τ 1 = τ 2 . Let p : X → Y = X/ τ 〈 〉 be the quotient map. Since τ has order three, p is a ramified covering of degree 3 . Clearly, the two zeros of ω are branched points of p of order 3 . Moreover since dimker( τ -id) ≥ 1 , one has genus( Y ) ≥ 1 . These two facts, combined with the Riemann-Hurwitz formula implies that genus( Y ) = 1 and the two zeros of ω are the only branched points of p . Since τ ∗ ω = ω , the form ω descends to a holomorphic 1 -form ξ on Y i.e. ω = p ξ ∗ .
Now the subgroup of Aut( X ) generated by τ , namely { id , τ 1 ◦ τ , τ 2 2 ◦ τ 1 } , is clearly invariant by the conjugations by τ 1 and τ 2 . Therefore τ 1 and τ 2 induces two involutions, say ι 1 and ι 2 , on Y . Since τ i permutes the zeros of ω , the equality τ ∗ i ω = -ω reads ι ∗ i ξ = -ξ , and ι i exchanges the images of the zeros
of ω by p . Now Y is a torus: there exists only one such involution. Hence ι 1 = ι 2 . The lemma is then proved. /square
Proof of Theorem 3.1. From Lemma 3.4 it is sufficient to assume that ( X,ω ) ∈ H (2 2) , odd . For simplicity we continue with the notions of the previous lemma. Let P, Q ∈ X denote the zeros of ω . We claim that there exists a basis of H Y, 1 ( Z ) given by a pair of simple closed geodesics { α, β } that are invariant by ι and that do not contain p P ( ) and p Q ( ) . Indeed one can pick a fixed point of ι and take a pair of simple closed geodesics passing through this point and missing the point p P ( ) and p Q ( ) = ι ( p P ( )) .
The next goal is to construct a symplectic basis of X and a self-adjoint endomorphism of Prym ( X,τ 1 ) . We lift the curves α β , to X in the following manner. Let R ∈ Y be the (unique) intersection point of α with β . Hence ι ( R ) = R and R is a regular point of the covering p . Since p ◦ τ 1 = ι ◦ p , the involution τ 1 induces a permutation (of order two) of p -1 ( R ) = { R , R , R 0 1 2 } . Therefore τ 1 fixes some point, say R 0 .
/negationslash
Let α 0 (respectively, β 0 ) be the pre-image of α (respectively, β ) passing through R 0 . For j = 1 2 , let α j = τ j ( α 0 ) , β j = τ j ( β 0 ) , where τ = τ 1 ◦ τ 2 . By construction:
Hence ( α , β 0 0 , α 1 , β 1 , α 2 , β 2 ) is a symplectic basis of H X, 1 ( Z ) . This allows us to construct a symplectic basis ( a , b 0 0 , a 1 , b 1 ) of H X, 1 ( Z ) -= ker( τ 1 +id) as usual:
$$\begin{cases} \ \langle \alpha _ { i }, \alpha _ { j } \rangle = \langle \beta _ { i }, \beta _ { j } \rangle = 0 & \text{for} \quad i, j \in \{ 0, 1, 2 \}, i \neq j, \\ \ \langle \alpha _ { i }, \beta _ { j } \rangle = \delta _ { i j } & \text{for} \quad i, j \in \{ 0, 1, 2 \}. \end{cases} \\ \alpha _ { i } \ \beta _ { i } \ \alpha _ { i } \ \beta _ { i } \ \rangle \text{is a symmlectic basis of } H _ { i } ( X \ \mathbb { V } ) \ \text{This allows us to construct}$$
$$\begin{cases} \ a _ { 0 } = \alpha _ { 0 } & b _ { 0 } = \beta _ { 0 } \\ \ a _ { 1 } = \alpha _ { 1 } + \alpha _ { 2 } & b _ { 1 } = \beta _ { 1 } + \beta _ { 2 } \end{cases}$$
The intersection form is given by the matrix ( J 0 0 2 J ) . One can normalize by using GL (2 + , R ) so that ξ α ( ) = 1 and ξ β ( ) = ı ∈ C . Then ω (viewed as an element of H 1 ( X, C ) -) is the vector (in the basis dual to ( a , b 0 0 , a 1 , b 1 ) )
$$( \omega ( a _ { 0 } ), \omega ( b _ { 0 } ), \omega ( a _ { 1 } ), \omega ( b _ { 1 } ) ) = ( 1, i, 2, 2 i ).$$
It is straightforward to check that ( X,ω ) coincide with S (2 2) , (1 1 , , -1) ∈ Ω E 9 (2 2) , .
$$T = \left ( \begin{smallmatrix} - \text{id} _ { 2 } & 2 \cdot \text{id} _ { 2 } \\ \text{id} _ { 2 } & 0 \end{smallmatrix} \right ) _ { ( a _ { 0 }, b _ { 0 }, a _ { 1 }, b _ { 1 } ) } \\ T \text{ ie cell} _ { \text{adioint with recent to the r} }$$
Let us consider the matrix
It is straightforward to check that T is self-adjoint with respect to the restriction of the intersection form on H M, 1 ( Z ) -. Moreover (1 , ı, 2 2 ) , ı · T = (1 , ı, 2 2 ) , ı thus ω is an eigenform for T : hence ( X,ω ) ∈ Ω E D (2 2) , odd . Since T 2 = -T +2 , the endomorphism T generates a proper subring of End(Prym( M,τ 1 )) isomorphic to O D where D = 1 + 4 · 2 = 9 . By Lemma 3.3, this subring is unique.
The same argument shows that End(Prym( M,τ 2 )) also contains a unique self-adjoint proper subring isomorphic to O 9 , for which ω is an eigenform. The proof of the theorem is now complete. /square
## 4. NON CONNECTEDNESS OF Ω E D ( κ )
In this section we will show that when D ≡ 1 mod 8 , the number of components of Ω E D (2 2) , odd and Ω E D (1 1 2) , , is at least two. It is worth noticing that this is not true in genus two, namely, Ω E D (2) has two connected components, while Ω E D (1 1) , is connected (see [Mc05]).
Theorem 4.1. For any D ≥ 9 satisfying D ≡ 1 mod 8 , the loci Ω E D (2 2) , odd and Ω E D (1 1 2) , , are not connected.
Proof of Theorem 4.1. First of all by Corollary 2.5, Ω E D ( κ ) is non-empty. Before going into the details, we first explain why Ω E D (4) is not connected when D ≡ 1 mod 8 (see [LN11, Theorem 6.1]). For every surface ( X,ω ) ∈ Ω E D (4) we denote by S the subspace of H 1 ( X, R ) -= ker( τ + Id) ⊂ H 1 ( X, R ) generated by { Re( ω , ) Im( ω ) } and by S ′ the orthogonal complement of S with respect to the intersection form in H 1 ( X, R ) -. By definition there is an injective ring morphism i : O D → End( Prym ( X,τ )) such that i ( O D ) is a self-adjoint proper subring of End( Prym ( X,τ )) , and for any T ∈ i ( O D ) , S is an eigenspace of T . It turns out that an element T ∈ Im( ) i is uniquely determined by its minimal polynomial and by the eigenvalue of its restriction to S . Indeed, the minimal polynomial of T has degree two; thus if T | S = λ id S then T | S ′ = λ ′ id S ′ , where λ and λ ′ are the roots of the minimal polynomial of T . Therefore T is given by the matrix ( λ 0 0 λ ′ ) in the decomposition H 1 ( X, R ) -= S ⊕ S ′ . In [LN11, Section 6] for each D ≡ 1 mod 8 , with D ≥ 17 , we constructed two surfaces ( X , ω i i ) ∈ Ω E D (4) , i = 0 1 , , where the corresponding generators of the order T 0 ∈ Im( i 0 ) and T 1 ∈ Im( i 1 ) satisfies:
- · T 0 | S 0 = λ id S 0 and T 1 | S 1 = λ id S 1 ,
- · T 0 and T 1 have the same minimal polynomial,
- · 〈 〉 , | Rg( T 0 ) = 0 mod 2 and 〈 , 〉 | Rg( T 1 ) = 0 mod 2 .
/negationslash
/negationslash
Now if ( X ,ω 0 0 ) and ( X ,ω 1 1 ) belong to the same connected component ( i.e. A · ( X ,ω 0 0 ) = ( X ,ω 1 1 ) where A ∈ GL (2 + , R ) ) then there exists an isomorphism f : H X , 1 ( 0 Z ) -→ H X , 1 ( 1 Z ) -such that T ′ 1 := f -1 ◦ T 1 ◦ f defines an endomorphism of Prym( X ,τ 0 0 ) . By uniqueness of the Prym involution and the map i 0 : O D → End( X ,τ 0 0 ) , it follows that both T 0 and T ′ 1 belong to i 0 ( O D ) . By construction T 0 and T ′ 1 have the same minimal polynomial and T 0 | S 0 = ( T ′ 1 ) | S 0 = λ id S 0 . Thus in view of the above remark, T 0 = T ′ 1 . This is a contradiction since 〈 , 〉 | Rg( T 0 ) = 〈 , 〉 | Rg( T ′ 1 ) mod 2 .
We now go back to the proof of Theorem 4.1 and apply similar ideas. Let ( w,h,e ) ∈ Z 3 be as in Lemma 2.4 where D = e 2 +8 wh ≡ 1 mod 8 . Note that e is odd. We will show that the two surfaces
$$( X _ { 0 }, \omega _ { 0 } ) \coloneqq S _ { \kappa } ( w, h, - e ) \in \Omega E _ { D } ( \kappa ) \quad \text{ and } \quad ( X _ { 1 }, \omega _ { 1 } ) \coloneqq S _ { \kappa } ( w, h, e ) \in \Omega E _ { D } ( \kappa ) \quad \\ \cdot \quad \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \,.$$
do not belong to the same component. Recall that by construction, for j = 0 1 , , we have associated to ( X ,ω j j ) a generator T j of the order i j ( O D ) ⊂ End( Prym ( X ,τ j j )) . Recall that, in the symplectic basis ( a , b 0 0 , a, b ) of H X , 1 ( j Z ) given in Lemma 2.4, the endomorphism is given by the matrix
$$T _ { 0 } = \left ( \begin{array} { c c } - e \text{Id} _ { 2 } & \binom { 2 w } { 0 } _ { 0 } \\ \binom { h } { 0 } _ { w } ^ { 0 } \end{array} \right ) \quad \text{and} \quad T _ { 1 } = \left ( \begin{array} { c c } e \text{Id} _ { 2 } & \binom { 2 w } { 0 } _ { 2 h } \\ \binom { h } { 0 } _ { w } ^ { 0 } \end{array} \right ) \text{ respectively }. \\ \text{Let us assume that there is a continuous path $\gamma \, \colon [0, 1] \to \Omega E _ { D } ( \kappa ) \text{ such that $\gamma(i)= (X_{i},\omega_{i})$ for $i=0,1$},$$
Let us assume that there is a continuous path γ : [0 , 1] → Ω E D ( κ ) such that γ i ( ) = ( X , ω i i ) for i = 0 1 , , we will draw a contradiction. Let ˜ γ be a lift of γ to the vector bundle Ω T 3 over the Teichmüller space T 3 . We will denote by ( X ,ω s s ) the image of s ∈ [0 , 1] by ˜ γ . Let Σ be the base surface of the Teichmüller space. By construction the path ˜ γ induces a continuous map which sends every s ∈ [0 , 1] to a tuple ( J s , τ s , L s , i s , S s ) , where
- · J s is the complex structure of H 1 (Σ , R ) , induced by the complex structure of X s ,
- · L s is the lattice H 1 (Σ , Z ) ∩ ker( τ s +id) ,
- · τ s ∈ Sp(6 , Z ) is the matrix which gives the action of the Prym involution of X s on H 1 (Σ , R ) ,
- · i s : O D → End( H 1 (Σ , R ) -s ) , where H 1 (Σ , R ) -s = ker( τ s +id) ⊂ H 1 (Σ , R ) , is an injective ring morphism where i s ( O D ) is a self-adjoint proper subring of End( H 1 (Σ , R ) -s ) that preserves L s .
- · S s is the subspace of H 1 ( X, R ) -= ker( τ s +Id) generated by { Re( ω s ) , Rg( ω s ) } .
Remark that since ω is holomorphic, S s is invariant by J s . The action of GL (2 + , R ) preserves the subspace S s ⊂ H 1 (Σ , R ) -, and the kernel foliation leaves invariant [Re( ω )] and [Im( ω )] . Therefore S s is invariant along the path ˜ γ . Clearly, the matrix τ s is also invariant along the deformation ˜ γ . This implies that L s and i s are also invariant along ˜ γ . In particular S 0 = S 1 = S and i 0 = i 1 .
Now Rg( T 0 + e · Id) mod 2 is generated by { a, b } . The restriction of the intersection form 〈 , 〉 to this subspace is equal to 0 mod 2 . On the other hand the restriction of the intersection form to Rg( T ′ 1 ) does not vanish modulo 2 :
There exists an isomorphism f : H X , 1 ( 0 Z ) --→ H X , 1 ( 1 Z ) -satisfying f ( S 0 ) = S 1 and such that T ′ 1 = f -1 ◦ T 1 ◦ f belongs to i 0 ( O D ) ⊂ End( Prym ( X ,τ 0 0 )) . Remark that T ′ 1 and T 0 + e · Id have the same minimal polynomial X 2 -eX -2 wh . In addition the eigenvalues of T ′ 1 on S 0 = C · ω 0 , and the eigenvalue of T 0 + e · Id on C · ω 0 are both equal to λ = ( e + √ D / ) 2 . Hence T ′ 1 = T 0 + e · Id .
$$\langle T _ { 1 } ( a _ { 0 } ), T _ { 1 } ( b _ { 0 } ) \rangle & \equiv \langle a _ { 0 }, b _ { 0 } \rangle \equiv 1 \bmod 2. \\ \intertext { I the theorem follows}$$
This is a contradiction, and the theorem follows.
/square
In Section 5.4 we will give a topological argument for the non connectedness of Ω E 9 (2 2) , odd .
## 5. COLLAPSING ZEROS ALONG A SADDLE CONNECTION
In this section we describe a surgery on collapsing several zeros of Prym eigenforms together such that the resulting surface is still a non degenerate Prym eigenform. This can be thought as the converse of the surgery 'breaking up a zero' (see [KZ03]).
In what follows, all the zeros are labelled and all the saddle connections are oriented : if ( X,ω ) ∈ Ω E D (2 2) , odd , we label the zeros by P and Q and if ( X,ω ) ∈ Ω E D (1 1 2) , , , we label the simple zeros by R , R 1 2 and the double zero by Q . Let σ 0 be a saddle connection on X .
## Convention 1. We will always assume that:
odd to
Q
that is invariant by
τ
.
Observe that such saddle connections always exist on any ( X,ω ) ∈ Ω E D ( κ ) : for κ = (2 2) , odd , take σ 0 to be the union of a path of minimal length from a regular fixed point of τ to a zero of ω and its image by τ , for κ = (1 1 2) , , , take a path of minimal length from the set { R , R 1 2 } to { Q } .
- (1) If ( X,ω ) ∈ Ω E D (2 2) , then σ 0 is a saddle connection from P (2) If ( X,ω ) ∈ Ω E D (1 1 2) , , then σ 0 is a saddle connection from R 1 to Q .
## 5.1. Admissible saddle connections. We begin with the following definition.
## Definition 5.1. Let ( X,ω ) ∈ Prym ( κ ) be a Prym form.
/negationslash
/negationslash
- (1) κ = (2 2) , odd : we say that σ 0 is admissible if for any saddle connection σ = σ 0 from P to Q satisfying ω σ ( ) = λω σ ( 0 ) , with λ ∈ R + , one has λ > 1 .
Observe that by definition the subset consisting of surfaces having an admissible saddle connection is an open GL (2 + , R ) -invariant subset.
- (2) κ = (1 1 2) , , : we say that σ 0 is admissible if, for any saddle connection σ = σ 0 starting from R 1 and satisfying ω σ ( ) = λω σ ( 0 ) , with λ ∈ R + , either λ > 1 if σ ends at Q , or λ > 2 if σ ends at R 2 .
Lemma 5.2. Let κ ∈ { (2 2) , odd , (1 1 2) , , } . For any ( X,ω ) ∈ Ω E D ( κ ) and any σ 0 satisfying Convention 1, there exists in a neighborhood of ( X,ω ) a surface ( X , ω ′ ′ ) ∈ Ω E D ( κ ) with a saddle connection σ ′ 0 (corresponding to σ 0 and also satisfies Convention 1) such that any saddle connection σ ′ on X ′ in the same direction as σ ′ 0 , if it exists, satisfies
- (1) Case κ = (2 2) , odd : ω ′ ( σ ′ ) = ω ′ ( σ ′ 0 ) .
- (2) Case κ = (1 1 2) , , :
- (a) If σ ′ connects a simple zero to the double zero then ω ′ ( σ ′ ) = ω ′ ( σ ′ 0 ) .
- (b) If σ ′ connects two simple zeros then ω ′ ( σ ′ ) = 2 ω ′ ( σ ′ 0 ) .
Proof of Lemma 5.2. For any vector v ∈ R 2 small enough we denote by σ ′ 0 the saddle connection on ( X , ω ′ ′ ) = ( X,ω ) + v corresponding to σ 0 . Observe that the set
Slope( X,ω ) = { s ∈ R ∪ {∞} : s is the slope of ω γ ( ) = 0 , with [ γ ] ∈ H X, 1 ( Z ) }
/negationslash is countable. Hence there exists a vector v ∈ R 2 small enough so that the slope of ω ′ ( σ ′ 0 ) does not belong to the set Slope( X,ω ) = Slope( X , ω ′ ′ ) .
/negationslash
Case κ = (2 2) , odd . Let σ ′ be a saddle connection starting from P in the same direction as σ ′ 0 . If σ ′ ends at P then [ σ ′ ] ∈ H X , 1 ( ′ Z ) and 0 = ω ′ ( σ ′ ) has the same slope as ω ′ ( σ ′ 0 ) . This is a contradiction. Thus σ ′ ends at Q and [ γ ′ ] = [ σ ′ 0 ∗ ( -σ ′ )] ∈ H X ,ω 1 ( ′ ′ ) . If ω ′ ( γ ′ ) = 0 then we get again a contradiction. Therefore ω ′ ( γ ′ ) = 0 i.e. ω ′ ( σ ′ ) = ω ′ ( σ ′ 0 ) .
/negationslash
Case κ = (1 1 2) , , . Let σ ′ be a saddle connection starting from R 1 in the same direction as σ ′ 0 . If σ ′ ends at R 1 we run into the same contradiction. If σ ′ ends at Q then we also run into the same conclusion namely ω ′ ( σ ′ ) = ω ′ ( σ ′ 0 ) . Hence let us assume that σ ′ ends at R 2 . Thus σ ′ 0 ∗ τ ( σ ′ 0 ) ∗ ( -σ ′ ) is a closed path from R 1 to R 1 (through Q and R 2 ). Therefore [ γ ′ ] = [ σ ′ 0 ∗ τ ( σ ′ 0 ) ∗ ( -σ ′ )] ∈ H X ,ω 1 ( ′ ′ ) . The same contradiction shows that ω ′ ( γ ′ ) = 0 , i.e. ω ′ ( σ ′ ) = 2 ω ′ ( σ ′ 0 ) . The lemma is proved. /square
## 5.2. Non-admissible saddle connection and twin/double-twin. Lemma 5.2 leads to the following natural definition.
Definition 5.3. Let ( X,ω ) ∈ Ω E D ( κ ) and let σ 0 be a saddle connection on X satisfying Convention 1. If κ = (2 2) , odd : a saddle connection σ is a twin of σ 0 if σ joins P to Q and ω σ ( ) = ω σ ( 0 ) . If κ = (1 1 2) , , :
- (1) a saddle connection σ is a twin of σ 0 if it has the same endpoints and ω σ ( ) = ω σ ( 0 ) .
- (2) a saddle connection σ is a double-twin of σ 0 if σ joins R 1 to R 2 and ω σ ( ) = 2 ω σ ( 0 ) .
When ( X,ω ) ∈ Ω E D (2 2) , odd , since the angle between two twin saddle connections is a multiple of 2 π and the angle at P is 6 π , we see that each saddle connection σ 0 has at most two twins. When ( X,ω ) ∈ Ω E D (1 1 2) , , , the same remark shows that σ 0 has at most one twin or one double-twin. Moreover the midpoint of any double twin saddle connection is fixed by the Prym involution.
Non admissible saddle connection does not necessarily imply the existence of a twin or double twin (see Remark 5.6 below). However Lemma 5.2 shows that, under mild assumption, this dichotomy holds. As an immediate corollary, we draw
Corollary 5.4. Let κ ∈ { (2 2) , odd , (1 1 2) , , } . For any connected component C of Ω E D ( κ ) , there exist ( X,ω ) ∈ C and a saddle connection σ 0 on ( X,ω ) satisfying Convention 1 such that either σ 0 is admissible, or σ 0 has a twin or a double twin.
- 5.3. Collapsing admissible saddle connections. We have the following proposition that is a converse to the surgery 'breaking up a zero' (see [KZ03]). We prove the proposition only in the setting of Corollary 5.4. A more general statement holds but it is not needed in this paper.
Proposition 5.5. Let ( X,ω ) ∈ Ω E D ( κ ) be a Prym form and σ 0 a saddle connection satisfying Convention 1. We assume that σ 0 is admissible. Then one can collapse the zeros of ω along σ 0 by using the kernel foliation so that the resulting surface belongs to Ω E D (4) .
In particular, if there is no saddle connection in the same direction as σ 0 connecting two different zeros, then one can collapse σ 0 to get a surface in Ω E D (4) .
Proof. We first consider the case when ( X,ω ) ∈ Ω E D (2 2) , odd . The proof we describe is constructive. Set /lscript = | σ 0 | . As usual we assume that σ 0 is horizontal. By definition of admissible saddle connection, any other horizontal saddle connection from P to Q has length > /lscript .
For any horizontal geodesic ray emanating from a zero of ω we say that the ray is positive if it has direction (1 0) , and negative if it has direction ( -1 0) , . For instance by convention σ 0 is a positive ray for P and a negative ray for Q . Since the conical angle at P and Q is 6 π , there are two other positive horizontal rays emanating from P , say σ + P, 1 and σ + P, 2 , as well as two other negative rays for Q , say σ -Q, 1 and σ -Q, 2 . We parametrize each ray by its length to the zero where it emanates. Obviously, if a positive ray intersects a negative ray then it corresponds to a (horizontal) saddle connection.
We will first prove the proposition under a slightly stronger condition
( C ) any horizontal saddle connection other than σ 0 has length > /lscript 2 .
We will construct a set T which is a union of horizontal rays as follows: the first element of T is σ 0 . Now if a ray σ + P, 1 or σ + P, 2 intersects σ -Q, 1 or σ -Q, 2 by assumption, the associated saddle connection has length λ > 2 /lscript . Hence we can choose ε > 0 such that positive rays σ + P, 1 and σ + P, 2 at time /lscript + ε do not intersect σ -Q, 1 and σ -Q, 2 at time /lscript + ε in their interior. These are the next elements of T . Finally we consider the negative rays from P and positive rays from Q at time ε . By the assumption, the union T of all of these rays is an embedded tree in X (see Figure 3).
Wenow consider a neighborhood T ( δ ) = { p ∈ X ; h( p, T ) < δ } of T , where h is the distance measured in the vertical direction. For δ > 0 small enough, T is a retract by deformation of T ( δ ) . We can easily construct T ( δ ) from 10 Euclidian rectangles whose heights are equal to δ and widths are equal to /lscript +2 ε .
We will now change the flat metric of T ( δ ) without changing the metric outside of this neighborhood. Given any /lscript ′ ∈ ]0 , /lscript [ , by varying the points where the rectangles are sewn, we can obtain a new surface ( X , ω ′ ′ ) in Ω E D (2 2) , odd with a saddle connection invariant by the involution whose length is equal to /lscript ′ . Note that the surfaces obtained from this construction belong the same leaf of the kernel foliation as ( X,ω ) .
When /lscript ′ = 0 , we get a new closed surface ( X ,ω 0 0 ) ∈ H (4) sharing the same absolute periods as ( X,ω ) . Moreover there exists an involution τ 0 on X 0 such that τ ∗ 0 ω 0 = -ω 0 . Hence ( X ,ω 0 0 ) ∈ Ω E D (4) .
Let us now give the proof of the lemma without the additional condition ( C ) . Using ( 1 0 0 e t ) , t > 0 , we can assume that any non-horizontal saddle connection has length > /lscript 4 . Set
/lscript 0 = min {| σ , | σ is a simple horizontal geodesic loop at P or Q . }
Choose any δ ≤ 1 2min / { /lscript, /lscript 0 } , and consider B P,δ ( ) := { x ∈ X, d ( x, P ) < δ } and B Q,δ ( ) := { x ∈ X, d ( x, Q ) < δ } , where d is the distance defined by flat metric. By assumption, B P,δ ( ) and B Q,δ ( ) are two embedded topological disks in X which are disjoint. Therefore, the surface ( X,ω ) + ( -δ, 0) , which is
FIGURE 3. Collapsing two zeros along a saddle connection invariant by τ .
obtained by moving P by δ/ 2 to the right, and Q by δ/ 2 to the left, is well defined. In the new surface, any horizontal saddle connection from P to Q has length reduced by δ , but the lengths of all horizontal geodesic loops are unchanged since they are absolute periods of ω . Note also that if σ is another horizontal saddle connection joining P to Q , then | σ | - | σ 0 | is also unchanged. It follows that after finite steps, we can find a surface in the horizontal leaf of ( X,ω ) such that | σ 0 | < 1 2 / /lscript 0 and | σ 0 | < 1 2 / | σ | for any other horizontal saddle connection σ from P to Q . We can now apply the above arguments to conclude.
We now turn into the case when ( X,ω ) ∈ Ω E D (1 1 2) , , . The construction is similar, we keep the same convention: σ 0 is a positive ray for R 1 and a negative ray for Q . Note that τ ( σ 0 ) is a horizontal saddle connection from Q to R 2 , it is a positive ray for Q and negative ray for R 2 . We denote by σ ± Q,i , i = 1 2 , , the two other positive/negative rays from Q , and σ + R 1 (resp. σ -R 2 ) the other positive (resp. negative) ray from R 1 (resp. from R 2 ).
We again prove the proposition with a slightly stronger assumption that for any other horizontal saddle connection σ , one has | σ | > 4 | σ 0 | . We will construct a set T which is a union of positive/negative rays parametrized by the length to its origin. The first elements of T are σ 0 and τ ( σ 0 ) . We then add to T
- · the negative rays from R 1 and positive rays from R 2 at time ε .
- · the rays σ + R 1 and σ -R 2 a time 2 /lscript + ε
- · the positive and negative rays from Q other than σ 0 and τ ( σ 0 ) at time /lscript + ε .
We now consider a neighborhood T ( δ ) = { p ∈ X ; h( p, T ) < δ } of T where h is the distance measured in the vertical direction. For δ > 0 small enough, T is a retract by deformation of T ( δ ) . We can easily construct T ( δ ) from 10 Euclidian rectangles whose heights are equal to δ and widths are equal to 2 /lscript +2 ε . As above we can change the flat metric of T ( δ ) without changing the metric outside of this neighborhood. The rest of the proof follows the same lines as the case κ = (2 2) , odd . /square with ε > 0 small. Now if the ray σ + R 1 intersects any negative horizontal ray then, by assumption, the associated saddle connection has length > 4 /lscript . Hence we can choose ε > 0 such that positive ray σ + R 1 at time 2 /lscript + ε does not intersect any negative ray from Q at time /lscript + ε , nor any negative ray from R 1 or R 2 at time 2 /lscript + ε . Similar arguments apply for other positive rays. It follows that T is a tree.
5.4. Twins and non connectedness of Prym eigenform loci when D = 9 . In this section we give another elementary proof of the non connectedness of the loci Ω E 9 (2 2) , odd (Theorem 4.1).
Another proof of Theorem 4.1, case Ω E 9 (2 2) , . Set X 0 := S (2 2) , (1 1 , , -1) and X 1 := S (2 2) , (1 1 1) , , (see Lemma 2.4). For i = 0 1 , , let C i be the connected component of ( X , ω i i ) .
We claim that on any surface in C 0 , any saddle connection which connects the two zeros of ω 0 has exactly two twins. Since this property is not satisfied by ( X ,ω 1 1 ) (the longest horizontal saddle connection on ( X ,ω 1 1 ) connects the zeros of ω 1 and has no other twins) this will prove the theorem for Ω E 9 (2 2) , odd .
By construction the surface ( X ,ω 0 0 ) has three distinct Prym involutions, each of which preserves exactly one of the tori in the decomposition shown in Figure 2. By Lemma 3.4 there exists a ramified covering p : X 0 → N 0 of degree three (ramified at the zeros of ω 0 ) where N 0 is a torus. Hence any saddle connection on X 0 which connect the zeros of ω 0 has two other twins. For any surface in the kernel foliation leaf or in the GL (2 + , R ) -orbit of ( X ,ω 0 0 ) , this property is clearly preserved (since surfaces still have three Prym involutions). Thus ( X ,ω 1 1 ) does not belong to the component of ( X ,ω 0 0 ) . /square
Remark 5.6. On the surface S (2 2) , (1 1 1) , , , the longest horizontal saddle connection (the one that is contained in the boundary components of the bottom cylinder) satisfies Convention 1, and has no twins. It is not admissible since there are two other saddle connections in the same direction with smaller length. If we move this surface slightly in the kernel foliation leaf to break this parallelism, we will find a twin of this saddle connection (see Figure 4). This example shows that having a saddle connection satisfying Convention 1 which has no twins does not imply the existence of an admissible one. On the other hand, we know that S (2 2) , (1 1 1) , , belongs to Ω E 9 (2 2) , odd and Ω E 9 (4) = ∅ , so there exists no admissible saddle connection on S (2 2) , (1 1 1) , , .
FIGURE 4. On the left: ( X,ω ) = S (2 2) , (1 1 1) , , , on the right: ( X,ω ) + (0 , ε ) . In ( X,ω ) , σ 0 has no twin, but in ( X,ω ) + (0 , ε ) it has one.
- 5.5. Twins and triple tori. The next lemma provides a useful criterion to have triple tori from twin saddle connections (see Section 1.4 for the definition of triple tori).
Lemma 5.7. Let ( X,ω ) be a translation surface and let σ 0 be a saddle connection on X satisfying Convention 1. We assume that σ 1 is a twin of σ 0 that is not invariant by τ .
- (1) If ( X,ω ) ∈ Prym (2 2) , odd and σ 1 ∪ τ ( σ 1 ) is separating then the triple of saddle connections σ , σ 0 1 , τ ( σ 1 ) decomposes ( X,ω ) into a triple of flat tori.
- (2) If ( X,ω ) ∈ Prym (1 1 2) , , and σ 0 ∪ σ 1 is separating then the pairs ( σ , σ 0 1 ) and ( τ ( σ 0 ) , τ ( σ 1 )) decomposes ( X,ω ) into a triple of flat tori.
Proof of Lemma 5.7. As usual we assume first that ( X,ω ) ∈ Prym (2 2) , odd . Let σ 2 = τ ( σ 1 ) . Wefirst begin by observing that ( X,ω ) is the connected sum of a flat torus ( X ,ω 0 0 ) with a surface ( X , ω ′ ′ ) ∈ H (1 1) , ,
along σ 1 ∪ σ 2 . Indeed the saddle connections σ 1 and σ 2 determine a pair of angle (2 π, 4 π ) at P and Q . Since τ permutes P and Q , and preserves the orientation of X , it turns out that the angles 2 π at P and the angle 2 π at Q belong to the same side of σ 1 ∪ σ 2 .
The involution τ induces two involutions: τ 0 on X 0 and τ ′ on X ′ . The involution τ 0 is uniquely determined by the properties τ 0 ( ω 0 ) = -ω 0 and τ 0 permutes the endpoints of σ (namely, P and Q ). Hence τ 0 is the elliptic involution and has in particular 4 fixed points: the midpoint of σ and three fixed points of τ . Since τ has 4 fixed points, τ ′ has exactly 2 fixed points: the midpoint of σ 0 ( σ 0 is invariant by τ ) and the midpoint of σ ′ .
As subsurfaces of X X , 0 and X ′ have a boundary that consists of the saddle connections σ 1 and σ 2 . We can glue σ 1 and σ 2 together to obtain two closed surfaces that we continue to denote by X 0 and X ′ . We now have on X 0 (respectively, on X ′ ) a marked geodesic segment σ (respectively, a saddle connection σ ′ ) that corresponds to the identification of σ 1 and σ 2 . Note also that σ 0 is contained in X ′ .
/negationslash
$$\iota ( \sigma _ { 0 } ) = - \sigma _ { 0 } \quad & \text{or} \quad \iota ( \sigma _ { 0 } ) = - \sigma ^ { \prime }. \\ \iota \cdots = \iota \cdots \, \text{bounds} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text` \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iota \cdots \, \text{mean} \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iul \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \ioli \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \i ol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iot \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iota \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \ii \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \ieli \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iocl \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \irol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iolor \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iolate \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iolo \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iql \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iOl \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \ioll \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol \, \iol$$
Let ι be the hyperelliptic involution of X ′ . Since ι has 6 fixed points, we derive τ ′ = ι . We claim that ι ( σ 0 ) = -σ ′ . Indeed, ι ( σ 0 ) is a saddle connection such that ω ′ ( ( ι σ 0 )) = -ω ′ ( σ 0 ) . Hence
/negationslash
If ι ( σ 0 ) = -σ 0 then τ ′ ◦ ι ( σ 0 ) = σ 0 , hence τ ′ ◦ ι is the identity map in the neighborhoods of P and Q . Therefore τ ′ ◦ ι = id X ′ : this is a contradiction since we know that τ ′ = ι . Now the closed curve σ 0 ∗ ( -σ ′ ) is preserved by ι , hence it is separating. Cut X ′ along this closed curve we obtain two flat tori ( X ,ω 1 1 ) and ( X ,ω 2 2 ) . It is not difficult to see that X 1 and X 2 are exchanged by τ ′ . By construction, X is the connected sum of X 0 , X 1 , and X 2 which are glued together along the slits corresponding to σ , σ 0 1 , σ 2 .
The proof for the case ( X,ω ) ∈ Prym (1 1 2) , , is similar, it follows the same lines as the above discussion.
/square
## 6. COLLAPSING SURFACES TO Ω E D (4)
The goal of this section is to establish the following theorem, which is a key step in the proof of Theorem A.
Theorem 6.1. Let κ ∈ { (2 2) , odd , (1 1 2) , , } . Let C be a connected component of Ω E D ( κ ) . If for every surface ( X,ω ) ∈ C there is no admissible saddle connection on ( X,ω ) then D ∈ { 9 16 , } . More precisely, under this assumption, C contains one of the following surfaces (see Lemma 2.4):
$$S _ { \kappa } ( 1, 1, - 1 ), S _ { \kappa } ( 1, 1, 1 ) \in \Omega E _ { 9 } ( \kappa ), \quad o r \quad S _ { \kappa } ( 1, 2, 0 ), S _ { ( 2, 2 ) } ( 2, 1, 0 ) \in \Omega E _ { 1 6 } ( \kappa ). \\ \Xi \quad \dots \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot$$
Recall that as an immediate consequence we draw an upper bound on the number of connected components of Ω E D (2 2) , odd and Ω E D (1 1 2) , , (see Section 7).
6.1. Strategy of the proof of Theorem 6.1. Let ( X,ω ) be a Prym eigenform and let σ 0 be a saddle connection satisfying Convention 1. In view of Corollary 5.4 we can assume that σ 0 has a twin or a double twin, say σ 1 , otherwise the theorem is proved. Depending the strata, we will distinguish three cases as follows:
- · κ = (2 2) , odd :
Case B τ ( σ 1 ) = -σ 1 and σ 1 ∪ τ ( σ 1 ) is non-separating.
Case A τ ( σ 1 ) = -σ 1 and σ 1 ∪ τ ( σ 1 ) is separating.
/negationslash
Case C τ ( σ 1 ) = -σ 1 .
/negationslash
- · κ = (1 1 2) , , :
Case A σ 1 is a twin and σ 0 ∪ σ 1 is separating.
Case B σ 1 is a twin and σ 0 ∪ σ 1 is non-separating.
We will first prove Theorem 6.1 under the assumption of Case A . This case is simpler since Lemma 5.7 applies: ( X,ω ) admits a three tori decomposition. In the other two cases, by Lemma 6.2 D is a square. We then prove that Case B and Case C can be reduced to Case A : this corresponds, respectively, to Sections 6.3 and 6.4, respectively.
Case C σ 1 is a double twin ( τ ( σ 1 ) = -σ 1 ).
- 6.2. Proof of Theorem 6.1 under the assumption of Case A. From now on we will assume that for any ( X,ω ) ∈ C ⊂ Ω E D ( κ ) , there exists no admissible saddle connection. Let σ 2 be the image of σ 1 by the Prym involution τ . Thanks to Lemma 5.7 ( X,ω ) admits a three tori decomposition: X 0 is preserved and X ,X 1 2 are exchanged by the Prym involution τ .
Claim 1. There exists ( X,ω ) ∈ C such that the horizontal direction is periodic on the tori ( X ,X ,X 0 1 2 ) .
Proof of Claim 1. By moving in the kernel foliation leaf and using GL (2 + , R ) action, we can assume that the slits σ i are parallel to a simple closed geodesic in ( X ,ω 0 0 ) which is horizontal. Since ( X,ω ) is completely periodic in the sense of Calta (see [LN13a]), the claim follows. /square
In the sequel we assume that ( X,ω ) is decomposed into three horizontal cylinders, say C , C , C 0 1 2 , along the saddle connections σ , σ 0 1 , σ 2 , where C 0 is fixed, and C , C 1 2 are exchanged by the Prym involution. We let s = | σ i | . Wedenote by /lscript i , h i the width and the height of the cylinder C i . Obviously /lscript 1 = /lscript 2 and h 1 = h 2 (see Figure 1 for the notations).
6.2.1. Case κ = (2 2) , odd .
Claim 2. One of the following two equalities holds: h 0 = h 1 or h 0 = 2 h 1 .
Proof of Claim 2. Let δ be a saddle connection in C 0 joining P to Q which crosses the core curve of C 0 only once. Note that δ is anti-invariant by τ . Using U = { ( 1 t 0 1 ) , t ∈ R } , we can assume that δ is vertical. By assumption, δ is not admissible. Changing the length of the slits (the length of σ i ) if necessary and using the argument in Lemma 5.2, we can assume that δ has a twin δ ′ . Remark that δ ′ must intersect C 1 ∪ C 2 , therefore we have | δ ′ | = mh 1 + nh 0 with m ∈ Z , m ≥ 1 . The condition | δ ′ | = | δ | implies n = 0 . Thus h 0 = mh 1 .
Assume that h 0 > 2 h 1 . Let η 1 be a geodesic segment in C 1 joining P to the midpoint of σ 0 . Set η 2 = τ ( η 1 ) and η = η 1 ∪ η 2 . Observe that η is a saddle connection invariant by τ . Again, by using the subgroup U we assume that η is vertical. Hence | η | = 2 h 1 < h 0 . Clearly any other vertical (upward) geodesic ray starting from P must intersect C 0 . Thus, if there exists another vertical saddle connection η ′ joining P to Q , we must have | η ′ | ≥ h 0 > h 2 1 = | η | . Hence η is admissible and the claim is proved. /square
## Claim 3. If h 0 = h 1 then, either:
- (1) /lscript 0 = /lscript 1 and ( X,ω ) is contained in the same component as S κ (1 1 , , -1) ∈ Ω E 9 ( κ ) , or
- (2) /lscript 0 = 2 /lscript 1 and ( X,ω ) is contained in the same component as S κ (1 2 0) , , ∈ Ω E 16 ( κ ) .
Proof of Claim 3. Let h := h 0 = h 1 . Up to the action of the horocycle flow we assume that δ 0 is a saddle connection in C 0 joining P to Q such that ω δ ( 0 ) = (0 , h ) . Since δ 0 is not admissible, there exists a vertical saddle connection δ 1 joining P and Q such that ω δ ( 1 ) = (0 , λ ) , where 0 < λ ≤ h . Actually ω δ ( 1 ) = (0 , h ) . Since δ 1 cannot be contained in C 0 , δ 1 is contained in C 1 . Thus δ 2 = τ ( δ 1 ) is contained in C 2 . Let γ be the saddle connection contained in C 1 ∪ C 2 joining P to Q and passing through the midpoint of σ 0 as
shown in Figure 5. By assumption there exists another saddle connection γ ′ joining P to Q parallel to γ such that | γ ′ | ≤ | γ | . We claim that either γ ′ or τ ( γ ′ ) starts in C 0 . This is clear if γ ′ is not invariant by τ . If γ ′ is invariant by τ and starts from C 2 then it must end in C 1 (since τ ( C 2 ) = C 1 ). In particular γ ′ must cross C 0 at least once. Hence the vertical coordinate ω γ ( ′ ) is greater than 2 h . Therefore | γ ′ | > γ | | that is a contradiction.
0
1
0
1
FIGURE 5. Claim 3: h 1 = h 0 : the surfaces S κ (1 1 , , -1) and diag (1 , 1 2 ) · S κ (1 2 0) , , .
We can now suppose that γ ′ starts in C 0 . Observe that ω γ ( ) = ( s -2 /lscript 1 , 2 h ) .
If γ ′ is contained in C 0 (Figure 5, right), γ ′ must intersect twice the core curve of C 0 . Thus ω γ ( ′ ) = ( s -/lscript 0 , 2 h ) , from which we deduce /lscript 0 = 2 /lscript 1 and ( X,ω ) = diag (1 , 1 2 ) · S κ (1 2 0) , , . The claim is proved. /square
If γ ′ is not contained in C 0 = X 0 (Figure 5, left), γ ′ must end up in C 1 . Since | γ ′ | ≤ | γ | elementary calculation shows that ω γ ( ′ ) = ( s -/lscript 0 -/lscript 1 , 2 h ) . Now γ and γ ′ are parallel, thus /lscript 0 = /lscript 1 and ( X,ω ) = S κ (1 1 , , -1) .
## Claim 4. If h 0 = 2 h 1 then, either:
- (1) /lscript 0 = /lscript 1 and ( X,ω ) is contained in the same component as S κ (2 1 0) , , ∈ Ω E 16 ( κ ) , or
- (2) /lscript 0 = 2 /lscript 1 and ( X,ω ) is contained in the same component as S κ (1 1 1) , , ∈ Ω E 9 ( κ ) .
Proof of Claim 4. Let h := h 1 = h / 0 2 . Let δ 1 be a geodesic segment, contained in C 1 , joining the midpoint of σ 0 to P . Using the horocycle flow we assume δ 1 to be vertical. Set δ 2 = τ ( δ 1 ) and δ = δ 1 ∪ δ 2 . By construction δ is a saddle connection which is invariant under τ . By assumption, there exists another vertical saddle connection δ ′ joining P to Q such that | δ ′ | ≤ | δ | . It is easy to see that any other vertical geodesic ray emanating from P intersects the core curve of C 0 . Since ω δ ( ) = (0 2 , h ) and h 0 = 2 h δ , ′ is contained in C 0 .
Now let γ be the saddle connection in C 1 ∪ C 2 passing through the midpoint of σ 0 , joining P to Q such that ω γ ( ) = (2 /lscript 1 , 2 h ) (see Figure 6 below).
By assumption, there exists a saddle connection γ ′ in the same direction as γ such that | γ ′ | ≤ | γ | . As in Claim 3 either γ ′ or τ ( γ ′ ) starts in C 0 . The proof follows the same lines. Up to permutation, γ ′ starts in C 0 . Since | γ ′ | ≤ | γ | and γ is parallel to γ ′ , γ ′ is actually contained in C 0 . In particular ω γ ( ′ ) = ( k/lscript 0 , 2 h ) for some k ∈ Z , k ≥ 1 and ω γ ( ′ ) = ω γ ( ) . We draw 2 /lscript 1 = k/lscript 0 . Now we claim that the inequality
C
1
C
C
2
γ
γ
′
γ
′
0
h
0
= 2
h , /lscript
1
0
=
/lscript
1
0
1
0
1
FIGURE 6. Claim 4: h 0 = 2 h 1 : the surface on the left belongs to component of S κ (2 1 0) , , , and on the right belongs to the component of S κ (1 1 1) , , .
holds. Indeed there exists a horizontal saddle connection σ ′ 0 in X 0 such that σ 1 ∪ σ ′ 0 and σ 2 ∪ σ ′ 0 are the two boundaries components of the cylinder C 0 . Similarly, there exists a pair of horizontal saddle connections σ , σ ′ 1 ′ 2 where σ ′ i is contained in X i such that σ ′ i ∪ σ 0 is a boundary component of C i . By construction we have τ ( σ ′ 0 ) = -σ ′ 0 , τ ( σ ′ 1 ) = -σ ′ 2 , and
If /lscript 0 < /lscript 1 then | σ ′ 0 | < σ | ′ 1 | = | σ ′ 2 | . Hence σ ′ 0 is admissible, contradicting our assumption.
$$\ell _ { 0 } = | \sigma _ { 1 } | + | \sigma _ { 0 } ^ { \prime } | \quad \text{ and } \quad \ell _ { 1 } = | \sigma _ { 0 } | + | \sigma _ { 1 } ^ { \prime } | = | \sigma _ { 0 } | + | \sigma _ { 2 } ^ { \prime } |. \\ \text{non} \, | \tau ^ { \prime } | > | \tau ^ { \prime } | = | \tau ^ { \prime } | \text{ } \text{ } \text{$\mathfrak{Un}$} \quad \text{$\mathfrak{Multi}$} \quad \text{contradiction} \quad \text{or} \quad \text{common}.$$
In conclusion 2 /lscript 1 = k/lscript 0 ≥ /lscript 1 implies /lscript 0 = /lscript 1 , or /lscript 0 = 2 /lscript 1 . The corresponding surfaces are represented in Figure 6. It is not hard to check that those two surfaces belong to the same connected component that S κ (2 1 0) , , and S κ (1 1 1) , , , respectively. The claim is proved. /square
6.2.2. Case κ = (1 1 2) , , .
Claim 5. One of the following two equalities holds: h 0 = h 1 , or h 0 = 2 h 1 .
Proof of Claim 5.
The proof of this claim follows the same lines as the proof of Claim 2. /square
Claim 6. If h 0 = h 1 then, either
- (1) /lscript 0 = /lscript 1 , and ( X,ω ) is contained in the same component as S κ (1 1 , , -1) ∈ Ω E 9 ( κ ) , or
Proof of Claim 6. Set h := h 0 = h 1 . We first consider a saddle connection δ contained in C 0 , joining R 1 to Q and intersecting the core curve of C 0 only once. As usual we assume δ to be vertical (hence ω δ ( ) = (0 , h ) ). By assumption δ has a twin or a double-twin δ 1 (necessarily δ 1 starts in C 1 ). Clearly δ 1 is a twin: otherwise it must end in C 2 , hence it must cross the core curve of C 0 at least once. In particular its length satisfies | δ 1 | ≥ 3 h > 2 | δ | that is a contradiction.
- (2) /lscript 0 = 2 /lscript 1 , and ( X,ω ) is contained in the same component as S κ (1 2 0) , , ∈ Ω E 16 ( κ ) .
Let γ be the saddle connection contained in C 1 and joining R 1 to Q , as shown in Figure 7 below.
By assumption ω γ ( ) = ( /lscript 1 , h ) . Now there exists another saddle connection γ ′ starting from R 1 in the same direction as γ . Observe that γ ′ must start in C 0 . Using Lemma 5.2, we can assume that γ ′ is either a twin or a double-twin of γ .
If γ ′ is a twin of γ then ω γ ( ′ ) = ( /lscript 1 , h ) . Hence /lscript 0 = /lscript 1 : ( X,ω ) = S κ (1 1 , , -1) . ′ ′ ′
If γ is a double-twin of γ then ω γ ( ) = (2 /lscript 1 , 2 h ) . Hence γ crosses twice the core curves of C 0 implying that /lscript 0 = 2 /lscript 1 : ( X,ω ) = diag (1 , 1 2 ) · S κ (1 2 0) , , . This proves the claim. /square
Q
P
FIGURE 7. Claim 6: h 0 = h 1 : the surfaces S κ (1 1 , , -1) and diag (1 , 1 2 ) · S κ (1 2 0) , , .
Claim 7. If h 0 = 2 h 1 then /lscript 0 = 2 /lscript 1 . In addition ( X,ω ) belongs to the same connected component as S κ (1 1 1) , , ∈ Ω E 9 ( κ ) .
Proof. Set h = h 1 . Let δ 1 be a saddle connection contained in C 1 joining R 1 to Q and intersecting the core curve of C 1 only once. We can suppose that δ 1 is vertical. By assumption, there exists another vertical saddle connection δ starting from R 1 . Observe that δ must intersect the core curve of C 0 , thus we have | δ | ≥ h 0 = 2 | δ 1 | . By assumption, δ must be a double-twin of δ 1 , which means that δ joins R 1 to R 2 and is contained in C 0 .
Now, let γ be the saddle connection in C 1 joining R 1 to Q as shown in Figure 8.
1
FIGURE 8. Claim 7: h 0 = 2 h 1 : the surface S κ (1 1 1) , ,
By assumption, there exists a saddle connection γ ′ starting from R 1 and parallel to γ . The same argument as above shows that γ ′ is a double-twin of γ and is contained in C 0 . It follows that /lscript 0 = 2 /lscript 1 . Thus ( X,ω ) belongs to the component of S κ (1 1 1) , , . The claim is proved. /square
6.3. Reduction from Case B to Case A. Let ( X,ω ) ∈ Ω E D ( κ ) and let σ 0 be a saddle connection in X satisfying Convention 1. We suppose that σ 0 has a twin σ 1 . Moreover, if κ = (2 2) , odd we assume that
/negationslash
σ 2 := -τ ( σ 1 ) = σ 1 and σ 1 ∪ σ 2 is non-separating, and if κ = (1 1 2) , , we assume that σ , σ 0 1 is nonseparating. Our aim is to show that there exists in the component of ( X,ω ) a surface having a family of homologous saddle connections satisfying Case A (this is Lemma 6.3). We first show
/negationslash
Lemma 6.2. Let ( X,ω ) ∈ Ω E D ( κ ) , where κ ∈ { (2 2) , odd , (1 1 2) , , } . If there exists c ∈ H X, 1 ( Z ) -satisfying c = 0 and ω c ( ) = 0 then D is a square. In particular, up to rescaling by GL (2 + , R ) , all the absolute periods of ω belong to Q + ı Q .
/negationslash
Proof of Lemma 6.2. We can assume that c is primitive in H X, 1 ( Z ) , that is for any n ∈ N , n > 1 , 1 n c ∈ H X, 1 ( Z ) . Pick a symplectic basis ( α , β 1 1 , α 2 , β 2 ) of H X, 1 ( Z ) -with β 2 = c . Set µ i = 〈 α , β i i 〉 , where 〈 , 〉 is the intersection form of H X, 1 ( Z ) , and ω α ( 1 ) = x 1 + ıy 1 , ω ( β 1 ) = z 1 + ıt 1 , ω ( α 2 ) = x 2 + ıy 2 . Since
$$\text{Area} ( X, \omega ) = \mu _ { 1 } \det \left ( \begin{array} { c c } x _ { 1 } & z _ { 1 } \\ y _ { 1 } & t _ { 1 } \end{array} \right ) + \mu _ { 2 } \det \left ( \begin{array} { c c } x _ { 2 } & 0 \\ y _ { 2 } & 0 \end{array} \right ) = \mu _ { 1 } \det \left ( \begin{array} { c c } x _ { 1 } & z _ { 1 } \\ y _ { 1 } & t _ { 1 } \end{array} \right ) > 0$$
it follows that ( x 1 + ıy 1 , z 1 + ıt 1 ) is a basis of R 2 . Using GL (2 + , R ) we can assume that ( x , y 1 1 ) = (1 0) , and ( z , t 1 1 ) = (0 1) , . By [LN11, Proposition 4.2] there exists a unique generator T of O D which is given in the basis ( α , β 1 1 , α 2 , β 2 ) by a matrix of the form
$$T & = \left ( \begin{array} { c c } e \text{Id} _ { 2 } & \binom { a \ b } { c \ d } \\ \frac { \mu _ { 1 } } { \mu _ { 2 } } \binom { d \ - b } { - c \ a } & 0 \end{array} \right ) \\ \text{it} \, T ^ { * } \omega & = \lambda \omega \text{ and } \lambda > 0. \text{ Observe that } \text{tl}$$
with ( a, b, c, d, e ) ∈ Z 5 such that T ω ∗ = λω and λ > 0 . Observe that the discriminant D satisfies D = e 2 -4 µ 1 µ 2 ( bc -ad ) . Since Re( ω ) = (1 0 , , x 2 , 0) and Im( ω ) = (0 1 , , y 2 , 0) in the above coordinates, direct computations show that b = d = 0 . Hence D = e 2 and ( x , y 2 2 ) ∈ Q 2 . The lemma is proved. /square
Lemma 6.3. Assume that κ = (2 2) , odd . Then there exists in the component of ( X,ω ) a surface having a triple of homologous saddle connections γ , γ 0 1 , γ 2 , where γ 0 is invariant, γ 1 and γ 2 are exchanged by the involution.
/negationslash
Proof. Let c 0 , c 1 , and c 2 denote the simple closed curves σ 1 ∗ -( σ 2 ) , σ 0 ∗ -( σ 1 ) and σ 0 ∗ -( σ 2 ) respectively. Note that we have c 0 = c 2 -c 1 and τ ( c 1 ) = -c 2 . By assumption 0 = c 0 ∈ H X, 1 ( Z ) . If 0 = c 1 ∈ H X, 1 ( Z ) then c 2 = -τ ( c 1 ) = 0 ∈ H X, 1 ( Z ) , which implies that c 0 = 0 ∈ H X, 1 ( Z ) . Thus we can conclude that all of the curves c 0 , c 1 , c 2 are non-separating.
Cut X along σ , σ 0 1 , σ 2 , we obtain a connected surface whose boundary has three components corresponding to c 0 , c 1 , c 2 . Gluing the pair of geodesic segments in each boundary component together, we get a closed translation surface ( X , ω ′ ′ ) with three marked geodesic segments. Since the angle between two consecutive twin saddle connections is 2 π , we derive that ω ′ has no zeros, thus ( X , ω ′ ′ ) must be a torus. We denote the geodesic segments in X ′ corresponding to c 0 , c 1 , c 2 by c ′ 0 , c ′ 1 , c ′ 2 respectively. The involution τ of X induces an involution τ ′ on X ′ , which leaves c ′ 0 invariant and exchanges c ′ 1 and c ′ 2 . Let P ′ i and Q ′ i , i = 0 1 2 , , , denote the endpoints of c ′ i , where P ′ i (respectively, Q ′ i ) corresponds to P (respectively, to Q ).
Observe that as ( X,ω ) moves in the leaf of the kernel foliation, the surface ( X , ω ′ ′ ) is the same, only the segments c ′ i vary. Therefore we can assume that ( X , ω ′ ′ ) is the standard torus C / ( Z ⊕ ı Z ) , and the length of c ′ 0 is small. Let δ 1 (respectively, η 1 ) denote the geodesic segment of minimal length from P ′ 0 to P ′ 1 (respectively, from Q ′ 0 to Q ′ 1 ). Note that as ( X,ω ) moves in the kernel foliation leaf, ω ′ ( δ 1 ) and ω ′ ( η 1 ) are invariant. Therefore, we can assume that δ 1 and c ′ 0 are not parallel.
FIGURE 9. Cylinders decomposition in direction of α ± .
Since c ′ 0 and c ′ 1 are parallel and have the same length, we see that c ′ 0 ∪ δ 1 ∪ c ′ 1 ∪ η 1 is the boundary of an embedded parallelogram in X ′ . It follows in particular that δ 1 and η 1 are parallel and have the same length. Let δ 2 = τ ′ ( η 1 ) and η 2 = τ ′ ( δ 1 ) . We have δ = δ 1 ∪ δ 2 is a geodesic segment joining P ′ 1 to P ′ 2 , and η = η 1 ∪ η 2 is a geodesic segment joining Q ′ 1 to Q ′ 2 .
Recall that ( X,ω ) is obtained from ( X , ω ′ ′ ) by slitting along c , c ′ 0 ′ 1 , c ′ 2 , and regluing the geodesic segments in the boundary. By construction, we see that ( X,ω ) admits a decomposition into four cylinders in the direction of α ± as shown in Figure 9. Let C 0 denote the largest cylinder in this decomposition, then it is easy to see that there exist in C 0 three homologous saddle connections γ , γ 0 1 , γ 2 such that τ ( γ 0 ) = -γ 0 , τ ( γ 1 ) = -γ 2 , and γ 1 ∪ γ 2 is a separating curve as desired. /square
Since ω c ( 0 ) = 0 , By Lemma 6.2, for any c ∈ H X, 1 ( Z ) , we have ω c ( ) ∈ Q + ı Q . Therefore ω ′ ( δ ) = ω ′ ( η ) ∈ Q + ı Q . Since X ′ is the standard torus, there exists a pair of parallel simple closed geodesics α + , α -of X ′ such that δ ⊂ α + , and η ⊂ α -(see Figure 9). When | c ′ i | is small enough and non-parallel to α ± , the geodesics α + and α -cut X ′ into two cylinders, one of which contains all the segments c ′ 0 , c ′ 1 , c ′ 2 .
Lemma 6.4. Assume that κ = (1 1 2) , , and σ 0 has a twin σ 1 such that the curve σ 0 ∗ ( -σ 1 ) is nonseparating. Then there exists in the component of ( X,ω ) a surface having two pairs of homologous saddle connections ( σ , σ ′ 1 ′′ 1 ) and ( σ , σ ′ 2 ′′ 2 ) , where σ ′ i and σ ′′ i join the simple zero R i to the double zero Q , and { σ , σ ′ 2 ′′ 2 } = τ ( { σ , σ ′ 1 ′′ 1 } ).
Proof. We will use similar ideas to the proof of Lemma 6.3. Let c 1 = σ 0 ∗ ( -σ 1 ) and c 2 = τ ( c 1 ) . By the cutting-gluing construction along c 1 and c 2 (using the assumption that c 1 is non-separating), we get a flat torus ( X , ω ′ ′ ) with three marked geodesic segments c , c ′ 1 ′ 2 , c ′′ such that ω ′ ( c ′ 1 ) = ω ′ ( c ′ 2 ) = 1 / ω 2 ′ ( c ′′ ) (see Figure 10). For i = 1 2 , , we denote the endpoints of c ′ i by R ′ i and Q ′ i so that R ′ i corresponds to R i and Q ′ i corresponds to Q . We denote the endpoints of c ′′ by R ′′ 1 and R ′′ 2 such that R ′′ i corresponds to R i . The midpoint of c ′′ corresponds to Q , we denote it by Q ′′ . We denote the subsegment of c ′′ between Q ′′ and R ′′ i by c ′′ i . The Prym involution of X gives rise to an involution τ ′ of X ′ which satisfies τ ′ ( c ′ 1 ) = -c , τ ′ 2 ′ ( c ′′ 1 ) = -c ′′ 2 .
As ( X,ω ) moves in the leaf of the kernel foliation, the surface ( X , ω ′ ′ ) remains the same, only the segments c ′ 1 , c ′ 2 , c ′′ vary. Therefore we can assume that c ′ 1 , c ′ 2 , c ′′ are contained in three distinct parallel simple closed geodesics of X ′ . Changing the direction of c ′′ slightly, we see that there exist geodesic segments σ ′ i from Q ′ i to R ′ i , and σ ′′ i from Q ′′ to R , i ′′ i = 1 2 , , (see Figure 10) such that
- · τ ′ ( σ ′ 1 ) = -σ ′ 2 and τ ′ ( σ ′′ 1 ) = -σ ′′ 2 ,
- · σ ′ i ∗ c ′ i and σ ′′ i ∗ c ′′ i are homologous.
FIGURE 10
Reconstruct ( X,ω ) from ( X , ω ′ ′ ) we see that σ ′ i and σ ′′ i are homologous saddle connections, and the pairs ( σ , σ ′ 1 ′′ 1 ) and ( σ , σ ′ 2 ′′ 2 ) have the desired properties. /square
- 6.4. Reduction from Case C to Case A. Let ( X,ω ) be a Prym eigenform in Prym (2 2) , odd /unionsq Prym (1 1 2) , , , and let σ 0 be a saddle connection on X satisfying Convention 1. If ( X,ω ) ∈ Prym (2 2) , odd we suppose that σ 0 has a twin σ 1 which is also invariant by the Prym involution τ , and if ( X,ω ) ∈ Prym (1 1 2) , , we suppose σ 0 has a double twin σ 1 (which is invariant by τ ). Our aim is to show that there exists in the component of ( X,ω ) a surface having a family of saddle connections satisfying Case A for both κ = (2 2) , odd and κ = (1 1 2) , , . We first show
## Lemma 6.5. Define
- · c = σ 0 ∗ ( -σ 1 ) , if ( X,ω ) ∈ Prym (2 2) , odd ,
Then in both cases, we have c = 0 ∈ H X, 1 ( Z ) -, and ω c ( ) = 0 .
- · c = σ 0 ∗ τ ( s 0 ) ∗ ( -σ 1 ) , if ( X,ω ) ∈ Prym (1 1 2) , , .
Proof. From the definition of c , we have τ ( c ) = -c , hence c ∈ H X, 1 ( Z ) -. It is also clear that ω c ( ) = 0 . All we need to show is that c = 0 ∈ H X, 1 ( Z ) .
/negationslash
/negationslash
/negationslash
Wefirst consider the case ( X,ω ) ∈ Prym (2 2) , odd . Remark that the pair of angles at P and Q determined σ 0 and σ 1 is (2 π, 4 π ) . Since τ ( σ 0 ) = -σ 0 and τ ( σ 1 ) = -σ 1 we see that the angle 2 π at P and the angle 4 π at Q belong to the same side of c , and vice versa. Cutting X along c we get a surface whose boundary has two components, each of which is a union of two geodesic segments corresponding to σ 0 and σ 1 . Since σ 0 and σ 1 are twins, the two segments in each component has the same length, therefore we can glue them together to get a closed (possibly disconnected) translation surface ( X , ω ′ ′ ) with two marked geodesic segments η , η 1 2 . If the new surface is disconnected, then each component is a translation surface with only one singularity of angle 4 π . Since such a surface does not exist, we conclude that X ′ is connected, and hence c = 0 in H X, 1 ( Z ) .
For the case ( X,ω ) ∈ Prym (1 1 2) , , , by a similar construction, that is cutting along c , then closing the boundary components of the new surface (by gluing the path corresponding to σ 0 ∪ τ ( σ 0 ) and the segment corresponding to σ 1 ), we also get a translation surface ( X , ω ′ ′ ) having two singularities with cone angle 4 π . The same argument as above shows that this surface belongs to H (1 1) , , therefore c = 0 ∈ H X, 1 ( Z ) . /square
/negationslash
Lemma 6.6. Let ( X,ω ) ∈ Prym (2 2) , odd be a Prym eigenform having a twin σ 1 of σ 0 that is invariant by τ . Then one can find in the connected component of ( X,ω ) another surface having a triple of homologous saddle connections.
Proof of Lemma 6.6. Cutting X along c = σ 0 ∪ σ 1 and gluing the two segments of each boundary component together, we get a closed translation surface ( X , ω ′ ′ ) ∈ H (1 1) , (see the proof of Lemma 6.5). By construction there exist on X ′ a pair of disjoint geodesic segments η , η 1 2 such that ω ′ ( η 1 ) = ω ′ ( η 2 ) = ω σ ( 0 ) and η i joins a zero of ω ′ to a regular point. Let ( P , Q i i ) i =1 2 , denote the endpoints of η i , where P i (respectively, Q i ) corresponds to P (respectively, to Q ). The numbering is chosen so that P 1 and Q 2 are the zeros of ω ′ . The involution τ of X descends to an involution of X ′ exchanging η 1 and η 2 . We denote this involution by τ ′ . Remark that τ ′ has two fixed points in X ′ , none of which are contained in the segments η , η 1 2 . Note also that as ( X,ω ) moves in its leaf of the kernel foliation, ( X , ω ′ ′ ) also moves in its leaf of the kernel foliation in H (1 1) , (only the relative periods change).
/negationslash
We consider the tuple ( X ,ω ,P ′′ ′′ ′′ 1 , P ′′ 2 ) as an element in H (0 0) , , that is the moduli space of flat tori with two marked points. We first observe that as ( X,ω ) moves in its leaf of the kernel foliation, the corresponding surfaces ( X ,ω ,P ′′ ′′ ′′ 1 , P ′′ 2 ) are the same in H (0 0) , (only ω ′′ ( η ′′ 1 ) = ω ′′ ( η ′′ 2 ) change). Indeed all the coordinates of ( X ,ω ,P ′′ ′′ ′′ 1 , P ′′ 2 ) are determined by the absolute periods of ( X,ω ) .
Let ι be the hyperelliptic involution of X ′ . Since ι has six fixed points but τ ′ has two, we have ι = τ . Remark that ι ◦ τ is also an involution of X ′ satisfying ( ι ◦ τ ) ∗ ω ′ = ω ′ . The surface X ′′ = X / ι ′ 〈 ◦ τ 〉 is an elliptic curve. Let π is the branched covering π : X ′ → X ′′ , which is is ramified at P 1 and Q 2 . Then ω ′ descends to a holomorphic 1 -form ω ′′ on X ′′ so that ω ′ = π ω ∗ ′′ . For i = 1 2 , let P ,Q ,η ′′ i ′′ i ′′ i denote the images of P , Q , η i i i in X ′′ . Note that we have ω ′′ ( η ′′ 1 ) = ω ′′ ( η ′′ 2 ) = ω σ ( 0 ) .
Let α ′′ 1 be a simple closed geodesic of ( X ,ω ′′ ′′ ) which passes through P ′′ 1 and does not contain P ′′ 2 . Using GL(2 , R ) , we can assume that α ′′ 1 is horizontal. By moving in the kernel foliation leaf of ( X,ω ) , we can also assume that η ′′ i are parallel to α ′′ 1 ( ω ′′ ( η ′′ i ) = λω ′′ ( α ′′ 1 ) , with 0 < λ < 1 ). By construction, the surface ( X , ω ′ ′ ) admits a decomposition into cylinders in the horizontal direction. Note that X ′ must have three horizontal cylinders, otherwise there would be horizontal saddle connection joining P 1 to Q 2 , which is excluded since α ′′ 1 does not contain η ′′ 2 (see Figure 11).
Lemma 6.7. Let ( X,ω ) ∈ Prym (1 1 2) , , be a Prym eigenform having a double twin σ 1 of σ 0 . Then one can find in the connected component of ( X,ω ) a surface having two pairs of homologous saddle connections ( σ , σ ′ 1 ′′ 1 ) and ( σ , σ ′ 2 ′′ 2 ) that are exchanged by the Prym involution.
We can reconstruct ( X,ω ) from ( X , ω ′ ′ ) : one sees that ( X,ω ) also admits a decomposition into three horizontal cylinders (see Figure 11). Consider the surface ( ˜ X,ω ˜) = ( X,ω ) + (0 , -/epsilon1 ) , with /epsilon1 > 0 small as shown in Figure 11. We see that ( ˜ X,ω ˜) admits a decomposition into four horizontal cylinders, three of which are simple. It is easy to check that there exists a triple of twin saddle connections γ , γ 0 1 , γ 2 in the largest horizontal cylinder of ˜ X (which is preserved by the Prym involution) which satisfy τ ( γ 0 ) = -γ 0 , τ ( γ 1 ) = -γ 2 , and γ 1 ∪ γ 2 is a separating curve. This proves the lemma. /square
Proof. Let ( X , ω ′ ′ ) ∈ H (1 1) , be the surface obtained by the 'cutting-gluing' construction along c = σ 0 ∗ τ ( σ 0 ) ∗ ( -σ 1 ) (see Lemma 6.5). Note that we have on X ′ two marked and disjoint geodesic segments η 1 and η 2 (corresponding to c ) such that the midpoint of η i is a zero of ω ′ , and ω ′ ( η i ) = ω σ ( 1 ) = 2 ω σ ( 0 ) . We will consider η i as a slit with two sides η ′ i and η ′′ i , where η ′ i corresponds to σ 1 , and η ′′ i corresponds to the union σ 0 ∪ τ ( σ 0 ) . Remark that the Prym involution τ induces an involution τ ′ on X ′ which is not the hyperelliptic involution. It follows that X ′ is a double cover of a torus.
By construction, as ( X,ω ) moves in its kernel foliation leaf ( X , ω ′ ′ ) is fixed, only the marked geodesic segments (slits) η i are changed. Using GL (2 + , R ) , we can assume that X ′ is horizontally periodic, and it has three horizontal cylinders (one may use the fact that X ′ is a double cover of a torus to show that a periodic direction with three cylinder does exist). We can arrange so that the slits are also horizontal,
FIGURE 11. σ 0 and σ 1 are invariant by τ , X ′ admits two involutions: the hyperelliptic one ι which fixes each of the cylinders, and the involution τ ′ induced by τ which exchanges the pair simple cylinders and fixes the larger one. Observe that τ ′ exchanges η 1 = P Q 1 1 and η 2 = P Q 2 2 .
FIGURE 12. Prym (1 1 2) , , case C: σ 0 has a double-twin.
and η ′ i are contained in the boundary of the largest cylinder (see Figure 12). Reconstruct ( X,ω ) from ( X , ω ′ ′ ) , we see that ( X,ω ) has two pairs of homologous saddle connections ( σ , σ ′ 1 ′′ 1 ) and ( σ , σ ′ 2 ′′ 2 ) , such that τ ( σ ′ 1 ) = -σ , τ ′ 2 ( σ ′′ 1 ) = σ ′′ 2 , and σ ′ 1 ∗ σ 1 ∗ σ ′′ 2 is homologous to the core curve of the largest horizontal cylinder in X . The lemma is then proved. /square
## 7. PROOF OF THE MAIN RESULT
Let us now give the proof of our main theorem.
Proof of Theorem A. First of all D ∈ N , D ≡ 0 1 4 , , mod 8 , the loci Ω E D ( κ ) are non empty: this is Corollary 2.5.
/negationslash
We now consider the cases D = 0 1 4 , , mod 8 , D ≥ 9 , and D ∈ { 9 16 , } . By Theorem 6.1 and Corollary 5.4, any component of Ω E D ( κ ) contains a surface with an admissible saddle connection that collapse to a point in Ω E D (4) . Recall that Ω E D (4) is a finite collection of Teichmüller discs. By Proposition 2.6 for any connected component C of Ω E D (4) , there exists at most one component of Ω E D ( κ ) adjacent to C i.e. its closure contains C . Therefore, the number of connected components of Ω E D ( κ ) is bounded from above by the number of components of Ω E D (4) .
In particular, when D ≡ 0 4 , mod 8 , since Ω E D (4) is connected (see Theorem 2.2), so is Ω E D ( κ ) . For D ≡ 1 mod 8 , by Theorem 2.2 we know that Ω E D (4) has two components, so Ω E D ( κ ) has at most two components. On the other hand, Theorem 4.1 tells us that Ω E D ( κ ) cannot be connected. Thus we can conclude that Ω E D ( κ ) has exactly two connected components.
For D ≡ 5 mod 8 , if Ω E D ( κ ) is non-empty then again Theorem 6.1, Corollary 5.4 and Proposition 5.5 implies that Ω E D (4) is also non-empty, which contradicts Theorem 2.2.
We now consider the cases D ∈ { 9 16 , } . Since Ω E D (4) = ∅ , by Proposition 5.5, there exists no admissible saddle connection on any surface in Ω E 9 ( κ ) /unionsq Ω E 16 ( κ ) . By Theorem 6.1, we see that any surface in Ω E 9 ( κ ) /unionsq Ω E 16 ( κ ) belongs to the same component as one of the surfaces
(see Lemma 2.4 for the definition).
$$S _ { \kappa } ( 1, 1, - 1 ), S _ { \kappa } ( 1, 1, 1 ), S _ { \kappa } ( 1, 2, 0 ) & \text{or} \ S _ { ( 2, 2 ) } ( 2, 1, 0 ) \\ \shortparallel \Lambda \Lambda \Lambda \shortparallel \Lambda$$
It follows immediately that Ω E 16 (2 2) , and Ω E 9 ( κ ) has at most two components and Ω E 16 (1 1 2) , , is connected. The fact that Ω E 9 ( κ ) is not connected is proved in Theorem 4.1. Hence Ω E 9 ( κ ) has exactly two connected components.
It remains to prove that Ω E 16 (2 2) , is connected. It is sufficient to show that S (2 2) , (1 2 0) , , , S (2 2) , (2 1 0) , , ∈ Ω E 16 (2 2) , odd belong to the same component. We consider ( X ,ω ε ε ) = S (2 2) , (2 1 0) + (0 , , , ε ) , with ε > 0 small enough (see Figure 13).
FIGURE 13. Connecting S (2 2) , (2 1 0) , , to S (2 2) , (1 2 0) , , .
Observe that ( X ,ω ε ε ) admits a decomposition into four horizontal cylinders. Moving horizontally in the kernel foliation leaf of ( X ,ω ε ε ) we get a surface ( Y, η ) = ( X ,ω ε ε ) + v , with v ∈ R × { } 0 , which admits a decomposition into three vertical cylinders. It is not difficult to see that ( Y, η ) can be connected to S (2 2) , (1 2 0) , , by using the action of GL (2 + , R ) and moving in the kernel foliation leaves. The proof of Theorem A is now complete. /square
As a direct corollary we prove Theorem B i.e. the existence in any component of Ω E D ( κ ) of surfaces which admit three-tori decompositions.
Proof of Theorem B. Let ( w,h,e ) ∈ Z 3 be as in Lemma 2.4 where D = e 2 + 8 wh . We consider the corresponding surfaces ( X ,ω ± ± ) := S κ ( w,h, ± e ) . By Lemma 2.4 ( X ,ω ± ± ) ∈ Ω E D ( κ ) .
If D ≡ 1 mod 8 then by Theorem A, Ω E D ( κ ) is connected and ( X ,ω ± ± ) admits a three-tori decomposition.
/negationslash
- If D ≡ 1 mod 8 then by Theorem A, Ω E D ( κ ) has two connected components and from the proof of Theorem 4.1, ( X ,ω + + ) and ( X ,ω --) do not belong to the same connect component. This ends the proof of Theorem B. /square
## REFERENCES
[Cal04] K. CALTA - 'Veech surfaces and complete periodicity in genus two', J. Amer. Math. Soc. 17 (2004), no. 4, pp. 871-908. [EMZ03] A. ESKIN, H. MASUR and A. ZORICH - 'The Principal Boundary, Counting Problems and the Siegel-Veech Constants ', Publ. Math. Inst. Hautes Études Sci. 97 (2003), pp. 61-179.
[HMSZ] P. HUBERT, H. MASUR, T. SCHMIDT, and A. ZORICH -'Problems on billiards, flat surfaces and translation surfaces ', in collection Problems on Mapping Class Groups and Related Topics , pp.233-243, Proc. Sympos. Pure Math., 74 , Amer. Math. Soc., Providence, RI (2006).
[KZ03]
M. KONTSEVICHand A. ZORICH-'Connected components of the moduli spaces of Abelian differentials with prescribed singularities',
Invent. Math.
153
(2003), no. 3, pp. 631-678.
[LN11] E. LANNEAU and D.-M. NGUYEN -'Teichmueller curves generated by Weierstrass Prym eigen forms in genus three and genus four ', J. of Topology (to appear) , arXiv:1111.2299 (2011).
[LN13a] 'Complete periodicity of Prym eigenforms ', arXiv:1301.0783 (2013).
[LN13b] ' GL (2 + , R ) -orbits in Prym eigenform loci ', arXiv:1310.8537 (2013).
[MaTa02] H. MASUR, S. TABACHNIKOV-'Rational billiards and flat structures', Handbook of dynamical systems , 1A , NorthHolland, Amsterdam (2002), pp. 1015-1089.
[MZ08] H. MASUR and A. ZORICH - 'Multiple Saddle Connections on flat Surfaces and Principal Boundary of the Moduli Spaces of Quadratic Differentials ', Geom. Funct. Anal. 18 (2008), no. 3, pp. 919-987.
[Mc03] C. MCMULLEN - 'Billiards and Teichmüller curves on Hilbert modular surfaces', J. Amer. Math. Soc. 16 (2003), no. 4, pp. 857-885.
[Mc05]
, 'Teichmüller curves in genus two: Discriminant and spin', Math. Ann. 333 (2005), pp. 87-130.
[Mc06] , 'Prym varieties and Teichmüller curves ', Duke Math. J. 133 (2006), pp. 569-590.
[Mc07] , 'Dynamics of SL 2 ( R ) over moduli space in genus two', Ann. of Math.(2) 165 (2007), no.2, pp. 397-456.
[MW14] Y. MINSKY and B. WEISS - 'Cohomology classes represented by measured foliations, and Mahler's question for interval exchanges ', Annales de E.N.S (to appear) (2014).
[Möl06] M. Möller, 'Variations of Hodge structure of Teichmüller curves ', J. Amer. Math. Soc. 19 no. 2, (2006) 327-344.
[Möl11] , 'Prym covers, theta functions and Kobayashi curves in Hilbert modular surfaces ', Amer. J. Math. (to appear) , arXiv:1111.2624 (2011).
[Zor06] A. ZORICH - 'Flat surfaces', Frontiers in number theory, physics, and geometry 437-583, Springer, Berlin, 2006.
INSTITUT FOURIER, UNIVERSITÉ DE GRENOBLE I, BP 74, 38402 SAINT-MARTIN-D'HÈRES , FRANCE E-mail address : [email protected]
IMB BORDEAUX-UNIVERSITÉ DE BORDEAUX, 351, COURS DE LA LIBÉRATION, 33405 TALENCE CEDEX, FRANCE E-mail address : [email protected]
| null |
[
"Erwan Lanneau",
"Duc-Manh Nguyen"
] |
2014-08-05T18:42:12+00:00
|
2014-10-18T13:24:33+00:00
|
[
"math.GT",
"math.AG",
"math.DS"
] |
Connected components of Prym eigenform loci in genus three
|
This paper is devoted to the classification of connected components of Prym
eigenform loci in the strata H(2,2)^odd and H(1,1,2) in the Abelian
differentials bundle in genus 3. These loci, discovered by McMullen are
GL^+(2,R)-invariant submanifolds (of complex dimension 3) that project to the
locus of Riemann surfaces whose Jacobian variety has a factor admitting real
multiplication by some quadratic order Ord_D.
It turns out that these subvarieties can be classified by the discriminant D
of the corresponding quadratic orders. However there algebraic varieties are
not necessarily irreducible. The main result we show is that for each
discriminant D the corresponding locus has one component if D is congruent to 0
or 4 mod 8, two components if D is congruent to 1 mod 8, and is empty
otherwise.
Our result contrasts with the case of Prym eigenform loci in the strata
H(1,1) (studied by McMullen) that is connected for every discriminant D.
|
1408.1065v3
|
## Using real algebraic geometry to solve combinatorial problems with symmetries
## Erik Sj¨land o
## November 11, 2014
## Abstract
Many combinatorial problems can be formulated as a polynomial optimization problem that can be solved by state-of-the-art methods in real algebraic geometry. In this paper we explain many important methods from real algebraic geometry, we review several applications and discuss implementation and computational aspects.
## Contents
| 1 | Introduction to semidefinite programing | Introduction to semidefinite programing | 2 |
|-----|-------------------------------------------------------|------------------------------------------------------------------------|-----|
| 2 | Polynomial optimization using real algebraic geometry | Polynomial optimization using real algebraic geometry | 5 |
| 3 | Sums of squares and the Positivstellensatz | Sums of squares and the Positivstellensatz | 6 |
| 4 | The moment problem and duality | The moment problem and duality | 9 |
| 5 | Exploiting symmetries in semidefinite programming | Exploiting symmetries in semidefinite programming | 15 |
| | 5.1 | Lov´sz a ϑ -number . . . . . . . . . . . . . . . . . . . . . . . . . . | 17 |
| | 5.2 | Block codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . | 18 |
| | 5.3 | Crossing numbers . . . . . . . . . . . . . . . . . . . . . . . . . . | 19 |
| | 5.4 | Kissing numbers . . . . . . . . . . . . . . . . . . . . . . . . . . | 19 |
| | 5.5 | Quadratic assignment problems . . . . . . . . . . . . . . . . . . | 20 |
| | 5.6 | Counting monochromatic arithmetic progressions in a 2-coloring | 20 |
| | 5.7 | Counting arithmetic progressions in a fixed density set . . . . . | 22 |
| 6 | Implementation | Implementation | 23 |
## 1 Introduction to semidefinite programing
Special cases of Linear Programming (LP) dates back to Fourier, but linear programming was in its generality first studied by Leonid Kantorovich in 1939 to reduce costs for the Russian army during World War II. In 1947 George Danzig published a paper with the famous simplex method to solve LPs, which is usually extremely efficient in practice but has exponential worst case scenarios. Several other algorithms have been proposed including the ellipsoid algorithm , which was proposed by Shor in 1972. Even though the ellipsoid algorithm is not convenient in practice it is the first algorithm proven to solve linear programming problems in polynomial time by Khachiyan in 1979 [11]. Another class of algorithms to solve linear programs are the interior point methods . The original method was invented by John von Neumann around the same time as Danzig introduced the simplex algorithm, and was later popularized by an efficient algorithm to give an approximate solution to an LP problem by Narendra Karmarkar in 1984 [10]. Complexity of LP-algorithms is still an active field of research, and LP has since it was introduced been used in a tremendous range of applications.
Semidefinite programming (SDP) started evolving from Linear Programming in the 1960s, and can be used to solve a wider variety of problems. Any LP can be formulated as an SDP, which is why it is natural to first define LP and then let it serve as a bridge to and as a first example of an SDP. There are various algorithms for finding close to optimal solutions to SDPs in polynomial time. Interior point methods have been the most popular and there are several SDP solvers (including CSDP, SeDuMi, SDPT3, DSDP, SDPA) that have good implementations of these algorithms. For a more extensive overview we refer to [48, 46].
In an LP you want to minimize or maximize a linear function under linear constraints. Any LP can be rewritten in canonical form:
$$\inf _ { \substack { c ^ { T } x \\ \text{subject to} } } \, & \, A x \leq b, \\ & x \geq 0, \\ & x \in \mathbb { R } ^ { n },$$
where A ∈ R m n × , b ∈ R m , c ∈ R n . In many textbooks there are also equality constraints in the formulation of the LP, but these are redundant as they can be rewritten as inequalities by d T x = e ⇔{ d T x ≤ e and -d T x ≤ - } e .
An LP in canonical form is referred to as the primal problem , and we define the dual problem of an LP to be
$$\sup _ { \substack { b ^ { T } y \\ \text{subject to} \\ \text{$A^{T} y \geq c$,} \\ \text{$y \geq 0$,} \\ \text{$y \in \mathbb{R}^{m}$.} }$$
The dual is also a linear LP problem, and it plays a strong role of many of the algorithms for solving the original primal LP problem. Fundamental in duality
theory is that the dual of the dual problem is the original primal problem. Even more important for optimization theory is the concepts of weak duality and strong duality . Weak duality ensures that any value of the dual problem gives a lower bound to the primal problem, and strong duality ensures that if the solution to one of the problems is finite the primal and dual solutions are equal. To prove that weak and strong duality holds for linear programs is a good exercise, and part of any elementary optimization course.
We are now ready to take the step into semidefinite programming. The essential idea is that we want to replace the non-negativity condition on the vector x ∈ R n with an appropriate non-negativity condition on a symmetric matrix X ∈ R n × n . It is natural to replace the non-negativity condition with a linear matrix inequality (LMI), which is equivalent to requiring that the matrix X is positive semidefinite , which has several equivalent definitions:
Theorem 1.1. Let A be a symmetric n × n matrix. The following are equivalent
- (a) A is positive semidefinite.
- (b) x T Ax ≥ 0 for all x ∈ R n .
- (c) All eigenvalues of A are nonnegative.
- (d) There is a unique lower triangular matrix L ∈ R n × n with L jj ≥ 0 for all j ∈ { 1 , . . . , n } such that LL T = A .
- (e) All principal minors of A are nonnegative.
The trace is a linear function on matrices, which generalizes the linear polynomials in LP. The canonical form of a semidefinite program is:
$$\dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \ delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \dots \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \ delta \ delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta delta \ delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \dia \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \Delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta
\quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \\ \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \ \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \ttext{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$ } \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \cdots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \\ \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \\ \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} & \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$} \quad \text{$f$$$
where C ∈ R n × n , A i ∈ R n × n and b i ∈ R for 1 ≤ i ≤ m .
Observe that the feasible set { X symmetric real n × n -matrix : tr( A X T i ) ≤ b i for i = 1 , . . . , m, X /followsequal 0 } is convex, and so SDP is a special class of convex optimization problems.
Just as in the LP case the original problem is called primal problem, and there is an associated dual problem:
$$\sup _ { \substack { b ^ { T } y \\ \text{subject to} } } & \, C - \sum _ { \substack { i = 1 \\ y \in \mathbb { R } ^ { m } } } ^ { m } A _ { i } y _ { i } \succeq 0,$$
Weak duality holds just as for LP, but strong duality does not always hold as the following standard example shows:
Example 1.2. The primal problem
$$\inf _ { \substack { x _ { 1 } \\ \text{subject to} } } \left ( \begin{array} { c c c } 0 & x _ { 1 } & 0 \\ x _ { 1 } & x _ { 2 } & 0 \\ 0 & 0 & x _ { 1 } + 1 \end{array} \right ) \succeq 0, \\ x _ { 1 }, x _ { 2 } & \in \mathbb { R },$$
has the dual
$$\sup _ { \substack { - y _ { 2 } \\ \text{subject to} } } \left ( \begin{array} { c c c } y _ { 1 } & ( 1 - y _ { 2 } ) / 2 & 0 \\ ( 1 - y _ { 2 } ) / 2 & 0 & 0 \\ 0 & 0 & y _ { 2 } \end{array} \right ) \succeq 0,$$
In order for the matrices to be positive semidefinite we see that both x 1 and (1 -y 2 ) / 2 must be 0 , so the solution to the primal problem is 0 and the solution to the dual problem is -1 .
Even though strong duality does not always hold it turns out that it holds under a condition called Slater's condition, which is a requirement of positive definiteness in either the primal or dual. Because of the importance of strong duality we finish the introduction by giving the exact statement:
Theorem 1.3. Let (P) denote a primal problem in canonical form and (D) its dual. Let p ∗ and d ∗ denote the optimal values of (P) and (D) respectively. If p ∗ is finite and there exists a solution X to (P) with X /follows 0 , then (D) is feasible and p ∗ = d ∗ . Analogously, if d ∗ is finite and there exists a solution y to (D) for which C -∑ m i =1 A y i i /follows 0 , then (P) is feasible and p ∗ = d ∗ .
## 2 Polynomial optimization using real algebraic geometry
In both mathematical and real world applications we encounter problems where we need to find the optimal value of a polynomial under polynomial constraints:
$$\inf _ { \substack { f ( x ) \\ \text{subject to } } } & \ g _ { 1 } ( x ) \geq 0, \dots, g _ { m } ( x ) \geq 0, \quad \quad ( 1 ) \\ & x \in \mathbb { R } ^ { n }.$$
We discuss how to attack polynomial optimization problems using methods that have grown out from classical questions about sums of squares. Although some of the problems and results in this chapter dates back to Hilbert it is still not obvious how to find exact or approximative solutions to most polynomial optimization problem. This is still an active area of research, and there are still many interesting related open problems. We aim to keep the level of abstraction low in order to make it easier to follow, and we only discuss results in real algebraic geometry related to polynomial optimization. A more extensive survey is the chapter on the subject by Laurent [25]. It is important to note that once a final sum of squares certificate has been found using our proposed methods, one does not need polynomial optimization to verify that the solution is correct.
## Definition 2.1. Let f , . . . , f 1 s , g ∈ R [ x , . . . , x 1 n ] .
/negationslash
A subset of R n is basic semialgebraic if it is on the form { f 1 ≥ 0 , . . . , f s ≥ 0 , g = 0 } and semialgebraic if it is a finite union of basic semialgebraic sets. A subset of R n is basic closed semialgebraic if it is on the form { f 1 ≥ 0 , . . . , f s ≥ 0 } .
/negationslash
We mentioned earlier that we can state equalities as inequalities, f = 0 ⇔ f ≥ 0 , -f ≥ 0, and in a similar fashion we can replace strict inequalities: f > 0 ⇔ f ≥ 0 , f = 0. We see that the basic semialgebraic sets contains all sets defined by = , ≥ /negationslash , = , > , whereas the basic closed semialgebraic sets are limited to = and ≥ , and is more suitable for optimization.
The classical Nullstellensatz that dates back to Hilbert states that if f ∈ C [ x , . . . , x 1 n ] vanishes on the subset of C n defined by g 1 = g 2 = · · · = g m = 0, where g , . . . , g 1 m ∈ C [ x , . . . , x 1 n ], then some power of f lies in the ideal of C [ x , . . . , x 1 n ] generated by g , . . . , g 1 m . The Nullstellensatz is a fundamental result in classical algebraic geometry, which deals with subsets in C n defined by polynomial equations. Real algebraic geometry deals with subset of R n defined by polynomial equations and inequalities, and the two topics are in some respects similar and in others not. An essential property for the field C n is that it is algebraically closed; every non-constant polynomial f ∈ C [ x , . . . , x 1 n ] has a root in C n . This is not the case for R n as for example x 2 + 1 = 0 does not have any real solutions. Despite this major drawback for the real numbers, many results carry over from the analogous complex cases and similar methods and techniques can often be used in real algebraic geometry. One thing that is not immediately clear is how to translate the Nullstellensatz into the real
algebraic context, and a lot of research has been done in this direction. The analogue to the Nullstellensatz is called the Positivstellensatz, and there are several versions of it depending on, among other things, the properties of the set { g 1 ≥ 0 , . . . , g m ≥ 0 } , and they are all aiming to find the most suitable condition to guarantee positivity of f . Once we have a certificate of positivity we can find the minimum of a polynomial f ( x ) ∈ R [ x ] by finding the largest λ ∈ R such that f ( x ) -λ is positive for all x ∈ R n . In this section we explore the duality between positive polynomials and the theory of moments, and how it can be used to find solutions or bounds to polynomial optimization problems.
## 3 Sums of squares and the Positivstellensatz
To be able to state the form of the Positivstellensatz that is most suitable for our optimization purposes we first need to understand the basic connections between positivity and sums of squares. First some simplifying notation: We write R [ x ] = R [ x , . . . , x 1 n ], f ≥ 0 when f ( x ) ≥ 0 for all x ∈ R n and f > 0 when f > 0 for all x ∈ R n .
It is obvious that if f is a sum of squares, say f = f 2 1 + · · · + f 2 k , then f ( x ) = f 1 ( x ) 2 + · · · + f k ( x ) 2 ≥ 0 for all x ∈ R n whereas questions about the converse can be very difficult. There exist nonnegative polynomials that are not sums of squares of polynomials, which was known already by Hilbert. Motzkin provided the first concrete example of such a polynomial in 1967:
Proposition 3.1. The polynomial s x, y ( ) = 1 -3 x y 2 2 + x y 2 4 + x y 4 2 is positive on R 2 , but it is not a sum of squares in R [ x, y ] .
Proof. Positivity follows directly by plugging in a = 1, b = x y 2 4 , and c = x y 4 2 into the inequality of the arithmetic and geometric means
$$\frac { a + b + c } { 3 } \geq ( a b c ) ^ { 1 / 3 } ( \text{if } a, b, c \geq 0 ).$$
To show that s x, y ( ) is not a sum of squares, suppose to the contrary that s = f 2 1 + . . . f 2 n for some polynomials f i ∈ R [ x, y ]. If d = max deg( { f i ) | i ∈ { 1 , . . . , n }} , then s must be of degree 2 d ( ≤ 2 d is obvious, and for the other direction let f id be the homogeneous part of degree d of f i . The part of degree 2 d of f is f 2 1 d + · · · + f 2 nd out of which at least one is nonzero by definition of d , and ≥ 2 d follows). Since s is of degree 6 any f i can have degree at most 3 and must therefore be a linear combination of the monomials
$$1, x, y, x ^ { 2 }, x y, y ^ { 2 }, x ^ { 3 }, x ^ { 2 } y, x y ^ { 2 }, y ^ { 3 }.$$
If an x 3 -term would appear in any f i then an x 6 term would appear in s with a positive coefficient. There is no way to cancel that term, so there is no x 3 -term in any f i . By the same argument we can conclude that there are no terms x 2 , x , y 3 , y 2 and y either and so f , . . . , f 1 n are on the form
$$f _ { i } = a _ { i } + b _ { i } x y + c _ { i } x ^ { 2 } y + d _ { i } x y ^ { 2 }.$$
We get a contradiction by
$$\sum _ { i = 1 } ^ { n } b _ { i } ^ { 2 } = - 3.$$
Despite the negative result one could hope for general theorems when we change the settings slightly. In 1900 Hilbert posed a famous set of problems, of which the 17th asked was whether any positive polynomial can be written as a sum of squares of rational functions. Artin proved the general case of this problem in 1927:
Theorem 3.2. For any f ∈ R [ x , . . . , x 1 n ] it is true that f ≥ 0 on R n if and only if f is a sum of squares of rational functions.
We are now ready to state the first version of the Positivstellensatz, which has some similarities with Hilbert's 17th problem. The main ideas were due to Krivine in 1964 [21], and Stengle in 1974 [44]. We present a version from [29] that fits with our notation.
Theorem 3.3 (Positivstellensatz) . Let f, g 1 , . . . , g m ∈ R [ x , . . . , x 1 n ] , K = { x ∈ R n | g i ( x ) ≥ 0 , i = 1 , . . . , m } and
$$T = \{ \sum _ { e \in \{ 0, 1 \} ^ { m } } \sigma _ { e } g ^ { e } | \sigma _ { e } \text{ is a sum of squares for all } e \in \{ 0, 1 \} ^ { m } \}.$$
- (a) f > 0 on K if and only if there exists p, q ∈ T such that pf = 1 + q .
- (b) f ≥ 0 on K if and only if there exists an integer m ≥ 0 and p, q ∈ T such that pf = f 2 m + q .
- (c) f = 0 on K if and only if there exists an integer m ≥ 0 such that -f 2 m ∈ T .
- (d) K = ∅ if and only if -1 ∈ T .
Recall that we want to find a good way of minimizing f . This would be possible if we could find the largest λ such that (b) holds for f -λ . The problem is that there is no algorithmic way of doing this since we cannot control the degrees in this version of the Positivstellensatz. Although the theorem is very important theoretically it is not satisfactory for our application. To get a certificate for positivity that is more suitable we need to impose further conditions on K . In the case when K is compact Scm¨ udgen managed to prove this version of the Positivstellensatz in 1991 [37]:
Theorem 3.4 (Schm¨ udgen's Positivstellensatz) . Let K = { x ∈ R n | g 1 ( x ) ≥ 0 , . . . , g m ( x ) ≥ 0 } be compact. If f is strictly positive on K , then
$$f = \sum _ { e \in \{ 0, 1 \} ^ { m } } \sigma _ { e } g ^ { e }$$
where σ e is a sum of squares for all e ∈ { 0 1 , } m .
This version is a major improvement, but in terms of computational efficiency it could be improved further as it is exponential in terms of the number of boundary conditions g i . Two years later Putinar added an additional condition on the polynomials to make another version of the Positivstellensatz [35] that turns out to work well with semidefinite relaxations, and which is only linear in the number of boundary conditions g i . Let us introduce the condition, state the theorem and then explore when the condition is satisfied.
Definition 3.5. Given polynomials g , . . . , g 1 m ∈ R [ x , . . . , x 1 n ] , the quadratic module generated by g , . . . , g 1 m is defined by:
$$\text{QM} ( g _ { 1 }, \dots, g _ { m } ) = \{ \sigma _ { 0 } + \sum _ { i = 1 } ^ { m } \sigma _ { i } g _ { i } | \sigma _ { 0 }, \dots, \sigma _ { m } \ a r e \ s u m s \ o f \ s q u a r e s \}.$$
Definition 3.6. A quadratic module QM( g , . . . , g 1 m ) is Archimedean if
$$N - \sum _ { i = 1 } ^ { n } x _ { i } ^ { 2 } \in Q M ( g _ { 1 }, \dots, g _ { m } )$$
for some N ∈ N .
Theorem 3.7 (Putinar's Positivstellensatz) . Let
$$K = \{ x \in \mathbb { R } ^ { n } | g _ { 1 } ( x ) \geq 0, \dots, g _ { m } ( x ) \geq 0 \}$$
be compact. If f is strictly positive on K and the associated quadratic module QM( g , . . . , g 1 m ) is Archimedean, then f = σ 0 + ∑ m i =1 σ g i i where σ , . . . , σ 1 m are sums of squares.
Putinar asked the question Assuming K is compact, is it true that QM is Archimedean? which would then allow us to remove this additional condition from the theorem. It is true and easy to show when m = 1, and is highly nontrivial but still true when m = 2, which was proven by Jacobi and Prestel in 2001 [8]. It is also true if the ring R [ x ,...,x 1 n ] M ∩-M has Krull dimension ≤ 1, but in general it is not true if m ≥ 3 and the Krull dimension is 2 or more. The different cases are discussed in further detail starting at page 97 of ?? .
For the general case we provide the following conditions due to Schm¨dgen u in 1991 [37] that are equivalent to the Archimedean condition:
Theorem 3.8. The following are equivalent:
- (1) QM( g , . . . , g 1 m ) is Archimedean.
- (2) There exist finitely many polynomials t 1 , . . . , t m ∈ QM( g , . . . , g 1 m ) such that the set
$$\{ x \in \mathbb { R } ^ { n } | t _ { 1 } ( x ) \geq 0, \dots, t _ { m } ( x ) \geq 0 \}$$
$$i s \, \text{compact and} \, \prod _ { i \in I } t _ { i } \in \text{QM} ( g _ { 1 }, \dots, g _ { m } ) \text{ for all } I \subset \{ 1, \dots, m \}$$
(3) There exists a polynomial p ∈ QM( g , . . . , g 1 m ) such that { x ∈ R n | p x ( ) ≥ 0 } is compact.
We have introduced enough state-of-the-art tools to find sum of squares based certificates to check positivity of polynomials. Recall that our goal is to solve polynomial optimization problems:
$$\inf _ { \substack { f ( x ) \\ \text{subject to } } } \, \underbrace { g _ { 1 } ( x ) \geq 0, \dots, g _ { m } ( x ) \geq 0, } _ { \substack { x \in \mathbb { R } ^ { n }. } }$$
Let
$$K = \{ x \in \mathbb { R } ^ { n } \colon g _ { j } ( x ) \geq 0, j = 1, \dots, m \}$$
denote the feasible set to the polynomial optimization problem. Sometimes we impose the following technical assumption:
Assumption 3.9. Let us assume that there exists a polynomial σ ∈ R [ x , . . . , x 1 n ] on the form
$$\sigma = \sigma _ { 0 } + \sum _ { j = 1 } ^ { m } \sigma _ { j } g _ { j },$$
where σ i is a sum of squares for i = 1 , . . . , m , such that the set { x ∈ R n | σ x ( ) ≥ 0 } is compact.
As the assumption is equivalent to part (3) of Theorem 3.8 it guarantees that quadratic module generated by polynomials g , . . . , g 1 m is Archimedean, and thus that Putinar's Positivstellensatz can be applied. When assumption 3.9 holds it is clear that the following problem has the same optimal value as our original polynomial optimization problem:
$$\sup _ { \begin{array} { c } \lambda \\ \text{subject to} \end{array} } \lambda & f ( X ) - \lambda = \sigma _ { 0 } + \sum _ { i = 1 } ^ { m } \sigma _ { i } g _ { i } \\ & \sigma _ { 0 }, \dots, \sigma _ { m } \text{ are sums of squares.}$$
In particular, λ is now the variable and X a formal indeterminate. Next we will explore the duality between these problems, and discuss how we can relax them to find lower bounds for polynomials on K .
## 4 The moment problem and duality
To fully understand how to solve polynomial optimization problems we need to introduce a duality theory, and to introduce a duality theory we first need to study the moment problem. The problem we are interested in is a special version of the Generalized Moment Problem (GMP). In its general form, a lot of problems from applied mathematics, statistics, probability theory, economics,
engineering, physics and operation research can be stated in terms of a moment problem. Although the GMP provides a nice theoretical framework and has great modeling power, it cannot in its full generality be solved numerically. We do not state the theorems in their most general form, but rather in our context of minimizing a polynomial over a closed basic semialgebraic set. We refer to [23] for the reader who is interested in a more complete background.
In our context f, g 1 , . . . , g m are polynomial from R n to R and K = { g 1 ≥ 0 , . . . , g m ≥ 0 } a compact closed basic semialgebraic set. This is a particular instance of the GMP (Theorem 1.1 page 6, [23]).
Theorem 4.1. Let M ( K ) + be the space of finite Borel measures µ on K . The following two problems are equivalent in the sense that f ∗ = ρ mom .
$$\begin{smallmatrix} f ^ { * } = \inf & f ( x ) \\ \text{subject to} & x \in K. \end{smallmatrix}$$
$$\rho _ { \text{mom} } = \inf _ { \mu \in \mathcal { M } ( K ) + } \ \int _ { K } f d \mu \\ \text{subject to} \ \int _ { K } d \mu = 1. \\.. \ \quad \ e \.. \ e \.. \ e \.. \ e \.. \ e \..$$
Since (3) is a linear program we find its dual through the standard procedure in linear programming:
$$\begin{smallmatrix} \rho _ { \text{pop} } = \sup & \lambda \\ \text{subject to} & f ( x ) - \lambda \geq 0 \text{ for all } x \in K. \end{smallmatrix}$$
In general strong duality holds when we Slater's condition is satisfied. In our case, when we are minimizing a polynomial over a compact feasible set K , we do not need Slater's condition for strong duality to hold (Special case of Theorem 1.3 page 8, [23]):
Theorem 4.2. Suppose that K is compact, that f ∈ R [ x ] and that we have the primal and dual pair (3) and (4) . Then ρ mom = ρ pop and if (3) has a feasible solution then it has an optimal solution.
Let v r ( x ) = [1 , x 1 , . . . , x n , x 1 x , . . . , x 2 2 n , . . . , x 1 · · · x , . . . , x r r n ] T be the vector of all monomials up to degree r (order first reversed graded lexicographically within groups with same set of exponentials, then the groups are ordered reversed graded lexicographically by leading terms in the groups). Let similarly v x ( ) be the infinite sequence of monomials v x ( ) = v r →∞ ( x ). Let y = ( y α ) ⊂ R be an infinite sequence indexed by α ∈ N n ordered the same way as the monomials in v r , and define L y : R [ x ] → R by
↦
$$f ( x ) = \sum _ { \alpha \in \mathbb { N } ^ { n } } f _ { \alpha } x ^ { \alpha } \mapsto L _ { y } ( f ) = \sum _ { \alpha \in \mathbb { N } ^ { n } } f _ { \alpha } y _ { \alpha }.$$
Let M r ( y ) be the moment matrix defined by
$$\mathbf M _ { r } ( y ) ( \alpha, \beta ) = L _ { y } ( x ^ { \alpha } x ^ { \beta } ) = y _ { \alpha + \beta }, \text{ for all } \alpha, \beta \in \mathbb { N } _ { r } ^ { n }.$$
Equivalently M r ( y ) = L y ( v r ( x v ) r ( x ) ), ′ which is why we say that the rows and columns of M r ( y ) are labeled by v r ( x ).
Given a polynomial u ∈ R [ x ] with coefficient vector u = { g } we define the localizing matrix to be the matrix M r ( uy ), obtained from M r ( y ) by:
$$\mathbf M _ { r } ( u y ) ( \alpha, \beta ) = L _ { y } ( u ( x ) x ^ { \alpha } x ^ { \beta } ) = \sum _ { g \in \mathbb { N } ^ { n } } u _ { g } y _ { g + \alpha + \beta }, \text{ for all } \alpha, \beta \in \mathbb { N } ^ { n } _ { r }.$$
To understand the notation let us consider two examples.
Example 4.3. When n = 2 , r = 1 and u x ( ) = a -x 2 1 + bx 2 2 we have
$$v _ { r } ( x ) = ( 1, x _ { 1 }, x _ { 2 } ),$$
$$y = ( y _ { 0 0 }, y _ { 1 0 }, y _ { 0 1 }, y _ { 1 1 }, y _ { 2 0 }, y _ { 0 2 }, \dots ), \\ \mathbf M _ { 1 } ( y ) = \left [ \begin{array} { c c c } y _ { 0 0 } & y _ { 1 0 } & y _ { 0 1 } \\ y _ { 1 0 } & y _ { 2 0 } & y _ { 1 1 } \\ y _ { 0 1 } & y _ { 1 1 } & y _ { 0 2 } \end{array} \right ],$$
and
$$M _ { 1 } ( u y ) = \left [ \begin{array} { c c c c } a y _ { 0 0 } - y _ { 2 0 } + b y _ { 0 2 } & a y _ { 1 0 } - y _ { 3 0 } + b y _ { 1 2 } & a y _ { 0 1 } - y _ { 2 1 } + b y _ { 0 3 } \\ a y _ { 1 0 } - y _ { 3 0 } + b y _ { 1 2 } & a y _ { 2 0 } - y _ { 4 0 } + b y _ { 2 2 } & a y _ { 1 1 } - y _ { 3 1 } + b y _ { 1 3 } \\ a y _ { 0 1 } - y _ { 2 1 } + b y _ { 0 3 } & a y _ { 1 1 } - y _ { 3 1 } + b y _ { 1 3 } & a y _ { 0 2 } - y _ { 2 2 } + b y _ { 0 4 } \end{array} \right ], \\ - \quad \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \end{array}.$$
Example 4.4. When n = 2 , r = 2 and u = ax 3 1 we have
$$v _ { r } ( x ) = ( 1, x _ { 1 }, x _ { 2 }, x _ { 1 } x _ { 2 }, x _ { 1 } ^ { 2 }, x _ { 2 } ^ { 2 } ),$$
y = ( y 00 , y 10 , y 01 , y 11 , y 20 , y 02 , y 21 , y 12 , y 30 , y 03 , y 22 , y 31 , y 13 , y 40 , y 04 , . . . ) ,
$$\mathbf M _ { 2 } ( y ) = \left [ \begin{array} { c c c c c } y _ { 0 0 } & y _ { 1 0 } & y _ { 0 1 } & y _ { 1 1 } & y _ { 2 0 } & y _ { 0 2 } \\ y _ { 1 0 } & y _ { 2 0 } & y _ { 1 1 } & y _ { 2 1 } & y _ { 3 0 } & y _ { 1 2 } \\ y _ { 0 1 } & y _ { 1 1 } & y _ { 0 2 } & y _ { 1 2 } & y _ { 2 1 } & y _ { 0 3 } \\ y _ { 1 1 } & y _ { 2 1 } & y _ { 1 2 } & y _ { 2 2 } & y _ { 3 1 } & y _ { 1 3 } \\ y _ { 2 0 } & y _ { 3 0 } & y _ { 2 1 } & y _ { 3 1 } & y _ { 4 0 } & y _ { 2 2 } \\ y _ { 0 2 } & y _ { 1 2 } & y _ { 0 3 } & y _ { 1 3 } & y _ { 2 2 } & y _ { 0 4 } \end{array} \right ] \\ \Gamma \, \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_ \_$$
and
$$M _ { 2 } ( u y ) = \left [ \begin{array} { c c c c c } a y _ { 3 0 } & a y _ { 4 0 } & a y _ { 3 1 } & a y _ { 4 1 } & a y _ { 5 0 } & a y _ { 3 2 } \\ a y _ { 4 0 } & a y _ { 5 0 } & a y _ { 4 1 } & a y _ { 5 1 } & a y _ { 6 0 } & a y _ { 4 2 } \\ a y _ { 3 1 } & a y _ { 4 1 } & a y _ { 3 2 } & a y _ { 4 2 } & a y _ { 5 1 } & a y _ { 3 3 } \\ a y _ { 4 1 } & a y _ { 5 1 } & a y _ { 4 2 } & a y _ { 5 2 } & a y _ { 6 1 } & a y _ { 4 3 } \\ a y _ { 5 0 } & a y _ { 6 0 } & a y _ { 5 1 } & a y _ { 6 1 } & a y _ { 7 0 } & a y _ { 5 2 } \\ a y _ { 3 2 } & a y _ { 4 2 } & a y _ { 3 3 } & a y _ { 4 3 } & a y _ { 5 2 } & a y _ { 3 4 } \end{array} \right ].$$
The following theorem, which is due to Haviland and Riesz, is one of the main building blocks to many proofs in duality theory (Theorem 3.1 page 53, [23]):
Theorem 4.5. Let y = ( y α α ) ∈ N n ⊂ R be an infinite sequence, and let K ⊂ R n be closed. There exists a finite Borel measure µ on K such that
$$\int _ { K } x ^ { \alpha } d \mu = y _ { \alpha }, \text{ for all } \alpha \in \mathbb { N } ^ { n }$$
if and only if L y ( p ) ≥ 0 for all polynomials p ∈ R [ x ] that are nonnegative on K .
One of the theorems based on Theorem 4.5 is the following (Theorem 3.8 page 63, [23]):
Theorem 4.6. Let y = ( y α α ) ∈ N n be a given infinite sequence in R , introduce the polynomials g , . . . , g 1 m ∈ R [ x ] and let K ⊂ R n be the closed basic semi-algebraic set
$$K = \{ x \in \mathbb { R } ^ { n } | g _ { i$$
Assume that K is compact.
- (a) There exists a finite Borel measure µ on K such that ∫ K x dµ α = y α for all α ∈ N n if and only if M r ( g J y ) /followsequal 0 for all J ⊆ { 1 , . . . , m } and all r ∈ N .
- (b) Assume that Assumption 3.9 holds, then there exists a finite Borel measure µ on K such that ∫ K x α dµ = y α for all α ∈ N n if and only if M r ( g y j ) /followsequal 0 and M r ( y ) /followsequal 0 for all j ∈ { 1 , . . . , m } and all r ∈ N .
$$\text{If we let } y _ { \alpha } = \int _ { K } x ^$$
y 0 + ∑ α ∈ N /integerdivide n 0 0 y α = y 0 , and can restate (3) as
$$\rho _ { \text{mom} } = \inf _ { y } & \quad L$$
If we let v j = deg( g j ) and apply Theorem 4.6 we get
$$\begin{array} {$$
When assumption 3.9 holds we get the dual problem through (4) and Putinar's Positivstellensatz:
$$\rho _ { \text{pop} } = \sup _ { \lambda, \{ \$$
$$\begin{array} { c } \lambda, \{ \sigma _ { i } \}$$
That σ i is a sum of squares is equivalent to that σ i = v x ( ) T Q v x i ( ) for some positive semidefinite matrix Q i , hence we get the equivalent formulation
$$\begin{array} {$$
Since the equality has to hold for every monomial separately we can rewrite the problem in order to get the problem on the form of a semidefinite program. If we use the notation h X ( ) = ∑ α [ h X ( )] α X α , we get the equivalent problem where [ f ( X ) -v X ( ) T Q v X 0 ( )+ ∑ m i =1 v X ( ) T Q v X g i ( ) i ( X )] g is a linear polynomial in the entries of the matrices Q i . Since the objective function is linear and constraints either linear or linear matrix inequalities this is indeed a semidefinite program.
$$\begin{array} { l l } \text{we use the notation } h ( X ) =$$
The primal and dual problems (8) and (9) have the same optimal values as the original pair (3) and (4), hence strong duality holds by Theorem 4.2. As you might have noticed we have worked our way to some infinite-dimensional semidefinite programs both in the primal and dual. What the primal and dual problems also share is that they are easy to relax. In order to find a lower bound to the primal problem (8) we only use moment matrices and localization matrices of degree r , and to relax the dual problem (10) we can limit the degrees by replacing v x ( ) with v r ( x ). Let s be a positive integer, we have the following hierarchy of primal and dual relaxations, commonly referred to as the Lasserre hierarchy:
$$\rho _ { s } ^ { * } = \inf _ { y } \quad L _$$
and
$$\text{and} \quad \rho _ { s } = \sup _ { \lambda,$$
The following important convergence result by Lasserre (similar to Theorem 3.4 page 805 [22], and Theorem 4.1 page 79 [23]) holds:
Theorem 4.7. Let f, g 1 , . . . , g m ∈ R [ x ] , let K = { g 1 ≥ 0 , . . . , g m ≥ 0 } be compact, and assume assumption 3.9 holds. Let ρ mom be the optimal value of (8) , assumed to be finite, let ρ pop be the optimal value of (9) , and consider the sequence of primal and dual semidefinite relaxations defined in (11) and (12) with optimal values ( ρ i ) and ( ρ ∗ i ) respectively. Then ρ ∗ i → ρ mom and ρ i → ρ mom when i →∞ .
Proof. By Theorem 4.2 we know that ρ pop = ρ mom . Furthermore we know that ρ mom ≥ ρ ∗ i since any feasible solution to the primal is a feasible solution to its relaxation, ρ pop ≥ ρ i since every solution to the dual relaxation is a solution to the dual, and ρ i ≤ ρ ∗ i because of weak duality in semidefinite programming. Thus we have
$$\rho _ { i } \leq \rho _ { i } ^ { * } \leq \rho _ { \text{pop} } = \rho _ { \text{mom} }, \text{ for all } i \geq \max _ { j } ( \deg ( g _ { j } ) ).$$
Let /epsilon1 > 0 and let λ be a feasible solution to (9) satisfying
$$\rho _ { \text{pop} } - \epsilon = \rho _ { \text{mom} } - \epsilon \leq \lambda \leq \rho _ { \text{pop} } = \rho _ { \text{mom} }.$$
Since λ is a feasible solution
$$f - \lambda \geq 0, \text{ on } K.$$
Let ¯ = λ λ -/epsilon1 be a new solution, it is indeed feasible since f -¯ λ ≥ /epsilon1 > 0 on K . Using Putinar's Positivstellensatz we get
$$f - \bar { \lambda } = \sigma _ { 0 } + \sum _ { j = 1 } ^ { m } \sigma _ { j } g _ { j },$$
where σ , . . . , σ 0 m are sums of squares. This implies that ¯ is a feasible solution λ to the relaxation (12) if 2 i ≥ max j σ g j j , which in turn implies that
$$\rho _ { \text{mom} } - 2 \epsilon \leq \bar { \lambda } \leq \rho _ { i },$$
where we got the first equality by removing /epsilon1 from both sides of the first inequality in (13). Since we have
$$\rho _ { \text{mom} } - 2 \epsilon \leq \rho _ { i } \leq \rho _ { i } ^ { * } \leq \rho _ { \text{mom} }$$
and since we picked /epsilon1 > 0 arbitrary the result follows.
It is possible to check whether the sequence of semidefinite relaxations has converged at relaxation i [7], but there is no guarantee that it converges in a finite number of steps. For many practical purposes the semidefinite program becomes huge and not possible to solve for computational reasons already for quite small values of i . Even when there is no hope for an optimal solution to the original problem the relaxation provides a lower bound, and the higher the order of the relaxation the better the lower bound.
## 5 Exploiting symmetries in semidefinite programming
In this section we explain how one can exploit symmetries in semidefinite programming. These methods or similar methods have been used in one way or another in the applications that are introduced in this section. For more details we refer to the references in each of the applications.
Let C and A , . . . , A 1 m be real symmetric matrices and b 1 , . . . , b m real numbers. In this section we provide the tools to reduce the order of the matrices in the semidefinite programming problem
$$\max \{ \text{tr} ( C X ) \ | \ X \text{ positive semidefinite,} \, \text{tr} ( A _ { i } X ) = b _ { i } \text{ for } i = 1, \dots, m \}$$
when it is invariant under a group acting on its variables.
Inspired by [18] and [15], we are using a ∗ -representation in order to reduce the dimension of the problem. For an introduction to ∗ -representations we refer to the book by Takesaki [45]. This method as well as other state-of-the art methods for invariant semidefinite programs are discussed in [1]. Other important recent contributions include [9, ? , 47, 28, 32, 36].
Definition 5.1. A matrix ∗ -algebra is a collection of matrices closed under addition, scalar and matrix multiplication, and transposition.
Let G be a finite group acting on a finite set Z and define a homomorphism h : G → S | Z | where S | Z | is the group of all permutations of Z . For every element g ∈ G we have a permutation h g = h g ( ) of Z , for which h gg ′ = h h g g ′ and h g -1 = h -1 g . For every permutation h g , define the corresponding permutation matrix M g ∈ { 0 1 , } | Z |×| Z | element-wise by
$$( M _ { g } ) _ { i, j } = \begin{cases} \ 1 & \text{if $h_{g}(i)=j$,} \\ 0 & \text{otherwise} \end{cases}$$
for all i, j ∈ Z . Let the span of these matrices define the matrix ∗ -algebra
$$\mathcal { A } = \left \{ \sum _ { g \in G } \lambda _ { g } M _ { g } \ | \ \lambda _ { g } \in \mathbb { R } \right \}.$$
The matrices X satisfying XM g = M X g for all g ∈ G are the invariant matrices of G . The collection of all such matrices,
$$\mathcal { A } ^ { \prime } = \{ X \in \mathbb { R } ^ { n \times n } | X M = M X \text{ for all } M \in \mathcal { A } \},$$
is the commutant of A , which again is a ∗ -algebra. Let d = dim A ′ be the dimension of the commutant.
A ′ has a basis of { 0 1 , } -matrices E , . . . , E 1 d such that ∑ d i =1 E i = J , where J is the matrix of size | Z | × | Z | with all ones.
For every i = 1 , . . . , d we normalize E i to
$$B _ { i } = \frac { 1 } { \sqrt { t r ( E _ { i } ^ { T } E _ { i } ) } } E _ { i }, \\ \intertext { there } \delta _ { z z } \text{ is the Kronecker} \, d.$$
i,j so that tr( B B T i j ) = δ i,j where δ i,j is the Kronecker delta. We define the multiplication parameters λ k by
$$B _ { i } B _ { j } = \sum _ { k = 1 } ^ { d } \lambda _ { i, j } ^ { k } B _ { k }$$
for i, j, k = 1 , . . . , d .
The multiplication parameters are used to define d × d -matrices L , . . . , L 1 d by
$$( L _ { k } ) _ { i, j } = \lambda _ { k, j } ^ { i }$$
for k, i, j = 1 , . . . , d . The matrices L , . . . , L 1 d spans the linear space
$$\mathcal { L } = \{ \sum _ { i = 1 } ^ { d } x _ { i } L _ { i } \colon x _ { 1 }, \dots, x _ { d } \in \mathbb { R } \}$$
Theorem 5.2 ([18]) . The linear function φ : A → ′ R d × d defined by φ B ( i ) = L i for i = 1 , . . . , d is a bijection. Additionally, the linear function also satisfies φ XY ( ) = φ X φ Y ( ) ( ) and φ X ( T ) = φ X ( ) T for all X,Y ∈ A ′ .
Corollary 5.3 ([18]) . ∑ d i =1 x B i i is positive semidefinite if and only if ∑ d i =1 x L i i is positive semidefinite.
Given that it is possible to find a solution X ∈ A ′ we can use Corollary 5.3 to reduce the order of the semidefinite constraint.
Lemma 5.4. There is a solution X ∈ A ′ to a G -invariant semidefinite program max tr( { CX ) | X positive semidefinite , tr( A X i ) = b i for i = 1 , . . . , m } .
Proof. Let C, A , . . . , A 1 m be | Z |×| Z | matrices commuting with M g for all g ∈ G . If X is an optimal solution to the optimization problem then the group average,
X ′ = 1 | G | ∑ g ∈ G M XM g T g , is also an optimal solution: It is feasible since
$$\mathfrak { z } \, M _ { g } X \, M _ { g } ^ { \prime } \,, \, \text{ is also an optimal solution: It is teasib} \\ \text{tr} ( A _ { j } X ^ { \prime } ) \ \ & = \text{tr} ( A _ { j } \frac { 1 } { | G | } \sum _ { g \in G } M _ { g } X M _ { g } ^ { T } ) \\ \ & = \text{tr} ( \frac { 1 } { | G | } \sum _ { g \in G } M _ { g } A _ { j } X M _ { g } ^ { T } ) \\ \ & = \text{tr} ( \frac { 1 } { | G | } \sum _ { g \in G } A _ { j } X ) \\ \ & = \text{tr} ( A _ { j } X ), \\ \text{ used that the well-known fact that the trace is in}$$
where we have used that the well-known fact that the trace is invariant under change of basis. By the same argument tr( CX ′ ) = tr( CX ), which implies that X ′ is optimal. It is easy to see that X ′ ∈ A ′ .
All in all we get the following theorem:
Theorem 5.5 ([18]) . The G -invariant semidefinite program
$$\max \{ \text{tr} ( C X ) \ | \ X \succeq$$
has a solution X = ∑ d i =1 x B i i that can be obtained by
$$\max \{ \text{tr} ( C X ) \ | \ \sum _ { i =$$
Dimension reduction of invariant semidefinite programs is useful in a wide range of combinatorial problems including Lov´sz a ϑ number, crossing numbers, error-correcting codes and kissing numbers. One can also do dimension reduction in quadratic assignment problems, which for example has been used to find a new relaxation of the traveling salesman problem. We briefly explain these applications below, and a more extensive survey of these and other recent applications can be found in [12, 1]. The methods are also useful to reduce the dimension of the invariant semidefinite programs arising when counting arithmetic progressions, which will be seen in later chapters. Many of the results in this thesis relies on computations that would not be possible without this step.
## 5.1 Lov´ asz ϑ -number
The ϑ number of a graph G = ( V, E ) is the optimal value of the semidefinite program
/negationslash
$$\begin{array} {$$
One of the main reasons to study ϑ G ( ) is that it is 'sandwiched' between the independence number α G ( ), the size of the largest clique in G , and χ G ( ¯ ), the chromatic number of the complement of G . We have the following theorem from Lov´ asz in 1979 [27]:
Theorem 5.6 (Lov´ asz sandwich theorem) .
$$\alpha ( G ) \leq \vartheta ( G ) \leq \chi ($$
To compute α G ( ) and χ G ( ) are NP-complete problems, thus if one can find ϑ quickly it gives valuable bounds for the other two graph properties when they cannot be obtained. For a long exposition of the ϑ number we refer to the notes by Knuth on the subject [20].
When there are a lot of symmetries in the graph one can use the described techniques to reduce the dimension of the semidefinite program. This allows calculation of the ϑ -number for larger graphs than otherwise possible [5, 17].
## 5.2 Block codes
Let us fix an alphabet q = { 0 1 , , . . . , q - } 1 for some integer q ≥ 2. q n is called the Hamming space and it is equipped with a metric d ( · , · ) called Hamming distance , which is given by
/negationslash
$$d ( u, v ) = | \{ i \, \colon u _ { i } \$$
A subset C ⊆ q n is called a code of length n , and for a nonempty code we define the minimum distance of C to be
$$\min \{ d ( u, v ) \, \colon u, v \in C \text{ are distinct} \}.$$
With A q ( n, d ) we denote the maximum size of a code of length n and minimum distance at least d :
$$A _ { q } ( n, d ) = \max \{ | C | \, \colon C \in \mathbf q ^ { n } \text{ has minimum distance at least } d \}.$$
A q ( n, d ) is an important quantity in coding theory, and it is of interest to find good upper and lower bounds. Lower bounds are usually obtained by finding explicit constructions of codes, whereas upper bounds are usually obtained using other methods.
The Hamming graph G q ( n, d ) is constructed on the vertex set q n by connecting two words u, v ∈ q n if d u, v ( ) < d . Since the codes of minimum distance at most d correspond to independent sets in G q ( n, d ) we have
$$A _ { q } ( n, d ) = \alpha ( G _ { q } ( n, d ) )$$
where α is the independence number of a graph. Recall that Lovasz ϑ number is a bound for the independence number, and hence the previous subsection gives a first class of bounds.
It is also fairly easy to see that for the graph parameter
$$\vartheta ^ { \prime } ( G ) = \max \{ \langle X, J \rangle \, \colon \langle X, I \rangle = 1, X _ { u v } = 0 \text{ if } u v \in E ( G ), X \geq 0, X \succeq 0 \},$$
where I is the identity and J the all one matrix, we have α G ( ) ≤ ϑ ′ ( G ) ≤ ϑ G ( ). Thus one can try to find an upper bound A q ( n, d ) by solving the semidefinite
program ϑ ′ ( G q ( n, d )). The semidefinite program has an exponential size, but by symmetry reduction it can be solved more efficiently. It was found independently in [30] and [38] that the semidefinite program can be reduced to a linear program that gives the Delsarte bounds [3]. Invariant semidefinite programming methods have been used to find sharper bounds in several papers including [39, 6, 24, 25].
## 5.3 Crossing numbers
The crossing number cr( G ) of a graph G is the minimum number of pairwise intersection of edges when G is drawn in a plane. Exact crossing numbers are only known in few examples, and to find exact crossing numbers of families of graphs is still an active area of research.
Zarankiewicz claimed in 1954 [49] that the crossing number cr( K m,n ) of the complete bipartite graph equals the Zarankiewicz number, which is given by Z m,n ( ) = /floorleft m -1 2 /floorright/floorleft m 2 /floorright/floorleft n -1 2 /floorright/floorleft n 2 /floorright . Ringel and Kainen independently found a gap in the proof and Zarankiewicz's claim has been conjectured since. Zarankiewicz argument for the upper bound was correct, and thus it is known that
$$c x ( K _ { m, n } ) \leq Z ( m, n )$$
for any positive integers m,n , and so one usually tries to improve on the lower bound. It is also conjectured that the crossing number of the complete graph cr( K n ) equals Z n ( ) = /floorleft n 2 /floorright/floorleft n -1 2 /floorright/floorleft n -2 2 /floorright/floorleft n -3 2 /floorright . Using semidefinite programming it was found in [16] that for each m ≥ 9 it holds that
$$\lim _ { n \to \infty } \text{cr} ( K _ { m,$$
lim n →∞ cr( K n,n ) /Z n, n ( ) ≥ 0 83 . , lim n →∞ cr( K n ) /Z n ( ) ≥ 0 83 . .
By carefully reducing the semidefinite program by exploiting symmetries the bound for cr( K m,n ) was improved [18] to
$$\lim _ { n \to \infty } \text{cr} ( K _ { m,$$
for all m ≥ 9. The computations heavily relies on the symmetry reduction; without them one would need to solve a semidefinite program with 40320 × 40320 matrices, which is still far from possible with current SDP-solvers. In [15] the matrices in the semidefinite program were further block-diagonalized, which significantly improved computation time.
## 5.4 Kissing numbers
The kissing number is the maximum number τ n of unit spheres with no pairwise overlap that can simultaneously touch the unit sphere in n -dimensional Euclidean space.
For the first non-trivial dimension, n = 3, Sch¨tte and van der Waerden [40] u found that τ 3 = 12. It is also known that τ 4 = 24 [33], and that τ 8 = 240 and
τ 24 = 196560 [34, 26]. For other dimensions only bounds are known. A way to find upper bounds using linear programming was developed by Delsarte [3], and Delsarte, Goethals and Seidel [4]. These bounds were improved by Bachoc and Vallentin [2], and Mittelmann and Vallentin [31] for all n ≤ 24 using a semidefinite program that strengthens the linear program. The symmetry of the sphere is essential to obtain a finite SDP relaxation.
## 5.5 Quadratic assignment problems
A quadratic assignment problem is a problem on the form
$$\min _ { X \in \Pi _ { n } } \text{tr} ( A X$$
where Π n is the set of n × n permutation matrices, and A and B are symmetric n × n matrices. Solving the quadratic assignment problem is computationally difficult, it is known to be an NP-hard problem, and thus trying to find bounds using relaxations is the best one can do when n is not very small.
From the quadratic assignment problem we can form the following SDP relaxation [50, 13]:
$$\begin{array} {$$
where ⊗ is the Kronecker product, I the identity matrix and J the all one matrix. It is a relaxation since Y = vec( X )vec( X ) T and y = diag( Y ) is a feasible solution if X ∈ Π . n Depending on the structure of A and B the sizes of the semidefinite constraint can sometimes be reduced significantly by the methods developed in this section. In [13, 14] the methods are discussed in the context of the quadratic assignment problem, and a lot of specific instances of the problem are discussed and solved. Computation times are improved in many instances and many bounds, especially for larger problems, are improved.
The traveling salesman problem is the problem of finding a Hamiltonian circuit of minimum length in a graph, and it is easy to show that it is a special case of the quadratic assignment problem. Using the SDP relaxation and symmetry reduction of the quadratic assignment problem one gets the current strongest SDP relaxation of the traveling salesman problem [19].
## 5.6 Counting monochromatic arithmetic progressions in a 2-coloring
One of the most difficult problems in Ramsey theory is to determine the existence of monochromatic arithmetic progressions in a group when its elements
have been colored. The surveyed tools from real algebraic geometry makes it possible to count the progressions, a generalization of the existence problem, by stating it as a semidefinite program. The enumeration is done for the cyclic group in [41] and for any finite group in [42].
An arithmetic progression in a finite group G of length k ∈ Z + is a set of k distinct elements, { a, b · a, b · b · a, . . . , b k -1 · a } , for any a ∈ G and b ∈ G /integerdivide { 0 } . In other words { 1 2 3 , , } , { 1 3 2 , , } , { 2 1 3 , , } , { 2 3 1 , , } , { 3 1 2 , , } and { 3 2 1 , , } should be considered as the same arithmetic progression. Thus when we sum over all arithmetic progressions only one representative for every arithmetic progression is used.
Let χ : G → {-1 1 , } be a 2-coloring of the group G , and let x g = χ g ( ) for all g ∈ G . Furthermore, let x be the vector of all variables x g . For a, b, c ∈ G , let
$$p ( x _ { a }, x _ { b }, x _ { c } ) \ &$$
Hence p is one if a, b and c are of the same color and zero otherwise. Let R (3 , G, 2) denote the minimum number of monochromatic arithmetic progression of length 3 in a 2-coloring of the finite group G . It holds that:
$$R ( 3, G, 2 ) = \min _ { x \in \{ - 1, 1 \} ^ { n } } \sum _ { \{ a, b, c \} \text{ is an A.P. in } G } p ( x _ { a }, x _ { b }, x _ { c } ).$$
The problem can easily be relaxed to a semidefinite program with a lot of symmetries that can be exploited by the methods in this section. Using these techniques one can obtain the following results:
Theorem 5.7 ([41]) . Let n be a positive integer and let R (3 , Z n , 2) denote the minimal number of monochromatic 3 -term arithmetic progressions in any twocoloring of Z n . n / 2 8 -c 1 n + c 2 ≤ R (3 , Z n , 2) ≤ n / 2 8 -c 1 n + c 3 for all values of n , where the constants depends on the modular arithmetic and are tabulated in the following table.
| n mod 24 | c 1 | c 2 | c 3 |
|------------------------------------|-------|-------|--------|
| 1 , 5 , 7 , 11 , 13 , 17 , 19 , 23 | 1 / 2 | 3 / 8 | 3 / 8 |
| 8 , 16 | 1 | 0 | 0 |
| 2 , 10 | 1 | 3 / 2 | 3 / 2 |
| 4 , 20 | 1 | 0 | 2 |
| 14 , 22 | 1 | 3 / 2 | 3 / 2 |
| 3 , 9 , 15 , 21 | 7 / 6 | 3 / 8 | 27 / 8 |
| 0 | 5 / 3 | 0 | 0 |
| 12 | 5 / 3 | 0 | 18 |
| 6 , 18 | 5 / 3 | 1 / 2 | 27 / 2 |
Theorem 5.8 ([42]) . Let G be any finite group and let R (3 , G, 2) denote the minimal number of monochromatic 3 -term arithmetic progressions in any twocoloring of G . Let G k denote the set of elements of G of order k , N = | G | and N k = | G k | . Denote the Euler phi function φ k ( ) = |{ t ∈ { 1 , . . . , k } : t and k are coprime }| . Let K = { k ∈ { 5 , . . . , n } : φ k ( ) ≥ 3 k 4 } . For any G there are ∑ n k =4 N N · k 2 + N N · 3 24 arithmetic progressions of length 3. At least
$$R ( 3, G, 2 ) \geq \sum _ { k \in K } \frac { N \cdot N _ { k } } { 8 } ( 1 - 3 \frac { k - \phi ( k ) } { \phi ( k ) } )$$
of them are monochromatic in a 2-coloring of G .
## 5.7 Counting arithmetic progressions in a fixed density set
The famous theorem by Szemer´ edi states that there exist arithmetic progressions of any length in any set of the integers of positive density. After Szemer´ edi's result there has been a lot of research on how to improve the bounds on the number N δ, k ( ) for which any subset of density at least δ of the cyclic group Z N δ,k ( ) contains an arithmetic progressions of length k . Using real algebraic geometry one can study the more general question of counting the number of arithmetic progressions of length k in any subset of Z N of fixed density Nδ , denoted W (3 , Z N , δ ). It is easy to see that if D = δN we have
$$W ( 3, \mathbb { Z } _ { N }, D / N ) = \min \{ \sum _ { \{ i, j, k \} \text{ is an A.P. in } \mathbb { Z } _ { N } } x _ { i } x _ { j } x _ { k } \, \colon x _ { i } \in \{ 0, 1 \}, \sum _ { i = 0 } ^ { N - 1 } x _ { i } = D \},$$
which can be relaxed and formulated as an invariant semidefinite program. The best lower bounds for W (3 , Z N , D/N ) as well as a discussion on how to obtain a generalization of Szemer´di's theorem from this family of semidefinite programs e can be found in [43]. For example the following holds:
Theorem 5.9 ([43]) . Let p be a prime number. A lower bound for the minimum number of arithmetic progressions of length 3 among all subsets of Z p of cardinality D ,
$$W ( 3, \mathbb { Z } _ { p }, D / p ) = \min \{ \sum _ { \{ i, j, k \} \ A. P. \ i n \ \mathbb { Z } _ { p } } x _ { i } x _ { j } x _ { k } \, \colon x _ { i } \in \{ 0, 1 \}, \sum _ { i = 0 } ^ { p - 1 } x _ { i } = D \},$$
is
$$\lambda = \frac { D ^ { 3 } - ( \frac { p + 3 } { 2 } ) D ^ { 2 } + ( \frac { p + 3 } { 2 } - 1 ) D } { p - 1 }.$$
A certificate for the lower bound is given by:
/negationslash
$$\sum _ { \{ i, j, k \} \ A. P. \ i n \ Z _ { p } } X _ { i } X _ { j } X _ { k } - \lambda = \sum _ { i = 0 } ^ { p - 1 } \sigma _ { 1, i } X _ { i } + \sum _ { i = 0 } ^ { p - 1 } \sigma _ { 2, i } X _ { i } + \sigma _ { 3 } ( D - \sum _ { i = 0 } ^ { p - 1 } X _ { i } ^ { 3 } ) \\ + \sigma _ { 4 } ( \sum _ { i \neq j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ),$$
where
$$\sigma _ { 1, i } \ = \frac { 1 } { p - 1 } \sum _ { 0 < j < k < ( p - 1 ) / 2 } ( X _ { j + i } - X _ { j + k + i } - X _ { n - j - k + i } + X _ { n - j + i } ) ^ { 2 }$$
$$\sigma _ { 2, i } \ = \frac { 1 } { p - 1 } ( D X _ { i } - \sum _ { j = 0 } ^ { p - 1 } X _ { j } ) ^ { 2 }$$
$$\sigma _ { 3 } \ = \frac { ( D - 1 ) ^ { 2 } } { n - 1 }$$
$$\sigma _ { 4 } \ = \frac { 4 D - p + 3 } { 2 ( p - 1 ) }.$$
$$& \frac { 1 } { p - 1 } \sum _ { 0 < j < k < ( p - 1 ) / 2 } \,, \\ & \frac { 1 } { p - 1 } ( D X _ { i } - \sum _ { j = 0 } ^ { p - 1 } X _ { i } ) \\ & \frac { ( D - 1 ) ^ { 2 } } { p - 1 } \\ & \frac { 4 D - p + 3 } { 2 ( p - 1 ) }. \\ &. \quad.$$
## 6 Implementation
The procedure is similar for any polynomial optimization problem, so for simplicity let p be prime and let us consider the problem of counting the number of arithmetic progressions in a fixed density set:
$$\min \{ \sum _ { \{ i, j, k \} \ A. P. \text{ in } \mathbb { Z } _ { p } } x _ { i } x _ { j } x _ { k } \, \colon x _ { i } \in \{ 0, 1 \}, \sum _ { i = 0 } ^ { p - 1 } x _ { i } = D \},$$
which is invariant under affine transformations; if { i, j, k } is an arithmetic progression and ( a, b ) ∈ Z /multicloseright Z ∗ , then also { a + bi, a + bj, a + bk } is an arithmetic progression. We know from representation theory that we can split all possible 3-sets into orbits;
$$\bigcup _ { 0 \leq i < j < k < n } \{ i, j, k \} = S _ { 1 } \cup \cdots \cup S _ { D }.$$
In other words, if and only if both { i 1 , j 1 , k 1 } and { i 2 , j 2 , k 2 } are in S t it holds that { i 1 , j 1 , k 1 } = { a + bi 2 , a + bj 2 , a + bk 2 } for some ( a, b ) ∈ Z/multicloseright Z ∗ .
The optimization problem can be bounded from below using a degree 3 relaxation. Relaxing the integer constraints we have 0 ≤ x i ≤ 1, and the constraint ∑ p -1 i =0 x i = D implies that ∑ p -1 i =0 x 2 i = ∑ p -1 i =0 x 3 i = D , ∑ i<j x x i j =
D D ( -1) / 2, ∑ i,j x x 2 i j = D D ( -1). We get:
$$D ( D - 1 ) / 2, \sum _ { i, j } x _ { i } ^ { 2 } x _ { j } = D ( D - 1 ). \, \text{We get:} \\ \rho _ { 3 } = \sup _ { \lambda, \{ \sigma _ { i } ^ { + }, \sigma _ { i } ^ { - } \} } \lambda \\ \text{subject to} \quad f - \lambda = \sigma _ { 0 } + \sum _ { i = 1 } ^ { m } \sigma _ { i } ^ { + } X _ { i } + \sum _ { i = 1 } ^ { m } \sigma _ { i } ^ { - } ( 1 - X _ { i } ) + c _ { X _ { 0 } } ( \sum _ { i = 0 } ^ { p - 1 } X _ { i } ^ { 3 } - D ) \\ + c _ { X _ { 0 } ^ { 3 } X _ { 1 } } ( \sum _ { i < j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) / 2 ) \\ + c _ { X _ { 0 } ^ { 3 } X _ { 1 } } ( \sum _ { i, j } X _ { i } ^ { 2 } X _ { j } - D ( D - 1 ) ) \\, \quad \, \, \,$$
Let v 1 = v 1 ( X ) = [1 , X 0 , . . . , X n -1 ] T and let us recall the notation h X ( ) = ∑ α [ h X ( )] α X α . We can rewrite the relaxation as
$$\sum _ { \alpha } [ h ( X ) ] _ { \alpha } X ^ { \alpha }. \, \text{ We can rewrite the relaxation as} \\ \rho _ { 3 } = \sup _ { \lambda, \{ Q _ { i } \} } \lambda _ { \quad \text{$\quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\quad \text{subject to} \quad [ f ( X ) - v _ { 1 } ^ { T } Q _ { 0 } v _ { 1 } + \sum _ { i = 0 } ^ { n - 1 } v _ { 1$$
Because of the invariance under the affine group we can restrict the conditions to α ∈ { X ,X ,X X ,X ,X X ,X 0 2 0 0 1 3 0 2 0 1 s 1 , . . . , X s D } , where s i ∈ S i and X s t = X X X i j k if s t = { i, j, k } . It is also enough to require that Q , Q 0 + 1 , Q -1 /followsequal 0 as the other Q -matrices are just permutations of these. To solve this optimization problem for a fixed p we need to do the following:
- 1. Generate a basis of { 0 1 , } -matrices { E 1 1 , , . . . , E 1 ,d 1 , E 2 1 , , . . . , E 2 ,d 2 , E 3 1 , , . . . , E 3 ,d 3 } , where { E 1 1 , , . . . , E 1 ,d 1 } forms a { 0 1 , } -basis for Q 0 , { E 2 1 , , . . . , E 2 ,d 2 } forms a { 0 1 , } -basis for Q + 1 and { E 3 1 , , . . . , E 3 ,d 3 } forms a { 0 1 , } -basis for Q -1 . Let B i,j be the matrix E i,j normalized.
- 1b. (optional) Generate a basis { L 1 1 , , . . . , L 1 ,d 1 , L 2 1 , , . . . , L 2 ,d 2 , E 3 1 , , . . . , L 3 ,d 3 }
- 2. Find variables C 1 ,j,α , measuring how much matrix E 1 ,j contributes to the coefficient-part of the monomial α when multiplying out v T 1 E 1 ,j v 1 . Similarly for C 2 ,j,α and C 3 ,j,α , but have to calculate the total contribution of
$$\begin{smallmatrix} \text{the sum} \sum _ { k = 0 } ^ { n - 1 } v _ { 1, k } ^ { T } E _ { 2, j } v _ { 1, k } X _ { k } \text{ and } \sum _ { k = 0 } ^ { n - 1 } v _ { 1, k } ^ { T } E _ { 3, j } v _ { 1, k } ( 1 - X _ { k } ) \text{ respectively} \\ \text{where } v _ { 1, k } = [ 1, X _ { k }, X _ { k + 1 }, \dots, X _ { n - 1 }, X _ { 1 }, \dots, X _ { k - 1 } ]. \end{smallmatrix}$$
- 3. Solve the following optimization problem using a solver
```
variables g(1,1),...,g(1,d_1) g(2,1),...,g(2,d_2),
g(3,1),...,g(3,d_3) lambda
maximize(lambda)
subject to
```
sum\_{i,j} C(i,j,1)*g(i,j)+c(X\_0)D+c(X\_0^2)D+...
...+c(X\_0^3)D+...+c(X\_0X\_1)D(D-1)/2+...
...+c(X\_0^2X\_1)D(D-1) == lambda;
sum\_{i,j} C(i,j,X\_0)*g(i,j) + c(X\_0) == 0;
sum\_{i,j} C(i,j,X\_0^2)*g(i,j)+ c(X\_0^2) == 0;
sum\_{i,j} C(i,j,X\_0X\_1)*g(i,j)+ c(X\_0X\_1) == 0;
sum\_{i,j} C(i,j,X\_0^3)*g(i,j)+ c(X\_0^3) == 0;
sum\_{i,j} C(i,j,X\_0^2X\_1)*g(i,j)+ c(X\_0^2X\_1) == 0;
sum\_{i,j} C(i,j,X^{s\_1})*g(i,j) == 0;
.
.
.
$$\sum _ { \{ i, j \} \, C ( i, j, X ^ { \prime } \{ s _ { \, D } \} ) * g ( i, j ) \, = \, 0 \, ;$$
g(1,1)*B(1,1)+...+g(1,d\_1)*B(1,d\_1) is PSD
g(2,1)*B(2,1)+...+g(2,d\_2)*B(2,d\_2) is PSD
g(3,1)*B(3,1)+...+g(3,d\_3)*B(3,d\_3) is PSD
Obtain the optimal value λ and numerical certificates Q i = g i, 1 B i, 1 + ... + g i,d i B i,d i as well as constants c α .
Each of the three steps takes a considerable amount of time to implement through hundreds of lines of codes. The same procedure is possible even for higher degree relaxation (the author has implemented a degree 5 relaxation for counting arithmetic progressions using the procedure above). Since the number of orbits 5-sets gets partitioned into is much larger than the number of orbits 3-sets gets partitioned into it gets extremely technical and difficult to write the code in part 1, 2 and 3.
The degree 3 relaxation above was implemented in Matlab and the software CVX was used to solve the semidefinite program using the SeDuMi solver. The
semidefinite program could easily be solved for all primes p up to 300 for all D ∈ { 0 1 , / 20 , . . . , p } on a macbook air, and then numerical difficulties started in the solver. The code was parallelized, so that it was solved for different values of D at the same time, and uploaded to a super computer to get less accurate solutions for p up to 613. For a specific prime we were still able to calculate all values of D in less than a day by using 100 computer nodes, but at this point the solutions were so inaccurate that solving for higher p seemed pointless.
The degree 5 relaxation was similarly implemented, and could be used to find improved lower bounds for all primes p ≤ 19. To solve for all D at p = 19 took approximately 3 days using 100 computer nodes, and solving the problem for all D at p = 23 seemed impossible with the computer at hand. For further discussion on this problem we refer to [43].
## Acknowledgements
I would like to thank Alexander Engstr¨m for our discussions on how symo metries can be used in combinatorial problems. I also want to thank Markus Schweighofer and Cynthia Vinzant for their corrections and remarks.
## References
- [1] Christine Bachoc, Dion C. Gijswijt, Alexander Schrijver and Frank Vallentin. Invariant semidefinite programs. Chapter of 'Handbook on Semidefinite, Conic and Polynomial Optimization'. Springer, Berlin, 2012. 219269.
- [2] Christine Bachoc and Frank Vallentin. New upper bounds for kissing numbers from semidefinite programming. J. Amer. Math. Soc. 21 (2008), no. 3, 909-924.
- [3] Philippe Delsarte. An algebraic approach to the association schemes of coding theory. Doctoral dissertation, Universite Catholique de Louvain, 1973. vi+97 pp.
- [4] Philippe Delsarte, Jean-Marie Goethals and Johan Jacob Seidel. Spherical codes and designs. Geom. Dedicata 6 (1977), no. 3, 363-388.
- [5] Igor Dukanovic and Franz Rendl. Semidefinite programming relaxations for graph coloring and maximal clique problems. Math. Program. 109 (2007), no. 2-3, 345-365.
- [6] Dion Gijswijt, Alexander Schrijver and Hajime Tanaka. New upper bounds for nonbinary codes based on the Terwilliger algebra and semidefinite programming. J. Combin. Theory Ser. A 113 (2006), no. 8, 1719-1731.
| [7] | Didier Henrion and Jean Bernard Lasserre. Detecting global optimality and extracting solutions in GloptiPoly. Chapter of 'Positive polynomials in con- trol'. Springer, Berlin, 2005. 293-310. |
|-------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [8] | Thomas Jacobi and Alexander Prestel. Distinguished representations of strictly positive polynomials. J. Reine Angew. Math. 532 (2001) 223-235. |
| [9] | Yoshihiro Kanno, Makoto Ohsaki, Kazuo Murota and Naoki Katoh. Group symmetry in interior-point methods for semidefinite program. Optim. Eng. 2 (2001), no. 3, 293-320. |
| [10] | Narendra Karmarkar. Anew polynomial-time algorithm for linear program- ming. Combinatorica 4 (1984), no. 4, 373-395. |
| [11] | Leonid Genrikhovich Khachiyan. A polynomial algorithm in linear pro- gramming. Comput. Math. Math. Phys. 244 (1979), 1093-1096. |
| [12] | Etienne de Klerk. Exploiting special structure in semidefinite programming: A survey of theory and applications. European J. Oper. Res. 201 (2010), no. 1, 1-10. |
| [13] | Etienne de Klerk and Renata Sotirov. Exploiting group symmetry in semidefinite programming relaxations of the quadratic assignment prob- lem. Math. Program. 122 (2010), no. 2, 225-246. |
| [14] | Etienne de Klerk and Renata Sotirov. Improved semidefinite programming bounds for quadratic assignment problems with suitable symmetry. Math. Program. 133 (2012), no. 1-2, 75-91. |
| [15] | Etienne de Klerk, Cristian Dobre and Dmitrii V. Pasechnik. Numerical block diagonalization of matrix ∗ -algebras with application to semidefinite programming. Math. Program. 129 (2011), no. 1, 91-111. |
| [16] | Etienne de Klerk, John Maharry, Dmitrii V. Pasechnik, R. Bruce Richter and Gelasio Salazar. Improved bounds for the crossing numbers of K m,n and K n . SIAM J. Discrete Math. 20 (2006), no. 1, 189-202. |
| [17] | Etienne de Klerk, Mike W. Newman, Dmitrii V. Pasechnik and Renata Sotirov. On the Lov´sz a ϑ -number of almost regular graphs with application to Erd˝os-R´nyi o e graphs. European J. Combin. 30 , (2009), no. 4, 879-888. |
| [18] | Etienne de Klerk, Dmitrii V. Pasechnik and Alexander Schrijver. Reduc- tion of symmetric semidefinite programs using the regular ∗ -representation. Math. Program. 109 (2007), no. 2-3, Ser. B, 613-624. |
| [19] | Etienne de Klerk, Dmitrii V. Pasechnik and Renata Sotirov. On semidefi- nite programming relaxations of the traveling salesman problem. SIAM J. Optim. 19 (2008), no. 4, 1559-1573. |
| [20] | Donald E. Knuth. The sandwich theorem. Electron. J. Combin. , 1 (1994), |
- [20] Donald E. Knuth. The sandwich theorem. Electron. J. Combin. , 1 (1994), A1.
| [21] | Jean-Louis Krivine. Anneaux pr´ordonn´s. e e J. Analyse Math. 12 (1964), no. 1, 307-326. |
|--------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [22] | Jean Bernard Lasserre. Global optimization with polynomials and the prob- lem of moments. SIAM J. Optim. 11 (2001), no. 3, 796-817. |
| [23] | Jean Bernard Lasserre. Moments, positive polynomials and their applica- tions. Imperial College Press Optimization Series, 1. Imperial College Press, London, 2010. 361 pp. |
| [24] | Monique Laurent. Strengthened semidefinite programming bounds for codes. Math. Program. 109 (2007), no. 2-3, 239-261. |
| [25] | Monique Laurent. Sums of squares, moment matrices and optimization over polynomials. Chapter of 'Emerging applications of algebraic geometry'. Springer, New York, 2009. 157-270. |
| [26] | Vladimir I. Levenshtein On bounds for packing in n-dimensional Euclidean space Dokl. Akad. Nauk 245 (1979), no. 6, 1299-1303. |
| [27] | L´szl´ a o Lov´sz. a On the Shannon capacity of a graph. IEEE Trans. Inform. Theory 25 (1979), no. 1, 1-7. |
| [28] | Takanori Maehara and Kazuo Murota. A numerical algorithm for block- diagonal decomposition of matrix ∗ -algebras with general irreducible com- ponents. Jpn. J. Ind. Appl. Math. 27 (2010), no. 2, 263-293. |
| [29] | Murray Marshall. Positive polynomials and sums of squares. Vol. 146. of Mathematical Surveys and Monographs. AMS, Providence, RI, 2008. 187 pp. |
| [30] | Robert J. McEliece, Eugene R. Rodemich and Howard C. Rumsey Jr. The Lov´sz a bound and some generalizations. J. Comb. Inf. Syst. Sci. 3 (1978), no. 3, 134-152. |
| [31] | Hans D. Mittelmann and Frank Vallentin. High-accuracy semidefinite pro- gramming bounds for kissing numbers. Exp. Math. 19 (2010), no. 2, 175- 179. |
| [32] | Kazuo Murota, Yoshihiro Kanno, Masakazu Kojima and Sadayoshi Ko- jima. A numerical algorithm for block-diagonal decomposition of matrix ∗ - algebras with application to semidefinite programming. Jpn. J. Ind. Appl. Math. 27 (2010), no. 1, 125-160. |
| [33] | Oleg R. Musin. The kissing number in four dimensions. Ann. of Math.(2) 168 (2008), no. 1, 1-32. |
| [34] | Andrew M. Odlyzko and Neil J. A. Sloane. New bounds on the number of unit spheres that can touch a unit sphere in n dimensions. J. Combin. |
| [35] | Mihai Putinar. Positive polynomials on compact semi-algebraic sets. Indi- ana Univ. Math. J. 42 (1993), no. 3, 969-984. |
|--------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [36] | Cordian Riener, Thorsten Theobald, Lina Jansson Andr´n e and Jean B. Lasserre. Exploiting symmetries in SDP-relaxations for polynomial opti- mization. Math. Oper. Res. 38 , (2013), no. 1, 122-141. |
| [37] | Konrad Schm¨dgen. u The K -moment problem for compact semi-algebraic sets. Math. Ann. 289 (1991), no. 1, 203-206. |
| [38] | Alexander Schrijver. 'Association schemes and the Shannon capacity: Eberlein-polynomials and the Erdos-Ko-Rado Theorem.' In: Algebraic Methods in Graph Theory (L. Lov´sz a and V, T, Sos, eds.) (1981): 671- 688. |
| [39] | Alexander Schrijver. New code upper bounds from the Terwilliger algebra and semidefinite programming. IEEE Trans. Inform. Theory 51 , (2005), no. 8, 2859-2866. |
| [40] | Kurt Sch¨tte u and Bartel Leendert van der Waerden. Das problem der dreizehn Kugeln. Math. Ann. 125 (1952), no. 1, 325-334. |
| [41] | Erik Sj¨land. o Enumeration of monochromatic three term arithmetic progressions in two-colorings of cyclic groups. Preprint available at http://arxiv.org/abs/1408.1058. |
| [42] | Erik Sj¨land. o Enumeration of monochromatic three term arithmetic pro- gressions in two-colorings of any finite group. Preprint available at http://arxiv.org/abs/1408.1088. |
| [43] | Erik Sj¨land. o Enumeration of three term arithmetic progressions in fixed density sets. Preprint available at http://arxiv.org/abs/1408.1063. |
| [44] | Gilbert Stengle. A Nullstellensatz and a Positivstellensatz in semialgebraic geometry. Math. Ann. 207 (1974), no. 2, 87-97. |
| [45] | Masamichi Takesaki. Theory of operator algebras I . Encyclopaedia of Math- ematical Sciences, 124. Operator Algebras and Non-commutative Geome- try, 5. Springer, Berlin, 2002. 525 pp. |
| [46] | Michael J. Todd. Semidefinite optimization. Acta Numer. 10 (2001), 515- 560. |
| [47] | Frank Vallentin. Symmetry in semidefinite programs. Linear Algebra Appl. 43 (2009), no. 1, 360-369. |
| [48] | Lieven Vandenberghe and Stephen Boyd. Semidefinite programming. SIAM rev. 38 (1996), no. 1, 49-95. |
| [49] | Kazimierz Zarankiewicz. On a problem of P. Tur´n a concerning graphs. Fund. Math. 41 , (1954), 137-145. |
[50] Quing Zhao, Stefan E. Karisch, Franz Rendl and Henry Wolkowicz. Semidefinite programming relaxations for the quadratic assignment problem. J. Comb. Optim. 2 (1998), no. 1, 71-109.
| null |
[
"Erik Sjöland"
] |
2014-08-05T18:44:16+00:00
|
2014-11-10T19:17:39+00:00
|
[
"math.CO"
] |
Using real algebraic geometry to solve combinatorial problems with symmetries
|
Many combinatorial problems can be formulated as a polynomial optimization
problem that can be solved by state-of-the-art methods in real algebraic
geometry. In this paper we explain many important methods from real algebraic
geometry, we review several applications and discuss implementation and
computational aspects.
|
1408.1066v1
|
## A Generalized Landauer Approach for Electron Transport in a Single Molecule
Augusto C. L. Moreira and Celso P. de Melo* Departamento de Física, Universidade Federal de Pernambuco 50.670-901 Recife, PE, Brazil
## ABSTRACT
We present a quaternion-inspired formalism specifically developed to evaluate the electric current that traverses a single molecule subjected to an externally applied voltage. The molecule of interest is covalently connected to two small metallic clusters, forming an 'extended molecule' complex. The quaternion approach allows for an integrated treatment of the charge transport in single molecules where both ballistic and co-tunneling (coherent) mechanisms are taken on equal footing, although only in the latter case the presence of eventual transient charged states of the system needs to be considered. We use a Dyson series to obtain a generalized Fermi golden rule, from which we derive an expression for the net current the two electrodes: in doing this, we take into account all possible transitions between electronic states localized at the electrodes and levels in the 'extended molecule' complex. In fact, one can apply the method to the entire range of coupling regimes, not only in the weak or strong cases, but also in intermediate situations, where ballistic and co-tunneling processes compete with each other. We also discuss initial results of the application of this formalism to the description of the electronic transport in two small organic molecules representative of two different limit situations. In the first case, a conjugated molecule (where spatially delocalized molecular orbitals favor ballistic contributions) is considered, and in the second the current traverses a saturated hydrocarbon (whose structure should contain more localized molecular orbitals). In both cases, we fully describe the field-induced self-adjustment of the electronic levels of the extended molecule complex at an ab initio quantum chemical level, using density functional theory.
## I. INTRODUCTION
Since Aviram and Ratner first proposed the concept of a molecular rectifier, a 1 crescent amount of theoretical effort has been dedicated to the understanding of electronic transport processes occurring at the molecular scale. However, after so many 2 years of hard and varied effort, the calculation of the current that flows through a single molecule connected to metal electrodes under an externally applied voltage remains one of the most difficult theoretical problems to be properly considered.
Molecular electronics is expected to present many advantages when compared to present day silicon based technology, such as the possibility of preparing devices of smaller size, higher speed, lower cost and novel functionalities. In principle, a single 2 molecule can play different active roles in a device, but a few general conclusions can be safely inferred, such as: i) long molecules have lower conductance than short ones, ii) conjugated systems have higher conductance that non-conjugated molecules of similar size, and iii) molecules with structural asymmetry exhibit asymmetric currentvoltage characteristics. The fact that one can relate the intrinsic molecular structure of the system to the measured current and conductance suggests that it is feasible the theoretical modeling of these two latter quantities.
## II. CONDUCTANCE THROUGH SINGLE-MOLECULES
## A. Transmission function
A first theoretical insight into the question of the mechanisms involved in the molecular transport appeared with Landauer's introduction of the concept of
'conductance from transmission'. The electric current 3 I of a very small conductor is associated to the coherent probability of transmission T(E)
$$I _ { c o h } = \frac { e } { \pi \hbar } \prod _ { - \infty } ^ { \infty } d E \, T ( E ) \left ( f _ { L } ( E ) - f _ { R } ( E ) \right ) \quad,$$
through the internal transversal modes, which are supposed to be limited in number when compared to the infinite range of those possible in the external electrodes and in the Fermi distribution functions ( , ( ) L R E f ) at the two (left/right) semi-infinite electrodes. 4
Different suggestions on how to calculate T(E) have been introduced, the most successful of them based on scattering theory. For instance, by use of the LippmannSchwinger (LS) equation 4 ˆ ˆ G V ψ φ φ + = + , one can construct the exact scattering states ψ with basis in the knowledge of the original unperturbed state φ and the potential V that couples the device region to the contacts, through the retarded Green·s function ˆ G + . In this case, the transmission T(E) is related to the scattering matrix 5 S that appears when we take the square of the modulus of ψ . Another possibility is to adopt the so-called Non Equilibrium Green·s Function (NEGF) formalism. 4, 6
In the former type of approach, the molecular coupling to the two terminal contacts (which are labeled as L and R , respectively) is described by an effective device Hamiltonian ˆ ˆ ˆ ˆ eff L R H H → + Σ + Σ that incorporates each electrode by taking into account its corresponding self-energy matrix ˆ i Σ (i = L , R ). Once this effective Hamiltonian is known, one can calculate the overall current through the system. In the non-interacting limit case of the NEGF, the transmission can be expressed as 4
$$T ( E ) = & T r \left \{ \hat { \Gamma } _ { L ^ { ( E ) } } \hat { G } ^ { + } ( E ) \hat { \Gamma } _ { R ^ { ( E ) } } \hat { G } ^ { - } ( E ) \right \}.$$
In Eq. (2), ˆ G + and ˆ G -are the retarded and the advanced Green's function of the coupled system, respectively, while ˆ i Γ , the broadening contribution of the molecular levels, is associated to the anti-hermitian part of ˆ i Σ .
Please note that the presence of an explicitly 'transmission function' through the molecular system remains valid as long as the electron transport is of a 'ballistic' nature, with no evidence of scattering at the molecular internal structure. In fact, this latter phenomenon is typically associated to the presence of spatially localized molecular states. On the contrary, ballistic transport of electrons must necessarily 3 involve a 'hot electron' scattered by an unoccupied molecular orbital that should be spatially delocalized throughout the entire system. 7
## B. Strong, weak and intermediate coupling regimes
The ballistic transport occurs whenever the interaction between the molecule and the metallic leads is strong. In the strong coupling limit, sometimes referred to as the self-consistent field regime 6 (SCF), the molecular levels are broadened by their interaction with the contacts. The corresponding energy (of the order of B k T ) is comparable to the single-electron charging energy U0 , i.e., the energy separation between states of different spin orientation. In fact, the value of the charging energy itself is related to the extent of the spatial localization of the wave function: levels with well-delocalized wave functions have a very small U0 . 6, 8 Under these conditions, even small external bias may cause a detectable current to traverse the device.
It is worthwhile to note that one can also use the above-mentioned methods to describe non-coherent phenomena (such as electron-phonon coupling) by adding an interaction potential and then evaluating the corresponding coupling constants. As
shown by Datta, while in the case of NEGF method this procedure would be equivalent 3 to attach an extra 'contact' to the device with a self-energy ˆ S Σ and an additional coupling ˆ S Γ representing the involved states, in the LS equation one makes the replacement ˆ ˆ ˆ V V U → + . In both cases, the complete wave function must now take into account both individual sub-spaces associated to phonon and electrons.
In the usual treatment of molecular electronic devices, 9-11 even after inclusion of non-coherent contributions the transport through the system remains confined to a single electronic potential surface; as a consequence, the active component (the single molecule, in our case) must remain unaltered in its original electronic state (be it its neutral form or a specific anion or cation). However, this picture is not valid if the coupling between the molecule and the metallic leads is sufficiently small; in this limit, the charge transport is sequential and we must allow for the presence of transient charged states of the molecular species considered. One now faces the regime of weak 6 coupling, where charging effects such as Coulomb blockade, 12 for example, may occur. In this case, either the multi electron master equation (MEME) approach must be employed or one could resort to NEGF methods out of the non-interactive limit, adding Coulomb electronic repulsion energy term in the Hamiltonian and employing the full expression for the electric current (see, for instance, Ref. [4], p. 242). In this weakly coupled regime (sometimes referred to as the Coulomb blockade (CB) limit), the charging energy U0 is much bigger than both B k T and the coupling between the contacts and the molecular levels. Now, due to the high strength of 0 U , very little current would flow when a small external bias is applied to the device. Once again, the charging energy is related to the extent of the wave function spatial localization: in the Coulomb blockade regime, one should expect to encounter more wave functions of a highly
localized nature. 6, 8, 13 Actually, in realistic nanodevices as is the case of a single molecule attached to metallic electrodes, delocalized and well-localized wave functions will coexist 13 and, hence, both Coulomb blockade and ballistic processes may contribute to the observed current. In this case, however, one should note that, although NEGF methods can be used to describe the charge flow in all (i.e., weak, strong and intermediate) coupling cases, they do not take into account a self-consistent charge distribution during the tunneling; consequently, the Green·s function of the complete problem is generally unknown. 14 Hence, when interactions beyond mean field approximation are included, it is not a simple task to compute the current in actual systems within the NEGF formalism. Therefore, one usually employs the orthodox MEME picture only when broadening effects are absent.
It is especially important to note that the issue of deciding a priori on whether a given molecular system will exhibit a strong rather than a weak coupling to a pair of existing metallic electrodes is far from settled. A criterion usually adopted is to assume 6 that a strong coupling will be developed through the molecular system if the corresponding molecular orbitals are sufficiently delocalized to present no vanishing electronic densities at the opposite sides of the extended molecule. 15 Also note that this assumption precludes the existence of strong coupling - and therefore of the occurrence of ballistic transport -in saturated molecules, where spatial localization of the electronic density at each extremity usually occurs. 1, 16
However, a logical quagmire rapidly develops if one stretches the argument further on to consider larger and larger extended molecular entities. While, on one hand, a progressive increase in the number of metallic atoms in the terminal clusters of the EM should improve the quality of the calculated results, 16 at the other, the larger the size of the EM, less likely it becomes to find states that are truly delocalized throughout
the extended molecule region. To compound to these difficulties, if the theoretical treatment of the problem allows for a new ab initio calculation of the molecular system every time that the external electric field is adjusted to a new value, the spatial localization of the frontier molecular orbitals can change accordingly. Consequently, the very nature of the transport regime will vary along the actual calculation of the current profile.
So, a more complete formalism for describing quantum transport through a single molecule connected to two terminal electrodes has to account for the occurrence of not only the two limiting regimes previously described in a separated manner; rather, it must consider the possibility that both regimes exist at once and treat them on equal a priori footing. One might also expect that the case of an intermediate coupling regime 6 would automatically be included in these more elaborated theoretical approaches. Note that, according to the intensity of the externally applied electric field, in the intermediate regime either ballistic processes or alternative charge transfer pathways involving transient charged molecular states (some of them corresponding to non-coherent tunneling events at the two electrode-molecule junctions 17 ) may dominate the overall charge transport.
## III. THE QUATERNIONIC FORMALISM
## A. Molecular charge states as independent quaternionic sub-spaces
Suppose that an electric field of varying intensity exists between two semi-infinite leads connected to opposite ends of the extended molecule. Then, once the three possible pathways are independently considered for each fixed value of the applied potential, both ballistic and co-tunneling terms will arise naturally in the resulting expression for the overall electric current. To be able to do this in a simple and elegant
manner, we will recur to the use of a quaternion -inspired formalism. Quaternions is a theoretical concept originally introduced by W. R. Hamilton in the XIX Century 18 that, in modern times, has been adapted by Adler 19 to quantum mechanical problems.
A quaternion is defined as a hyper complex number q a i b jc kd = + + + , where a , b , c , and d are real numbers and i , j and k are imaginary units that observe the cyclic properties 2 2 2 1 j k i j k i = = = = -and , , i j ji k jk kj ki i k j = -= = -= -= . Note that by taking advantage of three non-commuting distinct imaginary units, one can write three time-dependent Schrödinger equations, each one corresponding to a different unperturbed hermitian Hamiltonian, as
$$i \hbar { \partial } _ { t } \left | \phi _ { n _ { 1 } } ( t ) \right \rangle & = \hat { H } _ { 0 1 } \left | \phi _ { n _ { 1 } } ( t ) \right \rangle \quad \rightarrow \quad \hat { H } _ { 0 1 } \left | \phi _ { n _ { 1 } } ( t ) \right \rangle = E _ { n _ { 1 } } \left | \phi _ { n _ { 1 } } ( t ) \right \rangle \\ j \hbar { \partial } _ { t } \left | \phi _ { n _ { 2 } } ( t ) \right \rangle & = \hat { H } _ { 0 2 } \left | \phi _ { n _ { 2 } } ( t ) \right \rangle \quad \rightarrow \quad \hat { H } _ { 0 2 } \left | \phi _ { n _ { 2 } } ( t ) \right \rangle = E _ { n _ { 2 } } \left | \phi _ { n _ { 2 } } ( t ) \right \rangle \\ k \hbar { \partial } _ { t } \left | \phi _ { n _ { 3 } } ( t ) \right \rangle & = \hat { H } _ { 0 3 } \left | \phi _ { n _ { 3 } } ( t ) \right \rangle \quad \rightarrow \quad \hat { H } _ { 0 3 } \left | \phi _ { n _ { 3 } } ( t ) \right \rangle = E _ { n _ { 3 } } \left | \phi _ { n _ { 3 } } ( t ) \right \rangle \quad \quad.$$
It is easy to show that these three Hamiltonian operators do not commute and therefore they do not possess simultaneous time-dependent eigenstates. In this manner, each one of the three Hamiltonians must exhibit its own set of eigenstates. In the end, each eigenstate can be made orthogonal to all the others. We can take advantage of this fact, and write the time evolution of these Hamiltonians in terms of a single matrix expression, in the form
$$\text{passion}, \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \ text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \\ \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{In} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \,
\text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{ in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{ini} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \,\text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in } \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \tau \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \tag \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \tt{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in}
\text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{In } \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text {in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{i} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text{in} \, \text$$
where the x γ (x = 1, 2 and 3) normalization factors must necessarily obey the condition 2 2 2 1 2 3 1. We define the (unperturbed) quaternionic wave function and matrices, γ γ γ + + = respectively, by
$$\left | \mathbb { X } _ { 0, n } ( t ) \right \rangle \equiv & \begin{pmatrix} \gamma _ { _ { 1 } } \left | \phi _ { _ { n _ { 1 } } } ( t ) \right \rangle \\ \gamma _ { _ { 2 } } \left | \phi _ { _ { n _ { 2 } } } ( t ) \right \rangle \\ \gamma _ { _ { 3 } } \left | \phi _ { _ { n _ { 3 } } } ( t ) \right \rangle \end{pmatrix} \quad, \\ \nearrow \cdot \quad \omega \quad \wedge$$
$$\text{and } \mathbb { E } _ { 0, n } \equiv \begin{pmatrix} E _ { 0, n _ { 1 } } & 0 & 0 \\ 0 & E _ { 0, n _ { 2 } } & 0 \\ 0 & 0 & E _ { 0, n _ { 3 } } \end{pmatrix} \quad, \\ \text{a comment form was have}$$
$$\left | \mathbb { X } _ { 0, n } ( t ) \right \rangle \equiv \left ( \begin{array} { c c c } \gamma _ { _ { 1 } } \left | \phi _ { _ { n } } ( t ) \right \rangle \\ \gamma _ { 2 } \left | \phi _ { _ { n } } ( t ) \right \rangle \\ \gamma _ { 3 } \left | \phi _ { _ { n }, \, } ( t ) \right \rangle \end{array} \right ) \quad, \quad & ( 5. 1 ) \\ q \equiv \left ( \begin{array} { c c c } i & 0 & 0 \\ 0 & j & 0 \\ 0 & 0 & k \end{array} \right ) \\ \mathbb { H } _ { 0 } \equiv \left ( \begin{array} { c c c } \hat { H } _ { 0 1 } & 0 & 0 \\ 0 & \hat { H } _ { 0 2 } & 0 \\ 0 & 0 & \hat { H } _ { 0 3 } \end{array} \right ) \quad, \quad & ( 5. 3 ) \\ \text{and} \mathbb { E } _ { 0, n } \equiv \left ( \begin{array} { c c c } E _ { 0, n _ { 1 } } & 0 & 0 \\ 0 & E _ { 0, n _ { 2 } } & 0 \\ 0 & 0 & E _ { 0, n _ { 3 } } \end{array} \right ) \quad, \quad & ( 5. 4 ) \\ \text{must not form, you have}$$
so that, in a more compact form, we have
$$\mathfrak { q h } \partial _ { t } \left | \mathbb { X } _ { 0, n } ( t ) \right \rangle = \mathbb { H } _ { 0 } \left | \mathbb { X } _ { 0, n } ( t ) \right \rangle \ \rightarrow \ \mathbb { H } _ { 0 } \left | \mathbb { X } _ { 0, n } ( t ) \right \rangle = \mathbb { E } _ { 0, n } \left | \mathbb { X } _ { 0, n } ( t ) \right \rangle \ \.$$
We then have three orthogonal Hilbert spaces, each comprising one of the mutually non-commutative Hamiltonians. We will now discuss what happens when a time independent perturbation is present. Consider that a generic time independent perturbation like
$$\mathbb { V } = \begin{pmatrix} \hat { V } _ { 1 1 } & \hat { V } _ { 1 2 } & \hat { V } _ { 1 3 } \\ \hat { V } _ { 2 1 } & \hat { V } _ { 2 2 } & \hat { V } _ { 2 3 } \\ \hat { V } _ { 3 1 } & \hat { V } _ { 3 2 } & \hat { V } _ { 3 3 } \end{pmatrix}$$
is introduced in the system, so that one can write 0 = + H H V . As usual, we assume that each of its elements is much smaller than the characteristic energies of the unperturbed
system. We can write a LS equation for this perturbed quaternionic system and so define a corresponding S matrix (and, consequently, a T matrix). To see this, assume that the eigenvalues of 0 H are known, such that 0 0, 0, 0, n n n = H X E X , and consider the Hamiltonian 0 = + H H VGLYPH<2> . In the usual treatment, one associates 0 H to a free particle, while V is a scattering potential. The corresponding LS equation can be obtained if one admits that each unperturbed solution 0, n X in the continuous spectrum is mapped into a solution n P of the complete (i.e., perturbed) problem, which also belongs to the continuous spectrum and has the same energy 0, n n = E E . Therefore, implicitly one assumes that the total energy remains constant during the entire process. So, in this manner,
$$\mathbb { H } | \mathbb { P } _ { n } \rangle = \left ( \mathbb { H } _ { 0 } + \mathbb { V } \right ) | \mathbb { P } _ { n } \rangle = \mathbb { E } _ { n } \left | \mathbb { P } _ { n } \right \rangle \quad.$$
We can express the relationship between 0, n X and n P in a more explicit manner by rewriting the previous equations in the form
$$\left | \mathbb { P } _ { n } ^ { \pm } \right \rangle = \left | \mathbb { X } _ { 0, n } \right \rangle + \mathbb { G } _ { 0 } ^ { \pm } \mathbb { V } \left | \mathbb { P } _ { n } ^ { \pm } \right \rangle = \left | \mathbb { X } _ { 0, n } \right \rangle + \mathbb { G } _ { n } ^ { \pm } \mathbb { V } \left | \mathbb { X } _ { 0, n } \right \rangle = \widehat { \Omega } ^ { \pm } \left | \mathbb { X } _ { 0, n } \right \rangle \quad,$$
which is the quaternionic version of the Lippmann-Schwinger equation. Here we have introduced the Moller hyper-operator 20 0 ± ± ± ± Ω = + = + /slurabove I G VGLYPH<2> I G T GLYPH<2> , with the transition operator ± T being defined as ± ± = + T V VG V , where 0 0 ± ± ± ± = + G G G VG is the total
Green·s function. Note that 0 ± G can be written as
$$\begin{pmatrix} \hat { G } _ { 0 1 } ^ { \pm } & 0 & 0 \\ 0 & \hat { G } _ { 0 2 } ^ { \pm } & 0 \\ 0 & 0 & \hat { G } _ { 0 3 } ^ { \pm } \end{pmatrix} \begin{pmatrix} E - \hat { H } _ { 0 1 } \pm i \epsilon & 0 & 0 \\ 0 & E - \hat { H } _ { 0 2 } \pm j \epsilon & 0 \\ 0 & 0 & E - \hat { H } _ { 0 3 } \pm k \epsilon \end{pmatrix} = \begin{pmatrix} I & 0 & 0 \\ 0 & I & 0 \\ 0 & 0 & I \end{pmatrix} \quad, \ \ ( 9. 1 )$$
or, in a more compact manner,
$$\mathbb { G } _ { 0 } ^ { \pm } ( \mathbb { E } ) \left ( \mathbb { E } - \mathbb { H } _ { 0 } \pm q \varepsilon \right ) = \mathbb { I } _ { 3 } \quad.$$
/slurabove
Thus, exactly as in usual Quantum Mechanics 20 , ± Ω transforms each unperturbed basis vector into the corresponding perturbed vector. With the above definition, we are now ready to introduce the scattering matrix S for quaternionic systems. A scattering event can be described as a transformation of an initial unperturbed immerging state 0, i X into a final (emerging) state 0, f X , by effect of a scattering operator S in the form 0, 0, f i = X S X . In this manner, the amplitude of a given state 0, n X will be given by 0, 0, 0, 0, f n n i ni = = X X X X S S . If one takes advantage of the LS equation, one can write 0, n n ± ± = Ω /slurabove P X , so that the S matrix can be written as † -+ = Ω Ω /slurabove /slurabove S , since † 0, 0, 0, 0, n i n i n i --+ + = Ω Ω = /slurabove /slurabove P P X X X X S . Thus, the transition probability between two (quaternionic) states will be given by the square of the modulus of the corresponding S matrix element, i.e., 2 ni ni = ℘ S .
This general approach allows the application of the quaternionic formalism to different physical situations, such as charge transport in nanoscopic devices and electron transfer in complex systems. For a system that upon a perturbation can evolve alternatively through up to three distinct and non-commutative Hamiltonians, the quaternion formalism permits the knowledge of the total Green·s function, from which one can derive any physical quantity of interest.
To see this, let's write the total Green·s function ± G in the matrix form
$$\mathbb { G } ^ { \pm } = \begin{pmatrix} \hat { G } ^ { \pm } _ { 1 1 } & \hat { G } ^ { \pm } _ { 1 2 } & \hat { G } ^ { \pm } _ { 1 3 } \\ \hat { G } ^ { \pm } _ { 2 1 } & \hat { G } ^ { \pm } _ { 2 2 } & \hat { G } ^ { \pm } _ { 2 3 } \\ \hat { G } ^ { \pm } _ { 3 1 } & \hat { G } ^ { \pm } _ { 3 2 } & \hat { G } ^ { \pm } _ { 3 3 } \end{pmatrix} & & ( 1 0. 1 ) \\ \text{dard definitions}$$
or, using the standard definitions,
$$\Phi _ { \mathbb { G } ^ { \pm } } & = \begin{pmatrix} \hat { G } _ { 1 1 } ^ { \pm } & \hat { G } _ { 1 2 } ^ { \pm } & \hat { G } _ { 1 3 } ^ { \pm } \\ \hat { G } _ { 2 1 } ^ { \pm } & \hat { G } _ { 2 2 } ^ { \pm } & \hat { G } _ { 2 3 } ^ { \pm } \\ \hat { G } _ { 3 1 } ^ { \pm } & \hat { G } _ { 3 2 } ^ { \pm } & \hat { G } _ { 3 3 } ^ { \pm } \end{pmatrix} \begin{pmatrix} E - \hat { H } _ { 0 1 } - \hat { V } _ { 1 } \pm i \varepsilon & - \hat { V } _ { 1 2 } & - \hat { V } _ { 3 } \\ - \hat { V } _ { 2 1 } & E - \hat { H } _ { 0 2 } - \hat { V } _ { 2 2 } \pm j \varepsilon & - \hat { V } _ { 2 3 } \\ - \hat { V } _ { 3 1 } & - \hat { V } _ { 3 2 } & E - \hat { H } _ { 0 3 } - \hat { V } _ { 3 3 } \pm k \varepsilon \end{pmatrix} \quad ( 1 0. 2 ) \\ & = \begin{pmatrix} I & 0 & 0 \\ 0 & I & 0 \\ 0 & 0 & I \end{pmatrix} &, \\ \end{pmatrix} \begin{matrix} \begin{matrix} 0 \\ 0 \\ 0 & I & 0 \\ 0 & 0 & I \end{matrix} \end{matrix} \end{pmatrix}$$
or, finally, in a more compact form,
$$\mathbb { G } ^ { \pm } ( \mathbb { E } - \mathbb { H } \pm q \varepsilon ) = \mathbb { G } ^ { \pm } ( \mathbb { E } ) \left ( E \mathbb { I } _ { 3 } - \mathbb { H } \pm q \varepsilon \right ) = \mathbb { I } _ { 3 } \quad.$$
Let·s define the Green·s function of the uncoupled perturbed subspaces ˆ x g ± as ( ) ( ) ˆ ˆ x x x E E H I q g ε ± = -± , with x = 1, 2 or 3 (and q1 , q2 , q3 → i, j, k ) . In a matrix form, this yields
$$\begin{pmatrix} \hat { g } _ { 1 } ^ { \pm } & 0 & 0 \\ 0 & \hat { g } _ { 2 } ^ { \pm } & 0 \\ 0 & 0 & \hat { g } _ { 3 } ^ { \pm } \end{pmatrix} \begin{pmatrix} E - \hat { H } _ { 0 1 } - \hat { V } _ { 1 1 } \pm i \epsilon & 0 & 0 \\ 0 & E - \hat { H } _ { 0 2 } - \hat { V } _ { 2 2 } \pm j \epsilon & 0 \\ 0 & 0 & E - \hat { H } _ { 0 3 } - \hat { V } _ { 3 3 } \pm k \epsilon \end{pmatrix} = \begin{pmatrix} ( 1 1 ) \\ & & \\ & & & = \begin{pmatrix} I & 0 & 0 \\ 0 & I & 0 \\ 0 & 0 & I \end{pmatrix} \end{pmatrix} \.$$
For convenience, from now on we will drop the symbols indicating the advanced and retarded Green·s function. Then, using the above definitions, we can construct the system of equations
$$+ \hat { G } _ { 1 1 } \left ( E - \hat { H } _ { 0 1 } - \hat { V } _ { 1 1 } \right ) - \hat { G } _ { 1 2 } \hat { V } _ { 2 1 } - \hat { G } _ { 1 3 } \hat { V } _ { 3 1 } = I$$
$$- \hat { G } _ { 1 1 } \hat { V } _ { 1 2 } + \hat { G } _ { 1 2 } \left ( E - \hat { H } _ { 0 2 } - \hat { V } _ { 2 2 } \right ) - \hat { G } _ { 1 3 } \hat { V } _ { 3 2 } = 0$$
$$- G _ { 1 1 } V _ { 1 3 } - G _ { 1 2 } V _ { 2 3 } + G _ { 1 3 } \left ( E - H _ { 0 3 } - V _ { 3 3 } \right ) = 0$$
that allows to determine all elements of ( ) . For instance, from (12.1), we have E G 11 1 12 21 1 13 31 1 ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ G g G V g G V g = + + , which can be substituted in (12.2), so that 12 1 12 2 13 31 1 12 2 32 2 21 1 12 2 1 ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ( ) ( ) G gV g G V gV g V g I V gV g - = + + - .
Now, one can obtain 13 G by direct substitution,
$$& \hat { G } _ { 1 3 } \left [ \left ( I - \hat { V } _ { 3 1 } \hat { g } _ { 1 } \hat { V } _ { 1 3 } g _ { 3 } \right ) + ( \hat { V } _ { 3 1 } \hat { g } _ { 1 } \hat { V } _ { 1 2 } \hat { g } _ { 2 } + \hat { V } _ { 3 2 } \hat { g } _ { 2 } ) ( I - \hat { V } _ { 2 1 } \hat { g } _ { 1 } \hat { V } _ { 1 2 } \hat { g } _ { 2 } ) ^ { - 1 } \right ] = \\ & \quad = \hat { g } _ { 1 } \hat { V } _ { 1 3 } \hat { g } _ { 3 } + \hat { g } _ { 1 } \hat { V } _ { 1 2 } \hat { g } _ { 2 } ( I - \hat { V } _ { 2 1 } \hat { g } _ { 1 } \hat { V } _ { 1 2 } \hat { g } _ { 2 } ) ^ { - 1 } \left ( \hat { V } _ { 2 1 } \hat { g } _ { 1 } \hat { V } _ { 1 3 } \hat { g } _ { 3 } + \hat { V } _ { 3 } \hat { g } _ { 3 } \right ) \\ & \wedge \quad r \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \ \end{$$
$$\hat { G } _ { 1 3 } = & \left [ \hat { g } _ { 1 } \hat { V } _ { 1 3 } \hat { g } _ { 3 } + \hat { g } _ { 1 } \hat { V } _ { 1 2 } \hat { g } _ { 2 } ( I - \hat { V } _ { 2 1 } \hat { g } _ { 1 } \hat { V } _ { 1 2 } \hat { g } _ { 2 } ) ^ { - 1 } \left ( \hat { V } _ { 2 1 } \hat { g } _ { 1 } \hat { V } _ { 1 3 } \hat { g } _ { 3 } + \hat { V } _ { 2 3 } \hat { g } _ { 3 } \right ) \right ] \times \\ & \times \left [ \left ( I - \hat { V } _ { 3 1 } \hat { g } _ { 1 } \hat { V } _ { 1 3 } \hat { g } _ { 3 } \right ) + \hat { ( V } _ { 3 1 } \hat { g } _ { 1 } \hat { V } _ { 1 2 } \hat { g } _ { 2 } + \hat { V } _ { 3 2 } \hat { g } _ { 2 } ) ( I - \hat { V } _ { 2 1 } \hat { g } _ { 1 } \hat { V } _ { 1 2 } \hat { g } _ { 2 } ) ^ { - 1 } \left ( \hat { V } _ { 2 1 } \hat { g } _ { 1 } \hat { V } _ { 1 3 } \hat { g } _ { 3 } + \hat { V } _ { 2 3 } \hat { g } _ { 3 } \right ) \right ] ^ { - 1 }$$
Once 13 ˆ G and 12 ˆ G are known, 11 ˆ G can be also determined, and, in fact, one can apply the same reasoning to calculate all other terms that appear in ( ) . No approximations E G are involved, and note that all expressions depend only of the exact Green·s function of the uncoupled quaternionic sub-spaces ( ˆ x g ), which admits the usual expansion in a Dyson series of the type ˆ 0 0 ˆ ˆ ˆ ˆ x x x xx x G G V g g = + .
However, by retaining terms up to only a certain order in ˆ x g , one can obtain simpler expressions corresponding to non-exact expressions for ( ) that are useful for E G practical reasons. For example, in a treatment similar to that adopted in the usual first Born approximation 4 for ˆ x g , we can introduce a 'hyper' first Born approximation for ( ) by preserving only terms that involve the exact Green·s function of the uncoupled E G quaternionic sub-spaces
$$\mathbb { G } ( \mathbb { E } ) \approx \begin{pmatrix} \hat { g } _ { 1 } & 0 & 0 \\ 0 & \hat { g } _ { 2 } & 0 \\ 0 & 0 & \hat { g } _ { 3 } \end{pmatrix} \quad.$$
We will now show that in this case the relationship between S and T is similar to that found in usual Quantum Mechanics.
## B. Relationship between the scattering hypermatrix
## and the transition matrix
A formal relationship between the S and T matrices can be directly obtained in the form
$$\mathbb { S } _ { f i } \ & = \ & \left \langle \mathbb { P } _ { f } ^ { - } \left | \mathbb { P } _ { i } ^ { + } \right \rangle = \left \langle \mathbb { P } _ { f } ^ { - } \left | \widehat { \Omega } ^ { + } ( E _ { i } ) \left | \mathbb { X } _ { o, i } \right \rangle \\ & = \ & \left \langle \mathbb { P } _ { f } ^ { - } \left | \mathbb { I } + \mathbb { G } ^ { + } ( E _ { i } ) \mathbb { V } \left | \mathbb { X } _ { o, i } \right \rangle \\ & = \ & \left \langle \mathbb { P } _ { f } ^ { - } \left | \mathbb { X } _ { o, i } \right \rangle + \left \langle \mathbb { P } _ { f } ^ { - } \left | \mathbb { G } ^ { + } ( E _ { i } ) \mathbb { V } \left | \mathbb { X } _ { o, i } \right \rangle \\ & = \ \left \langle \mathbb { X } _ { o, f } \left | \mathbb { X } _ { o, i } \right \rangle + \left \langle \mathbb { X } _ { o, f } \left | \mathbb { T } ^ { + } ( E _ { f } ) \mathbb { G } _ { o } ^ { + } ( E _ { f } ) \left | \mathbb { X } _ { o, i } \right \rangle + \left \langle \mathbb { P } _ { f } ^ { - } \left | \mathbb { G } ^ { + } ( E _ { i } ) \mathbb { V } \left | \mathbb { X } _ { o, i } \right \rangle \ \.$$
$$& = \left \langle \mathbb { X } _ { o, j } \left | \mathbb { X } _ { o, j } \right \rangle + \left \langle \mathbb { X } _ { o, j } \right | \mathbb { T } ^ { - } ( E _ { j } ) \mathbb { G } _ { o } ^ { - } ( E _ { j } ) \left | \mathbb { X } _ { o, j } \right \rangle + \left \langle \mathbb { P } _ { j } ^ { - } \left | \mathbb { G } ^ { * } ( E _ { j } ) \mathbb { V } \right | \mathbb { X } _ { o, j } \right \rangle \,. \\ & \quad \text{$w$}, \text{ if } \, \text{ we use the identities} \ \mathbb { I } = \left | \mathbb { X } _ { o, s } \right \rangle \mathbb { X } _ { o, s } \right | = \left ( \begin{matrix} 1 _ { 1 } & 0 & 0 \\ 0 & 1 _ { 2 } & 0 \\ 0 & 0 & 1 _ { 3 } \end{matrix} \right ) \, \right \rangle \\ \mathbb { I } = \left | \mathbb { P } _ { j } ^ { - } \left \rangle \mathbb { X } _ { o, s } \right |, \, \text{ we have} \\ \mathbb { S } _ { f i } \ = \quad \begin{matrix} \delta _ { f i } + \mathbb { X } _ { o, j } \left | \mathbb { T } ^ { - } ( E _ { j } ) \left | \mathbb { X } _ { o, j } \right \rangle \mathbb { X } _ { o, j } \left | \mathbb { G } _ { o } ^ { * } ( E _ { j } ) \left | \mathbb { X } _ { o, j } \right \rangle + \left \langle \mathbb { P } _ { j } \left | \mathbb { G } ^ { * } ( E _ { j } ) \right | \mathbb { P } _ { j } \left | \mathbb { V } \right | \mathbb { V } _ { o, j } \right \rangle \\ & = \ \delta _ { f i } + \mathbb { X } _ { o, j } \left | \mathbb { T } ^ { - } ( E _ { j } ) \left | \mathbb { X } _ { o, j } \right \rangle \mathbb { X } _ { o, j } \left | \mathbb { G } _ { o } ^ { * } ( E _ { j } ) \left | \mathbb { X } _ { o, j } \right \rangle + \left \langle \mathbb { P } _ { j } \left | \mathbb { P } _ { j } \left | \mathbb { V } \right | \mathbb { V } _ { o, j } \right \rangle \\ & = \ \delta _ { f i } + \mathbb { X } _ { o, j } \left | \mathbb { T } ^ { - } ( E _ { j } ) \left | \mathbb { X } _ { o, j } \right \rangle \mathbb { X } _ { o, j } \left | \mathbb { G } _ { o } ^ { * } ( E _ { j } ) \left | \mathbb { X } _ { o, j } \right \rangle + \left \langle \mathbb { P } _ { j } \left | \mathbb { P } _ { j } \left | \mathbb { V } _ { o, j } \right \rangle \\ & = \ \delta _ { f i } + \mathbb { X } _ { o, j } \left | \mathbb { T } ^ { - } ( E _ { j } ) \left | \mathbb { X } _ { o, j } \right \rangle \mathbb { X } _ { o, j } \left | \mathbb { G } _ { o } ^ { * } ( E _ { j } ) \left | \mathbb { X } _ { o, j } \right \rangle + \left \langle \mathbb { P } _ { j } \left | \mathbb { G } ^ { * } ( E _ { j } ) \left | \mathbb { X } _ { o, j } \right \rangle + \left \langle \mathbb { P } _ { j } \left | \mathbb { T } ^ { - } ( E _ { j } ) \left | \mathbb { X } _ { o, j } \right \rangle \\ & \quad \text{$w$}. \\ & \quad \text{$in the above expression, we note that while the matrix} \ \mathbb { G } _ { o } ^ { - } \text{ is diagonal, } \mathbb { G } ^ { * } \text{ has a more}$$
In the above expression, we note that while the matrix † 0 -G is diagonal, + G has a more complicated structure. However, within the ( ) approximation (see Eqs. (13) and + E G (14)) the relationship between S and T assumes the form
$$S _ { f i } \equiv \ \widehat { \delta } _ { i } + \langle \mathbb { X } _ { 0, f } | \mathbb { T } ^ { - + } ( E _ { f } ) | \tilde { \mathbb { X } } _ { 0, \bar { f } } \rangle \langle \tilde { \mathbb { X } } _ { 0, \bar { f } } | \mathbb { G } ^ { - + } ( E _ { f } ) | \mathbb { X } _ { 0, i } \rangle + \\ + \langle \mathbb { P } ^ { - } _ { f } | \mathbb { G } ^ { + } ( E _ { i } ) | \tilde { \mathbb { P } } ^ { - } _ { f } \rangle \langle \tilde { \mathbb { X } } _ { 0, \bar { f } } | \mathbb { T } ^ { - + } ( E _ { f } ) | \mathbb { X } _ { 0, i } \rangle \\ \equiv \ \widehat { \delta } _ { f i } + \langle \mathbb { X } _ { 0, f } | \mathbb { T } ^ { - + } ( E _ { f } ) | \tilde { \mathbb { X } } _ { 0, \bar { f } } \rangle \langle \tilde { \mathbb { X } } _ { 0, \bar { f } } | \mathbb { G } ^ { - + } ( E _ { f } ) \mathbb { T } | \tilde { \mathbb { X } } _ { 0, i } \rangle + \\ + \langle \tilde { \mathbb { P } } ^ { - } _ { f } | \mathbb { Y } \mathbb { G } ^ { + } ( E _ { i } ) | \tilde { \mathbb { P } } ^ { - } _ { f } \rangle \langle \tilde { \mathbb { X } } _ { 0, \bar { f } } | \mathbb { T } ^ { - + } ( E _ { f } ) | \mathbb { X } _ { 0, i } \rangle \\ \equiv \ \widehat { \delta } _ { f i } + \langle \mathbb { X } _ { 0, f } | \mathbb { T } ^ { - + } ( E _ { f } ) \mathbb { Y } | \tilde { \mathbb { X } } _ { 0, \bar { f } } \rangle \langle \tilde { \mathbb { X } } _ { 0, \bar { f } } | \mathbb { G } ^ { - + } ( E _ { f } ) | \tilde { \mathbb { X } } _ { 0, i } \rangle + \\ + \langle \tilde { \mathbb { P } } ^ { - } _ { f } | \mathbb { G } ^ { + } ( E _ { i } ) | \tilde { \mathbb { P } } ^ { - } _ { f } \rangle \langle \tilde { \mathbb { X } } _ { 0, f } | \mathbb { Y } \mathbb { T } ^ { - + } ( E _ { f } ) | \mathbb { X } _ { 0, i } \rangle \\ + \langle \mathbb { X } _ { 0, f } | \mathbb { T } ^ { - + } ( E _ { f } ) | \mathbb { X } _ { 0, i } \rangle \delta _ { f } \left ( \mathbb { G } ^ { - + } ( E _ { f } ) \right ) _ { f } \tilde { f } \\ \equiv \ \widehat { \delta } _ { f i } + \langle \mathbb { X } _ { 0, f } | \mathbb { T } ^ { - + } ( E _ { f } ) | \tilde { \mathbb { X } } _ { 0, i } \rangle \left [ \left ( \mathbb { G } ^ { - + } _ { 0 } ( E _ { f } ) \right ) _ { i i } + \left ( \mathbb { G } ^ { + } ( E _ { i } ) \right ) _ { f f } \right ]$$
where we have used
$$| \mathbb { X } _ { _ { 0, n } } ( t ) \rangle \equiv \begin{pmatrix} \gamma _ { 1 } | \phi _ { _ { n _ { 1 } } } ( t ) \rangle \\ \gamma _ { 2 } | \phi _ { _ { n _ { 2 } } } ( t ) \rangle \\ \gamma _ { 3 } | \phi _ { _ { n _ { 3 } } } ( t ) \rangle \end{pmatrix} = \begin{pmatrix} \gamma _ { 1 } & 0 & 0 \\ 0 & \gamma _ { 2 } & 0 \\ 0 & 0 & \gamma _ { 3 } \end{pmatrix} \begin{pmatrix} | \phi _ { _ { n _ { 1 } } } ( t ) \rangle \\ | \phi _ { _ { n _ { 2 } } } ( t ) \rangle \end{pmatrix} = \Upsilon | \widetilde { \mathbb { X } } _ { _ { 0, n } } ( t ) \rangle \quad.$$
If we consider the term inside the brackets in an explicit manner,
$$\left [ \left ( \mathbb { G } _ { 0 } ^ { - \dagger } ( E _ { f } ) \right ) _ { i i } + \left ( \mathbb { G } ^ { + } ( E _ { i } ) \right ) _ { f f } \right ] =$$
$$= & T r \left [ \begin{pmatrix} \frac { 1 } { E _ { j 1 } ^ { - } E _ { i 1 } + i \eta } & 0 & 0 \\ 0 & \frac { 1 } { E _ { j 2 } ^ { - } E _ { i 2 } + j \eta } & 0 \\ 0 & 0 & \frac { 1 } { E _ { j 3 } ^ { - } E _ { i 3 } + k \eta } \end{pmatrix} - \left ( \begin{pmatrix} \frac { 1 } { E _ { j 1 } ^ { - } E _ { i 1 } - i \eta } & 0 & 0 \\ 0 & \frac { 1 } { E _ { j 2 } ^ { - } E _ { i 2 } - j \eta } & 0 \\ 0 & 0 & \frac { 1 } { E _ { j 3 } ^ { - } E _ { i 3 } ^ { - } k \eta } \end{pmatrix} \right ]$$
$$= - 2 \pi \left [ i \delta ( E _ { f 1 } - E _ { i 1 } ) + j \delta ( E _ { f 2 } - E _ { i 2 } ) + k \delta ( E _ { f 3 } - E _ { i 3 } ) \right ]$$
$$\cdot$$
we obtain
$$\mathbb { S } _ { f 1 } \cong \widehat { \delta } _ { f 1 } - 2 \pi \Big [ i \delta ( E _ { f 1 } - E _ { i 1 } ) + j \delta ( E _ { f 2 } - E _ { i 2 } ) + k \delta ( E _ { f 3 } - E _ { i 3 } ) \Big ] \Big ( \mathbb { T } ^ { \dagger } ( E _ { f } ) \Big ) _ { f i } \,.$$
It is reassuring to note that if we take the limit where 1 1 and γ = 2 3 0 in the above γ γ = = expression, we obtain the correct result for the case where a single subspace is
considered, i.e., ( ) † ( ) 2 ( ) f f E i f i f i f i E E S i T π δ δ --= -. Also, when f ≠ i and the initial and final states belong to the same continuous spectrum of energy, we have 1 2 3 f f f f E E E E = = = and 1 2 3 i i i i E E E E = = = . In this manner, finally we have [ ] ( ) † ( ) 2 ( ) f f E i f i f i E E i j k πδ ---+ + ≅ T S , whose modulus squared corresponds to the transition probability 2 f i f i = ℘ S , i.e.,
$$\left | \mathbb { S } _ { f i } \right | ^ { 2 } = 1 2 \pi ^ { 2 } \delta ^ { 2 } ( E _ { f } - E _ { i } ) \left | \left ( \mathbb { T } ^ { \dagger } ( E _ { f } ) \right ) _ { f i } \right | ^ { 2 } = 1 2 \pi ^ { 2 } \delta ( 0 ) \delta ( E _ { f } - E _ { i } ) \left | \left ( \mathbb { T } ( E _ { f } ) \right ) _ { f i } \right | ^ { 2 }.$$
However, since (0) 2 lim T T δ π →∞ = /planckover2pi , we can write the transition rate between the final and initial states ( f i R ) as
$$\frac { d \wp _ { f i } } { d T } = \mathbb { R } _ { f i } = 6 \pi \hbar { ^ { - 1 } } \, \delta ( E _ { f } - E _ { i } ) \left | \left ( \mathbb { T } ( E _ { f } ) \right ) _ { f i } \right | ^ { 2 } \quad.$$
We can easily recognize the above expression as Fermi golden rule generalized for quaternionic systems.
## IV. AN EXPRESSION FOR THE ELECTRIC CURRENT
The generalized Fermi Golden Rule can be used to estimate the electric current and, consequently, the conductance of nanoscopic systems. For this, let us consider that the continuous spectrum of initial [final] states as associated to the right ( i → R ) [left ( f → L )] electrode. We also have that each quaternionic subspace represents a possible charge state (i.e., neutral, single cation, single anion) of the device. If we multiply the above expression by the charge e , and transform the sums over the final and initial
$$\text{states into integrals} \left ( \sum _ { \beta = L ( R ) } \rightarrow \int d E _ { \beta } \rho _ { ( E _ { B } ) }, \text{ where } \rho _ { ( E _ { B } ) } \text{ is the density of states with} \\ \\ \beta = L, R \right ), \text{ we have} \\ \\ I _ { R L } = 6 e \pi \hslash ^ { - 1 } \left \{ \int d E _ { L } d E _ { R } \left [ \rho _ { ( E _ { R } ) } \rho _ { ( E _ { L } ) } \delta ( E _ { L } - E _ { R } ) \left | \left ( \mathbb { T } ( E _ { R } ) \right ) _ { R L } \right | ^ { 2 } \right ] \quad. \quad \text{(3)} \\ \\ \text{$\quad$In the specific case of electronic transport in nanoscopic systems (such as when a} \\ \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \end{cases}$$
In the specific case of electronic transport in nanoscopic systems (such as when a single molecule is connected to two electrodes, for instance), it is common to consider the fact that no direct transport exists between the two electrodes, but only through the molecule. Hence, the right to left (R → L) transmission will be given by
$$\left | \left ( \mathbb { T } ( E _ { R } ) \right ) _ { R L } \right | ^ { 2 } & = \left ( \mathbb { T } ^ { + } ( E _ { R } ) \right ) _ { R L } \left ( \mathbb { T } ^ { + } ( E _ { R } ) \right ) _ { R L } ^ { \dagger } \\ & = \left ( \mathbb { X } _ { 0, L } \left | \mathbb { V } _ { L } \mathbb { G } ^ { + } ( E _ { R } ) \mathbb { V } _ { R } \left | \mathbb { X } _ { 0, R } \right > \left < \mathbb { X } _ { 0, R } \left | \mathbb { V } _ { R } ^ { \dagger } \mathbb { G } ^ { - } ( E _ { R } ) \mathbb { V } _ { L } ^ { \dagger } \left | \mathbb { X } _ { 0, L } \right > \right ) \right ) \,.$$
If we use again the identity 0, 0, N N N N = ∑ /tildenosp /tildenosp X X I and take into account the invariance of the trace, we obtain
$$I _ { R L } = & \frac { 6 e \pi } { \hbar } \sum _ { N } \left \{ d E _ { L } \, d E _ { R } \left [ \rho _ { ( E _ { R } ) } \rho _ { ( E _ { L } ) } \delta ( E _ { L } - E _ { R } ) \langle \mathbb { X } _ { 0, L } | \mathbb { V } _ { L } ^ { + } \mathbb { G } ^ { + } ( E _ { R } ) \mathbb { V } _ { R } | \mathbb { X } _ { 0, R } \rangle. \\ \cdot \langle \mathbb { X } _ { 0, R } | \mathbb { V } _ { L } ^ { + } \mathbb { G } ^ { - } ( E _ { R } ) | \tilde { \mathbb { X } } _ { 0, N } \rangle \langle \tilde { \mathbb { X } } _ { 0, N } | \mathbb { V } _ { L } ^ { + } | \mathbb { X } _ { 0, L } \rangle \right ] \\ = & \frac { 6 e \pi } { \hbar } \sum _ { N } \left \{ d E _ { L } \, d E _ { R } \left [ \rho _ { ( E _ { R } ) } \rho _ { ( E _ { L } ) } \delta ( E _ { L } - E _ { R } ) \langle \tilde { \mathbb { X } } _ { 0, N } | \mathbb { V } _ { L } ^ { + } | \mathbb { X } _ { 0, L } \rangle. \\ \cdot \langle \mathbb { X } _ { 0, L } | \mathbb { V } _ { L } ^ { + } \mathbb { G } ^ { + } ( E _ { R } ) \mathbb { V } _ { R } | \mathbb { X } _ { 0, R } \rangle \langle \mathbb { X } _ { 0, R } | \mathbb { V } _ { L } ^ { + } \mathbb { G } ^ { - } ( E _ { R } ) | \tilde { \mathbb { X } } _ { 0, N } \rangle \right ]$$
After integrating in EL , and considering ER = E , we get
$$I _ { R L } = & \, \frac { 6 \, e } { 4 \pi h } \sum _ { N } \int d E \left [ \left \langle \tilde { \mathbb { X } } _ { o, N } | \left ( 2 \pi \phi _ { ( e ) } \mathbb { V } _ { L } ^ { \dagger } | \mathbb { X } _ { o, L } \right ) \langle \mathbb { X } _ { o, L } | \mathbb { V } _ { L } \right ) \bullet \\ & \bullet \mathbb { G } ^ { + } ( E ) \left ( 2 \pi \phi _ { ( e ) } \mathbb { V } _ { R } | \mathbb { X } _ { o, R } \right ) \langle \mathbb { X } _ { o, R } | \mathbb { V } _ { R } ^ { \dagger } \right ) \mathbb { G } ^ { - } ( E ) | \tilde { \mathbb { X } } _ { o, N } \right \rangle \right ] \quad.$$
If we define the matrix: ( ) † ( ) 0, 0, 2 E w w w w w πρ = Λ / V X X V ( w = R, L), sum over the N states and consider that the incident electron goes from an occupied level belonging to the right electrode (and, hence, described by the Fermi occupation factor ( -) ) to an unoccupied level in the left electrode (described by the Fermi R E f µ factor ( -) ( -) ), we have 1 E E A f µ µ = -
L L
$$I _ { R L } = & \frac { 3 e } { h } \int d E \, \widehat { T } r \left \{ \mathbf \Lambda _ { L } \mathbb { G } ^ { + } ( E ) \mathbf \Lambda _ { R } \mathbb { G } ^ { - } ( E ) \right \} f ( E _ { \, \cdot \, \mu _ { _ { k } } } ) A ( E _ { \, \cdot \, \mu _ { _ { L } } } ) \quad,$$
where the symbol Tr /slurabove denotes the trace over a quaternionic function ( 0, N /tildenosp X ). In an analogous manner, the left to right current will be
.
$$I _ { L R } = & \frac { 3 e } { h } \left \{ d E \, \widehat { T r } \left \{ \mathcal { M } _ { L } \mathbb { G } ^ { + } ( E ) \mathcal { M } _ { R } \mathbb { G } ^ { - } ( E ) \right \} A ( E \cdot \mu _ { _ { R } } ) f ( E \cdot \mu _ { _ { L } } ) \quad.$$
Hence, the total current can be written as
$$I _ { T } = I _ { R L } - I _ { L R } = \frac { 3 e } { h } \left \{ d E \widehat { T r } \left \{ \mathcal { A } _ { L } \mathbb { G } ^ { + } ( E ) \mathcal { A } _ { R } \mathbb { G } ^ { - } ( E ) \right \} \left ( f ( E \cdot \mu _ { R } ) - f ( E \cdot \mu _ { L } ) \right ) \quad,$$
which is the quaternionic version of the Landauer formula. In addition, we can identify the transmission as
$$\Im _ { ( E ) } = \widehat { T } r \left \{ \mathbf \Lambda _ { L } \mathbb { G } ^ { + } _ { ( E ) } \mathbf \Lambda _ { R } \mathbb { G } ^ { - } _ { ( E ) } \right \} \quad.$$
Although the above expression is quite similar to that used to describe the ballistic transport (see Eq. (1)), the present formalism makes possible to consider three alternative charge states for the system, in such a way that the evolution of the transport process can naturally lead to the mixing of the three different quaternionic subspaces. This intrinsic feature of the treatment emerges because the off diagonal elements of the matrix quantity ( ) † ( ) 0, 0, 2 E w w w w w πρ = Λ / V X X V (where w = R, L) introduce the mixing
of the subspaces. (Please note that this occurs in spite of the approximation done for ( ) in Eq. (14), when only the (exact) diagonal terms of the total Green·s function of E G each subspace were considered.) Thus, for instance, one electron could enter a system that was in an initially neutral state, but leave it when the system is in a negatively charged state. Because the total energy is conserved, the two states involved (one at the neutral and the other at the charged species) must be identical in energy and readily accessible; at least, if and only if broadening effects are included, nearby energy states should be available to participate of the transport process. This hypothetical, but possible, situation is depicted in Figure 1.
## V. APPLICATIONS: TRANSPORT IN CONJUGATED AND SATURATED SYSTEMS
As a stringent test of the range of applications of the theoretical formalism developed above, we examined two archetypical molecules of different kinds: benzenedi-thiol (BDT), a conjugated system, and octane-di-thiol (ODT), whose structure only contains saturated bonds. We have chosen these molecules not only because they represent canonical examples of strong and weak coupling regimes, respectively, 10, 21, 22 but also because experimental data for the current and conductance exist for them. 23-25 Here, we will briefly present the most relevant results (see Figs. 2-4), since a more detailed discussion can be found in Ref. [26].
In Figs. 2 and 3 we not only depict the corresponding extended molecules, where two opposite ends of the organic compound considered was coupled to a metallic cluster of 11 gold atoms, but also show some selected frontier molecular orbitals at some special values of the external potential that most affect them. This extended molecule was used to extract the coupling strengths at DFT level of calculation using the B3LYP
functional and the 6-31d basis set, for different values of an external electric field. The curves representing the behavior of the current and the conductance of these two molecules are shown in Figs. 4 and 5, where in each case we also indicate the individual contributions of the ballistic and co-tunneling terms to the total current traversing the molecule.
As one can see in Fig. 2b, in the case of the BDT molecule, up to an applied voltage appl V = 0.84 V the LUMO is completely localized in the left-hand side cluster. However, at this voltage an avoided-crossing situation 27 begins to occur for the HOMO and LUMO (as depicted in the inset of Fig. 4b), resulting in an exchange of the spatial localization of these two FMOs. Above this critical limit of the applied potential, the LUMO begins to act as a true 'conducting channel' for the system, a fact that is associated to the now increasing values of the up to then non-existent electric current. The most noticeable increase of the current occurs at an external bias of 1.1V, when the Fermi level crosses the energy of the LUMO+1, a completely delocalized orbital.
As expected, an entirely different physical situation is responsible for the transport in the saturated molecule ODT, for which no avoided-crossing situation (and, therefore, no consequent symmetry exchange) should be anticipated to occur. Since the LUMO remains localized in the left-hand side cluster for the entire range of examined potentials, it cannot contribute to a ballistic transport. This spatial localization precludes the transfer of electrons from the metal to the molecule. Hence, with no coupling between the cluster and the molecule itself, it is unlikely that an anion could be formed. However, as one can see in Fig. 3b, the HOMO - which is localized both in the organic molecule as in the right-hand side cluster - can lose one electron to the anode. As the cation is formed, the local density of states is changed and now the unoccupied molecular orbitals of the newly formed electron-deficient molecular system can accept
electrons from the cathode. Hence, a typical co-tunneling situation develops accordingly to what can be seen in Fig. 5b, where the contribution of the ballistic terms does not exist for the entire range of applied external field.
In general, the above results are in a good qualitative agreement with the experimental data available for these systems. 24, 25 The interested reader can find a more complete discussion of the application of the quaternionic formalism to the description of the transport in these two model systems in Ref. [26].
## VI. CONCLUSIONS AND PERSPECTIVES
We have presented the quaternionic method, a new theoretical approach specially developed to deal with the calculation of the electric current that flows through a molecule subject to an externally applied electric field, while connected to two terminal electrodes. A full ab initio self-consistent calculation can be implemented for the socalled extended molecule, which comprises the organic molecule of interest plus the two small metal clusters attached to it. The quarternionic approach allows for the possibility of concomitant ballistic and co-tunneling transport mechanisms, and the molecular conductance profile is particularly sensitive to the opening and closing of transport channels through the transient single-charged species. For this, we have chosen to follow a Dyson series and a generalized Fermi golden rule treatment within a quaternion formalism, where the evolution of an initially neutral molecule can proceed either preserving the neutral charge or by following the Hamiltonians that describe a single-charged anion or a single-charged cation. In this manner, one can go beyond a canonical quantum mechanical formalism of the separated subsystems, since we allow for the possibility of coupling between the three quaternionic subspaces (each one with
its own imaginary unit and corresponding to the three different charge states of the molecular system).
The main advantage of the present approach is that in the final expression for the transmission of an electron flowing through the molecule, the ballistic and co-tunneling terms appear naturally as complementary regimes of transport. Also, the treatment can take into account the possibility of describing strongly correlated situations (as those associated to the presence of spatially localized molecular orbitals) where the ballistic regime usually fails. The allowance for a natural 'competition' between the ballistic 3 and co-tunneling regimes gives rise to the possibility of application of the quaternionbased formalism to strong correlated systems where atomic-like localized molecular states play an important role.
As illustrative examples of the range of applications of the quaternionic approach, we have briefly discussed the transport in the organic molecules BDT and ODT, whose field-dependent electronic structures were determined using the B3LYP functional approximation and adopting the 6-31d basis set. For the first molecule, a conjugated system, we have found the ballistic mechanism to be dominant, as it should be expected. For the second, whose unoccupied states are spatially localized, the transport is made possible by the transfer of one electron from the molecule to the anode, with the frontier molecular orbitals of the corresponding cation playing an active role in the process.
We have shown that the present formalism allows for a unified description of the transport phenomena in molecular systems. Besides the ballistic processes, a non-zero probability exists for the capture of an electron or a hole from the connected electrodes, so that a change the charge of the molecule would allow the charge transfer to proceed in a new potential surface. The coexistence of different mechanisms not only gives
opportunity to new possibilities of approaching and interpreting the molecular conductance problem, but also opens up several evident possibilities of extension of the present work.
From the point of view of the quaternion formalism, one can imagine its application to the description of other physical situations where more than one evolution pathway is possible, as it would be the case of describing co-existing molecular tautomers or isomers. Furthermore, in case where more than three viable pathways for the time evolution of a given system can be foreseen, one could consider resorting to octanions, as to accommodate a larger gamut of different possibilities.
## References
- 1. A. Aviram and M. A. Ratner, Chemical Physics Letters 29 (2), 277-283 (1974).
- 2. C. Durkan, Current at the nanoscale: an introduction to nanoelectronics. (Imperial College Press; Distributed by World Scientific Pub., London, 2007).
- 3. S. Datta, Electronic transport in mesoscopic systems. (Cambridge University Press, Cambridge; New York, 1995).
- 4. M. Di Ventra, Electrical transport in nanoscale systems. (Cambridge University Press, Cambridge, 2008).
- 5. E. S. C. C. Juan, Molecular Electronics: An Introduction to Theory and Experiment (Word Scientific, 2010).
- 6. S. Datta, Quantum transport: atom to transistor. (Cambridge University Press, Cambridge, UK ; New York, 2005).
- 7. C. K. Wang, Y. Fu and Y. Luo, Physical Chemistry Chemical Physics 3 (22), 50175023 (2001).
- 8. M. Lundstrom, Nanoscale transistors: device physics, modeling and simulation. (Springer, New York, 2006).
- 9. S. Datta, M. A. Ratner and Y. Xue, Chemical Physics 281, 20 (2002).
- 10. K. Stokbro, J. Taylor, M. Brandbyge and H. Guo, in Introducing Molecular Electronics, edited by T. L. N. O. Physics (Springer, Berlin, 2005), Vol. 680, pp. 34.
- 11. M. A. Ratner, A. Nitzan and M. Galperin, Abstracts of Papers of the American Chemical Society 230, U2775-U2775 (2005).
- 12. D. K. Ferry and S. M. Goodnick, Transport in nanostructures. (Cambridge University Press, Cambridge, U.K.; New York, 1997).
- 13. R. A. Moreira and C. P. de Melo, submitted to The Journal of Chemical Physics (2014).
- 14. M. D. Ventra, Electrical Transport in Nanoscale Systems. (Cambridge University Press, New York, 2008).
- 15. F. Remacle and R. D. Levine, Chemical Physics Letters 383 (5-6), 537-543 (2004).
- 16. R. Li, S. M. Hou, J. X. Zhang, Z. K. Qian, Z. Y. Shen and X. Y. Zhao, J. Chem. Phys. 125 (19) (2006).
- 17. D. V. Averin and Y. V. Nazarov, Physical Review Letters 65 (19), 4 (1990).
- 18. W. R. Hamilton, Elements of Quaternions. (Dublin, 1866).
- 19. S. L. Adler, Quaternionic Quantum Mechanics and Quantum Fields. (Oxford University Press, 1995).
- 20. R. D. Levine, Quantum Mechanics of Molecular Rate Process. (Dover, New York, 1969).
- 21. M. Lundstrom and J. Guo, Nanoscale transistors: device physics, modeling and simulation. (Springer, New York, 2006).
- 22. F. Jiang, Y. X. Zhou, H. Chen, R. Note, H. Mizuseki and Y. Kawazoe, Physical Review B 72 (15), 155408 (2005).
- 23. C.-L. Ma, D. Nghiem and Y.-C. Chen, Applied Physics Letters 93 (22), 222111222113 (2008).
- 24. C. A. Martin, D. Ding, H. S. J. van der Zant and J. M. van Ruitenbeek, New Journal of Physics 10 (6), 065008 (2008).
- 25. J. Tomfohr, G. K. Ramachandran, O. F. Sankey and S. M. Lindsay, in Introducing Molecular Electronics, edited by G. Cuniberti, K. Richter and G. Fagas (Springer Berlin / Heidelberg, 2005), Vol. 680, pp. 301-312.
- 26. A. C. L. Moreira and C. P. de Melo, to be submitted (2014).
- 27. C. P. de Melo and A. C. L. Moreira, The Journal of Physical Chemistry C 116 (4), 3122-3131 (2011).
## Figure Captions
Figure 1: Two possible situations where only the subspaces 1 (left panel) and 2 (right panel) are considered. Each subspace (1 and 2) is composed by a two-level system (levels a and b ). The symbol σ µ represents the chemical potential of the σ - electrode ( , L R σ = ), xy ε the x -level of the y -subspace (i.e., x = a , b and y = 1, 2), and xy σ τ is the coupling between this energy level and the σ - electrode. Due to the coupling to the infinite electrodes, all levels ( a and b ) in all subspaces (1 and 2) are broadened.
Figure 2: (a) The extended molecule benzene-di-thiol where each metallic cluster contains 11 gold atoms. (b) Behavior of selected molecular orbitals as a function of the applied voltage.
Figure 3: (a) The extended molecule octane-di-thiol where each metallic cluster contains 11 gold atoms. (b) Behavior of selected molecular orbitals as a function of the applied voltage.
- Figure 4: Electric current and conductance (a) and ballistic and co-tunneling contributions to the current (b) for benzene-di-thiol as a function of the applied voltage. Inset: HOMO-LUMO crossing.
Figure 5: Electric current and conductance (a) and ballistic and co-tunneling contributions to the current (b) for octane-di-thiol as a function of the applied voltage.
Figure 1
02
SUBSPACE 2
TR
a2
HR
Figure 2
Figure 3
Figure 4
Figure 5
| null |
[
"Augusto C. L. Moreira",
"Celso P. de Melo"
] |
2014-08-05T18:44:45+00:00
|
2014-08-05T18:44:45+00:00
|
[
"cond-mat.mes-hall"
] |
A Generalized Landauer Approach for Electron Transport in a Single Molecule
|
We present a quaternion-inspired formalism specifically developed to evaluate
the electric current that traverses a single molecule subjected to an
externally applied voltage. The molecule of interest is covalently connected to
two small metallic clusters, forming an extended molecule complex. The
quaternion approach allows for an integrated treatment of the charge transport
in single molecules where both ballistic and co-tunneling (coherent) mechanisms
are taken on equal footing, although only in the latter case the presence of
eventual transient charged states of the system needs to be considered. We use
a Dyson series to obtain a generalized Fermi golden rule, from which we derive
an expression for the net current the two electrodes: in doing this, we take
into account all possible transitions between electronic states localized at
the electrodes and levels in the extended molecule complex. In fact, one can
apply the method to the entire range of coupling regimes, not only in the weak
or strong cases, but also in intermediate situations, where ballistic and
co-tunneling processes compete with each other. We also discuss initial results
of the application of this formalism to the description of the electronic
transport in two small organic molecules representative of two different limit
situations. In the first case, a conjugated molecule (where spatially
delocalized molecular orbitals favor ballistic contributions) is considered,
and in the second, the current traverses a saturated hydrocarbon (whose
structure should contain more localized molecular orbitals). In both cases, we
fully describe the field-induced self-adjustment of the electronic levels of
the extended molecule complex at an ab initio quantum chemical level, using
density functional theory.
|
1408.1067v1
|
## Spitzer SAGE-Spec: Near Infrared Spectroscopy, Dust Shells, and Cool Envelopes in Extreme Large Magellanic Cloud AGB Stars 1
R. D. Blum 2
S. Srinivasan, F. Kemper 3
B. Ling 3 4 ,
K. Volk 5
## ABSTRACT
K -band spectra are presented for a sample of 39 Spitzer IRS SAGE-Spec sources in the Large Magellanic Cloud. The spectra exhibit characteristics in very good agreement with their positions in the near infrared Spitzer colormagnitude diagrams and their properties as deduced from the Spitzer IRS spectra. Specifically, the near infrared spectra show strong atomic and molecular features representative of oxygen-rich and carbon-rich asymptotic giant branch stars, respectively. A small subset of stars were chosen from the luminous and red extreme 'tip ′′ of the color magnitude diagram. These objects have properties consistent with dusty envelopes but also cool, carbon-rich 'stellar ′′ cores. Modest amounts of dust mass loss combine with the stellar spectral energy distribution to make these objects appear extreme in their near infrared and mid infrared colors. One object in our sample, HV 915, a known post asymptotic giant branch star of the RV Tau type exhibits CO 2.3 µ m band head emission consistent with previous work that demonstrates the object has a circumstellar disk.
Subject headings: stars: , stars: carbon, stars: late-type, stars: mass-loss, (stars:) circumstellar matter, infrared: stars, (galaxies:) Magellanic Clouds
Facilities:
SOAR (), Spitzer()
## 1. Introduction
Meixner et al. (2006) presented Surveying the Agents of a Galaxy's Evolution (SAGE), a Spitzer IRAC and MIPS imaging survey of the Large Magellanic Cloud (LMC). The survey covered the main body of the LMC (7 × 7 degrees ) in all the IRAC and MIPS bands. 2 The resolved stellar populations in the LMC were particularly well covered on the giant and asymptotic giant branches (AGB) as well as on the sequences of more luminous red supergiants (RSG) (Blum et al. 2006). Kemper et al. (2010) selected a sample of ∼ 200 objects from the SAGE color-magnitude diagrams (CMD) to observe with the Spitzer IRS (see below); this IRS data set is known as the SAGE-Spec survey. See also Zijlstra et al. (2006) and Sloan et al. (2008) for Spitzer IRS observations of evolved stars in the LMC. In this paper, we present near infrared K -band spectra of the brighter IRS sources from the Kemper et al. (2010) sample. The IRS spectra have been classified with respect to their basic properties according to their mid and far-infrared spectral characteristics by Woods et al. (2011).
One of the primary goals of SAGE and SAGE-Spec is to quantify the mass loss return of evolved stars into the LMC interstellar medium. To this end, we have embarked on a program to study individual objects representing each class of mass losing star in the CMD to better understand the mass loss contributions as a function of position on the CMD (i.e. by evolutionary state, chemical composition, and mass). Knowledge of the type of 'stellar core ′′ embedded in the dusty envelope for any given source is an important model input and can help elucidate the nature of the object as well. Both Srinivasan et al. (2010) and Sargent et al. (2010) have modeled individual sources from the SAGE/SAGE-Spec sample. In these cases, photometry was used to constrain the stellar models at optical and near infrared wavelengths. Adding near infrared spectra can further constrain the stellar core
1 Based on observations obtained at the Southern Astrophysical Research (SOAR) telescope, which is a joint project of the Minist´rio da Ciˆncia, Tecnologia, e Inova¸˜o (MCTI) da Rep´blica Federativa do Brasil, e e ca u the U.S. National Optical Astronomy Observatory (NOAO), the University of North Carolina at Chapel Hill (UNC), and Michigan State University (MSU).
2 NOAO, 950 North Cherry Avenue, Tucson, Arizona, 85719
3 Academia Sinica, Institute of Astronomy and Astrophysics, 11F of Astronomy-Mathematics Building, NTU/AS, No. 1, Sec. 4, Roosevelt Rd, Taipei 10617, Taiwan, R.O.C.
4 University of Toronto and Academia Sinica Institute for Astronomy and Astrophysics
5 Space Telescope Science Institute, 3700 San Martin Drive, Baltimore, MD 21218, USA
properties (for example, leading to a more accurate T eff ).
The SAGE-Spec sample included 197 sources selected from the SAGE IRAC and MIPS photometry (Kemper et al. 2010). The sample was selected to cover the range of populations (young and old) expected based on the known star-formation history of the LMC and cover each area of the CMD. The present near infrared program concentrated on the brighter objects in the sample and those concentrated on the AGB, RSG, and extreme AGB (XAGB) as defined by Blum et al. (2006). These three populations are easily seen in Figure 1 and the sub-sample of objects observed is over plotted in the diagram. The RSG are the sequence with the brightest objects observed in this sample (e.g. object 19). The AGB stars include both oxygen (O-) rich and carbon (C-) rich stars above the red giant branch (the dense sequence of stars at J -[8.0] ≈ 1.0 and reaching [8.0] = 12). The XAGB starts near object 69 in this diagram and extends to redder colors (to object 144 and beyond). Galaxies inhabit the broad clump of fainter objects near J -[8.0] ≈ 5.0. The obvious sequences to the blue of the red giants and RSG are warmer evolved stars, hot stars, and foreground objects; see Blum et al. (2006) for details.
## 2. Observations
The brighter stars at 2 µ m ( K band) were chosen from the SAGE-Spec sample to be observed at the SOAR 4.1-m telescope on Cerro Pach´n in Chile. Practically speaking this o meant sources with K ≤ 12 mag. A sample of 39 stars was observed over a five night period in November, 2007. Spectra were obtained with the Ohio State InfraRed Imaging Spectrograph, OSIRIS. OSIRIS is described in detail by Depoy et al. (1993). The instrument was configured in the f /7 camera mode, which delivers 0.14 ′′ pixel -1 , an 0.45 ′′ (3 pixel) slit and a spectral resolution of R = λ/ ∆ λ = 3000.
Data were taken on the nights of 16-20 November 2007 (local time). The weather was good, mostly clear for the entire five night run as indicated by the GOES satellite imaging archive maintained at CTIO . The seeing was also good as measured on acquisition images. 6 16 November was outstanding with acquisition images recording delivered image quality (DIQ) under 0.5 ′′ most of the night. The other nights had DIQ near 0.5 ′′ in acquisition images most of the time. Excursions to 0.7 ′′ were observed as was one acquisition sequence as good as 0.31 . ′′ The objects were identified unambiguously in 2MASS images using the SAGE coordinates. In each case a bright K -band source was easily identified as the target in acquisition images which we compared to the 2MASS image. SAGE IRAC sources have
6 http://www.ctio.noao.edu/sitetests/GOES/
typical positional differences from 2MASS positions of ∼ < 0.3 ′′ (Meixner et al. 2006).
Each source was observed in the same way. Objects were acquired in imaging mode with OSIRIS and then placed on the center of the 72 ′′ × 0.42 ′′ slit. The telescope was then moved to dither the source along the slit obtaining six samples separated by 6 . ′′ Observations of hot stars (A-type) were made throughout each night every few hours and at airmass within 0.1 of each object to be used for correcting Telluric features in the object spectra.
Flat field images were obtained using the calibration lamp mounted on the same Instrument Support Box (ISB) that the instrument is mounted on. Flats were taken after each object sequence in the hope that this would more effectively remove fringing in the spectrum due to the instrument. However, this was not the case and in the end flats were averaged throughout the night.
## 3. Data Reduction
All basic data reduction steps were completed with IRAF . IRAF wrapper scripts from 7 the CTIO InfraRed Reduction (CIRRED ) package were used to execute the actual data 8 reduction steps.
The object images were flattened with the dome flats, and bad pixels corrected by interpolating across them in each dimension. The bad pixel mask was made using illuminated flat field images and dark images.
Sky images were made by median combining the on-source images themselves. Once corrected by the flat field and bad pixel mask, each object image was sky subtracted using the median sky image from that object sequence. One dimensional spectra were extracted from these images, and the six samples combined to form an object spectrum. The object spectra were calibrated in wavelength using the quadratic solution to emission lines from an Argon lamp in the ISB calibration source. The RMS error in the known positions of Ar lines was less than 0.1 pixels (1 pixel = 3.7 ˚ ). All the spectra were arbitrarily shifted to have a A CO 2-1 rotational vibrational band head position of 2.2935 µ m.
One dimensional spectra were extracted from the two dimensional images using IRAF
7 IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy (AURA) under cooperative agreement with the National Science Foundation.
8 http://www.ctio.noao.edu/instruments/ir instruments/datared.html
APEXTRACT. The average spectral flux differences between individual dithers (typically 10 -20 % but with occasional larger excursions) are dominated by slit positioning (due to tracking, flexure, and slit position angle orientation error) and variable cloud (in some cases). Each spectrum from individual dithers was normalized before median combining to enhance bad pixel rejection.
Each object spectrum was then divided by a hot star telluric ′′ standard ′′ star spectrum (HD or Hipparcos A-type stars) that had been reduced in just the same way. Small shifts in wavelength were made to align the hot star telluric features with those in the object spectrum (by cross correlation). The hot star chosen was always one that had been observed at an airmass within 0.1 of the object, and usually within 1 to 2 hours clock time.
As noted by Dewitt et al. (2013), there is an unknown fringe or other instrumental signature in the OSIRIS spectra taken at SOAR (see Figure 2). The pattern manifests itself as three main features approximately 60 pixels in width with amplitude 10% to 20% of the continuum. There are secondary features associated with each primary feature that have ∼ half the primary amplitude. Fortunately, these features are straightforward to remove from our spectra at a level which is sufficient for our science goals. The features appear at times, but not other times and affect both the objects and standards in the same way.
The features are removed as the last step of the reduction process. First, for each night, we built two feature or fringe maps. These were obtained by dividing hot star spectra of the same star taken at different times and then smoothing the resulting ratio by a ∼ 100 pixel moving average. The only structure left in such a ratio is the feature strength as a function of wavelength. Since we don't know if any particular object will have ′′ emission ′′ or ′′ absorption ′′ features after dividing by the telluric standard, we construct a map of both types. The telluric corrected object spectrum was reviewed and then the appropriate map chosen to use to correct any of the fringe-like features. The correction was made by using the IRAF task TELLURiC which allows the user to interactively shift a standard spectrum (in this case the feature map) relative to the object and also vary the strength of the features. Each object was corrected in this way and the spectra shifted and feature strength adjusted by eye until the best final spectrum was obtained (an RMS difference is given by the task TELLURIC which was also used to guide the choice of final correction). While the source of the features remains unclear, the correction is relatively straight forward and relies entirely on taking multiple spectra of the same object. An example of the features, uncorrected and corrected spectra is shown in Figure 2.
## 4. Results
The object properties are listed in Table 1. The objects are listed in order of increasing J -[8.0] color. The first column gives the object number as defined in our original observing list. The sources were independently numbered in the Spitzer IRS proposal. Each spectrum has an associated ID for the Spitzer IRS observations (Woods et al. 2011) as well, and this SSID is noted in the table. A 'type ′′ is assigned in the table based upon three distinct observations: the location in the CMD, the (present) near infrared spectrum, and the point source classification of Woods et al. (2011) for the associated IRS SSID.
The great majority of the sources presented herein are known variable stars; see the descriptions presented for each object by Woods et al. (2011) based on the variable classifications of Alcock et al. (1998); Groenewegen (2004); Ita et al. (2004); Fraser et al. (2005, 2008); Soszy˜ski et al. (2009) and Vijh et al. (2009). n Riebel and collaborators (private communication) have made a study a the photometric variability of SAGE sources in the IRAC bands. They find large amplitude variations (0.3 - 1 mag) for objects 3, 84, 87, 88, 94, and 144 in our sample in [3.6] and [4.5].
The object spectra are shown in Figures 3, 4, 5, 6, and 7. In the following, we discuss the near infrared features in these spectra that clearly identify their O-rich, C-rich, and 'extreme' character. Inspection and comparison of the individual spectra with the labels in Figure 1 for the J -[8.0] CMD shows the spectral morphology indicated is well matched to the position of the objects in the CMD. For example, C-rich stars do not appear on the lower part of the AGB in Figure 1 or on the RSG sequence. O-rich objects are generally confined to the lower part of the AGB, and the RSG sequence appears to be exclusively populated by O-rich stars. Some O-rich objects do fall on or near the XAGB, and we discuss these further below.
The objects shown in Figures 3 and 4 have the distinct features of normal M-type stars that are typical of stars on the RSG locus and O-rich AGB locus (Nikolaev & Weinberg 2000; Cioni et al. 2006; Blum et al. 2006). The features include strong CO absorption due to the 2-0 rotational-vibrational band head at 2.29 µ m, as well as atomic features such as Na I and Ca I at 2.21 µ m and 2.26 µ m, respectively (Kleinmann & Hall 1986). All these objects are independently identified as O-rich by Woods et al. (2011) from their Spitzer IRS spectra.
The C-rich stars (Figure 5) exhibit strong, unresolved, molecular absorption that results in so much distributed opacity the continuum appears 'noisy . ′′ But this is not the case (as is well known), the spectra are normal for these types; see, for example, Frogel & Hyland (1972), or more recently the compilation of Tanaka et al. (2007). The C-rich objects do still
show strong CO absorption at 2.29 µ m, but atomic features typical in the O-rich stars are not identifiable. Even a cursory comparison of Figures 3, 4, and 5 shows that the stars are easily distinguished by their near infrared spectra. All these objects are independently identified as C-rich by Woods et al. (2011) from their Spitzer IRS spectra.
The stars depicted in Figure 6 are so called 'extreme AGB (XAGB) ′′ stars as defined in Blum et al. (2006). These objects populate the plume of AGB stars that run to very red ('extreme ) colors in the ′′ J -[8.0] vs. [8.0] CMD (Figure 1). As defined by Blum et al. (2006) and based upon near infrared colors (Nikolaev & Weinberg 2000; Cioni et al. 2006), the XAGB stars are those with J -[8.0] > 3.1 mag. The sequence of spectra goes toward redder J -[8.0] from top to bottom. The bluest XAGB stars appear like normal C-rich objects. Those toward the red end of the sequence, however, become nearly featureless. There is still apparent CO 2.29 µ m absorption, though it is weaker as is the presumed normal distributed absorption seen in Figure 5. Frogel & Hyland (1972) studied Galactic C-stars with a range of colors and concluded veiling of the stellar photosphere by dust emission arising in the shell would explain the anti correlation of CO absorption strength and infrared color. None of the Frogel & Hyland (1972) stars appears as 'veiled ′′ as the reddest XAGB stars presented here or with as weak absorption, but the veiling mechanism is probably correct in the LMC case as well (see § 5).
The XAGB objects all show a strong feature not present in the normal C-rich stars, absorption at 2.26 µ m. This may be due to CN (Wallace & Hinkle 1996). Frogel & Hyland (1972) also identify numerous absorption features due to CN (some near 2.26 µ m) as well as features due to C H , and HCN, though no feature as strong as shown in Figure 6 is present 2 2 at 2.26 µ m in their spectra. Woods et al. (2011) classify the XAGB stars in Figure 6 as C-rich from their mid-infrared IRS spectra.
The remaining few stars are shown in Figure 7. These objects, discussed below, have spectra that are not necessarily expected for their location in the CMD. It is only with additional information (Woods et al. 2011, the Spitzer IRS point source characterization) that their nature is discovered. Object 170 is a galaxy according to Woods et al. (2011). Its spectral energy distribution (SED) places it in the galaxy region of the CMD (Blum et al. 2006). The remaining four objects turn out to be O-rich, though only object 7 is obviously so from the near infrared spectrum. Objects 7, 81, 88, and 90 are classified as post-AGB (PAGB) stars (Woods et al. 2011). Object 88 shows CO emission in the near infrared, while objects 81 and 90 show CO 2.3 µ m absorption.
## 5. Discussion
Blum et al. (2006) categorized most of the evolved stars in the LMC Spitzer and 2MASS CMDs based on colors and magnitudes and confirmed the basic characterization of objects in Figure 1 through association of position with what was known from spectroscopic followup in the literature. More recently, Boyer et al. (2011) did a similar classification for the AGB stars in the Small Magellanic Cloud which results in an an analogous set of C-rich and O-rich AGB stars. Boyer et al. (2011) adopted the same color cut to define XAGB stars in the J -[8.0] vs. [8.0] CMD as did Blum et al. (2006). As seen in Figure 1 (adopted from Blum et al. 2006), the XAGB clearly appears to be an extension of the AGB which itself had two distinct O- and C-rich components. The RSGs were classified as a luminous branch of more massive O-rich (i.e. K or M-type) stars and between the AGB and RSG sequence were bright or luminous O-rich AGB stars. But at colors with J -[8.0] ∼ > 3, little was known, and it was not clear what the nature of any particular XAGB star should be. The main objective of this work is to use near infrared spectra and the Spitzer IRS spectral characteristics to further explore the AGB with emphasis on the XAGB.
Figure 9 shows a blowup of the AGB, RSG, bright O-rich, XAGB region of the J -[8.0] vs. [8.0] CMD. The spectra presented in this work were selected to fall neatly on the sequences in this diagram. The O-rich stars (Figures 3 and 4) are distributed on the RSG sequence, the bright O-rich AGB sequence (between the RSGs and the normal C-star sequence), and at the tip of the main O-rich AGB above the red giant branch.
The normal C-rich stars (Figure 5) fall on the main C-rich sequence. XAGB objects 77 and 38 (Figure 6) are clearly normal C-rich stars (compare Figure 5). The IRS spectra are consistent with this classification. Of the remaining objects with colors that place them on or near the XAGB, most have rising continua to the red and weak features (objects 94, 87, 123, 142, 50, 144). One object, number 87 was not detected (we will discuss it briefly below). There are apparently three O-rich objects in the XAGB, and these are discussed in § 5.1.
## 5.1. Extreme AGB Stars
As noted above, several of the XAGB objects have near infrared spectra of normal Cstars. Upon closer examination, most of the XAGB stars with weak features have absorption band heads due to CO. Only object 94 and 142 do not (or the feature is very weak). All of the XAGB stars have evidence of an absorption feature at 2.26 µ m, except object 94. Above we speculated this absorption might be due to CN.
The objects in the XAGB also have the reddest colors in our sample. Their near infrared spectra rise steeply toward longer wavelengths. The nature of the continuum combined with the absorption due to CO and the fact that several are normal C-stars suggests the XAGB is largely populated with C-stars with varying amounts of dust emission. This emission causes the spectra to be rising in the near infrared and veils, or dilutes the otherwise normal Cstar spectrum. Frogel & Hyland (1972) used the anti-correlation of CO 2.29 µ m absorption strength and infrared color to argue this veiling mechanism was responsible for the systematic change in spectral morphology of a sample of Galactic C-stars.
To further test this idea, we took one of our normal C-stars and combined its near infrared spectrum with an equal contribution from a cool black body as a simple representation of emission from a dust shell (trial and error yielded a good match with a 400K blackbody). The result is shown in Figure 8. The combined spectrum is a remarkably good match to one of the XAGB stars, object 123. The unknown absorption feature is obviously not reproduced since it is not present in the original spectrum of the normal C-star, object 82. However, the veiling of the myriad absorption lines in the C-star including CO at band heads at 2.3 µ m and longer is well reproduced as is the overall shape of the continuum. The apparently good fit is fortuitous since there is also reddening in the circumstellar environment which will cause the near infrared spectrum to be more red. The 400K BB overcompensates (i.e. it is cooler than it should be in the absence of reddening) for the required slope, and so we might expect the real temperature to be higher (this is indeed the case as shown by detailed modeling; see below).
It is tempting to assume the reddest objects in the present sample must have the strongest mass-loss. However, the Spitzer IRS spectra suggest weak dust emission features for the XAGB objects observed at SOAR as a whole. The brighter O-rich stars in the present sample show more prominent emission features in the mid-infrared due to dust. The SOAR sample is limited by objects whose K -band magnitudes are brighter than 12, so the present sample, while very red in J -[8.0], does not contain the reddest XAGB stars (see for example Gruendl et al. 2008; Riebel et al. 2012).
In order to verify that the picture of the XAGB outlined above is correct, we ran a series of models to test whether modest amounts of dust could reproduce both the near infrared spectra of the XAGB stars (as suggested by Figure 8) as well as the mid-infrared Spitzer IRS spectra. Our models were created using the dust code 2Dust by Ueta & Meixner (2003). We represent the central star in these models with the Aringer et al. (2009) carbon-star models of LMC metallicity.
The present analysis expands on the work of Srinivasan et al. (2010) who did a detailed model of object 123 (but without specific knowledge of the near infrared spectrum). In
the present work, we test the nature of the XAGB by modeling one of the most 'extreme' cases, object 142 (Figure 6). Object 142 was defined as a variable star by Vijh et al. (2009) based on its SAGE epoch 1 and 2 IRAC photometry which varied significantly compared to photometric uncertainties. The IRAC epoch 1 data were obtained in 2005 July and epoch 2 followed about three months later. The MIPS data for each epoch were obtained in the week following the corresponding IRAC data. In addition to the SAGE photometry, near infrared photometry is available from 2MASS (Skrutskie et al. 2006) and the InfraRed Survey Facility (Kato et al. 2007, IRSF). Visible photometry (V and I) is available from Zaritsky et al. (2004). None of these observations was obtained simultaneously with another.
The dust properties were kept fixed, and were the same as in the GRAMS carbon-star grid (Srinivasan et al. 2011); in particular, the dust consisted of a mixture of amorphous carbon and 10% SiC by mass. Using a range of stellar cores with T eff between 3000 K and 4000 K, we constructed dust shell/envelope models with varying amounts of dust (controlled by the strength of the SiC feature at 11.3 µ m) and a few different values for the inner radius of the dust shell.
Models were constructed for two distinct cases for object 142. In the first case, we compared models only to the mean photometry. The results are shown in Figure 10. We find a range of models that are consistent with the photometry taking into account variability for the multiple epochs of the data. Models with T eff between 3000K and 4000K and with optical depth in the SiC 11.3 µ m line of 0.2 to 0.4 were constructed and compared to the data. Best fits to the mean photometry were obtained with T eff ≈ 3000K and optical depth of 0.3. Wfound that the inner radius was reasonably constrained by the mid-IR photometry alone (as was the case for Srinivasan et al. 2010). The best match occurs for an inner radius equal to three times the stellar radius (corresponding to inner radius dust temperatures near 1000 K). These models are normalized to the IRAC [8.0] mag flux.
The second case compared the model results to photometry, the SAGE-Spec IRS spectrum, and the present near infrared spectrum. The goal was to see if a model could reproduce the dusty IRS spectrum including the SiC 11.3 µ mfeature, the veiled near infrared spectrum and the photometry. We modified the near-infrared part of the input carbon-star photospheric models used with 2Dust by substituting the scaled K -band spectrum of object 82 for the model flux in this region. Object 82 is a normal C-star that is assumed to have negligible reddening from its circumstellar envelope or the interstellar medium of the LMC (at least in the near infrared). By adding in the normal C-star spectrum to the core star we can see how the emergent spectrum compares to the observed XAGB spectrum of object 142.
As a consistency check, we ran a set models to compare to the photometry and SAGE-
Spec spectrum of object 82. The object 82 data are best fit by models with similar T eff and luminosity as we found for the fits to the mean photometry for object 142 (3200K ± 100K vs. 3000K ± 100K and 51000 ± 600 L /circledot vs. 39000 ± 500 L /circledot for object 82 and object 142, respectively).
Initial comparisons of the models to the data showed that no models could fit the IRS data, the MIPS photometry and simultaneously the mean near infrared and visible photometry and produce a veiled near infrared spectrum at the same time. Completely veiling the near infrared spectral features required more dust than the models produced while still reproducing the (brighter) short wavelength fluxes. However, models could be obtained which compared well to the IRS spectrum, a sub-set of the near infrared, IRAC, and MIPS photometry and which simultaneously produced a suitably veiled near infrared spectrum.
In Figure 11, we show models compared to the IRAC and MIPS epoch 1 photometry (the more reddened epoch). The SAGE-Spec IRS spectrum is also shown as is the near infrared spectrum and photometry. The SAGE-Spec fluxes and MIPS photometry are all relatively similar. We scaled the IRS spectrum by a factor of 1.07 to match the epoch 1 MIPS 24 µ m photometry. With this correction, the IRS spectrum is a good match to the IRAC epoch 1 photometry at 6 and 8 µ m as well. Individual models were scaled to match the 24 µ m flux and 11.3 µ m SiC feature by eye.
The best match occurs for an optical depth of 0.45 in the SiC 11.3 µ mfeature; this model also matches the K -band and H -band photometry for 2MASS. No models do a good job of matching the near infrared, IRS spectral region and the shortest IRAC photometry point at 3.6 µ m. But the 0.45 optical depth model does a good job at matching the epoch 1 Spitzer photometry, IRS spectrum, the 2MASS photometry, and the morphology of the near infrared spectrum (strong veiling and slope). The resulting dust mass-loss rates are a few times 10 -9 M /circledot yr -1 .
The near infrared spectrum was scaled to the 2MASS data point in Figure 11. The actual spectrum clearly exhibits more veiling than the 0.45 optical depth model. This suggests that the near infrared spectrum was taken during an even lower brightness phase of the cycle for object 142 than the 2MASS and IRAC epoch 1 photometry. A model with optical depth of 0.5 (in the SiC feature) is shown in the inset and results in even more veiling, but lower near and mid infrared fluxes. Despite this inconsistency which results from the non-simultaneous observations, the models appear to give a plausible explanation of the XAGB stars.
The models are consistent with the picture that the XAGB stars, including those with the reddest colors, are mainly C-rich objects on the AGB with varying amounts of dust
production in their circumstellar environments. The near infrared and mid infrared spectra of XAGB stars are well matched by C-star core with a dust shell exhibiting modest mass loss. The SiC emission line at 11.3 µ m is well matched by the model, but there are several absorption features in the near and mid infrared that are not included in the model and hence not reproduced. The nature of the absorption at 2.26 µ m remains unknown because our models do not explicitly include the stellar atmosphere beyond a low resolution input SED. There are other absorptions in the IRS spectrum of XAGB stars. Srinivasan et al. (2010) showed how these could arise in the circumstellar envelope from warm gas near the inner dust radius. Detailed modeling of the molecular absorptions will be a subject of future work requiring more sophisticated models and perhaps higher resolution near infrared spectra.
The reader is reminded that objects become C-rich when sufficient carbon-rich material is dredged up as the star evolves along the AGB to produce a C/O ratio greater than unity (Iben & Renzini 1983) at the surface. The SAGE data set spans all ages and masses in the LMC central region. The bulk of the objects are old and low mass, hence they form the main RGB-OAGB-CAGB-XAGB sequence (the latter step inferred from this study and the arguments above). This is a ∼ monotonic sequence in luminosity as expected for a more or less uniformly old population ('single age'). There are more luminous O-rich sources than the XAGB stars, however. This is a result of their younger age and high mass. They evolve quickly before the carbin-rich material can be dredged up to the surface.
Object 94 may be an example of an XAGB C-star with maximal veiling such that even the CO band heads and 2.26 µ m feature are no longer present at the signal-to-noise of our spectrum. Unfortunately, the IRS spectrum for this object was not obtained (an error was made on positioning and another target was observed in its place). Thus we have no mid infrared spectrum to further characterize this source.
## 5.2. O-rich Sources in the Extreme AGB
Four objects (7, 81, 88, and 90) whose colors and magnitudes place them on or near the XAGB are classified as O-rich PAGB stars byWoods et al. (2011); see also Gruendl & Chu (2009) for object 7. The PAGB stage is an extremely short phase near the end of a low or intermediate mass star ( ∼ < 8 M /circledot ) lifetime (see the review by van Winckel 2003). This phase sees the bulk of the envelope of the star expanding away from the stellar 'core , but before ′′ the core ionizes the (detached) envelope as a planetary nebula (PN). The core is expected to show the results of nuclear processing in its atmosphere including s -process elements. Depending on the mass of the star and its evolution above the giant branch, it may produce
a C-rich or O-rich PAGB envelope (i.e. with or without a third dredge up, respectively). The C-rich or O-rich chemistry is evident in the mid infrared spectra of these stars, and this is in part used to determine the PAGB nature (Woods et al. 2011). The shaping of the imminent PN also begins in this phase, and a circumstellar disk may play an important role (Balick & Frank 2002). A detailed look at PAGB stars with Spitzer IRS spectra is given by Matsuura & et al. (2013).
While both C-rich and O-rich PAGB stars were observed by Spitzer as part of the SAGE-Spec sample (Woods et al. 2011), only O-rich PAGB were included in the present sample (the C-rich objects in the SAGE-Spec sample were below the brightness cut-off for the sample observed with SOAR/OSIRIS). Two of the four PAGB stars (objects 88 and 90) are classified as RV Tau variables (a subset of PAGB stars) based upon their MACHO light curves (Alcock et al. 1998).
Object 7 appears to have a normal M-type near infrared spectrum typical of a bright O-rich AGB star or M supergiant (see Figure 7). Indeed the location of this source is at the top or above the main XAGB locus consistent with an extension to redder colors of the bright O-rich AGB stars in Figure 1. Object 7 actually lies on a sequence of dusty mass loss stars noted by Blum et al. (2006); see their Figure 11. The analogous sequence is shown in the [8.0] -[24] vs. [8.0] CMD (Figure 12) and is populated by stars 15, 14, 122, 81, and 7 (and possibly 90).
Object 88 exhibits emission in the CO band head. CO band head emission has been associated with very dense and hot gas in a circumstellar disks in both young stellar objects and evolved stars (Scoville et al. 1983; McGregor et al. 1988). Circumstellar disks are also present in (at least) some PAGB stars (Sahai 1999). Gielen et al. (2009) presented a detailed study of object 88, also known as MACHO 79.5501.13. The object exhibits a double peaked SED with a strong near infrared excess in the K band; this SED is typical of a class of RV Tau stars with disks. The CO emission (see Figure 7) thus corroborates this picture nicely, though as far as we know there are no other O-rich PAGB stars with near infrared spectra showing such a disk signature. Object 90, also an RV Tau star, has a spectrum most similar to O-rich stars (Figures 3 and 4). However there is no strong Ca or Na exhibited in the spectrum. This may be a consequence of the lower signal-to-noise of our spectrum. Object 90 lies in a position below the normal C-star locus in Figure 1.
Object 81 also appears to be an O-rich star based upon its near infrared spectrum. It is very similar in appearance to the spectrum of object 90, but is not a known RV Tau star.
## 5.3. Mass Loss on the XAGB
As noted by Kemper et al. (2010), it appears that most XAGB stars are C-rich. Zijlstra et al. (2006) observed a sample of XAGB stars selected based upon their 2MASS colors. All but two turned out to be C-rich. To our knowledge, this is the most unbiased selection of a large sample of stars for follow-up spectroscopy to date. However, the Zijlstra et al. (2006) sample is only ∼ 30 objects compared to the 1100 XAGB stars in the central 7 ◦ × 7 ◦ of the LMC (Blum et al. 2006, i.e., 3%). Hundreds ( ∼ 300) of mid infrared spectra have been obtained in the LMC for objects in the XAGB region, but they are biased towards brighter magnitudes (Kastner et al. 2008; Buchanan et al. 2009) and/or were selected based on prior information to study targeted classes of objects (Sloan et al. 2008; Kemper et al. 2010). The large numbers of spectra covering most of the XAGB argue that most objects are indeed C-rich, especially if young stellar objects are excluded based on additional imaging and wavelength data (Gruendl & Chu 2009). This is further supported by analysis of multi wavelength photometry. Riebel et al. (2012) fit the photometry of the entire AGB population in the LMC and found that 97 % of their 1340 XAGB candidates were C-rich. The SED-based classification has good agreement ( > 80%) spectroscopic identification. Still, an unbiased spectroscopic assessment of the XAGB stellar content has not been done. See Figure 14 of Woods et al. (2011) for a recent J -[8.0] vs. [8.0] CMD with the combined sample of mid infrared spectroscopic sources over-plotted.
Figure 12 shows the SAGE objects in the [8.0] -[24] vs. [24] color-magnitude plane. The striking difference between this plot and Figure 1 is that the O-rich stars are redder in this new diagram where as the C-rich objects are generally redder in the other (for a given [8.0] magnitude. This is understood for the C-rich stars in J -[8.0] vs. [8.0] diagram because their SEDs are such that the J flux is on the Wien side of the black body curve, while the [8.0] flux is on the Raleigh-Jeans side. On the other hand, the O-rich stars have SEDs with J -band fluxes near the peak of the distribution.
Srinivasan et al. (2010) made a detailed mass loss calculation for object 123 and Sargent et al. (2010) did a similar analysis for two O-rich sources, one bright AGB star (HV5715) and one fainter AGB star (SSTISAGEMC J052206.92-715017.7). We did not observe either of the two O-rich stars analyzed by Sargent et al. (2010); however, objects 2 and 3 of our sample are similar to HV5715 and object 122 is similar to SSTISAGEMC J052206.92-715017.7. The O-rich stars in our sample similar in colors and magnitudes to HV5715 and SSTISAGEMC J052206.92-715017.7 appear to be normal late type stars with O-rich near infrared spectra. Object 123 analyzed by Srinivasan et al. (2010) is an XAGB star as discussed above.
Perhaps remarkably, the dust mass loss rates computed for these three objects and object 142 discussed above which span a large region in the CMD are quite similar. The
values range from 2 to 3 × 10 -9 M /circledot yr -1 . The XAGB stars, object 123 and 142 (C-rich) and SSTISAGEMC J052206.92-715017.7 are relatively low mass. The former are approximately 2 M /circledot and the latter 3 M /circledot based upon position in the Hertzprung-Russell diagram. SeeSrinivasan et al. (see 2010) and Sargent et al. (2010) for details. HV5715 is more massive, approximately 7 M /circledot (Sargent et al. 2010). SSTISAGEMC J052206.92-715017.7 sits on the reddest tip of of a feature attributed to high mass loss stars in the [8.0] -[24] vs. [24] CMD presented by Blum et al. (2006), so it may not be surprising that a relatively low mass 'normal ′′ O-rich AGB star has similar mass loss as an XAGB (C-rich) star.
## 6. Summary
We have presented a sample of 40 SAGE-Spec Spitzer IRS evolved stars in the Large Magellanic Cloud. These objects span the bright end of the color-magnitude diagram (CMD) from the tip of the normal asymptotic giant branch (AGB) to the top of the red supergiant (RSG) branch to the reddest colors on the extreme AGB (XAGB) as defined in the J -[8.0] vs. [8.0] CMD.
The near infrared spectra are an excellent match to the characterization provided by the IRS spectral properties. The C-rich spectral morphology owing to numerous CN and CO lines is unmistakable compared to the O-rich atomic and CO features.
We find the nature of the reddest XAGB stars is best explained by C-rich cores surrounded by dusty mass loss envelopes. We modeled one of these sources (142) with the dust radiative transfer code 2Dust (following Srinivasan et al. 2010) and find plausible models to explain the Spitzer IRS spectrum, IRAC and MIPS photometry as well as the observed near infrared spectral morphology. This latter result stems from adding normal C-star spectra to the 2Dust input SED and then comparing the resulting emergent spectrum which includes the effects of the overlying envelope emission and extinction to the observed near infrared spectrum. These quantitative results are in agreement with conclusions made for less obscured C-stars (Frogel & Hyland 1972) based on the anti-correlation of CO absorption strength and infrared color. We conclude that modest mass loss rates ( ∼ few by 10 -9 M /circledot yr -1 ) reproduce both the near infrared and mid infrared properties on the XAGB.
Four of the present sample of objects were selected as post AGB stars, a short lived phase of stellar evolution between the heavy mass loss of the AGB phase and the planetary nebula phase. These objects in our sample exhibit more O-rich like near infrared spectra. One of the sample (HV 915, object 88) has CO 2.3 µ m band head emission. Circumstellar emission is known in some other RV Tau (i.e. a PAGB subclass) stars; to our knowledge,
this is the first report of a near infrared spectrum with a disk like circumstellar signature for a RV Tau star. This emission and a disk origin are consistent with previous work showing the spectral energy distribution of object 88 is best explained by a disk.
The O-rich PAGB stars are overrepresented in our sample because Kemper et al. (2010) selected the parent sample based on properties of objects they wanted to investigate, not an unbiased sample of objects on the XAGB. Nevertheless, the picture that emerges (see Zijlstra et al. 2006; Kastner et al. 2008; Sloan et al. 2008; Woods et al. 2011) is that most of the XAGB is comprised of otherwise normal C-rich stars (typically a few solar masses) that have produced dusty shells due to mass loss. The O-rich PAGB stars and other sources (planetary nebulae, HII regions, young stellar objects, etc) projected onto the XAGB are interlopers or short lived, unrelated, phases of stellar evolution.
The authors would like to acknowledge Ken Hinkle (NOAO) for useful discussions about CN lines in the near infrared spectrum of evolved stars. The authors thank Masaaki Otsuka (ASIAA) for his assistance to B. Ling with near infrared spectral reductions. FK acknowledges generous support from the Taiwan National Science Council, grant number: NSC100-2112-M-001-023-MY3. The authors thank Jay Frogel for his careful reading of the manuscript as referee and his comments and suggestions which have improved our paper.
## REFERENCES
Alcock, C., Allsman, R. A., Alves, D. R., et al. 1998, AJ, 115, 1921
Aringer, B., Girardi, L., Nowotny, W., Marigo, P., & Lederer, M. T. 2009, A&A, 503, 913
Balick, B., & Frank, A. 2002, ARA&A, 40, 439
Blum, R. D., Mould, J. R., Olsen, K. A., et al. 2006, AJ, 132, 2034
Boyer, M. L., Srinivasan, S., van Loon, J. T., et al. 2011, AJ, 142, 103
Buchanan, C. L., Kastner, J. H., Hrivnak, B. J., & Sahai, R. 2009, AJ, 138, 1597
Cioni, M.-R. L., Girardi, L., Marigo, P., & Habing, H. J. 2006, A&A, 448, 77
D., r. 2014, Private Communication
Depoy, D. L., Atwood, B., Byard, P. L., Frogel, J., & O'Brien, T. P. 1993, in Society of PhotoOptical Instrumentation Engineers (SPIE) Conference Series, Vol. 1946, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, ed. A. M. Fowler, 667-672
- Dewitt, C., Bandyopadhyay, R. M., Eikenberry, S. S., et al. 2013, AJ, in press Fraser, O. J., Hawley, S. L., & Cook, K. H. 2008, AJ, 136, 1242 Fraser, O. J., Hawley, S. L., Cook, K. H., & Keller, S. C. 2005, AJ, 129, 768 Frogel, J. A., & Hyland, A. R. 1972, in Les Spectres des Astres dans l'Infrarouge et les Microondes, 111-120 Gielen, C., van Winckel, H., Reyniers, M., et al. 2009, A&A, 508, 1391 Groenewegen, M. A. T. 2004, A&A, 425, 595 Gruendl, R. A., & Chu, Y.-H. 2009, ApJS, 184, 172 Gruendl, R. A., Chu, Y.-H., Seale, J. P., et al. 2008, ApJ, 688, L9 Iben, Jr., I., & Renzini, A. 1983, ARA&A, 21, 271 Ita, Y., Tanab´, T., Matsunaga, N., et al. 2004, MNRAS, 353, 705 e Kastner, J. H., Thorndike, S. L., Romanczyk, P. A., et al. 2008, AJ, 136, 1221 Kato, D., Nagashima, C., Nagayama, T., et al. 2007, PASJ, 59, 615 Kemper, F., Woods, P. M., Antoniou, V., et al. 2010, PASP, 122, 683 Kleinmann, S. G., & Hall, D. N. B. 1986, ApJS, 62, 501 Matsuura, M., & et al. 2013, to be submitted in the MNRAS McGregor, P. J., Hyland, A. R., & Hillier, D. J. 1988, ApJ, 334, 639 Meixner, M., Gordon, K. D., Indebetouw, R., et al. 2006, AJ, 132, 2268 Nikolaev, S., & Weinberg, M. D. 2000, ApJ, 542, 804 Riebel, D., Srinivasan, S., Sargent, B., & Meixner, M. 2012, ApJ, 753, 71 Sahai, R. 1999, ApJ, 524, L125 Sargent, B. A., Srinivasan, S., Meixner, M., et al. 2010, ApJ, 716, 878 Scoville, N., Kleinmann, S. G., Hall, D. N. B., & Ridgway, S. T. 1983, ApJ, 275, 201 Skrutskie, M. F., Cutri, R. M., Stiening, R., et al. 2006, AJ, 131, 1163
```
- 18
<
< </
</ >
> | +
|
+ }
} )
-
```
Zijlstra, A. A., Matsuura, M., Wood, P. R., et al. 2006, MNRAS, 370, 1961
This preprint was prepared with the AAS LT E X macros v5.2. A
Fig. 1.J -[8.0] vs. [8.0] color-magnitude diagram (CMD) from the SAGE Survey (Meixner et al. 2006; Blum et al. 2006). The CMD uses the 2MASS near infrared J magnitude (Skrutskie et al. 2006). The SAGE-Spec sub-sample of red supergiants, O-rich AGB stars, C-rich AGB stars and XAGB stars observed here are labeled as black for O-rich and red for C-rich stars in the online version of the paper. The various sequences in this diagram are discussed in Blum et al. (2006) and the text. The density of objects has a maximum of 1062 sources per 0.06 mag × 0.06 mag bin. See Figure 9 for a blow up of the asymptotic giant branch (AGB), red supergiant sequence, and extreme AGB with the same labels.
- 20 -
Fig. 2.- Example object spectrum (obj4) with fringe-like features: solid line, object spectrum; dotted line, object corrected by telluric standard; dashed line, fringe-like feature map; solid line, final corrected spectrum. The feature map was constructed by dividing a spectrum of a hot star with a spectrum of itself obtained on the same night. See text.
Fig. 3.- Objects classified as O-rich based upon their near infrared spectra presented herein. Objects are ordered on J -[8.0] from blue to red (top to bottom, respectively). The spectra exhibit normal M-type or O-rich type features such as Na, Ca, and CO absorption; see text and Table 1. The difference between normal O-rich and normal C-rich stars is easily seen by comparing this Figure and Figure 4 to Figure 5, all plotted on the same scale.
Fig. 4.- Same as Figure 3 but for the remainder of the sample of O-rich stars. These are ordered by J -[8.0] from blue to red (top to bottom, respectively) and all with J -[8.0] more red than the objects in Figure 3; see Table 1.
Fig. 5.- Same as Figure 3 but for stars classified herein as C-rich. Positions of the ν = 2-4 and 3-5 band heads of CN are show (Thompson & Schnopper 1970; Wallace & Hinkle 1996).
- 24 -
Fig. 6.- Extreme asymptotic giant branch (XAGB) stars. Such stars have J -[8.0] > 3.1; see text. A feature, possibly due to CN, is seen in most of the objects near 2.26 µ m. The feature may arise in the extended envelopes of the mass losing stars; see text. Object 94 was selected for observing with the Spitzer IRS, but the observation resulted in a different object being observed by Spitzer . Thus no IRS counterpart exists. Object 94 is essentially featureless. The strong fringing in the central part of the spectrum is instrumental; see text.
- 25 -
Fig. 7.- Objects whose near infrared spectra are not typical of their position in the J -[8.0] vs. [8.0] CMD or are not typical asymptotic giant branch (AGB) stars. Object 7 is a normal O-rich or M-type star that sits above the XAGB sequence in Figure 1; see text. Objects 81 and 88 are O-rich post-AGB stars and 88 exhibits CO 2.29 µ m emission. Object 170 is an external (to the Large Magellanic Cloud) galaxy according to Woods et al. (2011), consistent with its position in the CMD (Blum et al. 2006).
- 26 -
Fig. 8.- One of the four most extreme AGB stars in our sample, object 123 ( black line). The continuum rises steeply to the red. Molecular features appear to be 'suppressed ′′ or veiled relative to more normal C-rich stars (see Figure 5). A simple combination of a normal C-rich star (object 82, blue line) and a black body of 400 K in equal proportion ( red line) match the extreme AGB star very well. The CO feature is veiled and the continuum slope is similar. There is a strong feature at 2.26 µ m that does not appear in the normal C-rich star. The line may be due to CN; see text. The simple model does not include reddening, so the blackbody is expected to be cooler than it otherwise would be (see text and Figure 11. This figure appears in color in the online version of the paper.
Fig. 9.- Blow up of the extreme asymptotic giant branch (XAGB) for the J -[8.0] vs. [8.0] color-magnitude diagram shown in Figure 1. Object labels are the same as Figure 1. The O-rich, C-rich, and XAGB sequences are indicated approximately in the Figure; see text. This figure appears in color in the online version of the paper.
Fig. 10.- 2Dust models (Ueta & Meixner 2003) compared to the observed mean photometry for object 142. The data points show the mean MIPS, IRAC, near infrared (2MASS and IRSF), and visible photometry (Zaritsky et al. 2004). The dashed red line is the input model photosphere, and the solid red line is the best fit to the mean photometry. Best fits to the mean photometry have optical depth in the SiC 11.3 µ m line of about 0.3. The range of optical depths represented by the grey lines in the figure is 0.2 - 0.4. These models have core photospheres with T eff = 3000K; see text.
Fig. 11.- 2Dust models (Ueta & Meixner 2003) and observed fluxes for object 142. Red data points and spectra are compared to the models. Black data points are for reference. The models were constructed with an input photosphere that incorporates the near-infrared spectrum of the normal C-star object 82 (see the Figure lower inset which shows the combined emergent spectrum in the near infrared). The squares show the MIPS, IRAC, 2MASS ( red ), and IRSF ( black ) photometry. The red and black MIPS/IRAC photometry correspond to SAGE epoch 1 and 2 data respectively. The red curves show the near infrared and SAGESpec (IRS) spectra. The MIPS, IRAC, near infrared photometry (2MASS and IRSF), near infrared spectrum and IRS spectrum were all obtained at different times, and thus do not uniquely constrain the models. The best matches to object 142 epoch 1 photometry, 2MASS photometry, and the IRS and near infrared spectra result from models with optical depth near 0.45 in the SiC 11.3 µ mfeature (Figure upper inset). The IRS spectrum has been shifted slightly to match the 24 µ m SAGE epoch 1 photometry and the near infrared spectrum was set to match the 2MASS K s photometry (2.159 µ m). See text.
Fig. 12.- [8.0] -[24] vs. [8.0] color-magnitude diagram (CMD) from the SAGE Survey (Meixner et al. 2006). The SAGE-Spec sub-sample of objects observed here are labeled as black for O-rich and red for C-rich stars. The density of objects has a maximum of 11 sources per 0.02 mag × 0.02 mag bin. This figure appears in color in the online version of the paper.
Table 1. SAGE-Spec Targets
| Object | K | J - [8.0] | [8.0] | [8.0]-[24] | [24] | SSID a | Other Name b | CMD Type c | NIR d | PSC e |
|----------|-------|-------------|---------|--------------|--------|----------|--------------------------------|--------------|------------|---------|
| 28 | 8.83 | 1.5 | 8.37 | 1.55 | 6.82 | 129 | NGC 2004 Wes 6-14 | RSG | O star | RSG |
| 108 | 8.41 | 1.55 | 7.87 | 1.06 | 6.8 | 122 | HV 5879 | RSG | O star | RSG |
| 40 | 10.23 | 1.71 | 9.79 | 1.26 | 8.54 | 143 | SSTISAGEMC J053343.98-705901.9 | B O* | O star | OAGB |
| 4 | 9.88 | 1.74 | 9.42 | 0.21 | 9.21 | 178 | SSTISAGEMC J054314.12-703835.1 | B O* | O star | OAGB |
| 5 | 10.68 | 1.79 | 10.11 | 1.08 | 9.02 | 197 | SSTISAGEMC J060053.62-680038.8 | B O* | O star | OAGB |
| 10 | 10.99 | 1.83 | 10.44 | 0.28 | 10.16 | 13 | SSTISAGEMC J045309.39-681710.8 | O* | O star | OAGB |
| 2 | 9.52 | 1.84 | 8.96 | 0.29 | 8.68 | 173 | WHO G 494 | B O* | O star | OAGB |
| 15 | 11.28 | 1.91 | 10.62 | 1.4 | 9.22 | 89 | SSTISAGEMC J052014.24-702931.0 | N O* | O star | OAGB |
| 25 | 8.74 | 1.95 | 7.89 | 1.53 | 6.36 | 170 | NGC 2100 Robb 4 | RSG | O star | RSG |
| 19 | 7.74 | 2.1 | 6.84 | 1.97 | 4.87 | 27 | MSX LMC 1271 | RSG | O star | RSG |
| 14 | 11.08 | 2.17 | 10.19 | 1.84 | 8.35 | 91 | SSTISAGEMC J052051.83-693407.6 | O* | O star | OAGB |
| 122 | 10.87 | 2.23 | 9.88 | 1.83 | 8.05 | 159 | SSTISAGEMC J053945.40-665809.4 | O* | O star | OAGB |
| 107 | 8.34 | 2.25 | 7.29 | 2.3 | 4.99 | 147 | HV 2700 | RSG | O star | RSG |
| 3 | 9.53 | 2.28 | 8.54 | 1.67 | 6.87 | 61 | SSTISAGEMC J051059.07-685613.7 | O* | O star | OAGB |
| 27 | 8.82 | 2.3 | 7.73 | 1.66 | 6.06 | 172 | W61 6-57 | RSG | O star | RSG |
| 21 | 8.05 | 2.38 | 6.88 | 1.65 | 5.23 | 135 | NGC 2011 SAGE IRS 1 | RSG | O star | RSG |
| 82 | 11.13 | 2.44 | 10.15 | 0.9 | 9.26 | 132 | KDM 4665 | C* | C star | CAGB |
| 73 | 10.4 | 2.51 | 9.34 | 0.48 | 8.86 | 120 | OGLE 052825.96-694647.4 | C* | C star | CAGB |
| 76 | 10.45 | 2.66 | 9.51 | 0.21 | 9.29 | 136 | KDM 4718 | C* | C star | CAGB |
| 1 | 8.78 | 2.69 | 7.44 | 2.17 | 5.28 | 165 | MSX LMC 947 | O* | O star | OAGB |
| 44 | 10.43 | 2.84 | 9.43 | 0.35 | 9.08 | 83 | SSTISAGEMC J051832.64-692525.5 | C* | C star | CAGB |
| 79 | 10.69 | 2.91 | 9.49 | -0.02 | 9.51 | 23 | KDM 1238 | C* | C star | CAGB |
| 84 | 11.42 | 2.94 | 10.05 | -0.07 | 10.12 | 57 | KDM 2187 | C* | C star | CAGB |
| 71 | 9.95 | 3 | 8.67 | 0.4 | 8.27 | 33 | LMC BM 12-14 | C* | C star | CAGB |
| 80 | 11.01 | 3.1 | 9.78 | -0.22 | 10 | 49 | SSTISAGEMC J050629.61-685534.9 | C* | C star | CAGB |
| 69 | 9.69 | 3.1 | 8.38 | 0.49 | 7.9 | 31 | KDM 1691 | B C* | C star | CAGB |
| 90 | 12.18 | 3.16 | 10.08 | 2.75 | 7.33 | 157 | HV 12631 | Below C* | O rich-- e | RV Tau |
| 37 | 10.03 | 3.27 | 8.75 | 0.3 | 8.45 | 12 | KDM 764 | C* | C star | CAGB |
- 31 -
Table 1-Continued
| Object | K | J - [8.0] | [8.0] | [8.0]-[24] | [24] | SSID a | Other Name b | CMD Type c | NIR d | PSC e |
|----------|-------|-------------|---------|--------------|--------|----------|--------------------------------|--------------|-----------------------------|-----------|
| 70 | 9.78 | 3.39 | 8.21 | -0.04 | 8.24 | 30 | KDM 1656 | B C* | C star | CAGB |
| 77 | 10.49 | 4.02 | 8.62 | 0.35 | 8.27 | 45 | SSTISAGEMC J050607.50-714148.4 | E XAGB | C star, CN? | CAGB |
| 7 | 10.82 | 4.11 | 7.79 | 3.45 | 4.34 | 177 | SSTISAGE1C J054310.86-672728.0 | Above XAGB | O star | OPAGB |
| 81 | 11.09 | 4.12 | 8.34 | 2.97 | 5.37 | 56 | SSTISAGEMC J050830.51-692237.4 | E XAGB | O rich? e | OPAGB |
| 38 | 10.06 | 4.97 | 7.67 | 0.12 | 7.55 | 51 | SHV 0507252-690238 | Mid XAGB | C star, CN? | CAGB |
| 94 | 9.91 | 5.23 | 7.24 | 0.45 | 6.79 | 145 | wrong star in IRS slit | Mid XAGB | C star | CAGB |
| 88 | 12.09 | 5.36 | 7.73 | 1.76 | 5.97 | 73 | HV 915 | Mid XAGB | CO Emission | O* RV Tau |
| 87 | 11.89 | 5.98 | 7.38 | 1.42 | 5.96 | 94 | HV942 | L XAGB | no source detected | R CrB |
| 170 | 14.16 | 6.02 | 9.87 | 3.06 | 6.81 | 193 | SSTISAGEMC J055143.27-684543.0 | galaxy | featureless | galaxy |
| 123 | 10.89 | 6.76 | 7.4 | 0.69 | 6.72 | 66 | OGLE J051306.52-690946.4 | L XAGB | C star, 'featureless', CN? | CAGB |
| 142 | 11.78 | 6.97 | 8.15 | 0.57 | 7.58 | 103 | SSTISAGEMC J052405.31-681802.5 | L XAGB | - C star, 'featureless' CN? | CAGB |
| 50 | 11.35 | 7.86 | 7.49 | 1.16 | 6.33 | 60 | SSTISAGEMC J051028.27-684431.2 | L XAGB | 32 C star, 'featureless' | CAGB |
| 144 | 11.81 | 8.24 | 7.55 | 0.83 | 6.72 | 98 | OGLE J052242.09-691526.2 | L XAGB | - C star, 'featureless' | CAGB |
a SSID is the Spitzer observation ID given in Kemper et al. (2010). See also Woods et al. (2011) for details of the object's Spitzer IRS spectrum and classification.
b Other names adopted from Woods et al. (2011).
c The expected object type based upon the position in the J -[8.0] vs. [8.0] color-magnitude diagram (CMD); see Figure 1. The type of object for a given position in the CMD is defined in Blum et al. (2006). O* = O-rich asymptotic giant branch (AGB). RSG = red supergiant. B O* = bright AGB (between the AGB and RSG sequences). B C* = bright C-star; XAGB = extreme AGB. E, Mid, and L = early (or blue), middle and late (or red) XAGB which roughly divides the XAGB sequence in three sections.
d Object type based upon morphology of the near infrared spectra presented herein. See text and Figures 3, 4, 5, 6, and 7. CN? = possible CN feature seen only in the XAGB stars; see text.
e Point Source Catalog type or characteristics based on Spitzer IRS spectra (Woods et al. 2011).
f Objects 90 and 81 appear to have a O-rich near infrared spectrum, but the signal to noise is not high (see Figure 7) and § 5.2.
|
10.1088/0004-6256/149/2/87
|
[
"R. D. Blum",
"S. Srinivasan",
"F. Kemper",
"B. Ling",
"K. Volk"
] |
2014-08-05T18:51:33+00:00
|
2014-08-05T18:51:33+00:00
|
[
"astro-ph.SR"
] |
Spitzer SAGE-Spec: Near Infrared Spectroscopy, Dust Shells, and Cool Envelopes in Extreme Large Magellanic Cloud AGB Stars
|
K-band spectra are presented for a sample of 39 Spitzer IRS SAGE-Spec sources
in the Large Magellanic Cloud. The spectra exhibit characteristics in very good
agreement with their positions in the near infrared - Spitzer color-magnitude
diagrams and their properties as deduced from the Spitzer IRS spectra.
Specifically, the near infrared spectra show strong atomic and molecular
features representative of oxygen-rich and carbon-rich asymptotic giant branch
stars, respectively. A small subset of stars were chosen from the luminous and
red extreme "tip" of the color magnitude diagram. These objects have properties
consistent with dusty envelopes but also cool, carbon-rich "stellar" cores.
Modest amounts of dust mass loss combine with the stellar spectral energy
distribution to make these objects appear extreme in their near infrared and
mid infrared colors. One object in our sample, HV 915, a known post asymptotic
giant branch star of the RV Tau type exhibits CO 2.3 micron band head emission
consistent with previous work that demonstrates the object has a circumstellar
disk.
|
1408.1068v1
|
## The Size of Software Projects Developed by Mexican Companies
## J. Aguilar , M. Sánchez , C. Fernández , E. Rocha , D. Martínez , and J. Figueroa 1 2 2 2 2 2
1 Departament of Engineering, Universidad Popular Autónoma del Estado de Puebla, Puebla, Pue., México 2 Institute of Computer Engineering, Universidad Tecnológica de la Mixteca , Huajuapan, Oax., México
Abstract -Currently, most software projects around the world are small rather than large. Despite this, there are more methodologies, tools, frameworks, processes, and so on, for developing and managing large software projects than for small ones. Small software projects are important because they generate considerable resources. For example: apps (small mobile applications) generate around $25 billion dollars of revenue. This paper shows our findings regarding the size of the projects built by Mexican software development companies. We surveyed 107 Mexican companies and found that 92% of their developed projects are micro and small, and 8% are medium or large. In addition, according to our research, 84.1% of companies in Mexico are micro or small businesses.
Keywords: Small projects, Software size classification, Companies size classification.
## 1 Introduction
There are many examples of tools, methodologies or, processes oriented to medium or large projects such as the following: CMMI[13], COCOMO [14], EPM [15], among others. These tools, methodologies and processes are highly accepted around the world. We need the same global acceptance for small software projects, tools, methodologies and processes, because micro and small software projects generate an important economic impact. For example, apps are mainly small projects, and the app stores run by Apple Inc, Google Inc, and so on, offer more than 700,000 apps each and, generate around $25 billion dollars of global revenue [7].
First of all, we need to identify the size of software projects. There are several factors, among them: i) Project development cost, ii) Number of people required to develop the software, and iii) Amount of software to be produced [6]. For example, a small software project has these factors: i) Cost between USD 7,500.00 and 192,000.00; ii) Team of less than 4 persons; and, iii) Size between 9 KLOCs and 38 KLOCs. This is our personal viewpoint about small software projects, although each author has his or her own viewpoint about software size classification.
In this paper we will show the findings of our research regarding: a) Size of software projects being developed in Mexico, and b) Company size in Mexico. We want to answer the following questions: What is the size of software projects developed by micro and small software companies in Mexico? How is software size to be classified? These questions are important to answer because there are currently many software process models oriented to large and medium projects, but micro and small projects have perhaps been forgotten.
## 2 Background
This section shows the main elements of our research.
## 2.1 Software Size Classification
First of all, we are going to show different viewpoints regarding software size classification. Fred Brooks [1] said that there are different kinds of software: i) Programs complete in themselves, ready to be run by the author on the system on which it was developed, ii) A programming product that can be run, tested, repaired and extended by anybody, iii) A Programming System, which is a collection of interacting programs, and iv) Programming Systems Product, which costs nine times as much as the other three kinds. Another author, Watts Humphrey [2], said that there are five stages of Software Product Size: Stage 0, very small program elements, written by programmers alone. Stage 1, small programs, or modules, designed, implemented, and tested by programmers alone; these programs typically range in size from only a few dozen to several hundred LOCs (Lines of Code). Stage 2, larger programs, or components, that typically involve teams of developers who develop and integrate multiple stage-1 modules into larger stage-2 component programs. Stage 3, very large projects that involve multiple teams controlled and directed by a central project management. Finally, Stage-4, massive multi-systems that involve many autonomous or loosely federated projects. The next author, Solvita Berzisa [3], said that project size is described with four attributes where three attributes -team size, budget and duration- are scalar values and such values should be grouped into intervals. The team size values have been divided into two intervals: less than seven and seven and more. Analyzed budget attribute values have been divided into five intervals: less than 10 000 USD, 10,000 to 50,000 USD, 50,000 to
100,000 USD, 100,000 to 500,000 USD, and more that 500,000 USD. Project duration values have been divided into four intervals: less than six months, six months to a year, from a year to two years and more than two years. These authors did their own software size classification.
Besides these viewpoints, The International Software Benchmarking Standards Group (ISBSG) is the global and independent source of data and analysis for the IT industry. There are 5,052 projects included in the ISBSG in June 1 2009, of which 76.1% are micro and small projects, from 0 to 399 function points, which corresponds to approximately 31,920 LOCs. We applied Backfiring[4] to convert function points to lines of code. Smaller projects are more common in this repository. The projects have been submitted from 24 countries. Projects of different size ranges have different key characteristics [5][8][9]. For example, we can use table 2.1 to find the ideal duration and staffing ranges for a project, based on its estimated effort.
Table 2.1 Time and effort in software projects according to their size.
| Project Size | Effort. Work hours of effort . | Months Duration | Ideal staffing |
|----------------|----------------------------------|-------------------|------------------------|
| Small | 8-360 | 0-3 | 1 person |
| Medium | 361-3600 | 3-6 | 3-7 persons |
| Large | 3601-24 000 | 6-12 | 7-24 full-time persons |
A software project can be considered to be small if one or more of the following criteria apply:
- a) Less than two man-years of development effort is needed.
- b) A single development team of five people or less is required.
- c) The amount of source code is less than 10 KLOCs.
The size of a particular project can vary greatly depending on the language chosen. For instance a 200,000 line Perl project generally has the functionality of a 1,200,000 line C project. This is from the estimate in Code Complete that states that Perl usually requires 1/6 the number of lines to do the same task as C. However, different size ranges entail different issues that involve major organizational differences. Specifically:
- 1. Small - reasonable size for one person to produce in a reasonable time.
1 The global and independent source of data and analysis for the IT industry. www.isbsg.org
- 2. Medium - reasonable size for a small team (max 8; beyond that size, communication issues arise that drop productivity and require reorganization to handle) to develop and entirely understand.
- 3. Large - A large team (significantly > 20 - note that the average team of 20 is about as productive as a team of 5-8 due to organizational issues).
- 4. Huge - The project is large enough that specialized tools are probably needed to help with project navigation so that team members can figure out what they need to understand to work on it.
As can be seen, there are a lot of viewpoints regarding software size classification. In this paper we try to take into account all these opinions and offer a proposal.
## 2.2 Economic impact of small software projects.
Basically, an app is a small software project. For instance, the average size of an Apple app is around 10 KLOCs [12]. For this paper, then, we consider an app as a small software project. Thus, we can see that there are millions of mobile devices sold around the world. For example, in 2012, there were 446 million Android devices, 199 million Apple devices and, 17 million Windows devices sold. Furthermore, there are millions of apps available [11]. The apps market was about USD 25 billion which is expected to be USD 155 billion in 2017 [10]. Obviously, the economic impact of small software projects is very important.
## 3 Our research
This section describes in detail our research; we only show our results about what kind of software are mainly implemented by Mexican software development companies. Section 3.1 will explain the methodology used for our work. Section 3.2 describes the results of our research, explaining the software size classification, company size classification and the size of Mexican software projects.
## 3.1 Methodology description
We began this research by doing a systematic search of the literature related to our project. So far, we have not found similar studies at least for Mexican companies. It is important to mention that this study is part of a larger project in which we want to identify good techniques and practices for developing small projects. The problem is that there are not enough tools, processes, methodologies and so on for small software projects. This problem was identified by some of our researchers, because they had been working directly in software development companies.
For this project, we wanted to have a better knowledge of Mexican companies working in software development projects and to gather data about characteristics of their small
software projects. In order to get this data, we contacted software companies from Mexico and offered a free SCRUM course to members from participating companies. We received answers from 107 Mexican companies. The results of our research are detailed in section 3.2.
## 3.2 Results
We analyzed data gathered from our survey, where 107 Mexican companies were asked about: a) Their size, and b) The size of the projects they developed. We found that 92% of their projects were micro and small projects, and 8% were medium or large projects. Additionally, we found 84.1% are micro and small companies and they barely use processes, methodologies or tools for developing and managing their small projects.
The survey was mainly focused on the size of projects being developed in Mexico and company size. We wanted to answer the question, what is the size of software projects being developed by micro and small software companies in Mexico?.
## 3.2.1 Kind of Projects Developed
With our research, we could answer this question: Q1. Since your company was founded, what percentage of projects (micro, small, medium, large) has your company developed? The companies were able to answer this question because we gave them our Project Size Classification Table (Table 3.2.1) in which they saw our classification of the sizes of different projects.
Table 3.2.1 Project Size classification
| Project Size | Features |
|----------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Micro | x Size (Lines of source code): from 1,600 to 9,600 x Development time (months): from 1 to3 x Development time (hours): from 160 to 960 x Team size (members): from 1 to 2 x Cost (USD): 1,200 to 4,800* |
| Small | x Size (Lines of source code): from 9,601 to 38,400 x Development time (months): from 3 to 6 x Development time (hours): from 961 to 3,840 x Team size (members): from 2 to 4 x Cost (USD): 7,501 to 192,000* |
| Medium | x Size (Lines of source code): from 38,401 to 960,000 x Development time (months): from 7 to 60 |
| | x Development time (hours): from 3,841 to 96,000 x Team size (members): from 5 to 10 x Cost (USD): from 192,050 to 4,800,00.00* |
|-----------|-----------------------------------------------------------------------------------------------------------------------------------|
| Large/Big | x Size (Lines of source code): from 960,000 to x Development time (months): from 61 to x Development time (hours): from 96,001 to x Team size (mem bers): from 11 to x Cost (USD): from 4,800,00.00 to * |
| Note | *The cost is different in each country. We took USD 50.00 per man-hour. |
Our data analysis showed us than Mexico is a country where mainly micro & small software projects are developed, consisting of 92% of software projects (Fig. 3.2.2.1). Medium & large software projects represent a much smaller percentage-- only 8%. We can see these results in figure 3.2.2.1.
Figure 3.2.2.1 Software projects developed in Mexico
We could also answer another question: Q2: What was the cost of your small projects?
Figure 3.2.2.2 Small software projects cost
The main activity of Mexican companies is to develop small projects, and the price of these small projects was mainly very low. 72% cost less than $3,850 USD ($ 50,000.00 Mexican pesos) (Fig. 3.2.2.2). Thus, we can infer: Mexican software development companies have developed small projects and its costs have been very low. Therefore, we need to give them mainly information to develop and manage micro & small projects.
Additionally, we posed this question: Q3: How long was needed to develop a small project?
Figure 3.2.2.3 Time to develop small software projects
We can see that Mexican software development companies need only a few weeks to develop a small project. 64% of projects required only one month or less to finish (Fig. 3.2.2.3). The question Q2 showed us the low cost to develop small projects. It may be that the reason for this low cost is the relatively little time necessary to develop this kind of project.
## 3.2.2 Company Size Classification.
We could answer another question: Q4. How many people work in your company?. The companies had four options: a) 1-5, b) 6-10, c) 11-50, d) 51 or more. When we analyzed the answers, we found that 84.1% were micro & small companies and 15.9% were medium and big (large) (table 3.2.2).
Table 3.2.2 Q2, Mexican Companies: Size classification.
| Size | Employees | Number Mexican companies. | Percentage |
|----------------------|-------------|-----------------------------|--------------|
| Micro | 1-10 | 62 | 57.9% |
| Small | 11-50 | 28 | 26.2% |
| Medium & Big (Large) | 51 or more | 17 | 15.9% |
| Total | 107 | 107 | 100% |
This table shows us that in Mexico there are more micro and small companies than medium and big or large (Fig. 3.2.2.4). Thus, we need to focus on developing tools, processes, methodologies, model life cycles, and so on, for the micro and small sector.
Figure 3.2.2.4 Mexican companies' size
## 4 Future Work
With our findings, we have established the situation regarding the size of software developed and companies' size in Mexico. Based on these results, we will need to research the following: What kind of tools, methodologies and processes are the companies currently using? Which are most useful for Mexican companies? What must a Mexican company do to successfully adopt a tool, methodology or process to support small projects' administration and development? Perhaps, we will need to make some proposals for new tools, new process or new methodologies focusing mainly on our Mexican context.
## 5 Conclusion
We have found useful information to develop a strategic plan to help Mexican companies develop quality small software projects. For example, we have found: a) Most of the companies surveyed have developed a considerably high number of small projects. We can thus infer that the main activity of these companies is to develop small software projects. b) The cost to develop small software projects is very low; this could be a problem when the companies need economic resources to buy software or hardware for their operations. c) The time required to develop small projects is only one month or less, although 36% of small projects needed more than one month to be finished. d) Mexico is a country where 84% of software development countries are micro or small. We have not shown what percentage of small projects fail, but in this research we can see: e) Failure percentage in small projects is directly related to softwarecompany size; the reasons for increasing failure in small projects are not clear. Finally we can say: In Mexico, mainly, software development companies need: tools, processes and methodologies to develop small software projects. The small
software projects have an important economic impact. It is very important to be aware of the size of companies in our country or region, and to know the size of the projects developed by them, in order to make plans to support them.
## 6 References
- [1] Frederick P. Brooks, Jr. ' The Mythical Man-Month : ' Essays on Software Engineering, Addison-Wesley, ISBN 0201-83595-9, 1995.
- [2] Watts S. Humphrey, ' A Discipline for Software Engineering , SEI Series in Software Engineering, Addison ' Wesley, ISBN 0201546108, 2005.
- [3] Solvita Berzisa, ' Project Management Knowledge Retrieval: Project Classification ' , Proceedings of the 8th International Scientific and Practical Conference. Vol II, ISBN 978-9984-44-071-2. 2011.
- [4] Capers Jones, ' Backfiring: converting lines of code to function points ' , Computer, Vol. 28, no. 11, pp 87-88, November 1995.
- [5] Stacy Goff, ' The Successful Project Profile: A monograph . Associates, Inc, the ProjectExperts, 2010 '
- [6] ESA BSSC. ' Guide to applying the ESA software Engineering Standards to small Software Projects ' . (1996).
- [7] Jessica E. Lessin and Spencer E. Ante, 'Apps Rocket Toward $25 billion in Sales, The Wall Street Journal . http://online.wsj.com/article/ SB10001424127887323293704578334401534217878.html (2013).
- [8] Mark C. Paulk. 'Using the Software CMM in Small Organizations'. Proceedings of the Pacific Northwest Software Quality Conference and the Eight International Conference on Software Quality , 1998.
- [9] Claude Y. Laporte, Frédéric Chevalier and Jean-Claude Maurice 'Improving project management for small projects'. ISO Focus+, www.iso.org/isofocus. 2013.
- [10] ' Transparency Market Research, Smartphone applications Market -Global Industry Size, Share, Trends, Analysis and Forecasts 2012-2018 ' .
- [11] ' Mobile Statistics', http://www.mobilestatistics.com, January 2014.
[12] ' Information is Beautiful', http://www.informationisbeautiful.net/visualizations/millionlines-of-code/ January 2014.
- [13] James Persse, 'Project Management Success with CMMI', Prentice Hall Computer, 2007.
- [14] Barry W. Boehm, Chris Abts, A. Winsor Brown, Sunita Chulani, Bradford K. Clark, Ellis Horowits, Ray Madachy, Donald Reifer, Bert Steece, 'Softwre Cost Estimation With COCOM O II', Prentice Hall PTR, 2000.
- [15] Value Prism Consulting, 'Microsoft Office Enterprise Project Management (EPM) Solution Delivers Strong Business Value', 2009.
| null |
[
"Jorge Aguilar",
"Moises Sanchez",
"Carlos Fernandez-y-Fernandez",
"Everth Rocha",
"David Martinez",
"Jose Figueroa"
] |
2014-08-05T18:53:59+00:00
|
2014-08-05T18:53:59+00:00
|
[
"cs.SE"
] |
The Size of Software Projects Developed by Mexican Companies
|
Currently, most software projects around the world are small rather than
large. Despite this, there are more methodologies, tools, frameworks,
processes, and so on, for developing and managing large software projects than
for small ones. Small software projects are important because they generate
considerable resources. For example: apps (small mobile applications) generate
around $25 billion dollars of revenue. This paper shows our findings regarding
the size of the projects built by Mexican software development companies. We
surveyed 107 Mexican companies and found that 92% of their developed projects
are micro and small, and 8% are medium or large. In addition, according to our
research, 84.1% of companies in Mexico are micro or small businesses.
|
1408.1069v1
|
## Electrostatic Binding and Hydrophobic Collapse of Peptide-Nucleic Acid Aggregates Quantified Using Force Spectroscopy
Joan Camunas-Soler, † ‡ , Silvia Frutos, † ‡ , Cristiano V. Bizarro, † ‡ @ , , Sara de Lorenzo, † ‡ , Maria Eugenia Fuentes-Perez, ¶ Roland Ramsch, § ‡ , Susana Vilchez, § ‡ , Conxita Solans, § ‡ , Fernando Moreno-Herrero, ¶ Fernando Albericio, ‖ , ‡ , ⊥ , # Ramón Eritja, § , ‖ , ‡ Ernest Giralt, ‖ , ‡ Sukhendu B. Dev, ‖ and Felix Ritort ∗ , † ‡ ,
Small Biosystems Lab, Departament de Física Fonamental, Universitat de Barcelona, Avda. Diagonal 647, 08028 Barcelona, Spain, CIBER de Bioingeniería, Biomateriales y Nanomedicina, Instituto de Salud Carlos III, Madrid, Spain, Centro Nacional de Biotecnología, CSIC, 28049 Cantoblanco, Madrid, Spain, Institut de Química Avançada de Catalunya, Consejo Superior de Investigaciones Científicas (IQAC-CSIC), 08034 Barcelona, Spain, Institute for Research in Biomedicine (IRB Barcelona), Barcelona Science Park, Baldiri Reixac 10-12, 08028 Barcelona, Spain, Department of Organic Chemistry, University of Barcelona, 08028. Barcelona, Spain, and School of Chemsitry and Physics, University of KwaZulu-Natal, 4001-Durban, South Africa
E-mail: [email protected]
## Abstract
Knowledge of the mechanisms of interaction between self-aggregating peptides and nucleic acids or other polyanions is key to the understanding of many aggregation processes underlying several human diseases ( e.g. Alzheimer's and Parkinson's diseases). Determining the affinity and kinetic steps of such interactions is challenging due to the competition between hydrophobic self-aggregating forces and electrostatic binding forces. Kahalalide F (KF) is an anticancer hydrophobic peptide which contains a single positive charge that confers strong aggregative properties with polyanions. This makes KF an ideal model to elucidate the mechanisms by which self-aggregation competes with binding to a strongly charged polyelectrolyte such as DNA. We use optical tweezers to apply mechanical forces to single DNA molecules and show that KF and DNA interact in a two-step kinetic process promoted by the electrostatic binding of DNA to the aggregate surface followed by the stabilization of the complex due to hydrophobic interactions. From the measured pulling curves we determine the spectrum of binding affinities, kinetic barriers and lengths of DNA segments sequestered within the KF-DNA complex. We find there is a capture distance beyond which the complex collapses into compact aggregates stabilized by strong hydrophobic forces, and discuss how the bending rigidity of the nucleic acid affects such process. We hypothesize that within an in vivo context, the enhanced electrostatic interaction of KF due to its aggregation might mediate the binding to other polyanions. The proposed methodology should be useful to quantitatively characterize other compounds or proteins in which the formation of aggregates is relevant.
∗ To whom correspondence should be addressed
† Small Biosystems Lab, Departament de Física Fonamental, Universitat de Barcelona, Avda. Diagonal 647, 08028 Barcelona, Spain
‡ ¶
CIBER de Bioingeniería, Biomateriales y Nanomedicina, Instituto de Salud Carlos III, Madrid, Spain
Centro Nacional de Biotecnología, CSIC, 28049 Cantoblanco, Madrid, Spain
§ Institut de Química Avançada de Catalunya, Consejo Superior de Investigaciones Científicas (IQAC-CSIC), 08034 Barcelona, Spain
‖ Institute for Research in Biomedicine (IRB Barcelona), Barcelona Science Park, Baldiri Reixac 10-12, 08028 Barcelona, Spain
⊥ Department of Organic Chemistry, University of Barcelona, 08028. Barcelona, Spain
# School of Chemsitry and Physics, University of KwaZulu-Natal, 4001-Durban, South Africa
@ Current Address: Centro de Pesquisas em Biologia Molecular e Funcional/PUCRS Avenida Ipiranga 6681, Tecnopuc, Partenon 90619-900, Porto Alegre, RS, Brazil
[keywords: DNA condensation | aggregation | single-molecule | force spectroscopy | DNApeptide complex | optical tweezers]
Understanding the driving forces by which self-aggregating molecules bind their targets inside the cell is of the utmost importance to elucidate their mechanisms of action. 1-7 Self-aggregating peptides bearing a definite charge are able to establish strong electrostatic interactions with oppositely charged polymers. In particular, recent bulk studies have shown that amyloid peptides with positive charges ( e.g. human lysozyme, A b 40, a -synuclein, histidine-leucine peptides) have a strong binding affinity to negatively charged polymers ( e.g. nucleic acids, polysaccharides, polylysines) stimulating aggregation and fibril formation. 1,8-13 Such ubiquituous interaction has triggered discussion on its connection to neurodegenerative diseases, and on the hypothetical role of aggregating peptides as scaffolds for polynucleotide assembly in early prebiotic life. 1,14 Despite their prevalence, and even if much progress has happened within the last few years, the complex and nonspecific nature of these interactions makes difficult to quantitatively determine key parameters such as their binding affinities and the kinetic steps during the interaction process.
Indeed, a full characterization of the interaction between hydrophobic peptides and polyanions is challenging due to the competition between peptide-peptide self-aggregating interactions and peptide-substrate electrostatic binding forces, as well as due to the transient and heterogenous nature of the formed complexes. 11 An excellent model to address such questions is the anticancer drug Kahalalide F (KF), a 14-residue cyclic depsipeptide originally isolated from the Hawaiian mollusk Elysia rufescens . 15 KF is a low solubility compound with a highly hydrophobic structure and a single positive residue (L-ornithine) (Figure 1), 16 which exhibits a potent cytotoxic activity against several tumor cell lines 17-19 causing the disruption of the plasma membrane due to the accumulation of peptide aggregates. 20 Although KF is a molecule with a strong tendency to aggregate it also has a single positive charge capable of establishing electrostatic interactions with negatively charged substrates such as DNA. Here we use optical tweezers, AFM imaging, and dynamic light scattering (DLS) to fully characterize the interaction of KF aggregates binding to DNA. Although the formation of KF-DNA complexes can be directly observed in AFM and DLS measurements, only force spectroscopy methods make possible to quantitatively determine the driving thermodynamic forces. In addition, by applying mechanical force to the ends of the
DNAit is possible to control and gain insight into the kinetic steps involved in the formation of the complex.
Figure 1: Kahalalide F structure.
We have found that binding of DNA to KF occurs in two kinetic steps: First, DNA binds KF particles due to the electrostatic attraction between the negatively charged DNA and the positively charged groups exposed on their surface (L-Orn). Electrostatic binding compacts DNA by sequestering DNA segments along the surface of the aggregate in a way reminiscent of a condensation process. This is followed by a slow remodeling of hydrophobic contacts and the irreversible entrapping of DNA within the KF-DNA complex. Modeling of the stretching curves yielded characteristic parameters of the interaction, such as the average length of DNA segments electrostatically bound to the aggregate, their affinity of binding and the barrier to unpeel them. DNA unzipping experiments show that KF also forms complexes with ssDNA. However the different mechanical (bending rigidity) and chemical (hydrophobicity) properties of the polyelectrolyte determine the kinetics of formation of the complex.
## Results
## KF compacts dsDNA
To study how KF binds DNA we stretched a single half l -DNA (24-kb) in the presence of KF in the optical tweezers set-up (Figure 2 a , Inset). First a DNA molecule was tethered between two beads and its elastic properties measured using the Worm-Like Chain (WLC) model (see
Methods). Next the DNA molecule was rinsed with 40 m M KF while it was maintained at an end-to-end distance of 6 m m. In this configuration the DNA can explore bended conformations due to thermal fluctuations, as the force remains below 0.4 pN at this extension. The flow was temporarily stopped after 5, 15 and 30 min in order to record a series of force-extension curves (Figure 2 a ). We collected measurements for at least ten different molecules finding a reproducible pattern (Figure S1, Supporting Information).
After flowing KF for 5 min, DNA maintained at low tension was compacted by the peptide. In order to stretch the compacted molecule, the KF-DNA complex must be unraveled, and therefore a sawtooth pattern with many force rips was observed (Figure 2 a , purple). This suggests that KF behaves as a DNA condensing agent, inducing kinks and loops on the DNA. The relaxation curves however, remained similar to those obtained for naked DNA (Figure 2 a , black) indicating that DNA compaction took place after the extension of the molecule was reduced. The relaxation curves were well described by the WLC model and showed a decrease of 25% in the persistence length (Figure 2 b , Inset). In experiments where KF was flowed for 5 or 15 min, the whole contour length of the DNA could be recovered after pulling up to 40 pN (Figure 2 c ). This reduction of the persistence length is likely due to the positively charged L-Orn residue that decreases the self-repulsion of the DNA phosphate-backbone.
However, after 15 to 30 min, many interactions could not be disrupted leading to an apparent shorter contour length, that was correlated to an increase of the persistence length. This phenomenology suggests that the KF-DNA complex started to collapse into a more stable and stiffer structure after 15 min. Remarkably, a repulsive negative force was detected after 30 min in most of the experiments if end-to-end distances lower than 4 m m were allowed (Figure 2 a , yellow), suggesting the formation of a thick KF-DNA aggregate of 1-3 m m of length. It was not possible to remove the bound KF by rinsing the molecule with peptide-free buffer for more than 45 min, reflecting the high stability of the final complex.
Binding of KF to DNA exhibits two regimes. First, there is a weak and fast regime apparently determined by the electrostatic attraction between the positively charged residues of the KF
Figure 2: KF binds to dsDNA. (a) DNA pulling curves before (black) and after flowing KF at different waiting times: 5 min (purple), 15 min (green) and 30 min (blue). The molecule is maintained at an extension of 6 m m (vertical dashed line), and the flow is temporarily stopped to perform pulling cycles between a minimum extension of 4.5 m m (dotted line) and a maximum force of 45 pN. The sawtooth pattern observed indicates that KF induces the compaction of DNA. Pulling cycles reaching end-to-end distances lower than 4 m mare shown in yellow. Data is filtered at 10 Hz bandwidth, v=500 nm/s. (Inset) Experimental set-up. (b,c) Persistence and contour length of five DNA molecules after flowing KF (black corresponds to naked DNA). The changes in the elastic parameters are a signature of the two regimes observed in KF-DNA complex formation: electrostatic binding and hydrophobic collapse. (d) Apparent contour length of a DNA molecule repeatedly pulled between a maximum force of 40 pN and a minimum extension that decreases in steps of 500 nm per pulling cycle (mean ± SD, N=10).
particles and the negatively charged backbone of DNA. According to our interpretation in this regime DNA binds to hydrophilic spots exposed on the surface of KF particles. Unpeeling of DNA segments requires forces typically lower than 20 pN. We will refer to this mode of binding as electrostatic binding (EB). This regime is observed in the first 15 min of the experiments shown in Figure 2 a , and is characterized by a constant contour length, the presence of force rips associated to unpeeling events and a reduced persistence length; the increased flexibility of the filament indicates a charge compensation that reduces self-repulsion of phosphates along the DNA backbone. There is a second stronger binding regime that occurs over longer timescales, that we attribute to the formation of an increasing number of stable hydrophobic contacts within the growing KF-DNA complex. Our interpretation is that in this regime DNA gets buried inside the bulk of the aggregate after being recruited by the hydrophilic spots exposed on its surface creating a stiff filament (as suggested by its increased persistence length). The slower timescales observed in this regime suggest that this mode of binding requires the remodeling and growth of a strongly hydrophobic complex. We will therefore refer to this second regime as hydrophobic collapse (HC). It is characterized by an irreversible decrease in contour length, an increase in persistence length, and the final collapse of the KF-DNA complex that is eventually compressed by pushing the two beads closer than 4 m m. The force required to disrupt such HC structure is above those accessible with our setup ( ∼ 100 pN), as suggested from previous AFM pulling experiments of single hydrophobically collapsed polymers. 21
What is the parameter that controls the prevalence of each binding regime? We expect that DNA bending fluctuations determine its binding to KF and the subsequent stabilization of the complex. Therefore, the molecular extension -or distance between beads- should be the parameter controlling the transition between both regimes. To verify this hypothesis we carried out experiments where the DNA was repeatedly pulled in the presence of KF between a maximum force of 40 pN and a minimum extension that progressively decreased from 8 m m to 2 m m in steps of 500 nm per pulling cycle. Such minimum extension controls the degree of compaction reached by the complex. For each cycle we then measured the apparent contour length at the maximum force
(40 pN). The results (Figure 2 d ) confirm the presence of the aforementioned regimes, which are separated by a threshold capture distance of 5.5 m m (corresponding to 66% of the contour length of the molecule). In the EB regime (relative extension ≥ 66%) the apparent contour length of the DNA fiber does not change whereas in the HC regime (relative extension ≤ 66%) it decreases linearly with the minimum distance between the two beads. A KF analog in which the ornithine residue was replaced by a negatively charged glutamic acid was investigated. No interaction between the KF analog and DNA was observed (Figure S2, Supporting Information), confirming that the positive charge of the ornithine residue is essential for electrostatic binding and providing further evidence that electrostatic interactions are key for the initial binding of the peptide to DNA. As well, a salt titration showed that the initial binding of KF to DNA and the subsequent DNA compaction is highly dependent on the ionic strength of the buffer, in agreement with the proposed mechanism (Figure S3, Supporting Information).
The formation of KF-DNA aggregates was directly observed by AFM imaging using a 2743bp DNA fragment (Figure 3 a-c ). Formation of blobs was observed at the initial time of mixing, their average size and number increased with time. Notably, after 20 min, a sharp decrease in the number of individual molecules bound to the mica surface was observed (Figure 3 d ). We attribute this to the formation of intermolecular complexes in which several DNA molecules are recruited into a single aggregate. As a consequence, no free DNA was observed after 30 minutes incubation time. These results were further confirmed with the use of longer DNA molecules ( l -DNA, 48-kb, Figure 3 e, f ). In the absence of DNA, large aggregates were not found on the images (Figure S4, Supporting Information). The observed aggregation of KF and binding to DNA was also characterized with DLS (Section S2, Supporting Information). The hydrodynamic radius of KF aggregates increased with time and remained constant after addition of DNA.
## KF-DNA affinity measurements in the EB regime
Pulling experiments of DNA in the presence of KF show a force-distance curve pattern with force rips and hysteresis even if the pulling is performed at very low speeds (Figure 4 a , blue). Low
Figure 3: AFM images of KF-DNA complexes. (a-c) AFM images of reactions of 1.65 ng linearized pGEM plasmid (2743-bp) and 100 m MKFobtained at 0, 20 and 30 min incubation times at room temperature, respectively. The number of free DNA molecules decreases with incubation time and large compaction blobs are observed. (d) DNA surface density at different incubation times, determined as the average number of free individual molecules per image of 9 m m 2 (mean ± SD, N ≥ 6). (e, f) Full l -DNA (48-kb) incubated with 100 m MKFfor 30 min. DNA condensation and formation of blobs are also seen for this larger DNA substrate. Bar scale is 600 nm (a, b, e) and 200 nm (c, f). Color scale (from dark to bright) is 0-2 nm in all AFM images.
pulling speeds are particularly useful to characterize the affinity of DNA binding to KF aggregates during the EB regime. In these experiments, the slope between two consecutive force rips reflects the elastic response of DNA with a given apparent contour length l 0. Each force rip is due to the unpeeling of a DNA segment that was electrostatically bound to the KF particle. A statistical analysis of force rips was used to determine the length of the DNA segments released during the unpeeling process (see Methods). 22 In this way, each experimental data point was associated to an apparent contour length l 0 (Figure 4 a , red left). A histogram of all the l 0 values showed a series of peaks that identify states that are stabilized by KF-DNA contacts (Figure 4 a , red right). The distance between two consecutive peaks is the length of the DNA segment released at every unpeeling event. The histogram was then fitted to a sum of Gaussians (Figure 4 a , bottom), and the distance between the mean of consecutive peaks was calculated. The experimental distribution of unpeeling events is broad (from a few nm to ∼ 400 nm) and follows an exponential distribution with mean size D l ∗ 0 = 31 ± 6 nm (Figure 4 b ). An exponential distribution of unpeeling lengths is known to correspond to the distribution of intervals expected in random partitioning a given contour length, in agreement with our hypothesis that DNA binds KF aggregates at hydrophilic spots in a random fashion.
Aforce vs. contour length representation (Figure 4 c ) emphasizes the release of DNA segments in a stepwise manner during the unpeeling process. The mechanical work performed at each unpeeling event ( W = F D x ) was then inferred from the rupture force value and released extension (see Methods). A histogram of the dissipated work (Figure S5, Supporting Information) shows an exponential distribution with average unpeeling energy of 13 5 . ± 5 kcal/mol. This value sets an upper limit to the free energy of binding of KF to DNA.
To gain a better understanding of the affinity of DNA binding to KF particles, we used a simple theoretical model that reproduces the experimental force-extension curves. We considered a model previously used to characterize the DNA-dendrimer condensation transition (see Methods). 23 The model reproduces the essential features of the experimental curves (Figure 4 d ) over a wide range of pulling speeds (30-500 nm/s). Despite the apparent large number of free parameters,
Figure 4: Modeling of DNA stretching experiments. (a) (Top, left) Blue curve shows a typical force-distance curve in a KF-DNA pulling experiment (20 Hz bandwidth, v=30 nm/s). Red curve shows the apparent contour length l 0 (right axis) for each experimental data point. The release cycle is fitted to the WLC model (black). (Top, right) Histogram of l 0 values. (Bottom) Detailed view of the histogram (red) and fit to a sum of Gaussians (blue). (b) Histogram of unpeeling segment lengths D l 0 and fit to an exponential distribution (green). Inset shows a log-normal plot (mean ± SD, N=435 events, 3 molecules). (c) Force vs. apparent contour length representation of the pulling experiment. Each color identifies a state (apparent contour length) temporally stabilized by KF-DNA contacts during the unpeeling process. (Inset) Sharp transitions between states are observed. Minimum and maximum forces of every state are indicated with crosses and diamonds respectively. (d) The black curve is an experimental pulling curve after 15 min interaction with KF (v=500 nm/s). A set of six simulations of the theoretical model is shown in red. (Inset) Scheme of the free-energy landscape of a two-states system at different forces (f). The main parameters describing the system are: the free energy difference ( D G) between the formed (F) and dissociated (D) conformations, the height of the barrier (B), the distance ( xi ) separating the two conformations, and the distance ( x † i ) from the transition state (TS) to the formed conformation. As the force is increased the free-energy landscape is tilted favoring the dissociated conformation once the critical force ( f c ) is reached.
only certain values in very specific ranges can reproduce these features (Section S3, Supporting Information). In brief, the experimental force-extension curves could be well described by assuming: (a) a low binding energy of DNA to KF aggregates ( D G ∼ 6 ± 2 kcal/mol); (b) a brittle unpeeling of the DNA segments ( x † i = 2 ± 1 nm, the barrier lying close to the formed conformation); and (c) a broad right-tailed distribution p B ( ) of high energy activation barriers given by p B ( ) = ( 1 / w ′ ) exp [ -( B -B 0 ) / w ′ ] with B ≥ B 0 = 89 kBT and w ′ = 5 kBT .
## DNA binds to KF aggregates at forces lower than 1pN
We followed the kinetics of DNA compaction by performing constant-force experiments at forces such that EB prevails (molecular extension ≥ 5.5 m m, Figure 2 d ). The DNA molecule was maintained at constant force using force-feedback, and we followed the time-evolution of the molecular extension while KF was flowed in.
At 1pN, a fast compaction took place (Figure 5 a ). The extension was reduced up to 40% in 20 min at a reproducible rate. This compaction is characterized by intermittent drops of extension that shorten the molecule by hundreds of nanometers in a few seconds (Figure 5 a , arrows). Pulling curves performed after this experiment (Figure 5 b ) showed again the characteristic sawtooth pattern.
In contrast, at 5pN the molecular extension remained constant within 100 nm after flowing KF for more than 30 min (Figure 5 ). c Still, intermittent large fluctuations on the order of tens of nanometers were often detected (Figure 5 c , arrows). These large fluctuations were never observed in controls without KF (Figure 5 c , gray), and we attribute them to individual binding events. Pulling cycles performed between 5-40 pN immediately after the peptide flow show a slight decrease in the persistence length and weak hysteresis effects suggesting very weak binding of DNA to KF (Figure 5 d ). Only by further decreasing the extension and force of the molecule full binding events were observed (Figure 5 d ).
Interestingly an overstretching transition was not always observed (Figure 5 d , yellow). We attribute this to the recruitment of DNA segments close to both ends of the tethered molecule
Figure 5: Kinetics of DNA binding to KF particles at a constant force. (a) DNA compaction at 1 pN. A control without peptide (gray) and two equivalent experiments at 40 m M KF (blue and yellow) are shown. The molecule was maintained at an initial extension A that relaxed at constant force down to a final value B. (b) Stretching of a DNA molecule before (black) and after (green, purple) the constant-force experiment at 1 pN. (c) KF does not compact DNA at 5 pN. A DNA molecule subjected at 5 pN is rinsed with KF and changes in the molecular extension are monitored. A control experiment without peptide (gray), and two independent experiments at 40 m MKF(blue and yellow) are shown. Large fluctuations indicative of individual binding events are observed (arrows) (d) Stretching of a DNA molecule only at forces higher than 5 pN immediately after the constant-force experiment at 5pN (purple). If the force was relaxed below 5 pN (green and yellow), the characteristic sawtooth pattern was immediately recovered. The force-extension curve of that molecule before flowing KF is shown in black. For all plots, raw data (1 kHz) is shown in light colors, and filtered data (1 Hz bandwidth for kinetic experiments, 10 Hz bandwidth for pulling experiments) is presented in dark colors. Pulling speed is 500 nm/s.
by KF particles that induce a torsionally constrained fiber, inhibiting the overstretching transition (Figure 5 d , Inset). 24 Otherwise, KF binding does not suppress or tilt the overstretching plateau as observed for DNA intercalators. 25 Moreover, the characteristic sawtooth pattern of KF remained visible after fully overstretching the DNA (Figure S6, Supporting Information).
## Unzipping experiments reveal different binding modes of KF to dsDNA and ssDNA
In a different set-up (Figure 6 a ), a 6.8-kb DNA hairpin was tethered and partially unzipped, maintaining at least half of the dsDNA stem open (Figure 6 a , dashed line) and then KF was flowed into the chamber. In this configuration the released ssDNA is long but rigid enough to severely restrict thermal fluctuations in the molecular extension (r m s d . . . . ∼ 20 nm). The advantage of this set-up is that the long separation between the hairpin and the beads ( ∼ 4 m m) inhibits any interaction between the beads and both the dsDNA region and linkers. The unzipping pattern of a DNA molecule is a fingerprint of its base-sequence 1 (Figure 6 a , gray), and changes of that pattern indicate a direct interaction between the peptide and DNA. Moreover, with this set-up we could explore the effect of KF on a DNA molecule maintained at zero force and forming a random-coil (the force stretches the linkers but not the hairpin).
After flowing KF for 3 min, the unzipping pattern substantially changed (Figure 6 a ) and forces up to 22 pN were needed to unzip the DNA. We attribute this to the increased force required to simultaneously break the base-pairing interactions and unpeel DNA segments from the KF particles. Consecutive unzipping curves show that ssDNA remains bound to KF particles at the maximum forces (25 pN). Surprisingly enough, the re-zipping trajectories overlapped with the re-zipping curves of naked DNA over a wide range of extensions ( ≥ 1500 nm). This is in agreement with a reannealing mechanism in which re-hybridization takes place first, followed by the formation of the complex. A similar phenomenology has been observed in the formation of amyloid nucleic acid fibers, in which the binding of amyloid peptides to oligonucleotides promotes their hybridization. 8
Note that in this experiment we only unzipped the region of the hairpin that remained in double-
stranded form while KF is flowed (right of the dashed line in Figure 6 b ). However, when we tried to rezip the region of the hairpin that remained as ssDNA during the peptide flow (left of the dashed line in Figure 6 b ), we could not recover the characteristic unzipping pattern of the molecule. This indicates that KF can bind ssDNA in a way that prevents re-hybridization of ssDNA strands. If the molecule was continuously submitted to unzipping/re-zipping cycles, the region of the DNA hairpin that previously re-zipped progressively looses that capability (Figure 6 c ) suggesting that KF is slowly binding to the stretched ssDNA (arrow in Figure 6 c ).
These results suggest that KF binds to the phosphate backbone in a configuration that does not interfere with base-pairing interactions when DNA is in its double-stranded form. However, when DNA is in its single-stranded form, KF can adopt configurations that interfere with the re-zipping of the molecule. These experiments also show that the interaction of the peptide with ssDNA is slow (in the order of minutes), as only the ssDNA regions of the hairpin that remained exposed for long times to the peptide were unable to re-hybridize.
## KF binds ssDNA
To characterize the interaction of ssDNA with KF aggregates we developed a simple method to generate a long ssDNA template (13-kb) for optical tweezers experiments (see Methods). By using this setup we could therefore measure the elastic response of the ssDNA down to forces as low as 1-2 pN (Figure 7 a ).
We then followed their molecular extension in the presence of KF. At 5 pN we observed a slow compaction ( ∼ 20-30 min) with an absolute reduction in extension close to 16% (Figure 7 b , upper panel), demonstrating that ssDNA binds KF. This phenomenology was reproducible within different experiments, and the slow kinetics agree with the results from unzipping experiments (Figure 6 ). We also measured the time evolution of the stiffness of the KF-ssDNA fiber by recording the c magnitude of the thermally induced fluctuations in the molecular extension (Section S4, Supporting Information). At 5 pN the molecule stiffened with time at a rate of (13 ± 5) 10 · -3 pN/(nm s) · (Figure 7 b , middle panel). However this change was only observed 10-15 min after compaction of
Figure 6: Unzipping experiments show that KF binds both dsDNA and ssDNA. (a) An unzipping pattern of the DNA hairpin before incubation with KF is fully represented in gray in each panel. The dashed line represents the position at which the molecule remained unzipped during the peptide flow. Consecutive pulling cycles of the dsDNA stem region after incubation with KF (blue, red) show a strong distortion of the unzipping pattern. However, the re-zipping of the hairpin remains unaffected indicating that the molecule can hybridize again. (Inset) A 6.8-kb hairpin is maintained partially unzipped leaving less than half of the dsDNA stem closed during the incubation with KF. This configuration prevents KF-DNA interactions mediated by the beads. (b) The region of the hairpin that remained as ssDNA during the peptide flow (left of dashed line) cannot hybridize again in contrast to what happens without KF (gray curve) or in the region maintained as dsDNA during the peptide flow (right of dashed line). Three pulling cycles are plotted in blue, red and yellow. (c) If the hairpin is rinsed with KF and then submitted to several consecutive pulling cycles, the non-hybridizing region increases with time. Four pulling cycles that reflect this trend are shown (red, blue, yellow and purple). Data is filtered at 10 Hz bandwidth, v=50 nm/s.
the fiber started, suggesting that stiffness changes are mostly due to the hydrophobic collapse of the KF-ssDNA aggregate rather than electrostatic binding of ssDNA to KF particles.
Pulling curves obtained after the peptide flow (Figure 7 c ) also show force rips in the stretching curves. However the sawtooth pattern was smoother than for dsDNA suggesting the occurrence of fewer events and higher unpeeling forces (Figure 7 c ). We attribute this to the increased hydrophobic forces that stabilize the KF-ssDNA complex, that also lead to a systematic shortening of the effective contour length of the ssDNA. At a higher stretching force of 10 pN, KF did not induce compaction of ssDNA though, but intermittent jumps in the extension were observed, indicative of individual binding events (Figure S7 a , Supporting Information). At this higher force, the stiffness of the molecule remained constant within the resolution of measurements (Figure S7 b , Supporting Information). However, KF-ssDNA compaction could be induced by lowering the force down to 5 pN (Figure S7 a , Supporting Information) reproducing the phenomenology reported in Figure 7 b .
Binding of ssDNA to KF aggregates was further corroborated by AFM. ssDNA molecules (2743-bp) were generated by heat denaturation and fast cooling down to 4 ◦ C(Figure 7 d ). Addition of KF to the ssDNA preparation triggered the formation of aggregation spots immediately after mixing (Figure 7 e ). Interestingly enough, longer incubation times yielded a reduction of free ssDNA molecules, together with the formation of large aggregates of KF surrounded by double stranded DNA (Figure 7 f-g ). This observation suggests that KF induces the rehybridization of ssDNA, as ssDNA molecules do not anneal after 30 min incubation at room temperature in the absence of KF. This is in agreement with the unzipping results (see previous section), and is likely due to the fact that KF traps and maintains close in space different ssDNA molecules that are occasionally able to rehybridize.
## Discussion
By combining single molecule techniques and bulk measurements we showed that KF forms particles that bind and compact DNA. Our measurements reveal that this process is characterized by
Figure 7: Kinetics of ssDNA binding to KF particles. (a) Method used to generate a long ssDNA template from a DNA hairpin. The specific binding of an oligonucleotide to the hairpin inhibits the hybridization of the molecule (blue) at forces lower than the average unzipping force. The full force-extension curve of the hairpin is plotted as a reference (gray). (b) Compaction and stiffening of ssDNA at 5 pN are representative of EB and HC respectively. (Top) Extension of a ssDNA molecule rinsed with KF at 5 pN. Three independent experiments at 40 m MKF(blue, red and yellow) and a control without peptide (gray) are shown (compaction starts at t=0). (Middle) Average stiffness of the ssDNA molecule at 5 pN during the peptide flow (red) and a control without peptide (black). Three individual experiments are shown in light red. The stiffness is measured from the fluctuations in the trap position. (c) Stretching curves of a ssDNA molecule before (gray) and after incubation with 40 m MKF at 3 pN for 25 min (purple, green, blue). Data is collected at 1 kHz (light colors) and filtered to 1 Hz bandwidth (dark colors). Pulling speed is 100 nm/s. (d) ssDNA molecules (1.8 nM molecules, 5 m M nucleotides) are adsorbed on a mica surface showing a much compact conformation than dsDNA due to its lower persistence length. (e) Immediately after mixing KF (100 m M) with ssDNA molecules (5 m Mnucleotides) we observe the formation of aggregation spots and a substantial decrease in the number of ssDNA molecules per image. (f) After 30 min incubation at room temperature, these effects are more evident as big aggregates are seen. (g) Histogram of ssDNA molecules at different incubation times (0 and 30 min) with and without KF, determined as the average number of free individual molecules per image of 4 m m 2 (mean ± SD, N ≥ 6).
two distinct phases controlled by the molecular extension of the DNA. First, there is a fast and weak binding regime determined by electrostatic binding (EB) to positive residues exposed on the surface of the KF particles (Figure 8 a ). This binding is triggered by spontaneous bending fluctuations along DNA. Upon reduction of the molecular extension a slow remodeling of the KF-DNA complex takes place; we propose that this new regime is led by the formation of new hydrophobic contacts that stabilize a hydrophobically collapsed (HC) structure. A capture distance separating both regimes is identified, corresponding to a relative extension of 66%. Remarkably enough, theoretical studies between spherical charged aggregates and oppositely charged polymeric chains predicted a capture distance leading to the irreversible adsorption of the chain to the aggregates. 27 The recruitment of DNA segments along the surface of the aggregate (EB) is driven by the electrostatic attraction between the negative charge of DNA and the positive charge of L-Orn residues that are most likely exposed on the surface of the particle forming hydrophilic spots (Figure 8 b ). This interpretation is supported by the following facts: (i) The persistence length of DNA is reduced during the initial binding of the peptide, suggesting a charge compensation that reduces self-repulsion of the DNA phosphate backbone (ii) A KF analog without a positive charge does not bind to DNA (Figure S2, Supporting Information), providing evidence that the positive charge is essential for binding (iii) A salt titration shows that binding of KF to DNA is inhibited at high salt condition, and that the strength of the interaction increases with decreasing ionic strength (Figure S3, Supporting Information) (iv) The zeta-potential value of KF, DNA, and KF-DNA complexes (Table S1, Supporting Information) indicate different surface charge densities depending on whether DNA is complexed with KF or not. In addition, the phenomenology observed during the EB regime cannot be understood from the isolated action of KF peptides (each bearing a single positive charge). It must be due instead to the concurrent action of several positive charges that are contained in each peptide aggregate, which are then able to electrostatically interact with DNA in a similar fashion that dendrimers or other polycationic agents do. Modeling of the experimental results shows that EB of DNA to KF particles is consistent with an exponential distribution of unpeeling lengths (average value D l 0 = 31 ± 6 nm), a low binding energy ( D G ∼ 6 ± 2 kcal/mol), a
brittle unpeeling process ( x † i = 2 ± 1 nm), and a disordered collection of activation barriers higher than B 0 = 89 kBT with an exponential right tail of width w ′ = 5 kBT . This is in contrast to results obtained for nucleosomal particles that show a single characteristic unpeeling length of 26 nm and a single energy barrier of 36-38 kBT . 28 The force at which we observe dsDNA compaction (1pN) is in the same range of forces that has been reported for other DNA compacting molecules such as histone-like FIS and HU proteins, 29,30 or polycationic condensing agents. 23,31,32
Unzipping experiments show that KF also forms complexes with ssDNA. Interestingly, unzipping experiments together with AFM show that KF does not inhibit dsDNA hybridization. These results agree with previous studies on the formation of amyloid nucleic acid fibers, which showed how charged surfaces of peptide complexes recruit oligonucleotides and promote their hybridization. 8 The two distinct phases observed in the formation of KF-dsDNA complexes (EB and HC) were also seen for ssDNA, showing that the mechanism of complexation is similar in both polyelectrolytes. Yet, our experiments suggest a stronger stabilization of the complex for ssDNA, which is likely due to its increased hydrophobicity and lower persistence length. Indeed, we have found that KF can compact ssDNA at higher forces than dsDNA does, indicating the role of spontaneous bending fluctuations to initiate EB. Finally, constant-force experiments with both DNA substrates show a correlation between the degree of compaction of the molecule and the effective rigidity of the tether (Figure 8 c ).
In relation to its biological activity, and as previously showed for amyloid fibrils, 1,9,13 the enhanced positive charge of KF aggregates could allow them to interact more effectively with other polyanionic molecules such as polysaccharides ( e.g. glycosaminoglycans). Whether or not the cytotoxic effects of KF are related to its interaction with polyanions, we hypothesize that these interactions could play a role in modulating the activity of the peptide either intracellularly or in the extracellular matrix. In particular KF might also interact with the phospholipids of the plasma membrane inducing the formation of pores and cell necrosis. 20
Figure 8: Model of KF-DNA complex formation. (a) Scheme of the two kinetic steps during the formation of the KF-DNA complex: Electrostatic binding (EB) and Hydrophobic collapse (HC). At low forces DNA (blue) binds to KF particles (yellow) due to the electrostatic attraction of the phosphate backbone to positive residues exposed on the aggregate surface (EB). This process is followed by the formation of new hydrophobic contacts between aggregates that form a larger collapsed structure (HC) (b) Pictorial representation of how bending fluctuations induce the EB of DNA (black) to electrophilic spots (blue) on the aggregate surface, and how hydrophobic interactions between peptides drive to a collapsed structure in which DNA becomes entrapped within the aggregated complex. (c) Phase diagram showing conditions that trigger DNA compaction due to EB of KF in constant force experiments. Average stiffness of the dsDNA (red) and ssDNA (green) molecules before flowing KF in the constant force experiments (2 ≤ N ≤ 9, mean ± SD). We only observe DNA compaction in the region highlighted in gray, suggesting that compaction depends directly on DNA bending fluctuations via template stiffness. A WLC model (red line) and FJC model (green line) with the parameters determined in the main text for dsDNA and ssDNA respectively are plotted as a reference.
## Conclusion
This study represents the first attempt to extract quantitative information about the binding affinity and kinetic steps involved in the interaction between a nucleic acid (DNA) and an anticancer selfaggregating peptide (KF) at the single molecule level. To date most studies of aggregation kinetics have been performed using ensemble techniques where the individual behavior of molecules cannot be distinguished. Using optical tweezers, we have shown that KF binds DNA in two kinetic steps (an initial electrostatic binding that is followed by an hydrophobic collapse of the peptideDNA complex), and characterized the spectrum of binding affinities, kinetics barriers and lengths of DNA segments sequestered within the KF-DNA complex. The proposed methodology is not limited to the characterization of amorphous aggregates. 33 Protein aggregation, a topic of major interest due to the role of the aggregation of misfolded proteins in neurodegenerative diseases, 34 might be well addressed using single molecule force spectroscopy. As well, AFM images of the nucleoid-associated proteins Dps 35,36 and of the drug cisplatin 37 are very similar to those we found here for KF-DNA, suggesting that many biochemical studies of protein complexes that face similar aggregation or compaction phenomena are susceptible of being studied with this approach. For instance, the aggregation of the splicing factor MBNL1 by mutant mRNA hairpins is at the core of Myotonic dystrophy type I. 38,39 Research on the formation of these RNA-protein aggregates, and of peptides that disrupt this interaction 40 could greatly benefit from the aforementioned approaches. Our study confirms force spectroscopy studies of single aggregates as potentially very useful to characterize the thermodynamic and mechanical properties of nucleic acid-peptide complexes. On a longer term, the study of the mechanical response of aggregates related to many relevant neurodegenerative diseases could also be approached by using a peptide template instead of a DNA molecule.
## Materials and Methods
## Optical tweezers set-up
A miniaturized dual beam optical tweezers instrument described in 1 has been used for the single molecule experiments. A single optical trap is created by focusing two counter-propagating laser beams ( l = 845 nm, P = 200 mW) into the center of a fluidics chamber mounted on a motorized stage. The optical trap can be displaced in a range of 12 m m by using a pair of piezoelectric actuators mechanically coupled to the laser optical fibers. For manipulation with the optical tweezers, DNA molecules are differentially end-labeled with digoxigenins and biotins so each end of the molecule can specifically bind to antidigoxigenin and streptavidin-coated beads, respectively (Supporting Information). A single DNA molecule is then tethered between two polystyrene beads. One is subjected on the tip of a micropipette, whereas the other one is confined in the optical trap (Figure 2 a , Inset). The molecule can be stretched by moving the trap relatively to the micropipette, and both the extension and force applied to it are determined in real time.
Force applied to the optically trapped bead is directly determined from the change in light momentum by measuring the deflection of the laser beams with position sensing detectors (PSD). 41 To measure the position of the optical trap, about 8% of the lasers light is split before entering the objective lenses and redirected to PSD's. In this way, trap displacements can be followed with sub-nanometer resolution. The relative molecular extension is inferred by subtracting the trap compliance (F/k) to the absolute displacement of the optical trap (trap stiffness = 70 pN/ m m). The extension and force applied to the molecule can be recorded at 1 kHz rate, and a resolution of 0.1 pN is achieved.
## Stretching experiments of dsDNA and ssDNA
The persistence length ( l p ) and stretch modulus ( S ) of each dsDNA molecule were determined before flowing KF with a fit to the Worm-Like Chain (WLC) model. Average values of l p = 44 7 . ± 2 0 . nm and S = 1419 ± 240 pN were obtained (N=10), in good accordance with the generally accepted
parameters. 7,43,44 A compatible value of l p = 43 0 . ± 2 0 nm was also found using the inextensible . WLCmodel. The stretching curves also showed the characteristic overstretching plateau at a force of 62 5 . ± 0 5 pN with an extension . ∼ 70% of the contour length. Molecules that showed an abnormally high hysteresis on the overstretching transition (generally attributed to highly nicked DNA molecules) were discarded. For the fits to the WLC model, the correction to the Marko-Siggia interpolation formula suggested in 45 was used, and a Levenberg-Marquadt algorithm was used for both minimizations. 46 Details of the experimental set-up and molecular synthesis are found in the Supporting Information.
To generate a 13-kb ssDNA molecule, a 6.8-kb DNA hairpin was fully unzipped in a buffer containing a 30-base oligonucleotide that binds to the loop and its flanking region due to base-pair complementarity. The high bending rigidity of the duplex at the loop region strongly stabilizes the ssDNA form over the dsDNA form at forces lower than the average unzipping force (14.5pN). Stretching curves of the ssDNA molecule were fitted to the Freely-Jointed Chain Model (Section S5, Supporting Information), finding a Kuhn length of b = 1 57 . ± 0 05 nm (N=5) in agreement . with previous results. 1,7
## Statistical analysis and simulations
For every data point ( x exp , f exp) of a force-distance curve, we determined its most probable apparent contour length ( l 0) by finding the theoretical WLC 45 that passes closest to that point at the force f exp:
$$\left | x _ { \exp } - x _ { W L C } \left ( l _ { 0 }, f _ { \exp } \right ) \right | = \min _ { l } \left ( \left | x _ { \exp } - x _ { W L C } \left ( l, f _ { \exp } \right ) \right | \right ).$$
The theoretical extension ( xWLC ( l 0 , f exp ) ) was determined using the elastic parameters ( l p , S ) obtained from a WLC fit to the relaxation curve.
To determine the mechanical work ( W = F D x ) performed to disrupt each KF-DNA contact in DNA stretching experiments, we determined the average rupture force and the extension of DNA released at every unpeeling event. This work is partially used to stretch the released DNA up to the
rupture force. The rest of work is dissipated into the solvent in the form of heat:
$$W _ { \text{dissipated} } = \frac { F \Delta F } { k } - \Delta G _ { \text{stretching} }$$
$$\Delta G _ { \text{stretching} } = \frac { \Delta L } { L } \int _ { 0 } ^ { x _ { \text{rup} } } F _ { \text{WLC} } ( x ) \, d x$$
where x rup is the molecular extension of the DNA fiber at the rupture force. In the above expressions we use the fact that the force-extension curve of the WLC model is a sole function of x / L .
The model considered to reproduce the KF-DNA stretching curves, simulates the contacts made between KF aggregates and DNA segments as a set of N non-interacting two-level systems. When force is applied to the molecule, each segment can yield an extension xi in a thermally activated process characterized by a critical force f c , and a dissociation rate kc . Each segment is described by its free energy of formation D Gi , activation barrier Bi , and distance to the transition state x † i . The released extension xi was assumed to follow the experimental distribution (Figure 4 b ), and we introduced some structural disorder by assuming that D G and B are also exponentially distributed. Simulation parameters that best describe the experimental curves are: p ( D G ) = ( 1 / w e ) -( D G -D G 0 ) / w with D G ≥ D G 0 = 10 kBT and w = 1 kBT ; x † i = 2 ± 1 nm; f c = 5 ± 3 pN; kc =0.5 s -1 ; p B ( ) = ( 1 / w ′ ) e -( B -B 0 ) / w ′ with B ≥ B 0 = 89 kBT and w ′ = 5 kBT . The contact-length distribution ( xi ) was assumed to follow the experimental distribution: p xi ( ) = ( 1 / w e ) -( xi -xi , 0 ) / w for xi ≥ xi , 0 ( p xi ( ) =0 otherwise) with xi , 0=8 nm and w =24 nm. Other specific parameters for the simulations shown in Figure 4 d : k trap =0.07 pN/nm, v=500 nm/s, l p =35 nm, DNA slack=6500 nm. Errors are an estimation of the range in which the features of the process are well reproduced by the model when each parameter is independently modified.
## Sample flow set-up
A syringe pump (PicoPump, KDScientific) and a glass syringe have been used for KF sample infusion. Polyethylene PE-10 tubing (BD Intramedic) is used to connect the syringe to the fluidics
chamber. For the experiments performed with a constant force protocol a buffer flow-rate of 3 m l/min is used to keep a low drag force on the bead. For the other experiments a higher flowrate of 9 m l/min was preferred (Supporting Information). The arrival of the KF solution into the experimental area can be monitored due to the slight change in the refractive index of the medium caused by the 2% DMSO content of the peptide buffer.
We were not able to establish DNA tethers between polystyrene beads with KF in the buffer, suggesting that strong condensation effects appear on untethered molecules. Therefore, experiments were always performed by flowing KF to DNA molecules that had been tethered in peptidefree buffer. Flow experiments performed with polystyrene beads without DNA do not show aggregation of KF onto the bead surfaces, nor an increase in stickiness between beads, ruling out aggregation on bead surfaces.
## AFMsample preparation and imaging
KF-DNA reactions included 1.65 ng of DNA (0.9 nM dsDNA molecules, pGEM3Z, 2743 bp) (Promega) linearized with Bam HI and 100 m Mof KF in 20 mM Tris-HCl (pH 7.5) and 100 mM NaCl. To facilitate adsorption of DNA in a buffer devoid of Mg 2 + ions, we pre-treated the mica surface with 100mM spermine tetrahydrochloride (S85610, Fluka, Sigma) dissolved in 10 mM Tris-HCl (pH 7.5). Pre-treatment with spermine consisted in deposition of 20 m l of 100 mM spermine tetrahydrochloride on a freshly cleaved mica surface, one minute adsorption, washing with Milli-Q water, and drying with nitrogen gas. Use of spermine at low concentrations allowed uniform adsorption of DNA molecules. Immediately after the spermine treatment, the mixture of KF and DNA incubated at room temperature and for the stated time was deposited on the mica. After 30 s, the mica surface was washed with filtered-MilliQ water and blown dry in a gentle stream of nitrogen gas. To study KF-ssDNA interactions, a stock of ssDNA was produced by heat denaturing the linearized pGEM plasmid at 95 C for 5 minutes and placing the tube quickly after on ice. ◦ ∼ 3 kb DNA molecules remain stable in its ssDNA form following this procedure as long as they remain at 4 ◦ C for at least one week. KF and ssDNA were mixed at the same proportions as for ds-
DNAand followed the sample preparation procedure described above. Samples were imaged in air at room temperature at identical conditions as previously described. 47 Standard image processing consisted of plane subtraction and flattening using WSxM freeware. 48
## Acknowledgement
The authors thank Pharmamar SA and G. Acosta for providing the peptides Kahalalide F and its analog. JCS acknowledges a grant associated to ICREA Academia 2008; FR is supported by the Human Frontier Science Program [Grant No. RGP55-2008], the Spanish Ministry of Economy and Competitiveness [Grant No. FIS2010-19342] and an ICREA Academia award; FMH acknowledges a grant from the European Research Council [Starting Grant number 206117], and the Spanish Ministry of Economy and Competitiveness [grant number FIS2011-24638]; MEFP was supported by a contract from CSIC [contract number 200920I123, associated to the ERC grant]. FA acknowledges CICYT [grant number: CTQ2012-30930], and SBD a invited professor position from AGAUR (Generalitat de Catalunya).
## Supporting Information Available
Additional methods, optical tweezers results, DLS measurements and AFM images. This material is available free of charge via the Internet at http://pubs.acs.org/.
## References
- 1. Calamai, M.; Kumita, J.; Mifsud, J.; Parrini, C.; Ramazzotti, M.; Ramponi, G.; Taddei, N.; Chiti, F.; Dobson, C. Nature and Significance of the Interactions between Amyloid Fibrils and Biological Polyelectrolytes. Biochemistry 2006 , 45 , 12806-12815.
- 2. Bucciantini, M.; Giannoni, E.; Chiti, F.; Baroni, F.; Formigli, L.; Zurdo, J.; Taddei, N.; Ram-
poni, G.; Dobson, C.; Stefani, M. Inherent Toxicity of Aggregates Implies a Common Mechanism for Protein Misfolding Diseases. Nature 2002 , 416 , 507-511.
- 3. Gsponer, J.; Vendruscolo, M. Theoretical Approaches to Protein Aggregation. Protein Pept. Lett. 2006 , 13 , 287-293.
- 4. Coan, K.; Shoichet, B. Stoichiometry and Physical Chemistry of Promiscuous Aggregate-Based Inhibitors. J. Am. Chem. Soc. 2008 , 130 , 9606-9612.
- 5. Puchalla, J.; Krantz, K.; Austin, R.; Rye, H. Burst Analysis Spectroscopy: A Versatile SingleParticle Approach for Studying Distributions of Protein Aggregates and Fluorescent Assemblies. Proc. Natl. Acad. Sci. U.S.A. 2008 , 105 , 14400-14405.
- 6. Feng, B.; Toyama, B.; Wille, H.; Colby, D.; Collins, S.; May, B.; Prusiner, S.; Weissman, J.; Shoichet, B. Small-Molecule Aggregates Inhibit Amyloid Polymerization. Nat. Chem. Biol. 2008 , 4 , 197-199.
- 7. Coan, K.; Maltby, D.; Burlingame, A.; Shoichet, B. Promiscuous Aggregate-Based Inhibitors Promote Enzyme Unfolding. J. Med. Chem. 2009 , 52 , 2067-2075.
- 8. Braun, S.; Humphreys, C.; Fraser, E.; Brancale, A.; Bochtler, M.; Dale, T. Amyloid-Asssociated Nucleic Acid Hybridisation. PLoS One 2011 , 6 , e19125.
- 9. Di Domizio, J.; Zhang, R.; Stagg, L.; Gagea, M.; Zhuo, M.; Ladbury, J.; Cao, W. Binding with Nucleic Acids or Glycosaminoglycans Converts Soluble Protein Oligomers to Amyloid. J. Biol. Chem. 2012 , 287 , 736-747.
- 10. Macedo, B.; Millen, T.; Braga, C.; Gomes, M.; Ferreira, P.; Kraineva, J.; Winter, R.; Silva, J.; Cordeiro, Y. Nonspecific Prion Protein-Nucleic Acid Interactions Lead to Different Aggregates and Cytotoxic Species. Biochemistry 2012 , 51 , 5402-5413.
- 11. Motamedi-Shad, N.; Garfagnini, T.; Penco, A.; Relini, A.; Fogolari, F.; Corazza, A.; Esposito, G.; Bemporad, F.; Chiti, F. Rapid Oligomer Formation of Human Muscle Acylphosphatase Induced by Heparan Sulfate. Nat. Struct. Mol. Biol. 2012 , 19 , 547-554.
- 12. Cherny, D.; Hoyer, W.; Subramaniam, V.; Jovin, T. Double-Stranded DNA Stimulates the Fibrillation of a -Synuclein in Vitro and Is Associated with the Mature Fibrils: an Electron Microscopy Study. J. Mol. Biol. 2004 , 344 , 929-938.
- 13. Cohlberg, J.; Li, J.; Uversky, V.; Fink, A. Heparin and Other Glycosaminoglycans Stimulate the Formation of Amyloid Fibrils from a -Synuclein in Vitro . Biochemistry 2002 , 41 , 15021511.
- 14. Dale, T. Protein and Nucleic Acid Together: a Mechanism for the Emergence of Biological Selection. J. Theor. Biol. 2006 , 240 , 337-342.
- 15. Hamann, M. T.; Scheuer, P. J. Kahalalide F: a Bioactive Depsipeptide from the Sacoglossan Mollusk Elysia Rufescens and the Green Alga Bryopsis sp . J. Am. Chem. Soc. 1993 , 115 , 58255826.
- 16. Lopez-Macia, A.; Jimenez, J. C.; Royo, M.; Giralt, E.; Albericio, F. Synthesis and Structure Determination of Kahalalide F. J. Am. Chem. Soc. 2001 , 123 , 11398-11401.
- 17. Garcia-Rocha, M.; Bonay, P.; Avila, J. The Antitumoral Compound Kahalalide F Acts on Cell Lysosomes. Cancer Lett. 1996 , 99 , 43-50.
- 18. Suarez, Y.; Gonzalez, L.; Cuadrado, A.; Berciano, M.; Lafarga, M.; Munoz, A. Kahalalide F, a New Marine-Derived Compound, Induces Oncosis in Human Prostate and Breast Cancer Cells. Mol. Cancer Ther. 2003 , 2 , 863-872.
- 19. Sewell, J. M.; Mayer, I.; Langdon, S. P.; Smyth, J. F.; Jodrell, D. I.; Guichard, S. M. The Mechanism of Action of Kahalalide F: Variable Cell Permeability in Human Hepatoma Cell Lines. Eur. J. Cancer 2005 , 41 , 1637-1644.
- 20. Molina-Guijarro, J.; Macías, Á.; García, C.; Muñoz, E.; García-Fernández, L.; David, M.; Núñez, L.; Martínez-Leal, J.; Moneo, V.; Cuevas, C. et al. Irvalec Inserts into the Plasma Membrane Causing Rapid Loss of Integrity and Necrotic Cell Death in Tumor Cells. PLoS One 2011 , 6 , e19042.
- 21. Li, I.; Walker, G. Signature of Hydrophobic Hydration in a Single Polymer. Proc. Natl. Acad. Sci. U.S.A. 2011 , 108 , 16527-16532.
- 22. Huguet, J.; Forns, N.; Ritort, F. Statistical Properties of Metastable Intermediates in DNA Unzipping. Phys. Rev. Lett. 2009 , 103 , 248106.
- 23. Ritort, F.; Mihardja, S.; Smith, S. B.; Bustamante, C. Condensation Transition in DNAPolyaminoamide Dendrimer Fibers Studied Using Optical Tweezers. Phys. Rev. Lett. 2006 , 96 , 118301.
- 24. Leger, J. F.; Romano, G.; Sarkar, A.; Robert, J.; Bourdieu, L.; Chatenay, D.; Marko, J. F. Structural Transitions of a Twisted and Stretched DNA Molecule. Phys. Rev. Lett. 1999 , 83 , 1066-1069.
- 25. Vladescu, I. D.; McCauley, M. J.; Nuñez, M. E.; Rouzina, I.; Williams, M. C. Quantifying Force-Dependent and Zero-Force DNA Intercalation by Single-Molecule Stretching. Nat. Methods 2007 , 4 , 517-522.
- 1. Huguet, J. M.; Bizarro, C. V.; Forns, N.; Smith, S. B.; Bustamante, C.; Ritort, F. SingleMolecule Derivation of Salt Dependent Base-Pair Free Energies in DNA. Proc. Natl. Acad. Sci. U.S.A. 2010 , 107 , 15431-15436.
- 27. Podgornik, R.; Jönsson, B. Stretching of Polyelectrolyte Chains by Oppositely Charged Aggregates. Europhys. Lett. 1993 , 24 , 501-506.
- 28. Brower-Toland, B.; Smith, C.; Yeh, R.; Lis, J.; Peterson, C.; Wang, M. Mechanical Disruption
of Individual Nucleosomes Reveals a Reversible Multistage Release of DNA. Proc. Natl. Acad. Sci. U.S.A. 2002 , 99 , 1960-1965.
- 29. Skoko, D.; Yan, J.; Johnson, R.; Marko, J. Low-Force DNA Condensation and Discontinuous High-Force Decondensation Reveal a Loop-Stabilizing Function of the Protein Fis. Phys. Rev. Lett. 2005 , 95 , 208101.
- 30. Van Noort, J.; Verbrugge, S.; Goosen, N.; Dekker, C.; Dame, R. T. Dual Architectural Roles of HU: Formation of Flexible Hinges and Rigid Filaments. Proc. Natl. Acad. Sci. U.S.A. 2004 , 101 , 6969-6974.
- 31. Todd, B.; Rau, D. Interplay of Ion Binding and Attraction in DNA Condensed by Multivalent Cations. Nucleic Acids Res. 2008 , 36 , 501-510.
- 32. Hormeño, S.; Moreno-Herrero, F.; Ibarra, B.; Carrascosa, J. L.; Valpuesta, J. M.; AriasGonzalez, J. R. Condensation Prevails over B-A Transition in the Structure of DNA at Low Humidity. Biophys. J. 2011 , 100 , 2006-2015.
- 33. Yoshimura, Y.; Lin, Y.; Yagi, H.; Lee, Y.; Kitayama, H.; Sakurai, K.; So, M.; Ogi, H.; Naiki, H.; Goto, Y. Distinguishing Crystal-Like Amyloid Fibrils and Glass-Like Amorphous Aggregates from their Kinetics of Formation. Proc. Natl. Acad. Sci. U.S.A. 2012 , 109 , 1444614451.
- 34. Ross, C.; Poirier, M. What is the Role of Protein Aggregation in Neurodegeneration? Nat. Rev. Mol. Cell Biol. 2005 , 6 , 891-898.
- 35. Ceci, P.; Cellai, S.; Falvo, E.; Rivetti, C.; Rossi, G.; Chiancone, E. DNA Condensation and Self-Aggregation of Escherichia Coli Dps are Coupled Phenomena Related to the Properties of the N-Terminus. Nucleic Acids Res. 2004 , 32 , 5935-5944.
- 36. Ceci, P.; Mangiarotti, L.; Rivetti, C.; Chiancone, E. The Neutrophil-Activating Dps Protein
of Helicobacter Pylori , HP-NAP, Adopts a Mechanism Different from Escherichia Coli Dps to Bind and Condense DNA. Nucleic Acids Res. 2007 , 35 , 2247-2256.
- 37. Hou, X.; Zhang, X.; Wei, K.; Ji, C.; Dou, S.; Wang, W.; Li, M.; Wang, P. Cisplatin Induces Loop Structures and Condensation of Single DNA Molecules. Nucleic Acids Res. 2009 , 37 , 1400-1410.
- 38. Dansithong, W.; Wolf, C.; Sarkar, P.; Paul, S.; Chiang, A.; Holt, I.; Morris, G.; Branco, D.; Sherwood, M.; Comai, L. et al. Cytoplasmic CUG RNA Foci Are Insufficient to Elicit Key DM1 Features. PLoS One 2008 , 3 , e3968.
- 39. Dickson, A. M.; Wilusz, C. J. Repeat Expansion Diseases: when a Good RNA Turns Bad. Wiley Interdiscip. Rev.: RNA 2010 , 1 , 173-192.
- 40. García-López, A.; Llamusí, B.; Orzáez, M.; Pérez-Payá, E.; Artero, R. In Vivo Discovery of a Peptide that Prevents CUG-RNA Hairpin Formation and Reverses RNA Toxicity in Myotonic Dystrophy Models. Proc. Natl. Acad. Sci. U.S.A. 2011 , 108 , 11866-11871.
- 41. Smith, S. B.; Cui, Y.; Bustamante, C. Optical-Trap Force Transducer that Operates by Direct Measurement of Light Momentum. Methods Enzymol. 2002 , 361 , 134-162.
- 7. Smith, S.; Cui, Y.; Bustamante, C. Overstretching B-DNA: the Elastic Response of Individual Double-Stranded and Single-Stranded DNA Molecules. Science 1996 , 271 , 795-799.
- 43. Wang, M. D.; Yin, H.; Landick, R.; Gelles, J.; Block, S. M. Stretching DNA with Optical Tweezers. Biophys. J. 1997 , 72 , 1335-1346.
- 44. Baumann, C. G.; Smith, S. B.; Bloomfield, V. A.; Bustamante, C. Ionic Effects on the Elasticity of Single DNA Molecules. Proc. Natl. Acad. Sci. U.S.A. 1997 , 94 , 6185-6190.
- 45. Bouchiat, C.; Wang, M.; Allemand, J.; Strick, T.; Block, S.; Croquette, V. Estimating the Persistence Length of a Worm-Like Chain Molecule from Force-Extension Measurements. Biophys. J. 1999 , 76 , 409-413.
- 46. Marquardt, D. W. An Algorithm for Least-Squares Estimation of Nonlinear Parameters. J. Soc. Ind. Appl. Math. 1963 , 11 , 431-441.
- 47. Fuentes-Perez, M. E.; Gwynn, E. J.; Dillingham, M. S.; Moreno-Herrero, F. Using DNA as a Fiducial Marker to Study SMC Complex Interactions with the Atomic Force Microscope. Biophys. J. 2012 , 102 , 839-848.
- 48. Horcas, I.; Fernández, R.; Gómez-Rodríguez, J. M.; Colchero, J.; Gómez-Herrero, J.; Baro, A. M. WSXM: a Software for Scanning Probe Microscopy and a Tool for Nanotechnology. Rev. Sci. Instrum. 2007 , 78 , 013705-013705.
Supporting Information: Electrostatic Binding and Hydrophobic Collapse of Peptide-Nucleic Acid Aggregates Quantified Using Force Spectroscopy
## S1 Supplementary Methods
## S1.1 Optical tweezers experiments
## S1.1.1 Stretching experiments with dsDNA
For the dsDNA stretching experiments, 1 m l of a ∼ 0 5 pmol/ml dilution of a 24-kb dsDNA stock . solution (see Section S1.2) was mixed and incubated for 30' with 5 m l antidigoxigenin-coated beads (0.5% w/v, 3.15 m m diameter) and 14 m l TE (Tris 10 mM, EDTA 1 mM, 0.01% NaN 3, pH7.5) 500 mM NaCl. The sample was then diluted to 1 ml in TE 100 mM NaCl, 0.1 mg/ml BSA (New England Biolabs). The sample was incubated at least 30 min with BSA before starting the experiments to passivate the bead surface. A sample of streptavidin-coated beads was also prepared by diluting 1 m l of beads (0.5% w/v, 1.87 m m diameter, Kisker Biotech) in 1 ml TE 100 mM NaCl, 0.1 mg/ml BSA. Beads were differentially flowed into the fluidics chamber through lateral channels, and a dsDNA tether between an optically trapped bead and a pipette-subjected bead was created as previously described.
The molecule was submitted to several stretch/relaxation cycles up to a maximum pulling force of 45 pN and a minimum molecular extension of ∼ 4 5 . m mto characterize its elastic properties. At least three consecutive pulling cycles were fitted to the Inextensible Worm-Like Chain (WLC) in the low-force regime (F ≤ 5 pN), and to the extensible WLC in the high-force regime (F ≤ 40 pN).
After flowing KF, the low-force region (F ≤ 5 pN) of the release cycles was still well described by the inextensible WLC model. The contour length and persistence length of the molecule were then obtained as an average of at least four pulling cycles. Higher force-data was not used for the fits, as data is not well described by the extensible WLC model after flowing KF.
For the experiments in which KF was flowed without keeping the molecule at a constant force,
the DNA was maintained at an end-to-end distance of 6 m m(flow-rate: 9 m l/min). This corresponds to a drag force on the trapped bead of ∼ 14 pN perpendicularly to the stretching direction. In these experiments the flow was stopped after 5, 15 and 30 min and stretching curves were collected at a pulling speed of 500 nm/s.
For the constant force experiments a lower flow-rate of 3 m l/min was used to maintain the overall tension on the DNA molecule similar to the force on the stretching direction. This buffer flow corresponds to a drag force of ∼ 4 pN perpendicular to the stretching direction. To accurately perform the constant force experiments at 1 pN, we flowed KF keeping the DNA molecule at a higher tension of 3 pN. Once the chamber was fully equilibrated the flow was stopped and the force set at 1 pN.
## S1.1.2 DNA unzipping experiments
For the DNA unzipping experiments a 6.8-kb DNA hairpin sample (see Section S1.2) was tethered by bringing the bead in the optical trap in close contact to the bead on the micropipette due to the short length of the hairpin handles. The DNA and beads incubations were prepared following the same protocol as for the DNA stretching experiments, the only difference being the specific DNA concentration required to optimize the tethering of a single DNA molecule between beads. The DNA hairpin was then fully unzipped in TE 100 mM NaCl at a pulling rate of 50 nm/s, and maintained at a position in which at least half of the hairpin stem remained unzipped. The molecule was then rinsed with KF for at least 3 min (flow-rate: 9 m l/min). The peptide flow was then stopped and unzipping/rezipping curves recorded.
## S1.1.3 Experiments with ssDNA
For the ssDNA experiments a 6.8-kb DNA hairpin was prepared and tethered as for the unzipping experiments. The tethered hairpin was fully unzipped in a buffer (TE 100 mM NaCl) that contains a 250 nM concentration of a 30-base oligonucleotide complementary to the loop and its flanking region (see Table S2 for oligonucleotide sequence, oligo name: Blockloop30). The annealing of an
oligonucleotide to the tethered hairpin stabilizes the ssDNA form over the dsDNA form down to forces close to 1 pN. The elastic response of the ssDNA molecules was then recorded at a pulling rate of 100 nm/s in the force range 5 pN < F < 20 pN and characterized by means of the FJC model (see Section S5). Then, the molecule was maintained at a constant force by means of the force-feedback while a freshly prepared 40 m MKFsolution was flowed (flow-rate: 3 m l/min).
## S1.2 DNA substrates preparation
For the DNA stretching experiments we prepared a 24508-bp DNA molecule with biotin and digoxigenin tags at the 3'-ends of the molecule. The DNA template was prepared by cleaving N6-methyladenine free l -DNA (New England Biolabs) with Xba I restriction enzyme, and purified using Wizard DNA clean-up system kit (Promega). The digoxigenin tag was prepared by annealing to the cosL end of l -DNA an oligonucleotide tailed at its 3'-end with digoxigenin-labeled dUTP's using terminal transferase (Roche). The biotin tag was prepared by annealing two complementary oligonucleotides designed to create an Xba I cohesive end at one side. One of the oligonucleotides was tailed at its 3'-end with multiple biotins with biotin-labeled dUTP's using terminal transferase (Roche). Oligonucleotides were purified after tailing steps using the QIAquick Nucleotide Removal Kit (Qiagen). The digoxigenin tag was annealed to the half l -DNA molecule by incubation for 10 min at 68 ◦ C in a 10-fold excess oligonucleotide. Then, the biotin tag was annealed by incubation for 1h at 42 ◦ C (using 20-fold excess oligonucleotides) followed by cooling down to room temperature. Ligation was performed as an overnight reaction at 16 ◦ C using T4 DNA ligase (New England Biolabs).
For the DNA unzipping experiments a 6838-bp DNA hairpin with a tetraloop at one end and two 29-bp dsDNA handles at the other end was prepared. The synthesis is based on a previously described method S1 but was modified to introduce double digoxigenin and biotin tags at each handle to enhance tether lifetimes (Figure S8). Briefly, N6-methyladenine free l -DNA (New England Biolabs) was digested with Bam HI, phosphorylated at its 5'-ends with T4 polynucleotide kinase (New England Biolabs), and purified using Wizard SV Gel and PCR clean-up system (Promega).
The 6770-bp restriction fragment contained between positions 41733 and 48502 (cosR end) was used as the stem of the DNA hairpin. To create the end-loop of the molecule, the stem was annealed to an oligonucleotide (BamHI-loop2) that self-assembles in a hairpin structure with a tetraloop at one end, and an overhang complementary to the Bam HI restriction site on the other end. To create the dsDNA handles, we used two partially complementary oligonucleotides (cosRlong and Bio-cosRshort3) that hybridize forming a protruding end complementary to the cosR end. The Bio-cosRshort3 oligonucleotide was purchased 5'-biotinylated, and the cosRlong oligonucleotide was tailed with multiple digoxigenins at its 3'-end as previously explained. To create the doubly biotynilated dsDNA handle, a third oligonucleotide (splint3) complementary to the unpaired regions of BIO-cosRshort3 was purchased and tailed at its 3' end with multiple biotins with biotinlabeled dUTP's using terminal transferase (Roche). To create the digoxigenin dsDNA handle we used a modified oligonucleotide (inverted-splint, Thermo Scientific) complementary to the unpaired region of cosRlong. This oligonucleotide contains two modifications: a C3 spacer at its 3' end to block this end in tailing reactions, and a polarity inversion at its 5' end using a 5'-5' linkage. This end can therefore be tailed with digoxigenin-labeled dUTP's using terminal transferase (Roche). In this way, both ends of the handle could be tailed with multiple digoxigenins (Figure S8). Oligonucleotides were purified after tailing steps using the Qiaquick Nucleotide Removal Kit (Qiagen). The oligonucleotides were annealed to the 6770-bp stem by incubation for 10 min at 70 ◦ C , followed by incubation for 10 min at 55 ◦ C and cooling down to room temperature. Ligation was performed as an overnight reaction at 16 ◦ C using T4 DNA ligase (New England Biolabs).
The sequence of the oligonucleotides used for the hairpin synthesis, together with the 30-base oligonucleotide used to generate the ssDNA template are specified in Table S2.
## S1.3 Kahalalide F sample preparation
Stock solutions containing 2 mM KF were prepared by dissolving a ∼ 1-2 mg sample of pure lyophylized Kahalalide F (gently provided by Pharmamar) in 100% DMSO (Sigma, Molecular Biology Grade). Stock solutions were then vortexed at low speed for 20 min and filtered with a
0.2 m m pore PTFE membrane filter (Millipore) that was pre-rinsed with 300 m l DMSO to reduce filter extractables. Sample concentration after stock filtration was verified by means of HPLC: An aliquot before and after filtering was collected and stored in ACN:H20 (50:50). These samples were sequentially processeÄd in an HPLC using a 5%-100% acetonitrile gradient. The elution peak was monitored at 220 nm, and the integrated area was compared between both samples. No sample loss happened during filtration (Figure S9). Ready-to-use Kahalalide F stock solutions were stored at -20 ◦ C as 20-60 m l aliquots. To prepare diluted working solutions, a stock solution aliquot was thawed and sequentially diluted in TE 100 mM NaCl ( e.g. 40 m M). The final DMSO concentration was always adjusted to 2%. Once thawed, stock solutions were not stored again to avoid the problems related to freeze-thaw cycles. S2 S3 , Samples were always prepared in glass vials, and a glass syringe used for sample filtering. Use of glass material was preferred to avoid material leaching from disposable plasticware, S4 and the good performance of glass infusion devices for low concentration samples of KF in clinical practice. S5
A mass spectrum of a 40 m M KF sample in 100 mM NH4OAc buffer pH7.0 (2% DMSO) is shown in Figure S10. The concentration of NH4OAc was set to obtain an ionic strength comparable to that used in the optical tweezers experiments.
## S1.3.1 Mass Spectrometry Methods
Sample are introduced using an Automated Nanoelectrospray. Triversa NanoMate (Advion BioSciences, Ithaca, NY, USA) sequentially aspirated the sample from a 384-well plate with disposable, conductive pipette tips, and infused it through the nanoESI Chip, which consists of 400 nozzles in a 20x20 array. Spray voltage was 1.75 kV and delivery pressure was 0.5 psi. Mass Spectrometer: Synapt HDMS (Waters, Manchester, UK). Samples were acquired with Masslynx software v.4 SCN 639 (Waters). MS Conditions for TOF results: NanoESI. Positive mode TOF. V mode. Sampling cone: 20 V. Source temperature: 20 C. Trap Collision Energy: 10. ◦ Transfer Collision Energy: 10. Trap Gas Flow: 8 ml/min. Vacuum Backing pressure: 5.89 mbar. m/z range: 300 to 5000, 500-15000. Instrument calibrated with CsI (external calibration). MSConditions for
Ion Mobility results: NanoESI. Positive mode Ion Mobility mode. V mode. Sampling cone: 20 V. Source temperature: 20 C. Trap Collision Energy: 10. Transfer Collision Energy: 10. Trap Gas ◦ Flow: 8 ml/min. IMS Wave Velocity: 300 m/s. IMS Wave Height: 9.5 V. Transfer Wave Velocity: 200 m/s. Transfer Wave Height: 8 V. Vacuum Backing pressure: 5.89 mbar. m/z range: 300 to 5000, 500-15000. Instrument calibrated with CsI (external calibration).
## S1.4 Dynamic Light Scattering Methods
A Photon Correlation Spectrometer (PCS) 3D from LS INSTRUMENTS was used for DLS measurements. The instrument is equipped with a He-Ne laser (632.8 nm). Measurements of at least 90 s were recorded at an angle of 90 . ◦ The hydrodynamic radius was calculated by a manual exponential fitting of the first cumulant parameter. Standard deviations were calculated from the second cumulant. The measurement temperature of 25 ◦ C was maintained by a decaline bath, which matches the refractive index of glass and does not therefore interfere with the measurement. The evolution of the hydrodynamic radius was observed during 1 hour. The time indicated corresponds to the minutes past after preparation of the sample and the beginning of the measurement.
Zeta potential measurements were carried out at 25 ◦ C with a Malvern Instrument Zetasizer Nano Z by laser Doppler electrophoresis. Disposable polystyrene cells were used. Solutions of peptide, DNA and the mixture were measured at the same concentration as the light scattering measurements were performed. Commercially available l -DNA (New England Biolabs) was used for the experiments.
## S2 DLS measurements of KF and KF-DNA complexes
To characterize the size of KF aggregates we performed DLS measurements. The hydrodynamic radius of KF particles one minute after sample preparation was 170 ± 20 nm (Figure S11 a , red) and showed a high polydispersity. The hydrodynamic radius increased linearly with time with a growing rate of 3.2 ± 0.6 nm/min, indicative of a rate of aggregation proportional to the surface
of the aggregate. We attribute such growth to the hydrophobic interactions between peptides. However, the hydrodynamic radius of KF-DNA mixtures remained constant within experimental errors during 60 min after sample preparation (Figure S11 a , black). This may be explained by the stabilizing effect induced by the added DNA. In fact, the average size of the particles could be stabilized by adding DNA at a latter time-stage (Figure S11 b , arrow). DNA as a strongly charged polyelectrolyte interacts with the positive charge of KF and might form an anionic, water-soluble shell around the peptide. We measured the zeta potential of KF particles finding a low positive value (Table S1) in agreement with the single positively charged residue of the peptide. A higher negative value was obtained for KF-DNA mixtures. KF oligomerization was also observed with size-exclusion HPLC (Figure S12).
## S3 Simulation of KF-DNA stretching experiments
As described in the main text, KF-DNA stretching curves are simulated using a set of N noninteracting two-states systems (Figure 4 d , Inset) that describe the contacts made between KF and DNA. Each contact can be in two conformations: formed or dissociated. The formed and dissociated states of contact i th have extensions corresponding to 0 and xi respectively. Initially the N contacts are found in the formed conformation. As the DNA is stretched and the force increased, the tilting of the free-energy landscape towards larger extensions favors the dissociated conformation, releasing an extension xi .
The parameters that best reproduce the experimental force-extension curves are described in the main text. A set of figures in which a different parameter is modified in each panel (the others being kept at the optimal value) is presented in Figure S13 a-d . In each panel, the optimal simulation is presented in black, whereas simulations with varying values of the modified parameters are presented in colors.
For D G 0 ≤ 5 kBT most of the contacts remain in the dissociated conformation (60% for D G 0 = 5 kBT ) at the end of a simulated pulling cycle (minimum extension of ∼ 4.5 m m). This high fraction
of dissociated contacts cannot explain the sawtooth pattern observed in the force-extension curve (FEC) in subsequent pulls. This suggests that a value of D G 0 ≤ 5 kBT is too low to reproduce the experimental curves. On the other hand, for D G 0 ∼ 10 kBT more than 90% of the contacts are formed again at the end of a pulling cycle (minimum extension of ∼ 4.5 m m), in agreement with the experimental results both at low (50 nm/s) and high pulling speeds (500 nm/s). Finally, for free energies greater than 15 kBT a sawtooth pattern is also observed in the relaxation curve, in disagreement with the releasing part of the observed experimental FEC for the same range of pulling speeds (Figure S13 a ).
Experimental force-extension curves show a broad distribution of rupture forces with most rupture events occurring at forces lower than 20 pN, but with a large rightmost tail (Figure S5 a ). This phenomenology can be well reproduced by considering: (i) that the transition state is close to the formed conformation (Figure S13 b ); and (ii) that the process is characterized by a disordered ensemble of barriers rather than a single valued barrier. This structural disorder is introduced in the form of an exponential distribution of barriers with a rightmost exponential tail of width w ′ ≥ 1 kBT (Figure S13 c ).
To model the experimental force-extension curves, we have assumed that the distribution of released lengths ( xi ) follows the experimental distribution shown in Figure 4 b . In Figure S13 d a simulation using the experimental distribution is compared to simulations in which the system is characterized by a single contact length ( xi ) rather than an exponential distribution. Finally the optimal distributions for the different parameters of the simulation are shown in Figure S13 e .
## S4 Measurement of tether stiffness from trap-distance fluctuations
A bead confined in an optical trap fluctuates around its equilibrium position due to thermal fluctuations. By the equipartition theorem the fluctuations of the bead position along the stretching direction are directly related to the effective stiffness of the system formed by the tethered molecule
and the optical trap: S6
$$< \delta y ^ { 2 } > = < y ^ { 2 } > - < y > ^ { 2 } = \frac { k _ { B } T } { k _ { \text{trap} } + k _ { \text{mol} } } \ \.$$
In an ideal force-feedback the stiffness of the trap vanishes and the bead fluctuations are only determined by the stiffness of the tether:
$$< \delta y ^ { 2 } > = \frac { k _ { B } T } { k _ { \text{mol} } } \ \.$$
In this case the fluctuations of the position of the optical trap are expected to match those of the bead ( < d y 2 > ). However, in our experimental set-up a finite frequency feedback of 1 kHz is used ( i.e. the force-feedback corrects the position of the optical trap by moving the piezoelectric actuators at a 1 kHz rate), and Eq. (S2) is not satisfied. To extract the value of k mol we followed a phenomenological approach that uses a modified version of Eq. (S2) containing a proportionality constant c :
$$< \delta y ^ { 2 } > = c \frac { k _ { B } T } { k _ { \text{mol} } } \ \.$$
The constant c includes all effects of the finite frequency of the feedback. The value of c has been obtained by fitting a set of measurements of ssDNA and dsDNA tethers at different average forces. The fluctuations in the position of the optical trap remain inversely proportional to k mol in the investigated range of stiffness (1 -10 · 10 -3 pN/nm) (Figure S14 a ), verifying the validity of the method based on Eq. (S3). This calibration method has been used to estimate the changes in stiffness that KF induces in ssDNA. The stiffness of the molecule during the peptide flow is measured in short time-windows to ensure reliable measurements of < d y 2 > .
Data is recorded at 1 kHz and low frequencies are filtered out to remove instrumental drift as experiments run for long times (30 min). Filtering low frequencies is also important to correct the changes in extension due to the compaction of the DNA with KF. A time-window of 3 s has been found to be the optimal value to reduce drift without affecting the measurements (vertical
dashed line in Figure S14 b ). Using a shorter time-window removes fluctuations that are relevant to determine the stiffness of the tether due to the large autocorrelation time shown by the data (Figure S14 ). c
## S5 Elastic properties of ssDNA before incubation with KF
Before flowing KF, pulling curves of the ssDNA molecule were recorded. The relative molecular extension was determined by subtracting the trap compliance (F/k) to the absolute displacement of the optical trap (trap stiffness: 70 pN/ m m). To obtain the absolute molecular extension, the elastic response of the tethered molecule was aligned to a reference unzipping trajectory in the range of forces from 15 to 20 pN (ssDNA response of the molecule). This method allowed to accurately calibrate the absolute extension of the molecule within less than 50 nm.
The method used to prepare the ssDNA template leaves a very short dsDNA stretch (88-bp) in relation to the length of the ssDNA chain (13650 bases). As the rigidity of dsDNA is ∼ 70-fold that of ssDNA, pulling curves mainly reflect the elastic response of the ssDNA region. At a given salt concentration, pulling curves of ssDNA can be modeled with the Freely-Jointed Chain (FJC) model: S1
$$x ( F ) = d _ { 0 } N _ { b } \left [ \frac { 1 } { \tanh ( \frac { F b } { k _ { b } T } ) } - \frac { k _ { b } T } { F b } \right ]$$
where: b : Kuhn length , d 0: Interphosphate distance, Nb : Number of bases.
Pulling curves were fitted to the FJC model finding average values of b = 1 57 . ± 0.05 nm and d 0=0 57 . ± 0 03 nm (N=5), compatible with previously reported results for the same ionic con-. dition. S1 S7 , Data was fitted on the force-range 5-25 pN, as experiments were performed at a salt concentration ([NaCl]=100 mM) at which secondary structure formation is not observed at low forces.
## S6 Supplementary Figures
Figure S1: Reproducibility of KF-DNA force-extension curves. ( a ) Pulling cycles of a 24kb DNA molecule before (black) and after flowing 40 m M KF at different waiting times: 5 min (purple), 15 min (green), 30 min (blue). ( b ) Pulling cycles of a 24-kb DNA molecule before (black) and after incubation with 40 m MKF at different waiting times of the interaction: 5 min (purple), 15 min (green), 30 min (blue). Similar trends to those reported in Figure 2 a are seen. Raw data was obtained at 1 kHz acquisition rate (light colors) and filtered to 10 Hz bandwidth (dark colors). Pulling speed is 500 nm/s.
Figure S2: The ornithine residue is essential for DNA binding. Stretching curves of a 24-kb DNA molecule before (black) and after (red) flowing a KF analog (40 m M) in which the ornithine residue has been replaced by a glutamic acid. The characteristic sawtooth pattern induced by KF is not observed, and compatible values for the elastic parameters are found if force-extension curves are fitted to the WLC model before and after flowing the analog. Data is filtered at 10Hz bandwidth.
Figure S3: Effect of ionic strength on KF-DNA interaction. (a) 10 mM NaCl. (b) 100 mM NaCl. (c) 250 mM NaCl. (d) 500 mM NaCl. In panels (a-d) we show representative experiments at each salt condition in which a DNA molecule is pulled before (black) and after flowing 10 m M KF (color). After flowing the peptide, the DNA molecule was repeatedly pulled between a maximum force of 45 pN and a minimum extension of 6 m m (purple), 4 m m (green) and 3 m m (blue). At the highest ionic strength (500 mM NaCl) we did not observe binding of the peptide to DNA, except for one experiment in which we obtained the results shown in panel d (green curve). Data is filtered at 50 Hz bandwidth, pulling speed v=500 nm/s. We performed at least 5 experiments at each condition. (e) Apparent contour length of the DNA molecule ( l 0 ∗ ) relative to its original extension ( l 0=8.3 m m), after being repeatedly pulled between a maximum force of 45 pN and a minimum extension of 4 m m. The degree of compaction of the DNA molecule increases with decreasing ionic strength (mean ± SD, N=5, except for 500 mM in which we only observed binding in one experiment).
Figure S4: AFM imaging of KF in the absence of DNA. ( a ) 100 m M KF immediately after dilution in aqueous buffer. ( b ) 100 m MKFincubated for 30 min at room temperature. Aggregation spots are occasionally observed on the surface at both incubation times. Bar scale is 600 nm. Color scale (from dark to bright) is 0-2 nm in all AFM images.
Figure S5: Analysis of KF-DNA unpeeling events ( a ) Average rupture force of KF-DNA contacts. The average rupture force is calculated as the mean of the average force immediately before and after an unpeeling event ( b ) Histogram of dissipated work in individual unpeeling events (calculated using equation (2) in main text). The histogram follows an exponential distribution of mean 23 ± 8 kBT (green). A rightmost tail corresponding to individual unpeeling events with W dissipated ≥ 150 kBT is observed. Insets show a log-normal plot (right) and the same plot with an enlarged range of W dissipated values (left). Unpeeling events with dissipated work as large as 400 kBT are observed. For both figures N=435 events from 3 molecules. Error bars are the statistical error measured between different molecules.
Figure S6: KFbinding does not change the overstretching transition. Four consecutive stretching curves (purple, green, blue and yellow) of a 24-kb DNA molecule after incubation with 50 m M KF. The molecule is fully overstretched at each pulling cycle. The sawtooth pattern at low extensions is clearly visible at each cycle. Data was filtered at 100 Hz bandwidth, v=1000 nm/s.
Figure S7: KF compacts ssDNA at 5 pN but not at 10 pN ( a ) A ssDNA molecule is maintained for more than 50 min at a constant force of 10 pN with a flow of KF, without observing a significant decrease in extension (blue). However, when the force is lowered to 5 pN (top arrow) a compaction equivalent to that reported in Figure 7 b is seen (bottom arrow). A negative control without peptide in the flowed buffer does not show DNA compaction (gray). Raw data is obtained at 1 kHz (light colors) and filtered at 1 Hz bandwidth (dark colors). ( b ) Average stiffness of ssDNA at 10 pN during the first 30 min of the peptide flow (red) compared to a negative control without peptide (black). Three individual experiments are shown in light red. In contrast to the results shown in Figure 7 b , the molecular stiffness remains constant in time.
Figure S8: Scheme of the DNA hairpin synthesis. The DNA hairpin is created by ligating a set of oligonucleotides (red, green, purple) to a 6.7-kb restriction fragment of l -DNA (black). The cosR end, and an Xba I cohesive end were respectively used to anneal the dsDNA handles and the end-loop to the l -DNA fragment. To create the biotinylated handle (green) one oligonucleotide was purchased 5'-biotinlylated (Bio-cosRshort3) and the other one was tailed with multiple biotins at its 3' end (splint3). To create a dsDNA handle with multiple digoxigenins at each strand we used an oligonucleotide containing a 5'-5' inversion and a blocked 3' end (inverted-splint). In this way the two oligonucleotides that create this handle (cosRlong, inverted-splint) could be tailed with multiple digoxigenins at the appropriate end.
Figure S9: KF stock filtration does not reduce sample concentration ( a ) Chromatogram of a freshly dissolved 2 mM KF stock solution before filtration (wavelength: 220 nm). The elution peak of KF is seen at t=6.4 min. ( b ) Chromatogram of a freshly dissolved 2 mM KF stock solution after filtration. The elution peak of KF is seen at t=6.5 min. The area and height of the filtered and non-filtered samples are comparable, indicating that KF concentration is not reduced due to filtration. The high peak at t=1 min corresponds to DMSO.
Figure S10: Mass spectrometry of Kahalalide F ( a ) Mass spectrum of a 40 m M KF sample. Peaks found at 1478 m/Z correspond to the singly charged ion and doubly charged dimer. The peaks at 967 m/Z and 511 m/Z correspond to a fragmentation reaction of the peptide at D-val/DPro. S8 ( b ) Zoom of the main peak of the mass spectrum from (a).
Figure S11: DLSmeasurements of KF and KF-DNA complexes. (a) (red) Hydrodynamic radius of KF particles in the buffer used for optical tweezers experiments (40 m MKF, 25 ◦ C , mean ± SD, N=9). KF forms nanometer-sized aggregates whose size grows linearly with time. The aggregation rate is obtained from a linear fit (red line). Significant differences are not seen between KF stock aliquots stored at -20 ◦ C or solutions freshly prepared from lyophilized KF. (black) Hydrodynamic radius of KF aggregates when DNA is added to the sample immediately after dilution (KF 40 m M, 48-kb l -DNA6.25 m g/ml, 25 ◦ C , mean ± SD, N=3). The size of the KF aggregates remains constant up to 1 hour after dilution, showing that DNA has a stabilizing effect on the size of the aggregates. A linear fit is shown in black. (b) The size of the KF aggregates (black) can also be stabilized by adding DNA (arrow) after 60 min (mean ± SD, N ≥ 3). A linear fit is shown to highlight the stabilization of the particle size after adding DNA.
Figure S12: Aggregation of KF dilutions. Size-exclusion chromatography of 40 m MKFdilutions at different waiting times: immediately after dilution (blue), after 30 min (green) and after 3 h (red). The width of the elution peak indicates peptide aggregation at all times.
Figure S13: Simulations of KF-DNA stretching experiments. ( a ) Simulations varying D G 0 (minimum value of the exponential distribution of D G values, width w = 1 kBT ) : D G 0=5 kBT (red), D G 0=10 kBT (black, optimal value), D G 0=15 kBT (green), D G 0=20 kBT (blue), D G 0=40 kBT (purple). ( b ) Simulations varying x † i : x † i =10 nm (red), x † i =8 nm (green), x † i =5 nm (blue), x † i =3 nm (purple), x † i =2 nm (black, optimal value), x † i =1 nm (yellow). ( c ) Simulations varying w ′ (exponential tail of the barrier B): w ′ =0 kBT (red), w ′ =1 kBT (green), w ′ =3 kBT (blue), w ′ =5 kBT (black, optimal value), w ′ =7 kBT (purple), w ′ =10 kBT (yellow) ( d ) Simulations varying xi (contact-length distribution): experimental distribution p xi ( ) = ( 1 / w ) exp [ -( xi -xi , 0 ) / w ] for xi ≥ xi , 0 ( p xi ( ) =0 otherwise) with xi , 0=8 nm and w =24 nm (black), xi = 10 nm (red), xi =30 nm (green), xi =100 nm (blue). Simulations are shifted by 1 m m for clarity. ( e ) Optimal distributions for the different parameters of the simulation (corresponding to the black lines in panels a-d).
Figure S14: Stiffness determination from trap-position fluctuations. ( a ) Phenomenological calibration of the molecular stiffness. The experimental variance of the trap position for a set of ssDNAand dsDNA molecules at different average forces was measured with the force-feedback protocol. The stiffness of the molecule (kmol) was determined as the derivative of the force-extension curve at each force. ( b ) Drift correction of the experimental data. The position of the trap is recorded at 1 kHz, and drift is removed by subtracting the average position using time-windows in the range 0.01-60 s (black). A window of 3 s (dashed line) is the optimal value to remove drift without distorting the measurement of the fluctuations. Data is fitted to an arctangent function (red). Measurements correspond to a 13-kb ssDNA molecule at 5 pN (N=5). ( c ) Autocorrelation function of the distance during two constant force experiments at 8 pN for a ssDNA molecule after subtracting the drift. Data with (red) and without (black) a buffer flow are shown.
## S7 Supplementary Tables
Table S1: Zeta potential measurements. Zeta pontential of KF (40 m M)and l -DNA(6.25 m g/ml, 0.2 nM) solutions as well as their mixtures in water at 25 ◦ C (mean ± SD, N ≥ 3). KF aggregates in water have a slighlty positive zeta potential -too low to prevent aggregation-, but suggesting that peptide aggregates might preferentially expose the positively charged residues on the surface. The addition of DNA to the sample shifts the zeta potential to negative values, further supporting that DNA binds to the peptide aggregates preventing their growth.
| | Zeta Potential (mV) | Electrophoretic Mobility ( m m · cm · V · s - 1 |
|--------|-----------------------|---------------------------------------------------|
| KF | 7 . 3 ± 0 . 3 | 0 . 6 ± 0 . 1 |
| DNA | - 46 . 1 ± 5 . 5 | - 3 . 4 ± 0 . 3 |
| DNA/KF | - 30 . 1 ± 1 . 2 | - 2 . 4 ± 0 . 1 |
Table S2: Oligonucleotides used for the DNA hairpin synthesis and to generate the ssDNA template.
| Oligonucleotide name | Oligonucleotide sequence |
|------------------------|---------------------------------------------------------------------------------|
| BamHI-loop II | 5'-GATCGCCAGTTCGCGTTCGCCAGCATCCG ACTACGGATGCTGGCGAACGCGAACTGGC-3' |
| cosRlong | 5'-Pho-GGGCGGCGACCTAAGATCTATTATATATGTG TCTCTATTAGTTAGTGGTGGAAACACAGTGCCAGCGC-3' |
| BIO-cosRshort3 | 5'-Bio-GACTTCACTAATACGACTCACTATAGGG AAATAGAGACACATATATAATAGATCTT-3' |
| splint3 | 5'-TCCCTATAGTGAGTCGTATTAGTGAAGTC-3' |
| Blockloop30 | 5'-TAGTCGGATGCTGGCGAACGCGAACTGGCG-3' |
## References
[S1] Huguet, J. M.; Bizarro, C. V.; Forns, N.; Smith, S. B.; Bustamante, C.; Ritort, F. SingleMolecule Derivation of Salt Dependent Base-Pair Free Energies in DNA. Proc. Natl. Acad. Sci. U.S.A. 2010 , 107 , 15431-15436.
[S2] Borchardt, R.; Kerns, E.; Lipinski, C.; Thakker, D.; Wang, B. Pharmaceutical Profiling in Drug Discovery for Lead Selection ; American Assoc. of Pharm. Scientists, 2005; Vol. 1.
- [S3] Kozikowski, B. A.; Burt, T. M.; Tirey, D. A.; Williams, L. E.; Kuzmak, B. R.; Stanton, D. T.; Morand, K. L.; Nelson, S. L. The Effect of Freeze/Thaw Cycles on the Stability of Compounds in DMSO. J. Biomol. Screening 2003 , 8 , 210-215.
- [S4] McDonald, G. R.; Hudson, A. L.; Dunn, S. M. J.; You, H.; Baker, G.; Whittal, R.; Martin, J.; Jha, A.; Edmondson, D. E.; Holt, A. Bioactive Contaminants Leach from Disposable Laboratory Plasticware. Science 2008 , 322 , 917.
- [S5] Nuijen, B.; Bouma, M.; Manada, C.; Jimeno, J. M.; Lazaro, L.; Bult, A.; Beijnen, J. Compatibility and Stability of the Investigational Polypeptide Marine Anticancer Agent Kahalalide F in Infusion Devices. Invest. New Drugs 2001 , 19 , 273-281.
- [S6] Forns, N.; De Lorenzo, S.; Manosas, M.; Hayashi, K.; Huguet, J.; Ritort, F. Improving Signal/Noise Resolution in Single-Molecule Experiments Using Molecular Constructs with Short Handles. Biophys. J. 2011 , 100 , 1765-1774.
- [S7] Smith, S.; Cui, Y.; Bustamante, C. Overstretching B-DNA: the Elastic Response of Individual Double-Stranded and Single-Stranded DNA Molecules. Science 1996 , 271 , 795-799.
- [S8] Stokvis, E.; Rosing, H.; Lopez-Lazaro, L.; Rodriguez, I.; Jimeno, J.; Supko, J.; Schellens, J. H. M.; Beijnen, J. H. Quantitative Analysis of the Novel Depsipeptide Anticancer Drug Kahalalide F in Human Plasma by High-Performance Liquid Chromatography under Basic Conditions Coupled to Electrospray Ionization Tandem Mass Spectrometry. J. Mass Spectrom. 2002 , 37 , 992-1000.
## Graphical TOC Entry
|
10.1021/nn4007237
|
[
"Joan Camunas-Soler",
"Silvia Frutos",
"Cristiano V. Bizarro",
"Sara de Loreno",
"Maria Eugenia Fuentes-Perez",
"Roland Ramsch",
"Susana Vilchez",
"Conxita Solans",
"Fernando Moreno-Herrero",
"Fernando Albericio",
"Ramon Eritja",
"Ernest Giralt",
"Sukhendu B. Dev",
"Felix Ritort"
] |
2014-08-05T19:00:15+00:00
|
2014-08-05T19:00:15+00:00
|
[
"physics.bio-ph",
"cond-mat.soft",
"q-bio.BM"
] |
Electrostatic Binding and Hydrophobic Collapse of Peptide-Nucleic Acid Aggregates Quantified Using Force Spectroscopy
|
Knowledge of the mechanisms of interaction between self-aggregating peptides
and nucleic acids or other polyanions is key to the understanding of many
aggregation processes underlying several human diseases (e.g. Alzheimer's and
Parkinson's diseases). Determining the affinity and kinetic steps of such
interactions is challenging due to the competition between hydrophobic
self-aggregating forces and electrostatic binding forces. Kahalalide F (KF) is
an anticancer hydrophobic peptide which contains a single positive charge that
confers strong aggregative properties with polyanions. This makes KF an ideal
model to elucidate the mechanisms by which self-aggregation competes with
binding to a strongly charged polyelectrolyte such as DNA. We use optical
tweezers to apply mechanical forces to single DNA molecules and show that KF
and DNA interact in a two-step kinetic process promoted by the electrostatic
binding of DNA to the aggregate surface followed by the stabilization of the
complex due to hydrophobic interactions. From the measured pulling curves we
determine the spectrum of binding affinities, kinetic barriers and lengths of
DNA segments sequestered within the KF-DNA complex. We find there is a capture
distance beyond which the complex collapses into compact aggregates stabilized
by strong hydrophobic forces, and discuss how the bending rigidity of the
nucleic acid affects such process. We hypothesize that within an in vivo
context, the enhanced electrostatic interaction of KF due to its aggregation
might mediate the binding to other polyanions. The proposed methodology should
be useful to quantitatively characterize other compounds or proteins in which
the formation of aggregates is relevant.
|
1408.1070v1
|
## ON THE EQUIVALENCE BETWEEN MV-ALGEBRAS AND l -GROUPS WITH STRONG UNIT
## EDUARDO J. DUBUC AND YURI A. POVEDA
Introduction. This note concerns the category M of MV-algebras and its full subcategory CM of totally ordered MV-algebras, for short, MV-chains , and the category G u of abelian l-groups with strong unit, for short l u -groups , with unital l-homomorphism as arrows, and its full subcategory CG u of totally ordered abelian l-groups with strong unit, for short l u -chains .
In [1] C.C. Chang proved that any MV -chain A was isomorphic to the segment A ∼ = Γ( A , u ∗ ) of a l u -chain A ∗ . He constructs A ∗ by the simple intuitive idea of putting denumerable copies of A on top of each other (indexed by the integers). Moreover, he also show that any l u -chain G can be recovered from its segment, since G ∼ = Γ( G,u ) ∗ , establishing an equivalence of categories. In [3] D. Mundici extended this result to arbitrary MV -algebras and l u -groups. He takes a representation of A as a sub-direct product of chains A i , and observes that A ↪ → ∏ i A ∗ i . Then he let A ∗ be the subgroup generated by A inside ∏ i A ∗ i . He proves that this idea works, and establish an equivalence of categories in a rather elaborate way by means of his concept of good sequences and its complicated arithmetics. In this note, essentially self-contained except for Chang's result, we give a simple proof of this equivalence taking advantage directly of the arithmetics of the product l -group ∏ i A ∗ i , avoiding entirely the notion of good sequence.
## 1. Preliminaries on l -groups and MV-algebras
In this section we refer the reader to the monograph [2] rather than to the original sources. We consider products ∏ i ∈ I X i to be the set of sections of the projection from the disjoint union ∐ i ∈ I X i π -→ I . In this way, the elements of the product are functions I σ -→ ∐ i ∈ I X i with values in the fibers, for i ∈ I, σ i ( ) ∈ X i , and the algebraic and order structures are the pointwise structures. De domain X of an injective morphism X ↪ → ∏ i ∈ I X i is a sub direct product of the family { X i } i ∈ I . An standard application of Zorn's lemma proves that abelian l -group and MV -algebras are subdirect products of the respective totally ordered structures. Given a l u -group ( G,u ) and MV -algebra A , we have the canonical sub direct product representations:
↦
$$\begin{array} {$$
↦
where Sp G ( ), Sp A ( ) are the sets of all prime l -ideals (resp. ideals). For each P we have the quotient ( G,u ) ρ -→ ( G/P,u P ), (resp. A ρ -→ A/P ) and ̂ x g ( P ) = ρ x ( ) = [ x ] P , (resp. ̂ a v ( P ) = ρ a ( ) = [ a ] P ). Note that ̂ u g > 0, but it
is not a strong unit for the product l -group, which in general will not be a l u -group (clearly when Sp G ( ) is infinite they are unbounded sections).
For G any abelian l -group, x ∈ G , set x + = 0 ∨ x, x -= 0 ∨-x . Then:
$$( 1. 2 ) \quad \ \ x = x ^ { + } - x ^ { - }$$
Given any l u -group ( G,u ), it is not straightforward but also not difficult to check that the segment Γ( G,u ) = [0 , u ] = { x ∈ G | 0 ≤ x ≤ u } is a MValgebra with operations defined as follows:
$$x \oplus y = u \wedge ( x + y ), \quad \neg x =$$
and that the lattice structure induced by G is the natural MV -algebra lattice structure. The segment Γ( G,u ) is defined on arrows by restriction and therefore determines a functor G u Γ -→ M . Recall now the following:
## 1.4. Theorem.
- a) ( [2, 7.2.4] ). For any l -ideal J of G J , ∩ [0 , u ] is an ideal of Γ( G,u ) , and the morphism below is well defined and an isomorphism.
↦
$$\Gamma ( G, u ) / ( J \cap [ 0, u ] ) \stackrel$$
- b) ( [2, 7.2.2, 7.2.3] ) For any l u -group ( G,u ) , the correspondence Sp G ( ) ∼ = -→ Sp (Γ( G,u )) , P -→ P ∩ [0 , u ] , is bijective.
↦
## Proof.
- a) For x ∈ G , 0 ≤ x ≤ u , x ∈ J ⇐⇒ x ∈ J ∩ [0 , u ]. This shows it is well defined and injective. For the surjectivity, take y ∈ G such that [ y ] J = [ x ] J . Then [ y ∧ u ] J = [ y ] J ∧ [ u ] J = [ x ] J ∧ [ u ] J = [ x ] J .
- b) This correspondence holds between the sets of all ideals, it is elementary but not straightforward, we refer the reader to [2, Theorem 7.2.2] for a detailed proof. It immediately follows from a) that J is prime if and only if ( J ∩ [0 , u ]) is prime. /square
We establish now the key result of this paper:
- 1.5. Theorem. Let G be any abelian l -group, u ∈ G,u > 0 , and let [0 , u ] be the segment [0 , u ] = { a ∈ G | 0 ≤ a ≤ u } . Let [0 , u ] ∗ ⊂ G be the subgroup generated by [0 , u ] . Then:
- a) [0 , u ] ⊆ { x ∈ G such that ∃ n ≥ 0 , | x | ≤ nu . }
- It follows that [0 , u ] is a l u -group with strong unit u
- ∗ ∗ .
- b) Given any sub MV -algebra A ⊂ [0 , u ] , let A ∗ ⊂ [0 , u ] ∗ be the subgroup generated by A , then A ∗ is a l u -group with strong unit u , and A = Γ( A , u ∗ ) .
- c) [0 , u ] ∗ = { x ∈ G such that ∃ n ≥ 0 , | x | ≤ nu . }
It follows that any l u -group is generated by its segment.
Proof. a): Let x ∈ [0 , u ] ∗ , x = ∑ n -1 k =0 ε k a k , a k ∈ [0 , u ] not necessarily distinct, and ε k = 1 or ε k = -1. Then | x | ≤ ∑ n -1 k =0 | ε k a k | = ∑ n -1 k =0 a k ≤ nu .
- b) It only remains to check that if 0 ≤ x = ∑ n -1 k =0 a k ≤ u , then x ∈ A . But this is clear since in this case x = ⊕ n -1 k =0 a k .
- c) It remains to prove the other inclusion ' ⊇ ' : Take a family { G i } i ∈ I of l -chains such that G is a subdirect product. We can safely assume that G is a subset G ⊂ ∏ G i . Let x ≥ 0 and n ≥ 0 be such that x ≤ nu . Let for
i ∈ I each i ∈ I n i ≥ 1 be such that ( n i -1) u i ( ) ≤ x i ( ) ≤ n u i i ( ). Clearly n i ≤ n for all i . Let a k = ( x -ku ) ∧ u ∨ 0, 0 ≤ k ≤ n -1. Then, ( ∑ n -1 k =0 a k )( ) = i ∑ n -1 k =0 ( x i ( ) -ku i ( )) ∧ u i ( ) ∨ 0( ) = ( i n i -1) u i ( )+ x i ( ) -( n i -1) u i ( )+0( ) = i x i ( ). Thus x = ∑ n -1 k =0 a k . But any x is the difference of two positive elements, x = x + -x -(1.2), so the result follows. /square
## 2. Chang's equivalence
Given any MV-chain A , Chang's construction consists in the set
$$A ^ { * } = \{ ( m, a ) \colon m \in \mathbb$$
ordered lexicografically, together with the following definitions:
$$( m + 1, 0 ) & = ( m, 1 ) \\ ( 2. 1 )$$
In [1, Lemma 5] the reader can find a careful and detailed proof that A ∗ is a l u -chain with strong unit u = (0 , 1). With this, the proof of the following is immediate:
2.2. Theorem. [1, Lemma 6] Given any MV-chain A and l u -chain ( G, u ) , the arrows:
$$\Gamma ( G, u ) ^ { * } \stackrel { \upsilon } { \$$
$$\colon \quad & A \stackrel { \imath } { \longrightarrow } \$$
are isomorphisms.
↦
Proof. That ı is an isomorphism is clear by construction of A ∗ . On the other hand, given x ∈ G , let n x be the integer such that n u x ≤ x ≤ ( n x + 1) u . Then, the function x -→ ( n , x x -n u x ), is an inverse for υ . /square
It is easy to verify that Chang isomorphisms actually determine an equivalence of categories.
- 2.3. Theorem. Chang's construction A ∗ can be defined on arrows and together with the segment Γ determine functors in such a way that ı , υ become natural transformations. Thus they establish an equivalence of categories:
/d47
$$\mathcal { C M } \xrightarrow { ( - ) ^ { * } } \math$$
/d47
/d47
/d47
Proof. Given A h -→ B in CM , define A ∗ h ∗ -→ B ∗ by h ∗ ( n, a ) = ( n, h a ( )). Clearly the equation ( h h 2 1 ) ∗ = h h ∗ 2 ∗ 1 holds. From definitions 2.1 it is straightforward to check that h ∗ is well defined and that it is a l u -morphism.
The naturality of ı and υ consists on the commutativity of the following diagrams:
/d47
/d47
/d31
/d127
/d47
/d47
/d47
/d47
/d15
$$\begin{array} {$$
/d15
/d15
/d15
/d15
/d47
/d15
/d15
/d15
/d47
/d31
/d127
/d47
/d47
/d15
/d15
/d47
/d47
that we check as follows: Let a ∈ A , then Γ( h ∗ ) ı ( a ) = Γ( h ∗ )(0 , a ) = h ∗ (0 , a ) = (0 , h ( a )) = ı h ( a . ) Let ( n, x ) ∈ Γ( G,u ) ∗ , then ϕυ n,x ( ) = ϕ nu ( + x ) = nu + ϕ x ( ) = υ n, ϕ x ( ( )) = υ n, ( (Γ ϕ )( x )) = υ (Γ ϕ ) ∗ ( n, x . ) /square
## 3. Extension of the functors, ( -) ∗ and Γ
Consider the following diagrams:
/d47
/d47
/d127
/d95
/d127
/d95
/d127
/d95
/d127
/d95
/d15
$$\begin{array} { c c c } \mathcal { C M } \xrightarrow { ( - ) ^ { * } } \mathcal { C G } _ { u }, & \quad \mathcal { C G } _ { u } \xrightarrow { \Gamma } \math$$
/d15
/d15
/d15
/d15
/d15
/d15
/d15
/d47
/d47
We will complete the diagram on the left by extending the Chang construction into a functor M -→ G ( -) ∗ u , and prove that this extends the equivalence in the first row into an equivalence in the second row.
Given any MV -algebra A we will define the l u -group A ∗ as a subdirect product of the family of l u -chains { ( A/P ) ∗ } P ∈ Sp A ( ) . We refer to the canonical subdirect product representation of A (1.1) and to Theorem 2.2. Consider the diagram:
/d47
/d47
/d47
/d47
/d47
/d47
/d47
/d47
/d19
/d19
/d68
/d68
/d48
$$A \underset { \underset { P \in S p ( A ) } { \overset { v } { \stackrel { v } { ( - ) } } } } { \overset { ( - ) } { \stackrel { v } { \stackrel { v } { ( - )$$
/d48
/d29
/d125
where A ◦ denotes the image subset of the composite arrow ı ̂ ( -) v , and we let A ∗ be the subgroup generated by A ◦ , and u = ı ̂ 1 . v For a ∈ A , ı a ̂ v ( P ) = ı ([ a ] P ), thus ı a ̂ v ≤ u , i.e. A ◦ ⊂ [0 , u ]. From Theorem 1.5 b) we have:
3.1. Theorem. A ∗ is a l u -group with strong unit u , and A ◦ = Γ( A , u ∗ ) . It follows that A ı ̂ ( -) v -→ Γ( A , u ∗ ) is an isomorphism. /square
- 3.2. Theorem. Chang isomorphisms υ in Theorem 2.2 on the fibers, restricted to Γ( G,u ) ∗ , determine an isomorphism Γ( G,u ) ∗ υ -→ ( G,u ) .
/d47
/d47
ON THE EQUIVALENCE BETWEEN MV-ALGEBRAS AND l -GROUPS WITH STRONG UNIT 5
Proof. Consider the diagram:
/d31
/d127
/d47
/d47
✤
/d47
/d47
$$& P r o o f. \text{ Consider the diagram:} \\ & \Gamma ( G, u ) ^ { * } \longleftarrow \prod _ { P \in S p ( G ) } ( \Gamma ( G, u ) / P \cap [ 0, u ] ) ^$$
✤
/d15
/d15
✤
/d47
/d15
/d15
/d47
/d47
/d47
The arrows on the top are justified by theorem 1.4 b) and a) respectively. We show now that going all the way right and down in the diagram factors through ( G,u ). It is enough to consider generators ı x ̂ v ∈ Γ( G,u ) ◦ , x ∈ Γ( G,u ). Then: ı x ̂ v ( P ) = ı ([ x ] P ∩ [0 ,u ] ), wich by 1.4 a) corresponds to ı ([ x ] P ) which in turn goes down to υ ı ([ x ] P ) = [ x ] P = ̂ x g ( P ). It is clear that Γ( G,u ) ∗ υ -→ ( G,u ) so determined is injective, and the surjectivity follows immediately from theorem 1.5, c).
$$( 3. 3 ) \ \text{Remark that for any } x \in \Gamma ( G, u ), \ v i \widehat { x } ^ { v } = x. \quad \begin{array} {$$
We show now that the construction of A ∗ can be defined on arrows in a functorial way. This follows by standard techniques in representation theory.
3.4. Theorem. Let A h -→ B a morphism of MV -algebras. Then there is a morphism of l u -groups A ∗ h ∗ -→ B ∗ such that the following diagram commutes:
/d47
/d47
/d31
/d127
/d47
/d47
/d31
/d127
/d47
/d47
/d31
/d127
/d47
/d47
✤
✤
✤
$$& \ m o r p h i s m o f { l } _ { u } \text{-groups } A ^ { * } \stackrel { \gamma } { \longrightarrow } B ^ { * } \text{ such that the following diagram commutes: } \\ & \quad \text$$
✤
✤
/d15
/d15
/d15
/d15
/d15
✤
/d15
/d15
/d15
✤
✤
/d47
/d15
/d47
/d31
/d127
/d47
/d47
✤
/d15
/d31
/d127
/d47
/d47
/d31
/d127
/d47
/d47
defines a functor M -→ G ( -) ∗ u in such a way that the isomorphisms ı ̂ ( -) v -→ Γ( A , u ∗ ) and Γ( G,u ) ∗ υ -→ ( G,u ) become natural transformations.
Proof. Let P ∈ Sp B ( ). Then h P ∈ SpA and there is a MV -algebra morphism A/h -1 P h | P -→ B/P well-defined by h | P ([ a ] h -1 P ) = [ h a ( )] P . From the left diagram in (2.4) we have a commutative diagram:
This A -1
/d47
/d47
/d15
/d15
/d47
$$\langle - \mu _ { \nu } \cdots \cdots - \mu _ { \nu } \cdots \cdots \cdots - \mu _ { \nu } \cdots \circ \mu _ { \nu } \cdots \\ A / h ^ { - 1 } P \xrightarrow { \imath } ( A / h ^ { - 1 } P ) ^ { * } \\ h | _ { P } \Big \downarrow \bigoplus _ { B / P } ^ { \imath } ( B / P ) ^ { * }$$
/d15
/d15
/d47
We define the right most vertical arrow in diagram (3.5) by h ∗ ( σ )( P ) = ( h | P ) ∗ ( σ h ( -1 P ), for any σ ∈ ∏ P ∈ Sp A ( ) ( A/P ) ∗ . It is straight- forward to check that h ∗ is a morphism of l -groups and that the equation ( h h 2 1 ) ∗ = h h ∗ 2 ∗ 1 holds. We prove now that it restricts to A ∗ -→ B ∗ . It is enough to check that it restricts to A ◦ -→ B ◦ , which amounts to show that the exterior of diagram 3.5 commutes: Let a ∈ A and P ∈ S p ( B ),
✤
✤
✤
✤
then h ∗ ( ı a ̂ v )( P ) = ( h | P ) ∗ ( ı a ̂ v ( h -1 P ) = ( h | P ) ∗ ( [ ı a ] h -1 P ) = ıh | P ([ a ] h -1 P ) = ı h a [ ( )] P = ıh a ̂ ( ) v ( P ), thus h ∗ ( ı a ̂ v ) = ı h ̂ ( a ) v . It only remains to show that υ is a natural transformation, that is, for any ( G,u ) ϕ -→ ( H,u ), υ (Γ ϕ ) ∗ = ϕυ , see the right diagram in (2.4). It is enough to check the equation on the generators ı x ̂ v ∈ Γ( G,u ) ◦ , x ∈ Γ( G,u ). Then: υ (Γ ϕ ) ∗ ı x ̂ v = υ ı ̂ (Γ ϕ x ) v = υ ı ϕx ̂ v = ϕx = ϕυıx ̂ v , this last two equalities by (3.3) in Theorem 3.2. /square
- 3.6 . A final comment. This finishes the proof that the functor G u -→ M establishes an equivalence of categories, by constructing explicitly a quasiinverse M-→G ( -) ∗ u directly related to Chang's original construction. Namely, we use the l -group determined pointwise by the Chang l -groups constructed on the fibers of the canonical subdirect product representation, and profit directly to be inside this large l -group in order to prove that the construction actually determines an equivalence of categories.
- Γ
3.7 . A word on the role of good-sequences. We refer to Theorem 1.5. Looking at the proof of item c), it is immediate to check by evaluating on each i ∈ I , that the sequence a k ∈ [0 , u ] that represents x as x = ∑ n -1 k =0 a k , a k = ( x -ku ) ∧ u ∨ 0, 0 ≤ k ≤ n -1, satisfies a k ⊕ a k +1 = a k . That is, it is a (canonical) good sequence attached to x , and it the unique such that represents x . This determines a bijection between the set of good sequences of A = [0 , u ] and the l u -group [0 , u ] ∗ = { x ∈ G such that ∃ n ≥ 0 , | x | ≤ nu } ⊂ G . This bijection transports the structure of G into a l u -group structure on the set of good sequences, which is cumbersome and alien to previous intuition. In our proof of the theorem we work directly in the l u -group A ∗ ⊂ G = ∏ P A ∗ P , P ∈ Sp A ( ), A P = A/P , instead of working with the good sequences and its complicated arithmetic. The subdirect representation theorem is anyway essential in both approaches as we know them today.
It would be interesting to find a constructive proof (no Zorn's lemma) of this equivalence. Note that the construction A ∗ = Gs A ( ) (set of good sequences) is there, and the definition of its arithmetics and strong unit is constructive. Probably we have an isomorphism of abelian groups Gs A ( ) = ∼ Free ( | A / | ) ∼ , the quotient of the free abelian group on the underlying set of A by the congruence generated by the pairs ( a, b ) ∼ ( a ⊕ b, a /circledot b ). In this way we actually have A ∗ , the challenge is to prove that it is an l -group, and that it is the required l u -group that yields the equivalence, without using that there are enough primes.
## References
- [1] Chang C.C., A new proof of the completenes completenesssof the Lukasiewicz axioms , Transactions of the American Mathematical Society, 88, p. 467-490.
- [2] Cignoli R., D'Ottaviano I., Mundici D., Algebraic Foundations of Many-valued Reasoning , Trends in Logic vol 7, Kluwer Academic Puplishers (2000).
- [3] Mundici D., Interpretation of AF C ∗ -algebras in Lukasiewicz sentential calculus , J. Funct. Anal. Vol. 65, 15-63 (1986).
| null |
[
"Eduardo J. Dubuc",
"Yuri A. Poveda"
] |
2014-08-05T19:05:31+00:00
|
2014-08-05T19:05:31+00:00
|
[
"math.LO"
] |
On the equivalence between MV-algebras and $l$-groups with strong unit
|
In "A new proof of the completeness of the Lukasiewicz axioms"} (Transactions
of the American Mathematical Society, 88) C.C. Chang proved that any totally
ordered $MV$-algebra $A$ was isomorphic to the segment $A \cong \Gamma(A^*, u)$
of a totally ordered $l$-group with strong unit $A^*$. This was done by the
simple intuitive idea of putting denumerable copies of $A$ on top of each other
(indexed by the integers). Moreover, he also show that any such group $G$ can
be recovered from its segment since $G \cong \Gamma(G, u)^*$, establishing an
equivalence of categories. In "Interpretation of AF $C^*$-algebras in
Lukasiewicz sentential calculus" (J. Funct. Anal. Vol. 65) D. Mundici extended
this result to arbitrary $MV$-algebras and $l$-groups with strong unit. He
takes the representation of $A$ as a sub-direct product of chains $A_i$, and
observes that $A \overset {} {\hookrightarrow} \prod_i G_i$ where $G_i =
A_i^*$. Then he let $A^*$ be the $l$-subgroup generated by $A$ inside $\prod_i
G_i$. He proves that this idea works, and establish an equivalence of
categories in a rather elaborate way by means of his concept of good sequences
and its complicated arithmetics. In this note, essentially self-contained
except for Chang's result, we give a simple proof of this equivalence taking
advantage directly of the arithmetics of the the product $l$-group $\prod_i
G_i$, avoiding entirely the notion of good sequence.
|
1408.1071v1
|
## D-Branes on Spaces Stratified Fibered Over Hyperbolic Orbifolds
A. A. Bytsenko ( a ) 1 , M. Chaichian ( b ) 2 , M. E. X. Guimar˜ aes ( c ) 3
- ( a ) Departamento de F´ ısica, Universidade Estadual de Londrina Caixa Postal 6001, Londrina-Paran´, Brazil a
- ( b ) Department of Physics, University of Helsinki P.O. Box 64, FI-00014 Helsinki, Finland
- ( c ) Instituto de F´ ısica, Universidade Federal Fluminense Av. Gal. Milton Tavares de Souza, s/n CEP 24210-346, Niter´i-RJ, Brazil o
## Abstract
We apply the methods of homology and K-theory for branes wrapping spaces stratified fibered over hyperbolic orbifolds. In addition, we discuss the algebraic K-theory of any discrete co-compact Lie group in terms of appropriate homology and AtiyahHirzebruch type spectral sequence with its non-trivial lift to K-homology. We emphasize the fact that the physical D-branes properties are completely transparent within the mathematical framework of K-theory. We derive criteria for D-brane stability in the case of strongly virtually negatively curved groups. We show that branes wrapping spaces stratified fibered over hyperbolic orbifolds carry charge structure and change the additive structural properties in K-homology.
Keywords: D-branes; methods of K-theory
Mathematics Subject Classification 2010: 12Gxx, 51Pxx, 81Txx, 81T30
1 [email protected]
2 [email protected]
3 [email protected]
## 1 Introduction
In this paper we will elucidate the observation that methods of algebraic K-theory and Khomology are appropriate tools in classyfy branes wrapping spaces stratified fibered over hyperbolic orbifolds. Generally speaking, D-branes are complicate objects and cannot be considered as just subspace in an ambient spacetime. Their dynamics require a more abstract mathematical tools (such as of a derived category [1], for example). Nevertheless in algebraic K-theory and K-homology D-branes can be realized naturally as an appropriate spacetime background with reveals various properties of their quantum dynamics that could not be otherwise detected if one classified the brane worldvolumes using only singular homology [2, 3].
This paper is outlined as follows. In Sect. 2 we select the requisite mathematical material used for investigation of spaces stratified fibered over orbifolds. Our purpose there is to discuss homology with twisted spectrum coefficients. The main attention we pay for an Atiyah-Hirzebruch type spectral sequence with it relation to (ordinary) homology.
Then we continue with algebraic K-theory of spaces stratified fibered over hyperbolic orbifolds. Before discussing D-branes we begin with some background material on stratified systems of (abelian) groups over hyperbolic manifolds, and homology groups in which the obstructions lie. We follow lines and conventions of [4] which apply throghout. We examine homology spectrum - Ω-spectrum, stratified system of abelian groups and virtually negatively curved groups.
In the course of paper we follow the lines of [2] and assume the folowing definitions: A Dbrane on background B is a triple [ X,F,ψ ], where X is a Spin C submanifold of B , interpreted as a D-brane in string theory, F is the Chan-Paton boundle on X and ψ : X ↪ → B is the embedding. A D-brane can be interpreted as the homology class of K-cycle [ X,F,ψ ] ∈ K ( B ); we assume that the base space B is a simply connected finite CW-complex (one can consider B to be only path-connected). There are many such situations in which one is interested in the classification of D-branes in X. In fact, many of the examples considered in literature fall into this category. For instance, both the lens spaces L p q , . . . , q ( ; 1 n ) and the real projective spaces R P 2 n +1 are circle bundles over C P n [2]. Furthermore, the very application of vector bundle modification identifies those D-branes such that one is a spherical fibration over the other.
In Sect. 3 we analyze the additive structure in the homology class of K-cycle and appropriate structural properties for D-branes wrapping spaces stratified fibered over hyperbolic orbifolds.
Finally in Sect. 4 we derive stability criteria for D-branes wrapping spaces stratified fibered over hyperbolic orbifolds trying to put physical notions in a mathematically rigorous setting as much as possible.
## 2 Branes on spaces stratified fibered over orbifolds
Let X be a closed (means compact and without boundary) connected Riemannian manifold with strictly negative sectional curvatures and let { X j } ∞ j =1 , X 1 ⊂ X 2 ⊂ X 3 ⊂ ... be a sequence of connected compact smooth manifolds. Let F be a finite group which acting on X via isometries and on each X j via smooth maps. (The action of F on X j need not to be free but on X is free.) Assume also the smooth embedding X j ⊂ X j +1 is both F -equivariant and j -connected. Let X ∞ = ⋃ ∞ j =1 X j and give X the direct limit topology. The induced action of F on X is free. Assuming that { M j } ∞ j =1 is the orbit space X × X j under the diagonal action of F , and
$$M ^ { \infty } = \varinjlim _ { j } M ^ { j } \,.$$
Let X be the orbit space X/F and { p j } ∞ j =1 : M j → X be the map induced from the canonical projection of X × X j onto X . The set { p j } is stratified systems of fibrations on X . Let as before G ( Y ) be the stable topolgical pseudo-isotopy Ω-spectrum associated to a topological space Y ; then we have to define a class of maps p ∞ .
Stratified system of abelian groups. With a stratified system of fibrations p : M →B one can associate a homology Ω-spectrum H ( B G ; ( p )) with tvisted spectrum coefficients and a map of Ω-spectra: H ( B G ; ( p )) -→ G ( M ). A functor G from spaces to spectra is said to be homotopy invariant provided a homotopy equivalence of spaces X → Y induces a homotopy equivalence of spectra G ( X ) →G ( Y ) [4]. A homology Ω-spectrum H ( B G ; ( p )) is defined and discussed in the lectures [5]. One can construct Atiyah-Hirzebruch type spectral sequence [4, 6]:
Suppose G is a homotopy invariant spectrum valued functor of spaces, and p : M → B is an stratified system of fibrations discussed above. Then the homotopy groups π q G ( p -1 ( x )) form stratified systems of groups over X , denoted by π q G ( p ), and there is a (homological) spectral sequence with
$$E _ { p, q } ^ { 2 } = H _ { p } ( \mathcal { B } ; \pi _ { q } \mathcal { G } ( p ) )$$
which abuts to π p + q H ( B ; π q ( G ( p ))), where π q G ( p x ( )) denotes the stratified system of abelian groups over B with the group π q G ( p -1 ( x )) corresponding to the point x ∈ B .
Let A be a stratified system of abelian groups. The homology groups H j ( B A ; ) can be calculated as follows. A triangulation of a compact manifold X consists of finite family of closed subsets {T } j n j =1 that cover X , and family of homeomorphisms ϕ j : T ′ j → T j , where each T ′ j is a Euclidean simplex. (In the case of compact surface B = S each T ′ j is a triangle in the plane R 2 , i.e., a compact subset of R 2 bounded by three distinct strait lines.) Let T be such a triangulation that each strata T j is a subcomplex. Then a stratified system A restricted to each open simplex is a constant system of coefficients and
$$\Phi _ { 1 } ^ { \quad } H ( \mathcal { B } ; \mathcal { A } ) = \bigoplus _ { j } ^ { \quad \lambda } H _ { j } ( X ; \mathcal { A } )$$
is the homology of finite chain complex
$$C \ \coloneqq \ C _ { 0 } \longrightarrow C _ { 1 } \longrightarrow \dots \longrightarrow C _ { n } \,, \\ C _ { j } \ \ = \ \bigoplus _ { s \in \mathcal { T } _ { j } } \mathcal { A } ( \widehat { s } ) \,.$$
Here ̂ s denotes the barycenter of s . In fact n -dimensional s -simplex can be decomposed on ( n +1)! small n -dimensional simplexes. Vertexes of new siplexes are the centers of gravity of faces of initial simplex. Set { x j } of centers is a set of vertexes of some simplex of barycentric subdivision if the corresponding faces can be composed to chain of consecutively enclosured faces.
The E 2 0 ,q -terms. These terms in spectral sequence (2.2) can be evaluated in the following manner. Let G be the extension of π 1 ( X ) determined by the action of F on X . In fact the group G is isomorphic to the factor group π 1 ( M /π ) 1 ( X ). Let F ( G ) be the category with objects the finite subgroups of G . For each element γ ∈ G we can determine a morphism G 1 → G 2 provided γG γ 1 -1 ⊆ G 2 . Then E 2 0 ,q -terms are isomorphic to the direct limit
$$E _ { 0, q } ^ { 2 } \xrightarrow { \text{isomorphism} } \lim _ { \stackrel { \longrightarrow } { G \in \mathcal { F } ( \mathfrak { G } ) } } \pi _ { q } ( \mathcal { G } ( X / p ( G ) ) ) \,,$$
where p : G → F is the canonical projection and X/p G ( ) is the orbit space. This isomorphism can be demonstrate by using of a basic result of E. Cartan :
- (i) First let us remind some conventions which will be required. Let A µ be the Yang-Mills connection associated with some quantum field model and let W be a (affine) space of all connections A µ . The action of group G on A µ is given by: A µ → gA g µ -1 + ∂ gg µ -1 , g ∈ G, where the infinite group G is the group of gauge transformations. Set of classes of equivalence of space W with respect to G action is called the orbit space . Remind also that if geodesic γ in X has ends belonging Y ⊂ X then it entirely lie in Y ; manifold Y satisfies this condition is called completely geodesic submanifold in X .
/negationslash
- (ii) Let us consider a map ϕ : G/ Γ → G, ϕ : g Γ → α g g ( ) -1 , where α -involutery authomorphism of group G satisfies: α 2 = Id, α = Id, ( G α ) 0 ⊂ Γ ⊂ G α . Here G α is a set of fixed points of the authomorphism α , and ( G α ) 0 is a component of unit of G α . First note that ϕ is the diffeomorphism G/ Γ on closed completely geodesic manifold X α : X α = { g ∈ G α g g | ( ) = Id } . A differential α in point Id , dα Id , is the involutery authomorphism of Lie algebra A of group G . The operator dα Id decomposes the space A on two subspaces P and Q associated with eigenvalues ± 1 of dα Id : P = {P ∈ A dα | Id ( P ) = P} , Q = {Q ∈ A dα | Id ( Q ) = -Q} .
It can be shown that the algebra Q of Lie group Γ is a P -orthogonal (with respect to scalar product on A ) to Q -subspace in A . The subspace P is invariant with respect to adjoint representation Ad (Γ) (see for detail [7]). The space X α is the closed submanifold in G , and dim X α = dim( G/ Γ). 4 Also X α is the completely geodesic submanifold in G . It is known
4 Indeed, the space X α is a kernel of the map ϕ . In order to find dim X α it is convenient to calculate the kernel of tangent map T ϕ g : T G g → T α g g ( ) G in a point g . Ker T ϕ g = gW g , where W ≡ W g = {P ∈ A dα | Id ( P ) + Ad( g ) P = 0 . It means that Ker } T ψ g = P , rank ϕ = dim A -dimKer ϕ = dim Q , and it is not depend on choice of a poit g . It follows that X α is the close submanifold in G and dim X α = dim( G/ Γ).
that geodesics (with respect to invariant metric) in space G/ Γ are orbits of one-parameter subgroups of group G [7]. Let γ be a geodesic in a group G tangent to X α in a point g : γ t = g e [ t P ], where e [ - ≡ ] exp( -), P ∈ A , and g · P ∈ T X g α = Ker T ϕ g = gWg . We have
$$& \alpha ( \gamma _ { t } ) \gamma _ { t } = \alpha$$
Let us proof that X α = Im ϕ . We have X α ⊃ Im ϕ ; it can be prooved that X α ⊂ Im ϕ . In X α there is a finite number N of points { g j } and geodesics { γ j } which joins g j -1 with g j , such that g N ∈ Im ϕ . Let g N -1 = g N e [ t P ], P ∈ W g N for a some number of t . Then we have:
$$& \alpha ( \mathbf e [ - ( 1 / 2 ) t \mathcal { P }$$
(iii) Finally repeating this procedure for g , j 0 ≤ j ≤ N -1 we get g 0 ∈ Im ϕ . Due to E. Cartan any orbit of a group G can be imbeded in G as a completely geodesic submanifold. For any compact subgroup Γ ⊂ G exists involutory authomorphism α , leaving subgroup Γ fixed, it follows that all orbits of group G can be realized as completely geodesic manifolds.
## 3 Strongly virtually negatively curved groups
Let X G B ( ; p ) be a cofiber (in category of spectra) which in fact is also an Ω-spectrum. In the case when X is a contractible the following result holds [6]:
$$\begin{array} {$$
$$K _ { n } ( \mathbb { Z } G _ { M } ) \otimes \$$
where π 1 ( M ) has been denoted by G M , G y = π 1 ( p -1 ( y )) for y ∈ B , Wh n is the Whitehead group and Wh n -/lscript ( G y ) ⊗ Q , K n -/lscript ( Z G y ) ⊗ Q are the suitable stratified systems of abelian groups over B . By definition a group G M is strongly virtually negatively curved if G M is isomorphic to π 1 ( M ), where M is constructed as before with constraint that X is contractible. Note that any discrete co-compact subgroup Γ of a Lie group G , where G is either O n, ( 1) , U n, ( 1) , Sp n, ( 1), or F 4 , is strongly virtually negatively curved.
Let B = G/ Γ be an irreducible rank one symmetric space and Γ ⊂ G a maximal compact subgroup. The following statement holds [6]: The algebraic K-theory of any discrete co-compact subgroup of a Lie group G , where G = O n, ( 1) , U n, ( 1) , Sp n, ( 1) , F 4 can be calculate in terms of the homology of the double coset space G M \B ≡ G M \ G/ Γ. As a consequence the following result follows:
$$K _ { n } ( \mathbb { Z } G _ { M } ) \otimes \$$
where p n is a stratified system of Q vector space over G M \ G/ Γ and the vector space p n ( G M g Γ) corresponding to the double coset G M g Γ is isomorphic to K /sharp ( Z ( G M ∩ g Γ g -1 )) ⊗ Q . In addition G M ∩ g Γ g -1 is a finite subgroup of G M . Any discrete co-compact subgroup of Lie group G , where G is either O n, ( 1) , U ( n, 1) , Sp n, ( 1) or F 4 , is strongly virtually negatively curved. Thus we obtain the additive structure in the homology class of K-cycle and appropriate structural properties for D-branes wrapping spaces stratified fibered over hyperbolic orbifolds.
## 4 Conclusions
The subject of this paper the intriguing relationship between D-branes on spaces stratified fibered over negatively curved orbifolds and algebraic K-theory. The mathematical methods are based in Atiyah-Hirzebruch type spectral sequence which we used for stratified system of (abelian) groups. By the use of this technique we obtain results in the structure of the K-homology of strongly virtually negatively groups. In summary we present discussion and main results obtained in this paper:
- (i) A D-brane wrapping homologically nontrivial cycle can nevertheless be unstable, if for some X j + k ⊂ X the following condition holds [8]: PD X ( j ⊂ X j + k ) = ω 3 ( X j + k ) + [ H ] | X j + k . In this equation the left hand side denotes the Poincar´ dual of e X j in X j + k . One can use a mathematical algorithm, the Atiyah-Hirzebruch spectral sequence (see Sect. 2), to determine which homology classes lift to K-theory classes (that is, to determine which D-branes are stable and which are not allowed) - it gives the rigorous stability criterion [2].
Suppose stratified system of fibrations on X extends through the spectral sequence as a non-trivial element of homology groups E 2 p,q . The following result is effective for the rational calculation of many K-groups (cf. Theorem 2 of [6]): If as before p : M →B is a stratified system of fibrations and X is aspherical with Wh n ( π X 1 × S 1 ) ⊗ Q = 0 for all integers n , then X G B ( ; p ) ⊗ Q = 0. The stratified system of fibrations could be extends through the spectral sequence as a non-trivial element of homology groups but it can have trivial lift to K-homology and it vanishes in E ∞ p,q .
- (ii) For each of the stratified systems of fibrations: p : M →B there is homotopy equivalence of Ω-spectra: G ( M ) = ∼ H ( B G ; ( p )) × X G B ( ; p . ) This equation is in conformity with an AtiyahHirzebruch type stectral sequence (2.2). The Atiyah-Hirzebruch spectral sequence keeps track of the possible obstructions for a homology cycle of starting from H p ( B ; π q ( G ( p ))) in (2.2) (the initial terms are given in (2.5)), to survive to E ∞ p,q . For a number of explicit examples, strongly virtually negatively curved G M , the groups G y = π 1 ( p -1 ( y )) are finite because they are isomorphic to subgroups of Γ. As a consequence one can use the extensive knowledge of the algebraic K-theory of finite groups [9, 10]. 5
5 For example we have the following result. If G M is a strongly virtually negatively curved then K n ( Z G M ) = 0, ∀ n < -1. In the case n = -1, K -1 ( Z G M ) /similarequal lim -→ G ∈F ( G M ) K -1 ( Z G M ) , K -1 ( Z G M ) is a finitely generated abelian group; it is generated by the images of K -1 ( Z G M ), as G varies over the finite subgroups of G M , under the map functorially induced by the inclusion of G into G M . (Perhaps this statement is true for a much larger class of groups G M . In particular, in [6] it has been conjectured that it is true for
- (iii) The homotopy groups π k G ( p -1 ) form stratified systems of groups π k G ( p ) over spaces X . D-branes wrapping spaces stratified fibered over hyperbolic orbifolds carry charge structure and change the additive structural properties in K-homology.
## Acknowledgments
AAB and MEXG would like to acknowledge the Conselho Nacional de Desenvolvimento Cient´ ıfico e Tecnol´gico (CNPq, Brazil) and Funda¸ao Araucaria (Parana, Brazil) for finano c cial support. The support of the Academy of Finland under the Projects No. 136539 and 140886 is gratefully acknowledged.
## References
- [1] M. R. Douglas, D-Branes, Categories and N = 1 Supersymmetry , J. Math. Phys. 42 (2001) 267-275; [arXiv:hep-th/0011017].
- [2] R. M. G. Reis and R. J. Szabo, Geometric K-Homology of Flat D-Branes , Commun. Math. Phys. 266 (2006) 71-122; [arXiv: hep-th/0507043].
- [3] A. A. Bytsenko, Homology and K-Theory Methods for Classes of Branes Wrapping Nontrivial Cycles , J. Phys. A: Math. Theor. 41 (2008) 045402; [arXiv:hep-th/0710.0305].
- [4] F. Quinn, Ends of Maps, II , Invent. Math. 68 (1982) 353-424.
- [5] A. K. Bousfield and D. M. Kan, Homotopy limits, completions, and localizations , Lecture Notes in Mathematics, 304 , Berlin-Heidelberg-New York, Springer 1972.
- [6] F. T. Farrell and L. E. Jones, Algebraic K-theory of spaces stratified fibered over hyperbolic orbifolds , Proc. Natl. Acad. Sci. USA, 83 (1986) 5364-5366.
- [7] S. Helgason, Differential Geometry and Symmetric Spaces , Pure and Applied Math. Ser. 12 , Academic Press 1962.
- [8] J. M. Maldacena, G. W. Moore and N. Seiberg, D-Brane Instantons and K-Theory Charges , JHEP 0111 (2001) 062; [arXiv:hep-th/0108100].
- [9] D. Quillen, Finite generation of the groups K i of rings of algebraic integers, Algebraic K-theory, I: Higher K-theories (Proc. Conf., Battelle Memorial Inst., Seattle, Wash., 1972), Lecture Notes in Math. 341 (1973) 179-198, Springer, Berlin.
- [10] A. Borel, Stable real cohomology of arithmetic groups , Ann. Sci. Ecole Norm. Sup. (4) 7 (1974) 235-272.
all finitely generated subgroups G M of GL n ( C )).
|
10.1142/S0217751X14501371
|
[
"A. A. Bytsenko",
"M. Chaichian",
"M. E. X. Guimarães"
] |
2014-08-05T19:06:37+00:00
|
2014-08-05T19:06:37+00:00
|
[
"hep-th",
"math-ph",
"math.MP"
] |
D-Branes on Spaces Stratified Fibered Over Hyperbolic Orbifolds
|
We apply the methods of homology and K-theory for branes wrapping spaces
stratified fibered over hyperbolic orbifolds. In addition, we discuss the
algebraic K-theory of any discrete co-compact Lie group in terms of appropriate
homology and Atiyah-Hirzebruch type spectral sequence with its non-trivial lift
to K-homology. We emphasize the fact that the physical D-branes properties are
completely transparent within the mathematical framework of K-theory. We derive
criteria for D-brane stability in the case of strongly virtually negatively
curved groups. We show that branes wrapping spaces stratified fibered over
hyperbolic orbifolds carry charge structure and change the additive structural
properties in K-homology.
|
1408.1072v1
|
## SOME NOTES ON ESAKIA SPACES
## DIRK HOFMANN AND PEDRO NORA
Dedicated to Manuela Sobral
Abstract. Under Stone/Priestley duality for distributive lattices, Esakia spaces correspond to Heyting algebras which leads to the well-known dual equivalence between the category of Esakia spaces and morphisms on one side and the category of Heyting algebras and Heyting morphisms on the other. Based on the technique of idempotent split completion, we give a simple proof of a more general result involving certain relations rather then functions as morphisms. We also extend the notion of Esakia space to all stably locally compact spaces and show that these spaces define the idempotent split completion of compact Hausdorff spaces. Finally, we exhibit connections with split algebras for related monads.
## Introduction
These notes evolve around the observation that Esakia duality for Heyting algebras arises more naturally when considering the larger category SpecDist with objects spectral spaces and with morphisms spectral distributors. In fact, as we observed already in [Hofmann, 2014], in this category Esakia spaces define the idempotent split completion of Stone spaces. Furthermore, it is well-known that SpecDist is dually equivalent to the category DLat ⊥ ∨ , of distributive lattices and maps preserving finite suprema and that, under this equivalence, Stone spaces correspond to Boolean algebras. This tells us that the category of Esakia spaces and spectral distributors is dually equivalent to the idempotent split completion of the category Boole ⊥ ∨ , of Boolean algebras and maps preserving finite suprema. However, the main ingredients to identify this category as the full subcategory of DLat ⊥ ∨ , defined by all co-Heyting algebras were already provided by McKinsey and Tarski in 1946.
In order to present this argumentation, we carefully recall in Section 1 various aspects of spectral spaces and Stone spaces which are the spaces occurring on the topological side of the famous duality theorems of Stone for distributive lattices and Boolean algebras. Special emphasis is given to the larger class of stably locally compact spaces and their relationship with ordered compact Hausdorff spaces. We also briefly present the extension of Stone's result to categories of continuous relations, an idea attributed to Halmos. These continuous relations and, more generally, spectral distributors, are best understood using the Vietoris monad which is the topic of Section 2. In particular, we identify adjunctions in the Kleisli category of the lower Vietors monad on the category of stably locally compact spaces and spectral maps which is then used to describe Esakia spaces as the idempotent split completion of Stone spaces. In Section 3 we use the facts presented in the previous section to deduce Esakia dualities using the technique of idempotent split completion. Moreover, we extend the notion of Esakia space to all stably locally compact spaces and deduce in Section 4 that the category of (generalised) Esakia spaces and spectral distributors is the idempotent split completion of the category of compact
Date : August 6, 2014.
2010 Mathematics Subject Classification. 03G05, 03G10, 18A40, 18C15, 18C20, 54H10 .
Key words and phrases. Boolean algebra, distributive lattice, Heyting algebra, dual equivalence, Stone space, spectral space, Esakia space, Vietoris functor, idempotent split completion, split algebra.
Partial financial assistance by Portuguese funds through CIDMA (Center for Research and Development in Mathematics and Applications), and the Portuguese Foundation for Science and Technology ('FCT - Funda¸ c˜o a para a Ciˆncia e a Tecnologia'), within the project PEst-OE/MAT/UI4106/2014, and by the project NASONI e under the contract FCOMP-01-0124-FEDER-028923 is gratefully acknowledged.
Hausdorff spaces and continuous relations. Finally, the idempotent split completion of Kleisli categories is ultimately linked to the notion of split algebra for a monad, which is the topic of Section 5.
## 1. Stone and Halmos dualities
The aim of this section is to collect some well-known facts about duality theory for Boolean algebras and distributive lattices and about the topological spaces which occur as their duals. As much as possible we try to indicate original sources.
Naturally, we begin with the classical Stone dualities stating (in modern language) that the category Stone of Stone spaces (= zero-dimensional compact Hausdorff topological spaces) and continuous maps is dually equivalent to the category Boole of Boolean algebras and homomorphisms (see [Stone, 1936])
$$\text{Stone} ^ { \text{op} } \simeq \text{Boo} ;$$
and that the category Spec of spectral spaces and spectral maps is dually equivalent to the category DLat of distributive lattices 1 and homomorphisms (see [Stone, 1938])
$$\text{Spec} ^ { \text{op} } \simeq D L a t.$$
We recall that a topological space X is spectral whenever X is sober and the compact and open subsets are closed under finite intersections and form a base for the topology of X . Note that in particular every spectral space is compact. A continuous map f : X → Y between spectral spaces is called spectral whenever f -1 ( A ) is compact, for every A ⊆ Y compact and open. A subset of a Stone space is compact if and only if it is closed, hence every Stone space is spectral and every continuous map between Stone spaces is spectral; that is, Stone is a full subcategory of Spec . Moreover, a spectral space X is a Stone space if and only if X is Hausdorff. Under the equivalence above, a spectral space X corresponds to the distributive lattice of compact and open subsets of X ordered by inclusion; if X is a Stone space, then the lattice of compact opens is actually the Boolean algebra of closed and open subsets. In the other direction, to a distributive lattice L one associates its prime spectrum spec L ; and spec L is Hausdorff if and only if L is a Boolean algebra. For a detailed presentation of these duality theorems and many of their consequences we refer to [Johnstone, 1986].
Another important aspect of spectral spaces is disclosed in [Hochster, 1969]: besides being the prime spectra of distributive lattices, spectral spaces are also precisely the prime spectra of commutative rings with unit. For a common study of lattice spectra and ring spectra we refer to [Simmons, 1980]. Hochster also constructs a right adjoint Spec → Stone to the inclusion functor Stone ↪ → Spec which associates to a spectral space X the topological space with the same underlying set and with the topology generated by the open subsets and the complements of the compact open subsets of X , this space is called the patch of X . Furthermore, in this paper we follow [Hochster, 1969] and consider the natural underlying order of a topological T0-space X defined as
$$x \leq y \text{ whenever } y \in \overline { \{ x \} },$$
which is equivalent to saying that the principal filter /squaresmallsolid x converges to y . This order relation is discrete if and only if X is T1.
For a general (not necessarily T0) topological space, this relation is still reflexive and transitive and leads to another important feature of the category Top of topological spaces and continuous maps: Top is a 2-category. Here, for continuous maps f, g : X → Y between topological spaces we write f ≤ g whenever f ( x ) ≤ g x ( ) for all x ∈ X , which defines the 2-cells in Top . Consequently, we consider also subcategories of Top as 2-categories; and note that this structure becomes trivial in CompHaus and Stone . The 2-categorical nature of Top leads us to consider the notion
1 We note that for us a lattice is an ordered set with finite suprema and finite infima, hence every lattice has a largest element /latticetop and a smallest element ⊥ .
of adjunction: for continuous maps f : X → Y and g : Y → X , we say that f is left adjoint to g , written as f /turnstileright g , if 1 X ≤ gf and fg ≤ 1 Y . Given f , there exists up to equivalence at most one such g , and in this case we call f a left adjoint continuous map.
Hochster also introduces a notion of dual space : for a spectral space X , the set X equipped with the topology generated by the complements of the compact open subsets of X is a spectral space whose underlying order is dual to the underlying order of X . Therefore we denote this space by X op , and it is not hard to see that ( X op ) op = X . Since every spectral map f : X → Y is also a spectral map of type X op → Y op , we obtain a 2-functor
$$( - ) ^ { \text{op} } \colon \text{Spec} \rightarrow \text{Spec} ^ { \text{co} }.$$
A different perspective on spectral spaces is offered by Priestley [1970, 1972]. Priestley showed that the category DLat is also dually equivalent to the full subcategory Priest of the category OrdCompHaus of ordered compact Hausdorff spaces (as introduced in [Nachbin, 1950]) and continuous monotone maps defined by all order-separated spaces (these spaces are nowadays called Priestley spaces ). We note that the compact Hausdorff topology of a Priestley space is necessarily a Stone topology. Hence, in an indirect way she showed that the categories Spec and Priest are equivalent,
## Spec /similarequal Priest .
A couple of years later, Cornish proved this fact directly in [Cornish, 1975] (see also [Fleisher, 2000]); in fact, both categories are shown to be isomorphic. Here a spectral space X corresponds to the Priestley space with the same underlying set, ordered by the underlying order of X , and equipped with the patch topology. In the other direction, a Priestley space X corresponds to the spectral space whose topology is given by all those opens of X which are also down-closed.
More generally, this construction does not only apply to Priestley spaces but indeed to all ordered compact Hausdorff spaces and defines an isomorphism between OrdCompHaus and the category StLocComp of stably locally compact spaces and spectral maps between them. Below we sketch this correspondence, for more information we refer to [Gierz et al. , 1980] and also to the more recent [Jung, 2004]. A topological space X is called stably locally compact if X is sober, locally compact and finite intersections of compact down-sets (with respect to the underlying order of X ) are compact. A continuous map f : X → Y between stably locally compact spaces is spectral whenever f -1 ( A ) is compact, for every A ⊆ Y compact and downclosed. Equivalently, a topological space X is stably locally compact if and only if X is T0, locally compact and every ultrafilter in X has a smallest convergence point with respect to the underlying order of X ; and a map f : X → Y between stably locally compact spaces is spectral (in particular continuous) if and only if f is monotone with respect to the underlying orders and, moreover, the diagram
/d47
/d47
/d15
/d15
/d15
/d15
/d47
/d47
commutes. Here U : Set → Set denotes the ultrafilter functor and the maps α : UX → X and β : UY → Y pick, for each ultrafilter, the smallest convergence point. Every spectral space is stably locally compact, and the two notions of spectral maps between spectral spaces are actually equivalent. We also point out that a continuous map f : X → Y between stably locally compact spaces which is left adjoint in Top is automatically spectral.
Every compact Hausdorff space is stably locally compact and every continuous map between compact Hausdorff spaces is spectral, which defines the inclusion functor
## CompHaus ↪ → StLocComp .
As for spectral spaces and Stone spaces, this functor has a right adjoint
## StLocComp → CompHaus
which sends a stably locally compact space X to the compact Hausdorff space with the same underlying set and the (generalised) patch topology, that is, the topology generated by the open subsets and the complements of the compact down-closed subsets of X . For X spectral, this topology coincides with the patch topology described above. Using this generalised patch topology, the correspondence between spectral spaces and Priestley spaces extends immediately: every stably locally compact space X defines an ordered compact Hausdorff space with the patch topology and the underlying order of X , and an ordered compact Hausdorff space X becomes a stably locally compact space where the topology is given by all down-closed opens of X . Clearly, OrdCompHaus is also a 2-category with the point-wise order on maps; and then the above described isomorphism
## OrdCompHaus /similarequal StLocComp
is an isomorphism of 2-categories. In terms of ordered compact Hausdorff spaces, the adjunction
/d40
/d40
/d104
/d104
/d40
/d40
/d104
becomes
/d104
Moreover, there is a 2-functor
$$( - ) ^ { \text{op} } \colon OrdCompHaus \rightarrow OrdCompHaus ^ { \text{co} }$$
which inverts the order relation of a compact Hausdorff space X ; and which induces a 2-functor
$$( - ) ^ { \text{op} } \colon StLocComp \rightarrow StLocComp ^ { \text{co} }$$
where X op turns out to be the space with the same underlying set as X and with the topology induced by the complements of the compact down-sets of X . If X is spectral, this notion of dual space coincides with the one of Hochster described above.
Another interesting generalisation of Stone's duality theorem is given in [Halmos, 1956]. Instead of continuous maps, Halmos considers continuous relations and shows that the category StoneRel of Stone spaces and continuous relations (called Boolean relations in [Halmos, 1956]) is dually equivalent to the category Boole ⊥ ∨ , of Boolean algebras with 'hemimorphisms', that is, maps preserving finite suprema but not necessarily finite infima (see also [Sambin and Vaccaro, 1988]). Similarly, the category SpecDist of spectral spaces and spectral distributors (respectively Priestley spaces and Priestley distributors, see [Cignoli et al. , 1991]) is dually equivalent to the category DLat ⊥ ∨ , of distributive lattices and maps preserving finite suprema.
We have not yet explained the meaning of continuous relation and spectral distributor, which is the subject of the next section.
## 2. Vietoris monads
Similarly to the fact that the category Rel of sets and relations can be seen as the Kleisli category of the power-set monad on Set , we will describe StoneRel and SpecDist as Kleisli categories of certain monads.
Before doing so, we recall the notion of 'monotone relation' between ordered sets. A relation r : X - → Y between ordered sets is called a distributor whenever, for all x, x ′ ∈ X and y, y ′ ∈ Y ,
↦
$$( x \, r \, y \, \& \, y \leq y ^ { \prime } ) \, \Rightarrow \, x \, r \, y ^ { \prime } \, \quad \text{ and } \quad \ ( x \leq x ^ { \prime } \, \& \, x ^ { \prime } \, r \, y ^ { \prime } ) \, \Rightarrow \, x \, r \, y ^ { \prime }.$$
Put differently, the corresponding map /rightanglenw r /rightanglene : X → PY from X into the powerset of Y has its image in the ordered set P ↑ Y of all up-closed subsets of Y (ordered by inverse inclusion), and the
restriction /rightanglenw r /rightanglene : X → P ↑ Y is monotone. We write r : X - → ◦ Y to indicate that r is a distributor. The relational composite of distributors is a distributor again, and the identity with respect to this composition law is the order relation on an ordered set X . We have thus described the category Dist of ordered sets and distributors which becomes a 2-category when considering the inclusion order of relations. We also note that Dist is isomorphic to the Kleisli category of the up-set monad on the category of ordered sets and monotone maps. A monotone map f : X → Y between ordered sets induces distributors f ∗ : X - → ◦ Y and f ∗ : Y - → ◦ X defined by
$$x \, f _ { * } \, y \, \text{ whenever } f ( x ) \leq y \quad \text{ and } \quad \ \ y \, f ^ { * } \, x \, \text{ whenever } y \leq f ( x )$$
respectively; that is, f ∗ = ≤ Y · f and f ∗ = f ◦ · ≤ Y . We also remark that f ∗ /turnstileright f ∗ in the ordered category Dist , in fact, every adjunction in Dist is of this form (see [Borceux and Dejean, 1986]).
Arguably, the topological counterpart to the up-set monad is the lower Vietoris monad /CE = ( V, e, m ) on Top which consists of the functor V : Top → Top sending a topological space X to the space
$$\ V X = \{ A \subseteq X \ | \ A \text{ is closed} \}$$
with the topology generated by the sets
$$U ^ { \hat { \cdot } } = \{ A \in V X \ | \ A \cap U \neq \varnothing \} \quad ( U \subseteq X \text{ open} ),$$
/negationslash and V f : V X → V Y sends A to f A [ ], for f : X → Y in Top ; and the unit e and the multiplication m of are given by
↦
/CE
$$\ m v \, \omega _ { 0 } \, \nu _ { 0 } \, \omega _ { 0 } \, \beta ^ { + } \nu _ { 0 } \, \omega _ { 0 } \\ e _ { X } \colon X \to V X, \, x \mapsto \overline { \{ x \} } \quad \text{ \quad \ and \quad \ } m _ { X } \colon V V X \to V X, \, \mathcal { A } \mapsto \bigcup \mathcal { A }$$
respectively. We note that the underlying order of V X is the opposite of subset inclusion, that is, A ≤ B if and only if A ⊇ B , for all A,B ∈ V X . We also note that V : Top → Top is a 2-functor. The following lemma describes the convergence in V X (see [Hofmann, 2014]).
Lemma 2.1. Let X be a topological space, A ∈ V X and p be an ultrafilter on V X . Then p → A in V X if and only if A ⊆ ⋂ A∈ p ⋃ A .
A continuous map f : X → Y between topological spaces is called down-wards open whenever, for every open subset A ⊆ X , the down-closure ↓ f A [ ] of f A [ ] is open in Y . Below we record some important properties of /CE , for more information we refer to Schalk [1993], [Escard´, 1998] o and Hofmann [2014].
- (2) Let f : X → Y be in Top . Then V f has a left adjoint if and only if f is down-wards open.
Proposition 2.2. (1) The monad /CE = ( V, e, m ) on Top is of Kock-Z¨berlein type, that is, o e V X ≤ V e X for every topological spaces X (see Kock [1995] and Z¨berlein [1976]). o
- (3) For every topological space X , if X is stably locally compact, then so is V X .
- (4) If X is stably locally compact, then e X : X → V X and m X : V V X → V X are spectral.
- (5) If f : X → Y is a continuous map between stably locally compact spaces, then V f : V X → V Y is spectral if and only if f : X → Y is spectral.
- (6) A stably compact space X is spectral if and only if V X is spectral.
Consequently, the monad /CE = ( V, e, m ) on Top restricts to Kock-Z¨ oberlein monads on StLocComp and on Spec , also denoted by = ( V, e, m ). Using the adjunction
/d40
/d40
/d104
$$\text{also denote by $\Varepsilon=(V,e,m)$. Using the ad;} \\ \text{StLocComp} \underset { \text{T} } { \underbrace { \text{T} } } \text{ CompHaus},$$
/d104
we can transfer the monad /CE on StLocComp to the Vietoris monad ̂ /CE = ( ̂ V , e, m ) on CompHaus . Hence, ̂ V X is the patch space of V X ; the topology of ̂ V X is generated by the sets
$$U ^ { \bigotimes } \ \ ( U \subseteq X \text{ open} ) \quad \text{ and } \quad \{ A \subseteq X \text{ closed } | \ A \cap K = \varnothing \} \ \ ( K \subseteq X \text{ compact} ).$$
We note that this is the topology on the set of closed subsets of a compact Hausdorff space originally considered by [Vietoris, 1922]. The unit e and the multiplication m are as above, but note that e X ( x ) = { x } since X is Hausdorff.
Proposition 2.3. A compact Hausdorff space X is a Stone space if and only if ̂ V X is a Stone space.
Therefore the monad ̂ /CE on CompHaus restricts to a monad on Stone which we also denote by ̂ = ( ̂ V , e, m ).
↦
/CE For a compact Hausdorff space X , a map r : X → ̂ V X is continuous if and only if r : X → V X is spectral, hence CompHaus ̂ /CE can be considered as a full subcategory of StLocComp /CE and consequently Stone ̂ /CE as a full subcategory of Spec /CE . A relation r : X - → Y between spectral spaces is called spectral distributor , indicated as r : X - → ◦ Y , whenever r corresponds to a morphism in StLocComp /CE , that is, the map /rightanglenw r /rightanglene : X → PY factors as X → V Y ↪ → PY and, moreover, X → V Y is spectral. Then r : X - → ◦ Y is also a distributor between the underlying ordered sets of X and Y , which justifies our nomenclature. Furthermore, relational composition of spectral distributors corresponds to composition in StLocComp /CE since the lower Vietoris functor on StLocComp 'behaves like the up-set functor', that is, for a spectral map f : X → Y and A ⊆ X , one has f A [ ] = ↑ f A [ ]. Therefore StLocComp /CE is isomorphic to the category StLocCompDist of stably locally compact spaces and spectral distributors, with relational composition and the identity on X given by the underlying order relation of X . The category StLocCompDist becomes a 2-category via the inclusion order of relations which is dual to the order in StLocComp , that is,
/CE
## StLocComp /similarequal StLocCompDist co ;
and we have a forgetful 2-functor
/CE
## StLocCompDist → Dist .
↦
For a compact Hausdorff space X , the underlying order is discrete and therefore we write CompHausRel to denote the full subcategory of StLocCompDist defined by compact Hausdorff spaces. We refer to the morphisms in CompHausRel as continuous relations , and write r : X - → Y in this case. Clearly, there is a canonical forgetful functor CompHausRel → Rel . Finally, we denote by SpecDist the full subcategory of StLocCompDist defined by all spectral spaces; likewise, StoneRel denotes the full subcategory of CompHausRel defined by all Stone spaces.
↦
↦
Below we give a characterisation of spectral distributors in terms of ultrafilter convergence (see [Hofmann, 2014]). Before doing so, we recall from [Barr, 1970] that the ultrafilter functor U : Set → Set extends to a functor U : Rel → Rel ; here, for a relation r : X - → Y , the relation Ur : UX - → UY is given by
$$\mathfrak { k } U r \, \mathfrak { y } \iff \forall A \in \mathfrak { k }. \{ y \in Y \ | \ x r \, y \text{ for some } x \in A \} \in \mathfrak { y }$$
for all x ∈ UX and y ∈ UY (see also [Clementino and Hofmann, 2004]).
↦
Proposition 2.4. Let X and Y be stably locally compact spaces with ultrafilter convergence a : UX - → X and b : UY - → Y respectively. Then a relation r : X - → Y is a spectral distributor r : X - → ◦ Y if and only if r is a distributor between the underlying ordered sets and the diagram of relations
↦
↦
✤
/d47
/d47
/d15
/d15
/d15
/d15
commutes.
✤
r
/d47
/d47
For a spectral map f : X → Y between stably locally compact spaces, the spectral distributor corresponding to the composite X f - → Y e Y -→ V Y is given by f ∗ : X - → ◦ Y , defined with respect to the underlying orders. Also note that the definition of f ∗ can be applied to any map f : X → Y , not only to monotone and spectral maps. However, we have:
Proposition 2.5. Let X and Y be stably locally compact spaces and f : X → Y be a map. Then f is spectral if and only if f ∗ is a spectral distributor.
↦
Proof. Clearly, if f is spectral, then f ∗ is a spectral distributor. Assume now that f ∗ is a spectral distributor. Then f is certainly monotone. The convergence a : UX - → X of X can be written as a = α ∗ , where α : UX → X is the monotone map which sends an ultrafilter x ∈ UX to its smallest convergence point; similarly, b = β ∗ with β : UY → Y being the monotone map sending an ultrafilter y ∈ UY to its smallest convergence point. Applying U : Rel → Rel to the order relation on X and Y gives a reflexive and transitive (but not necessarily anti-symmetric) relation on UX and UY respectively, and then Uf : UX → UY is a monotone map. From U ( ≤ Y · f ) = U ( ≤ Y ) · Uf it follows that ( Uf ) ∗ = U f ( ∗ ), and from f ∗ · α ∗ = β ∗ · ( Uf ) ∗ we deduce that the diagram
/d47
/d47
/d15
/d15
/d47
/d47
commutes.
/square
↦
Similarly, for arbitrary topological spaces X and Y we characterise those relations r : X - → Y which correspond to continuous maps of type /rightanglenw r /rightanglene : X → V Y , we call such relations continuous distributors.
↦
↦
Proposition 2.6. Let X and Y be topological spaces with ultrafilter convergence a : UX - → X and b : UY - → Y respectively. Then a relation r : X - → Y is a continuous distributor r : X - → ◦ Y if and only if r is a distributor between the underlying ordered sets and, moreover,
↦
✤
/d47
/d47
✤
/d47
/d47
/d15
/d15
/d15
/d15
/d15
/d15
/d15
/d15
✤
/d47
/d47
✤
/d47
/d47
↦
Proof. Clearly, if r : X - → Y is a continuous distributor, then r is also a distributor between the underlying orders. Let now x ∈ X and assume that y → y in Y and r x ( ) ∈ y . Since r x ( ) is closed, y ∈ r x ( ) and therefore x ≤ xr y . Let now x ∈ UX x , ∈ X and y ∈ Y with x → x and xry . Then U r /rightanglenw /rightanglene ( ) x → /rightanglenw r /rightanglene ( x ) and therefore, by Lemma 2.1,
$$y \in \overline { \{ y ^ { \prime } \in Y \, | \, x ^ { \prime } \, r \, y ^ { \prime } \text{ for some } x ^ { \prime } \in A \} }$$
for all A ∈ x . Consequently, there is some y ∈ UY with y → y and
$$\{ y ^ { \prime } \in Y \, | \, x ^ { \prime } \, r \, y ^ { \prime } \text{ for some } x ^ { \prime } \in A \} \in \mathfrak { y }$$
for all A ∈ x , hence x ( Ur ) y . To see the reverse implication, we show first that r x ( ) is closed, for all x ∈ X . In fact, if there is some y ∈ UY with r x ( ) ∈ y and y → y in Y , then there is some x ′ ∈ X with x ≤ x r y ′ and, since r is a distributor, xr y . To see that /rightanglenw r /rightanglene : X → V Y is continuous, assume that x → x in X . Then, for every y ∈ /rightanglenw r /rightanglene ( x ), there is some y ∈ UY with y → y and x ( Ur ) y , hence
$$y \in \overline { \{ y ^ { \prime } \in Y \, | \, x ^ { \prime } \, r \, y ^ { \prime } \text{ for some } x ^ { \prime } \in A \} }$$
for all A ∈ x . This proves U r /rightanglenw /rightanglene ( ) x → /rightanglenw r /rightanglene ( x ).
/square
/d15
/d15
The (order-theoretic) distributor f ∗ : Y - → ◦ X is not always a spectral distributor. In fact, in Proposition 2.2 we have already characterised those spectral maps f : X → Y where V f has a left adjoint. Since /CE is of Kock-Z¨ oberlein type, it is easy to see that such a left adjoint is necessarily an algebra homomorphism, hence:
Theorem 2.7. For a morphism f : X → Y in StLocComp , the following assertions are equivalent.
- (i) f is down-wards open.
↦
- (ii) The spectral distributor f ∗ : X - → Y has a right adjoint in StLocCompDist .
- (iii) the distributor f ∗ : Y - → ◦ X is a spectral distributor.
## 3. Esakia dualities
Besides Boolean algebras, another important class of distributive lattices is the class of Heyting algebras, which correspond under Stone (resp. Priestley) duality to certain spectral (resp. Priestley) spaces. The precise description of this correspondence dates back to 1974; in fact, quoting [Davey and Galati, 2003]: 'The description of the restricted Priestley duality for Heyting algebras was first worked out by M. Adams. The paper in which the description appeared was distributed to a number of those working on applications of Priestley duality but was never published . . . Esakia gives a duality for Boolean algebras with an additional closure operation and then indicates how to use this duality to obtain a duality for Heyting algebras. It should be noted that Esakia's result was obtained without reference to either distributive lattices or Priestley duality. Indeed, because the proof of his duality for Heyting algebras is indirect and missing many details, it is not immediately clear from the paper that Esakia's duality actually is the restricted Priestley duality.'. The result of Esakia mentioned above is published in [Esakia, 1974]. The Priestley spaces corresponding to Heyting algebras are often called Esakia spaces (the designation Heyting spaces is used in [Davey and Galati, 2003]), they are precisely those Priestley spaces X where the down-closure of every open subset of X is again open. Viewing X as a spectral space, X is an Esakia space precisely when, for every open subset A of the patch space X p of X , its down-closure ↓ A is open in X p ; and ↓ A is open in X p if and only if ↓ A is open in X .
In this section we wish to make the point that this duality for Heyting algebras arises more naturally when considering the larger category of spectral spaces and spectral distributors. For technical reasons we will consider here co-Heyting algebras, that is, distributive lattices L where L op is a Heyting algebra.
To start, we extend the notion of Esakia space to stably locally compact spaces on the obvious way.
Definition 3.1. A stably locally compact space X is called an Esakia space whenever, for every open subset A of the patch space X p of X , its down-closure ↓ A is open in X .
Bearing in mind Theorem 2.7, one obtains the following characterisation (see also [Hofmann, 2014]).
Theorem 3.2. For a stably locally compact space X , the following assertions are equivalent.
- (i) X is an Esakia space.
↦
- (ii) The spectral map i : X p → X, x → x is down-wards open.
- (iii) The spectral distributor i ∗ : X p - → ◦ X has a right adjoint (necessarily given by i ∗ ).
- (iv) X is a split subobject of a compact Hausdorff space Y in StLocCompDist .
If X is spectral, then the space Y in the last assertion can be chosen as a Stone space.
We write GEsaDist to denote the full subcategory of StLocCompDist defined by all Esakia spaces, and EsaDist stands for the full subcategory of GEsaDist defined by all spectral spaces. Recall that SpecDist /similarequal DLat op ⊥ ∨ , , and one easily sees that the category DLat ⊥ ∨ , is idempotent split complete. All told:
Corollary 3.3. The category EsaDist is the idempotent split completion of StoneRel .
The algebraic analogue to Theorem 3.2 is essentially proven in [McKinsey and Tarski, 1946]. For a distributive lattice L , we consider its Booleanisation j : L ↪ → B which is given by any epimorphic embedding in DLat of L into a Boolean algebra B (it is a completion in the sense of [Br¨mmer u et al. , 1992]). Translated to Spec , the homomorphism j corresponds to the spectral map i : X p → X . Furthermore, every lattice homomorphism f : L 1 → L 2 extends to a homomorphism f : B 1 → B 2 between the corresponding Boolean algebras.
Theorem 3.4. For a distributive lattice L , the following assertions are equivalent.
- (1) L is a co-Heyting algebra.
- (2) The lattice homomorphism j : L → B has a left adjoint in DLat ⊥ ∨ , j + : B → L .
- (3) L is a split subobject of a Boolean algebra in DLat ⊥ ∨ , .
If f : L 1 → L 2 be a lattice homomorphism between Heyting algebras, then f preserves the coHeyting operation if and only if the diagram
/d47
/d47
/d15
/d15
/d47
/d47
commutes.
Proof. The equivalence (1) ⇐⇒ (2) is shown in [McKinsey and Tarski, 1946], and (2) ⇒ (3) is obvious. To see (3) ⇒ (1), let s : L → B and r : B → L in DLat ⊥ ∨ , with rs = 1 L and B a co-Heyting algebra with co-Heyting operation a /arrowtailright b , for a, b ∈ B . For x, y ∈ H , put
$$x \rightarrow y = r ( s ( x ) \longmapsto s ( y ) ).$$
Then, for all x, y, z ∈ L ,
$$y \leq x \vee z & \iff s ( y ) \leq s ( x ) \vee s ( z ) \\ & \iff s ( x ) \mapsto s ( y ) \leq s ( z ) \\ & \implies x \to y = r ( s ( x ) \mapsto s ( y ) ) \leq r s ( z ) = z.$$
To conclude the missing implication, just observe that s x ( ) /arrowtailright s y ( ) ≤ sr s x ( ( ) /arrowtailright s y ( )). The second statement is clear. /square
We denote the full subcategory of DLat ⊥ ∨ , defined by all co-Heyting algebras by coHeyt ⊥ ∨ , .
Corollary 3.5. The category coHeyt ⊥ ∨ , is the idempotent split completion of Boole ⊥ ∨ , .
In the sequel Esa denotes the full subcategory of Spec defined by Esakia spaces, and coHeyt the full subcategory of DLat defined by co-Heyting algebras. From the discussion above we obtain the main result of this section.
Theorem 3.6. The equivalence SpecDist /similarequal DLat op ⊥ ∨ , restricts to an equivalence
$$E a D i s t \simeq \alpha \text{Hey} ^ { \text{op} } _ { \perp, \vee }$$
which, when restricted to maps and lattice homomorphisms, yields Esa /similarequal coHeyt op . Moreover, a morphism f : L 2 → L 1 in coHeyt preserves the co-Heyting operation if and only if the corresponding spectral map g : X 1 → X 2 makes the diagram of spectral distributors
/d47
/d47
/d15
/d15
/d15
/d15
/d15
/d15
/d47
/d47
commutative; element-wise: for all x ∈ X 1 and y ∈ X 2 with g x ( ) ≤ y , there is some x ′ ∈ X 1 with x ≤ x ′ and g x ( ′ ) = y .
## 4. Generalised Esakia spaces as idempotent split completion
With the results of the last section in mind, we would like to conclude that GEsaDist is the idempotent split completion of the category CompHausRel . This follows indeed from Theorem 3.2, as soon as we know that the category StLocCompDist is idempotent split complete. Similarly to the case of spectral spaces, it is easier to argue in the dual category; and the following result is essentially in [Jung et al. , 2001]. We write StContDLat ∨ , /lessmuch to denote the category of continuous distributive lattices where the way-below relation is stable under finite infima and maps preserving suprema and the way-below relation. Note that every continuous distributive lattice is a frame.
Theorem 4.1. The category StLocCompDist is dually equivalent to the category StContDLat ∨ , /lessmuch .
Proposition 4.2. The category StContDLat ∨ , /lessmuch is idempotent split complete.
Proof. Let e : L → L be an idempotent morphism in StContDLat ∨ , /lessmuch . Then e splits in the category of sup-lattices and sup-preserving maps, that is, there is a complete lattice M and sup-preserving maps r : L → M and s : M → L so that e = sr and rs = 1 M . Then M is certainly a distributive lattice, and, since the embedding s : M → L preserves suprema, for all x, y ∈ M one has
$$s ( x ) \ll s ( y ) \ \Rightarrow \ x \ll y.$$
Consequently, since e : L → L preserves the way-below relation, so does r : L → M . We show now that s : M → L preserves the way-below relation. To this end, let x /lessmuch y in M . Since L is a continuous lattice,
$$s ( y ) = \bigvee \{ b \in L \ | \ b \ll s ( y ) \},$$
and note that { b ∈ L | b /lessmuch s y ( ) } is directed. Hence, y = rs y ( ) is the directed supremum of { r b ( ) ∈ L | b /lessmuch s y ( ) } . Therefore there exist some b /lessmuch s y ( ) with x ≤ r b ( ) and, since e preserves the way-below relation, we obtain
$$s ( x ) \leq s r ( b ) = e ( b ) \ll e ( s ( y ) ) = s ( y ).$$
This shows that s preserves the way-below relation, and from that it follows that M is a continuous lattice. Finally, we prove that the way-below relation in M is stable under finite infima. Note that r ( /latticetop ) = /latticetop since r is surjective, therefore, since /latticetop /lessmuch /latticetop in L , we obtain /latticetop /lessmuch /latticetop in M . Let now x /lessmuch x ′ and y /lessmuch y ′ in M . Then
$$x \wedge y = r ( s ( x ) \wedge s ( y ) ) \ll r ( s ( x ^ { \prime } ) \wedge s ( y ^ { \prime } ) ) = x ^ { \prime } \wedge y ^ { \prime }. & & \Box$$
Corollary 4.3. The category GEsaDist is the idempotent split completion of CompHausRel .
## 5. Split algebras
In the previous section we have seen that the Kleisli category StLocComp /CE is idempotent split complete; consequently, the full subcategory of StLocComp /CE defined by the free algebras V Y is idempotent split complete. This seems to be a rare case, and in general the idempotent split completion of the free algebras for a monad defines an interesting class of algebras, called split algebras (see [Rosebrugh and Wood, 2004]). Most notably, for the up-set monad 2 on the category of ordered sets and monotone functions, the split algebras are precisely the completely distributive complete lattices (see Fawcett and Wood [1990] and [Rosebrugh and Wood, 1994]). In this section we will relate the free algebras for the lower Vietoris monad on StLocComp with the split algebras for other monads in topology.
To start, recall that the filter monad = ( F, e, m ) on Top consists of
2 equivalently, the down-set monad
/BY
- · the functor F : Top → Top where, for a topological space X FX , is the set of all filters on the lattices of opens of X equipped with the topology generated by the sets A # = { f ∈ FX | A ∈ } f ( A ⊆ X open), and, for f : X → Y , the map Ff sends a filter f on the opens of X to the filter { B ⊆ Y | f -1 [ B ] ∈ } f on the opens of Y ;
- · and the natural transformations e : 1 → F and m FF : → F are given by
$$e _ { X } ( x ) = \dot { x } = \{ A \subseteq X \ | \ x \in A \} \quad \text{ and } \quad m _ { X } ( \mathfrak { F } ) = \{ A \subseteq X \ | \ A ^ { \# } \in \mathfrak { F } \},$$
for all topological spaces X , F ∈ FFX and x ∈ X .
We also note that the filter monad on Top is of Kock-Z¨ oberlein type, dual to the case of the lower Vietoris monad: Fe X ≤ e FX for all topological spaces X . A filter f in the lattice of opens of X is called prime whenever, for A,B ⊆ X open, A ∪ B ∈ f implies A ∈ f or B ∈ f . We denote by /BY p = ( F , e, m p ) the submonad of /BY of defined by prime filters. For more information we refer to [Escard´, 1997]. o
$$\text{Top} ^ { \mathbb { F } } \simeq \text{CompHaus} ^ { \widehat { \mathbb { W } } }.$$
As shown in [Day, 1975], the category of Eilenberg-Moore algebras of the filter monad /BY on Top is equivalent to the category ContLat of continuous lattices and Scott-continuous and inf-preserving maps; this latter category is introduced in [Scott, 1972]. We also recall from [Simmons, 1982] that Top /BY p is the category StLocComp ; in particular, there is a forgetful functor Top /BY → StLocComp . In [Wyler, 1981] it is shown that the category of Eilenberg-Moore algebras for the Vietoris monad ̂ /CE on the category CompHaus of compact Hausdorff spaces and continuous maps is equivalent to ContLat , hence:
/BY /CE A slight generalisation of this result is presented in [Hofmann, 2014].
Proposition 5.1. The functor ( -) op : StLocComp → StLocComp restricts to an equivalence Top /similarequal StLocComp .
Following [Rosebrugh and Wood, 2004], for a monad /BW on a category C where idempotents split we consider the full subcategory Spl( C /BW ) of C /BW defined by the split algebras , that is, by those /BW -algebras X with algebra structure α : DX → X for which exists a homomorphism t : X → DX with αt = 1 X . If /BW is of Kock-Z¨ oberlein type, these splittings are adjoint to the algebra structure. Put differently, Spl( C /BW ) is the idempotent split completion of the full subcategory of C defined by the free algebras.
/BY /CE Since StLocComp /similarequal Top /BY p , we conclude that the filter monad /BY on Top is isomorphic to the composite monad of /BY p and /CE via a distributive law, and that FX = ( V ( F p ( X ) op )) op for each topological space X .
/BW The split algebras for various filter monads are studied in [Hofmann, 2013] where they are characterised by a disconnectedness condition. In particular:
Proposition 5.2. The split algebras for the prime filter monad are precisely those stably compact spaces X where, for every open subset of X , its closure in the patch topology is open in X . The split algebras for the filter monad are precisely the filter spaces of frames.
We note that every split algebra for /BY p is a spectral space. The following lemma is easy to prove (see [Hofmann, 2013], for instance).
Lemma 5.3. Let L be a distributive lattice. Then the following assertions are equivalent.
- (i) L is a coframe.
- (ii) L is a split subobject in DLat of the lattice C ( X ) of closed subsets, for some topological space X .
- (iii) L is a split subobject in DLat ⊥ ∨ , of the lattice C ( Y ) of closed subsets, for some topological space Y .
The result above allows us to recover the description of frames inside Priestley (resp. Stone) duality of [Pultr and Sichler, 1988]. Note that those spaces which are split algebras for /BY p are called f-spaces there.
Proof. First note that every split algebra X for /BY p is spectral since X is a subspace of a free algebra F Y p ( Y in Top ) in StLocComp /similarequal Top /BY p . Since Spec /similarequal DLat op , a spectral space X is a split subobject of some F Y p in Spec if and only if S X is a split subobject of S F p ( Y ) /similarequal O Y in DLat . /square
Corollary 5.4. A stably compact space X is a split algebra for /BY p if and only if X is spectral and its lattice S X of compact open subsets of X is a frame.
On the other side, the split algebras for the lower Vietoris monad on StLocComp are precisely the free algebras V Y since StLocComp /CE is idempotent split complete. It is also observed in [Hofmann, 2014] that the equivalence functor ( -) op : Top /BY → StLocComp /CE restricts to a functor ( -) op : Spl( Top ) → Spl( StLocComp );
/BY /CE but this functor is not an equivalence since X op is spectral for every X in Spl( Top /BY ), but V Y is spectral if and only if Y is spectral (see Proposition 2.2). This leaves us with the problem to characterise those spectral spaces X so that V X is in the image of the functor above.
Theorem 5.5. Let X be a spectral space. Then ( V X ) op is a split algebra for the filter monad on Top if and only if X op is a split algebra for the prime filter monad p on Top .
/d111
/BY /BY Proof. For a spectral space X , ( V X ) op is a split algebra for /BY if and only if there is a topological space Y and morphisms
$$( V X ) ^ { \text{op} } \xrightarrow { s } ( V ( ( F _ { p } Y ) ^ { \text{op} } ) ) ^ { \text{op} }$$
/d47
/d47
/d111
in Top /BY with rs = 1. Hence,
$$V X \xrightarrow { s ^ { \text{op} } } V ( ( F _ { p } Y ) ^ { \text{op} } )$$
/d47
/d47
/d111
/d111
/d47
Spec /CE , which corresponds to spectral distributors
$$X \xrightarrow { \varphi } \underset { \psi } { \longleftrightarrow } ( F _ { p } Y ) ^ { \text{op} }$$
/d47
/d111
/d111
with ψ · ϕ = 1 X in SpecDist . Since SpecDist /similarequal DLat op ⊥ ∨ , , this is equivalent to S X being a split subobject of ( S F Y p ) op /similarequal C Y in DLat op ⊥ ∨ , , which in turn is equivalent to S ( X op ) being a frame. /square
## References
Barr, M. (1970), Relational algebras, in Reports of the Midwest Category Seminar, IV , pages 39-55, Lecture Notes in Mathematics, Vol. 137. Springer, Berlin.
Borceux, F. and Dejean, D. (1986), Cauchy completion in category theory, Cahiers Topologie G´om. e Diff´rentielle Cat´g. e e 27 (2), 133-146.
Br¨ ummer, G. C. L. , Giuli, E. and Herrlich, H. (1992), Epireflections which are completions, Cahiers Topologie G´om. Diff´rentielle Cat´g. e e e 33 (1), 71-93.
Cignoli, R. , Lafalce, S. and Petrovich, A. (1991), Remarks on Priestley duality for distributive lattices, Order 8 (3), 299-315.
Clementino, M. M. and Hofmann, D. (2004), On extensions of lax monads, Theory Appl. Categ. 13 , No. 3, 41-60.
Cornish, W. H. (1975), On H. Priestley's dual of the category of bounded distributive lattices, Mat. Vesn., N. Ser. 12 , 329-332.
Davey, B. A. and Galati, J. C. (2003), A coalgebraic view of Heyting duality, Studia Logica 75 (3), 259-270.
Day, A. (1975), Filter monads, continuous lattices and closure systems, Canad. J. Math. 27 , 50-59. Esakia, L. (1974), Topological Kripke models, Dokl. Akad. Nauk SSSR 214 , 298-301.
Escard´ o, M. H. (1997), Injective spaces via the filter monad, in Proceedings of the 12th Summer Conference on General Topology and its Applications (North Bay, ON, 1997) , volume 22, pages 97100.
Escard´ o, M. H. (1998), Properly injective spaces and function spaces, Topology Appl. 89 (1-2), 75-120.
Fawcett, B. and Wood, R. J. (1990), Constructive complete distributivity. I, Math. Proc. Cambridge Philos. Soc. 107 (1), 81-89.
Fleisher, I. (2000), Priestley's duality from Stone's., Adv. Appl. Math. 25 (3), 233-238.
- Gierz, G. , Hofmann, K. H. , Keimel, K. , Lawson, J. D. , Mislove, M. W. and Scott, D. S. (1980), A compendium of continuous lattices , Springer-Verlag, Berlin, xx+371 pages.
Halmos, P. R. (1956), Algebraic logic. I. Monadic Boolean algebras, Compositio Math. 12 , 217-249. Hochster, M. (1969), Prime ideal structure in commutative rings, Trans. Amer. Math. Soc. 142
, 43-60.
Hofmann, D. (2013), A four for the price of one duality principle for distributive spaces, Order 30 (2), 643-655, arXiv:math.GN/1102.2605 .
Hofmann, D. (2014), The enriched Vietoris monad on representable spaces, J. Pure Appl. Algebra 218 (12), 2274-2318, arXiv:math.CT/1212.5539 .
- Johnstone, P. T. (1986), Stone spaces , volume 3 of Cambridge Studies in Advanced Mathematics , Cambridge University Press, Cambridge, xxii+370 pages, reprint of the 1982 edition.
- Jung, A. (2004), Stably compact spaces and the probabilistic powerspace construction, in J. Desharnais and P. Panangaden, editors, Domain-theoretic Methods in Probabilistic Processes , volume 87, 15pp.
- Jung, A. , Kegelmann, M. and Moshier, M. A. (2001), Stably compact spaces and closed relations, in MFPS 2001. Papers from the 17th conference on the mathematical foundations of programming semantics, Aarhus University, Aarhus, Denmark, May 23-26, 2001. , pages 209-231, Amsterdam: Elsevier.
- Kock, A. (1995), Monads for which structures are adjoint to units, J. Pure Appl. Algebra 104 (1), 41-59.
- McKinsey, J. and Tarski, A. (1946), On closed elements in closure algebras, Ann. Math. (2) 47 , 122-162.
- Nachbin, L. (1950), Topologia e Ordem , Univ. of Chicago Press, in Portuguese, English translation: Topology and Order, Van Nostrand, Princeton (1965).
- Priestley, H. A. (1970), Representation of distributive lattices by means of ordered stone spaces, Bull. London Math. Soc. 2 , 186-190.
Priestley, H. A. (1972), Ordered topological spaces and the representation of distributive lattices, Proc. London Math. Soc. (3) 24 , 507-530.
- Pultr, A. and Sichler, J. (1988), Frames in Priestley's duality, Cah. Topol. G´ eom. Diff´r. e Cat´ eg. 29 (3), 193-202.
Rosebrugh, R. and Wood, R. J. (1994), Constructive complete distributivity IV, Appl. Categ. Structures 2 (2), 119-144.
Rosebrugh, R. and Wood, R. J. (2004), Split structures, Theory Appl. Categ. 13 , No. 12, 172-183. Sambin, G. and Vaccaro, V. (1988), Topology and duality in modal logic, Ann. Pure Appl. Logic 37 (3), 249-296.
Schalk, A. (1993), Algebras for Generalized Power Constructions , Ph.D. thesis, Technische Hochschule Darmstadt.
Scott, D. (1972), Continuous lattices, in Toposes, algebraic geometry and logic (Conf., Dalhousie Univ., Halifax, N. S., 1971) , pages 97-136. Lecture Notes in Math., Vol. 274, Springer, Berlin.
Simmons, H. (1980), Reticulated rings, J. Algebra 66 , 169-192.
Simmons, H. (1982), A couple of triples, Topology Appl. 13 (2), 201-223.
Stone, M. H. (1936), The theory of representations for Boolean algebras, Trans. Amer. Math. Soc. 40 (1), 37-111.
Stone, M. H. (1938), Topological representations of distributive lattices and Brouwerian logics, ˇ asopis C pro pˇstov´n´ ı matematiky a fysiky e a 67 (1), 1-25, http://dml.cz/handle/10338.dmlcz/124080 .
Vietoris, L. (1922), Bereiche zweiter Ordnung, Monatsh. Math. Phys. 32 (1), 258-280.
Wyler, O. (1981), Algebraic theories of continuous lattices, Continuous lattices, Proc. Conf., Bremen 1979, Lect. Notes Math. 871, 390-413 (1981).
Z¨ oberlein, V. (1976), Doctrines on 2-categories, Math. Z. 148 (3), 267-279.
Center for Research and Development in Mathematics and Applications, Department of Mathematics, University of Aveiro, 3810-193 Aveiro, Portugal
E-mail address : [email protected]
E-mail address :
[email protected]
| null |
[
"Dirk Hofmann",
"Pedro Nora"
] |
2014-08-05T19:12:26+00:00
|
2014-08-05T19:12:26+00:00
|
[
"math.CT"
] |
Some notes on Esakia spaces
|
Under Stone/Priestley duality for distributive lattices, Esakia spaces
correspond to Heyting algebras which leads to the well-known dual equivalence
between the category of Esakia spaces and morphisms on one side and the
category of Heyting algebras and Heyting morphisms on the other. Based on the
technique of idempotent split completion, we give a simple proof of a more
general result involving certain relations rather then functions as morphisms.
We also extend the notion of Esakia space to all stably locally compact spaces
and show that these spaces define the idempotent split completion of compact
Hausdorff spaces. Finally, we exhibit connections with split algebras for
related monads.
|
1408.1074v1
|
## Appell F 1 and Conformal Mapping
## Steven Finch
August 5, 2014
Abstract. This is the last of a trilogy of papers on triangle centers. A fairly obscure 'conformal center of gravity' is computed for the class of all isosceles triangles. This calculation appears to be new. A byproduct is the logarithmic capacity or transfinite diameter of such, yielding results consistent with Haegi (1951).
Before discussing triangles, let us give both a review of [1] and a preview involving a simpler region in the plane.
Let Ω = { x + iy ∈ C : y > 0, x 2 + y 2 < 1 } , the interior of the upper half-disk of unit radius. Let ∆ denote the (full) disk of unit radius and Σ denote the infinite horizontal strip of width π . Define a function /lscript : Ω → Σ by [2]
$$\ell ( z ) = \ln \left ( \frac { ( 1 + z ) ^ { 2 } } { ( 1 - z ) ^ { 2 } } \right ). \\. \ \Lambda \ \dots \ \Lambda \..$$
The conformal map f w : Ω → ∆ given by
$$f _ { w } ( z ) = \frac { \exp ( \ell ( z ) ) - \exp ( \ell ( w ) ) } { \exp ( \ell ( z ) ) - \exp ( \overline { \ell ( w ) } ) } \\ \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \colon$$
satisfies f w ( w ) = 0; it is well-known that ln | f w ( z ) | is Green's function for Ω. We deduce that
∣ ∣ ∣ ∣ ∣ where /lscript ′ denotes the derivative of /lscript . Restricting attention to the y -axis only, we have
$$h ( w ) = \lim _ { z \to w } \left | \frac { f _ { w } ( z ) } { z - w } \right | = \left | \frac { \exp ( \ell ( w ) ) \ell ^ { \prime } ( w ) } { \exp ( \ell ( w ) ) - \exp ( \overline { \ell ( w ) } ) } \right | \\ \text{notes the derivative of $\ell$.\ \ Restricting attention to the $y$-axis onl} \\ \cdot \. \. \. \. \ \.$$
$$h ( i y ) = \left | \frac { \frac { 2 ( 1 + i y ) } { ( 1 - i y ) ^ { 2 } } + \frac { 2 ( 1 + i y ) ^ { 2 } } { ( 1 - i y ) ^ { 3 } } } { \frac { ( 1 + i y ) ^ { 2 } } { ( 1 - i y ) ^ { 2 } } - \frac { ( i + y ) ^ { 2 } } { ( - i + y ) ^ { 2 } } } \right | = \left | \frac { \left ( 1 + i y \right ) ^ { 2 } } { 2 y \left ( 1 - y ^ { 2 } \right ) } \right | = \frac { 1 + y ^ { 2 } } { 2 y \left ( 1 - y ^ { 2 } \right ) } \\ ) < y < 1. \ \text{Minimizing this expression, it follows that}$$
∣ ∣ ∣ ∣ for 0 < y < 1. Minimizing this expression, it follows that
$$i y _ { 0 } = i \sqrt { - 2 + \sqrt { 5 } } = ( 0. 4 8 5 8 6 8 2 7 1 7 5 6 6 4 5 6 7 8 1 8 2 8 6 3 8 \dots ) i \\ \overline { \text{Copyright $\circledast}$ 2014 by Steven R. Finch. All rights reserved.}$$
0 Copyright c © 2014 by Steven R. Finch. All rights reserved.
is the inner conformal center (what was called the least capacity point in [1]) for Ω. Also,
$$\frac { 1 } { h ( i y _ { 0 } ) } = \sqrt { - 2 2 + 1 0 \sqrt { 5 } } = 0. 6 0 0 5 6 6 2 1 2 0 0 1 5 5 5 2 1 5 7 7 3 3 8 9 4 \dots \\. \quad. \quad. \quad.. \quad. \quad.. \quad.. \quad.. \quad.. \quad.$$
is the maximum inner radius of Ω [3, 4]. This concludes our review.
Let Ω c denote the complement of the closure of Ω. Inverting a function [5]
$$\frac { 1 } { z } & = \frac { \left ( \frac { \sqrt { 3 } } { 2 } - \frac { i } { 2 } \right ) \left ( 1 + \frac { 1 } { w } \right ) ^ { 2 / 3 } + \left ( \frac { \sqrt { 3 } } { 2 } + \frac { i } { 2 } \right ) \left ( 1 - \frac { 1 } { w } \right ) ^ { 2 / 3 } } { \left ( 1 + \frac { 1 } { w } \right ) ^ { 2 / 3 } - \left ( 1 - \frac { 1 } { w } \right ) ^ { 2 / 3 } } \\ \text{of } w \text{ yields}$$
in terms of w yields
$$\frac { \left ( 1 + \frac { 1 } { w } \right ) ^ { 2 / 3 } } { \left ( 1 - \frac { 1 } { w } \right ) ^ { 2 / 3 } } = \frac { \frac { \sqrt { 3 } } { 2 } + \frac { i } { 2 } + \frac { 1 } { z } } { - \frac { \sqrt { 3 } } { 2 } + \frac { i } { 2 } + \frac { 1 } { z } } = m ( z ). \\ \mathbf p \ g \colon & \Delta \to \Omega ^ { c }$$
The conformal map g : ∆ → Ω c
$$g ( z ) & = \frac { 1 + 2 m ( z ) ^ { 3 / 2 } + m ( z ) ^ { 3 } } { - 1 + m ( z ) ^ { 3 } } = \frac { 4 } { 3 \sqrt { 3 } } \frac { 1 } { z } + \frac { 2 i } { 3 \sqrt { 3 } } + O \left ( z \right ) \\ \nu \right ) \quad \cdot \quad \cdot \quad \cdot \quad \nu \right ) \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot$$
satisfies g (0) = ∞ , has positive leading Laurent coefficient, and is unique in this regard. The constant term of the series expansion
$$\frac { 2 i } { 3 \sqrt { 3 } } = ( 0. 3 8 4 9 0 0 1 7 9 4 5 9 7 5 0 5 0 9 6 7 2 7 6 5 8 \dots ) i$$
is the outer conformal center , which clearly lies in Ω but is not the same as the inner conformal center. A motivating feature is
$$\frac { 1 } { 2 \pi } \underset { 0 } { \overset { 2 \pi } { \int } } g \left ( ( 1 - \varepsilon ) e ^ { i t } \right ) d t = \frac { 2 i } { 3 \sqrt { 3 } } \\. \quad \cdot \quad \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \, \cdots \end{array}$$
for ε > 0, but the literature is small [6, 7, 8, 9, 10]. The leading coefficient (of 1 /z ) is the outer radius of Ω [3, 11]:
$$\frac { 4 } { 3 \sqrt { 3 } } = 0. 7 6 9 8 0 0 3 5 8 9 1 9 5 0 1 0 1 9 3 4 5 5 3 1 7 \dots$$
which is also known as the logarithmic capacity or transfinite diameter of Ω. An alternative definition is [12]
$$\lim _ { n \to \infty } \max _ { z _ { 1 }, z _ { 2 }, \dots, z _ { n } \in \Omega } \left ( \prod _ { j < k } | z _ { j } - z _ { k } | \right ) ^ { \frac { 2 } { n ( n - 1 ) } }, \\ \prod _ { j \to \infty } \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \end{array}$$
that is, the maximal geometric mean of pairwise distances for n points in Ω, in the limit as n → ∞ . This constitutes a fascinating collision of ideas from potential theory; replacing a geometric mean by an arithmetic mean seems to be an open computational issue.
As a conclusion to our preview, the Appell F 1 function can be written as a definite integral [13, 14]
$$F _ { 1 } ( a, b, b ^ { \prime }, c ; x, y ) = \frac { \Gamma ( c ) } { \Gamma ( a ) \Gamma ( c - a ) } \underset { 0 } { \overset { 1 } { \int } } s ^ { a - 1 } ( 1 - s ) ^ { c - a - 1 } ( 1 - s x ) ^ { - b } ( 1 - s y ) ^ { - b ^ { \prime } } d s,$$
$$\int _ { \dots } | x | < 1, \quad | y | < 1, \quad \text{Re} ( c ) > \text{Re} ( a ) > 0$$
as well as a double hypergeometric series
$$F _ { 1 } ( a, b, b ^ { \prime }, c ; x, y ) & = \sum _ { m = 0 n = 0 } ^ { \infty } \frac { 1 } { m! n! } \frac { \Gamma ( a + m + n ) } { \Gamma ( a ) } \frac { \Gamma ( b + m ) } { \Gamma ( b ) } \frac { \Gamma ( b ^ { \prime } + n ) } { \Gamma ( b ^ { \prime } ) } \frac { \Gamma ( c ) } { \Gamma ( c + m + n ) } x ^ { m } y ^ { n }. \\ \intertext { t-summation to m a t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i nt i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t i n t junit }$$
Its numerical implementation in Mathematica is crucial to everything that follows.
## 1. Isosceles Triangles
Let
$$T _ { \theta } = \left \{ x + i y \in \mathbb { C } \, \colon 0 < x < \cos ( \theta / 2 ), \ y < \tan ( \theta / 2 ) x, \ y > - \tan ( \theta / 2 ) x \right \},$$
the interior of an isosceles triangle with apex angle 0 < θ < π located at the origin. Two sides of unit length meet there; the third (vertical) side has x -intercept cos( θ/ 2) and length 2 sin( θ/ 2).
To construct a conformal map g : ∆ → T c θ requires two steps. First, define
$$f ( z ) = \underset { z _ { 0 } } { \overset { z } { \int } } \frac { ( \zeta - a _ { 1 } ) ^ { \mu _ { 1 } } ( \zeta - a _ { 2 } ) ^ { \mu _ { 2 } } ( \zeta - a _ { 3 } ) ^ { \mu _ { 3 } } } { \zeta ^ { 2 } } d \zeta$$
on ∆, where a 1 , a 2 , a 3 are prevertices of the unit circle mapping onto vertices of T θ , and π (1 + µ 1 ), π (1 + µ 2 ), π (1 + µ 3 ) are exterior angles at the corresponding vertices.
Also, z 0 is some point of ∆ other than 0, and the integral is taken along any curve in ∆ joining z 0 to z not passing through 0 (it does not matter which). Clearly
$$\mu _ { 1 } = \frac { \pi - \theta } { \pi }, \quad \mu _ { 2 } = \mu _ { 3 } = \frac { \pi + \theta } { 2 \pi }.$$
The choice of point a 1 is arbitrary; here let a 1 = -1. The remaining two points a 2 , a 3 must satisfy the constraint [15, 16]
$$\mu _ { 1 } / a _ { 1 } + \mu _ { 2 } / a _ { 2 } + \mu _ { 3 } / a _ { 3 } = 0$$
in order that f (0) = ∞ . Thus
$$\ddots \nmid \times \nmid \, \substack { \cdots \, \cdots \, \cdots \, \\ a _ { 2 } = \frac { \pi - \theta + 2 i \sqrt { \pi \theta } } { \pi + \theta }, \quad a _ { 3 } = \frac { \pi - \theta - 2 i \sqrt { \pi \theta } } { \pi + \theta }$$
work for our purposes. We further choose z 0 = -1, so that
$$f ( z ) = & \underset { - 1 } { \int } \frac { z ^ { 2 } } { ( \zeta + 1 ) ^ { 1 - \theta / \pi } ( \zeta ^ { 2 } - 2 \frac { \pi - \theta } { \pi + \theta } \zeta + 1 ) ^ { ( \pi + \theta ) / ( 2 \pi ) } } { \zeta ^ { 2 } } d \zeta. \\.. \ e c \quad &. \ e c \quad \int \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 } \Delta _ { 0 }.$$
The image of { a , a 1 2 , a 3 } under f evidently lies in the left half plane - needing rotation by π - plus rescaling so that the vertical triangle side has the proper length. This second step is achieved by defining
$$g ( z ) = - \frac { 2 \sin ( \theta / 2 ) } { \text{Im} \, f ( a _ { 2 } ) - \text{Im} \, f ( a _ { 3 } ) } f ( z ). \\ _ { - \, / \alpha \, \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \colon \end{array}$$
For the scenario θ = π/ 2, it is true that the coefficient
$$& \alpha \ v a v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v o c u n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n v e c u a n } \\ & \quad - \frac { \sqrt { 2 } } { \text{Im} \, f ( a _ { 2 } ) - \text{Im} \, f ( a _ { 3 } ) } = 0. 4 7 5 6 3 4 4 4 3 8 7 9 9 8 1 9 3 2 0 5 6 7 5 7 0 \dots = \frac { 3 ^ { 3 / 4 } } { 2 ^ { 7 / 2 } \pi ^ { 3 / 2 } } \Gamma \left ( \frac { 1 } { 4 } \right ) ^ { 2 } = \kappa \\ & \quad \ast \text{In which communication communication } \text{ This conversion of this or other problem of an occolates with}$$
to high numerical precision. This expression (the outer radius of an isosceles right triangle) is well-known and is a special case of a more general formula due to Haegi [3, 17]. More on this will be given soon.
Our key result is that the function f ( z ) possesses an exact representation. Let
$$\text{Our key result is that the function } f ( z ) \text{ possesses an exact representation. } \text{ Let} \\ \xi ( z ) = \frac { \left ( \sqrt { \pi } - i \sqrt { \theta } \right ) ( z + 1 ) } { 2 \sqrt { \pi } }, \quad \eta ( z ) = \frac { \left ( \sqrt { \pi } + i \sqrt { \theta } \right ) ( z + 1 ) } { 2 \sqrt { \pi } }, \\ \varphi ( z ) = \left [ \frac { \pi \left ( z - 1 \right ) ^ { 2 } + \theta ( z + 1 ) ^ { 2 } } { \pi + \theta } \right ] ^ { ( \pi + \theta ) / ( 2 \pi ) },$$
$$\psi ( z ) = \left [ \frac { \pi \left ( z - 1 \right ) ^ { 2 } + \theta ( z + 1 ) ^ { 2 } } { \pi + \theta } \right ] ^ { ( \pi - \theta ) / ( 2 \pi ) },$$
$$\delta ( z ) & = \left [ - \sqrt { \pi } \left ( z - 1 \right ) - i \sqrt { \theta } ( z + 1 ) \right ] ^ { \left ( \pi - \theta \right ) / ( 2 \pi ) } \left [ - \sqrt { \pi } \left ( z - 1 \right ) + i \sqrt { \theta } ( z + 1 ) \right ] ^ { \left ( \pi - \theta \right ) / ( 2 \pi ) }. \\ \text{Then we have}$$
$$\psi ( z ) = \left \lfloor \frac { \pi \, ( z - 1 ) ^ { 2 } + \theta ( z + 1 ) ^ { 2 } } { \pi + \theta } \right \rfloor ^ { \vee \cdot \lambda ^ { \prime / \dots \lambda ^ { \prime } } }, \\ \delta ( z ) = \left [ - \sqrt { \pi } \, ( z - 1 ) - i \sqrt { \theta } ( z + 1 ) \right ] ^ { ( \pi - \theta ) / ( 2 \pi ) } \left [ - \sqrt { \pi } \, ( z - 1 ) + i \sqrt { \theta } ( z + 1 ) \right ] ^ { ( } \\ \text{Then we have}$$
Then we have
$$f ( z ) & = ( z + 1 ) ^ { 1 - \theta / \pi } \left \{ - \frac { \varphi ( z ) } { z } + \frac { 2 ^ { \theta / \pi } \pi ^ { ( \pi + \theta ) / ( 2 \pi ) } } { 2 \pi ^ { 2 } + \pi \theta - \theta ^ { 2 } } \frac { \delta ( z ) } { \psi ( z ) } \cdot \\ & \left [ - 2 ( 2 \pi - \theta ) F _ { 1 } \left ( 1 - \frac { \theta } { \pi }, \frac { \pi - \theta } { 2 \pi }, \frac { \pi - \theta } { 2 \pi }, 2 - \frac { \theta } { \pi } ; \xi ( z ), \eta ( z ) \right ) + \\ & ( \pi + \theta ) ( z + 1 ) F _ { 1 } \left ( 2 - \frac { \theta } { \pi }, \frac { \pi - \theta } { 2 \pi }, \frac { \pi - \theta } { 2 \pi }, 3 - \frac { \theta } { \pi } ; \eta ( z ), \xi ( z ) \right ) \right ] \right \}$$
$$+ \, 1 ) F _ { 1 } \Big ( 2 - \frac { \theta } { \pi }, \frac { \pi - \theta } { 2 \pi }, \frac { \pi - \theta } { 2 \pi }, 3 - \frac { \theta } { \pi } ; \eta ( z ), \xi ( z ) \Big ) \Big ] \Big \} \Big. \quad \\ \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \, \quad \$$
as can be easily proved after-the-fact by differentiation. (Our before-the-fact technique consisted of examining rational multiples of θ in Mathematica, seeking recognizable patterns.) This integral evaluation appears to be new.
Returning to the θ = π/ 2 scenario,
$$g ( z ) = \frac { \kappa } { z } + \lambda + O \left ( z \right )$$
as z → 0, where
$$\lambda = \kappa \frac { 2 ^ { 5 / 4 } } { 3 ^ { 3 / 4 } } \left [ 2 F _ { 1 } \left ( \frac { 1 } { 2 }, \frac { 1 } { 4 }, \frac { 1 } { 4 }, \frac { 3 } { 2 } ; \frac { 2 - i \sqrt { 2 } } { 4 }, \frac { 2 + i \sqrt { 2 } } { 4 } \right ) - F _ { 1 } \left ( \frac { 3 } { 2 }, \frac { 1 } { 4 }, \frac { 1 } { 4 }, \frac { 5 } { 2 } ; \frac { 2 + i \sqrt { 2 } } { 4 }, \frac { 2 - i \sqrt { 2 } } { 4 } \right ) \right ]$$
= 0 5045039334500261012764068 . ...
is the outer conformal center of T π/ 2 . We wonder whether this expression for λ can be simplified, for example, as a ratio of gamma or Gauss hypergeometric function values.
As a corollary, let ˜ = T { x + iy ∈ C : x > 0, y > 0, x + y < 1 } , the initial triangle examined in [1, 18]. The outer conformal center of ˜ is simply T
$$\left ( \frac { 1 + i } { \sqrt { 2 } } \right ) \lambda = ( 0. 3 5 6 7 3 8 1 5 2 4 7 7 8 0 0 1 4 0 6 7 5 1 3 0 7 \dots ) ( 1 + i ) \\. \quad \. \quad \. \quad \. \quad \. \quad \.$$
which is not the same as the inner conformal center (0 301 . ... )(1 + i ).
We mention finally that the outer conformal center of T π/ 3 (an equilateral triangle) is 1 / √ 3, that is, it coincides with the centroid of T π/ 3 . No other scenarios with such recognizable λ have been found.
## 2. Haegi's Formula
An arbitrary triangle with sides a , b , c and opposite angles
$$\alpha = \arccos \left ( \frac { b ^ { 2 } + c ^ { 2 } - a ^ { 2 } } { 2 b c } \right ),$$
$$\beta = \arccos \left ( \frac { a ^ { 2 } + c ^ { 2 } - b ^ { 2 } } { 2 a c } \right ),$$
$$\gamma = \arccos \left ( \frac { a ^ { 2 } + b ^ { 2 } - c ^ { 2 } } { 2 a b } \right )$$
has area, circumradius and logarithmic capacity given by
$$&, c \, \text{and opposite angles} \\ & \underset { \substack { \colon } \, } { c } \left ( \frac { b ^ { 2 } + c ^ { 2 } - a ^ { 2 } } { 2 b c } \right ), \\ & \underset { \substack { \colon } \, } { c } \left ( \frac { a ^ { 2 } + c ^ { 2 } - b ^ { 2 } } { 2 a c } \right ), \\ & \underset { \substack { \colon } \, } { c } \left ( \frac { a ^ { 2 } + b ^ { 2 } - c ^ { 2 } } { 2 a b } \right ) \\ & \text{nic capacity given by}$$
$$A = \sqrt { \frac { a + b + c } { 2 } \frac { - a + b + c } { 2 } \frac { a - b + c } { 2 } \frac { a + b - c } { 2 } },$$
$$R = \frac { a b c } { \sqrt { ( a + b + c ) ( b + c - a ) ( c + a - b ) ( a + b - c ) } }, \\ \kappa = \frac { A } { 4 \pi ^ { 2 } q ( \alpha / \pi ) q ( \beta / \pi ) q ( \gamma / \pi ) R }$$
where
Under the special circumstances that a = b = 1 and c = 2 sin( θ/ 2), we have
$$q ( x ) = \frac { 1 } { \Gamma ( x ) } \sqrt { \frac { x ^ { x } } { ( 1 - x ) ^ { 1 - x } } }. \\ \text{istances that } a = h = 1 \text{ and } c = 2 \text{ s}$$
$$\cdot \\ \kappa ( \theta ) = \frac { \sqrt { \pi + \theta } } { 8 \pi ^ { 5 / 2 } } \left ( \frac { \pi + \theta } { 4 \theta } \right ) ^ { \theta / ( 2 \pi ) } \frac { \sin ( \theta ) ^ { 2 } } { \sin ( \theta / 2 ) } \Gamma \left ( \frac { \theta } { \pi } \right ) \Gamma \left ( \frac { \pi - \theta } { 2 \pi } \right ) ^ { 2 } \\.. \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \end{cases}$$
for the isosceles triangles T θ . Over such triangles, the one with maximal κ has θ = 2 5360873621 . ... , which seems not to have been noticed before. Over the family of all triangles with fixed A , the one with minimal κ is equilateral, as proved by P´lya o & Szeg¨ o [3, 19, 20]. If we fix perimeter rather than area, then (to the contrary) the equilateral triangle provides maximal κ .
## 3. Addendum: 30 ◦ -60 ◦ -90 ◦ Triangle
$$\mu _ { 1 } = \frac { 1 } { 2 }, \quad \mu _ { 2 } = \frac { 2 } { 3 }, \quad \mu _ { 3 } = \frac { 5 } { 6 }.$$
Define T = { x + iy ∈ C : x > 0, y > 0, √ 3 x + y < √ 3 } . Proceeding as before, we obtain
The choice of point a 1 is arbitrary; here let a 1 = 1. From µ /a 1 1 + µ /a 2 2 + µ /a 3 3 = 0, we deduce that
$$a _ { 2 } = i, \quad a _ { 3 } = - \frac { 3 } { 5 } - \frac { 4 } { 5 } i \\ \cdot \quad \cdot \cdot \cdot \, e \, \mathfrak { m } \, \quad \cdot \,.$$
work for our purposes. Choosing z 0 = 1, it follows that
$$\mu _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { 2 } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu \nu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu. } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu, \mu. } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu. } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu} ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } \beta _ { \mu } ^ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu. } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta ^ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \nu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu} \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ { \mu } \beta _ {$$
An exact representation for f ( z ) in terms of the Appell F 1 function is possible. From this, we can verify the outer radius expression [3, 17]
$$\kappa = \frac { 5 ^ { 5 / 1 2 } } { 2 ^ { 1 0 / 3 } \pi ^ { 2 } } \Gamma \left ( \frac { 1 } { 3 } \right ) ^ { 3 } = 0. 3 7 7 9 1 3 7 4 2 9 7 0 9 5 5 8 3 2 1 0 2 4 8 8 2 \dots \\ \text{sh numerical position but how not not determined to not conformed} \, \text{$a$}$$
to high numerical precision, but have not yet determined the outer conformal center of T .
## 4. Addendum: 6 9 13 --Triangle
The Schwarz-Christoffel toolbox for Matlab [21, 22] makes numerical computations of a conformal map feasible. For the triangle with vertices
$$\begin{matrix} \colon \text{$10$ the unique with $\colon$} \\ 0, & 6, & - \frac { 1 3 } { 3 } + \frac { 4 \sqrt { 3 5 } } { 3 } i \end{matrix}$$
the following code:
$$<_Python_> p = polygon([0 6 -13/3+(4*sqrt(35)/3)*i])
f = extermap(p,scmapopt('Tolerance',1e-18))
p = parameters(f)
format long
p.prevertex$$
gives
$$\mu _ { 1 } = 0. 6 5 9, \quad \mu _ { 2 } = 0. 2 0 7, \quad \mu _ { 3 } = 0. 1 3 2, \\ a _ { 1 } = 1, \quad a _ { 2 } = 0. 0 1 6 3 - 0. 9 9 9 8 i, \quad a _ { 3 } = - 0. 4 0 6 9 + 0. 9 1 3 4 i. \\ \text{Index-form extensions for these environments and preventices are possible yet, comm}$$
Closed-form expressions for these exponents and prevertices are possible yet cumbersome. The same is true for the outer radius κ = 3 805336. . Determining the outer conformal center (even approximately) remains open. Figures 1, 2, 3 provide conformal map images of ten evenly-spaced concentric circles in the disk; orthogonal trajectories are also indicated. We leave the task of exploring whether outer conformal centers belong in Kimberling's database [23] to someone else.
Figure 1: Images of ten concentric circles, center at 0 356 + (0 356) . . . i
Figure 2: Images of ten concentric circles, center unknown.
Figure 3: Images of ten concentric circles, center unknown.
## 5. Acknowledgements
I am grateful to Thomas Ransford [12, 16] for explaining Schwarz-Christoffel prevertex calculations for the scenario θ = π/ 2.
## References
- [1] S. Finch, Least capacity point of triangles, http://arxiv.org/abs/1407.4105.
- [2] J. Kevorkian, Partial Differential Equations. Analytical Solution Techniques , 2 nd ed., Springer-Verlag, 2000, pp. 129-130; MR1728947 (2000i:35001).
- [3] G. P´ olya and G. Szeg¨, o Isoperimetric Inequalities in Mathematical Physics , Princeton Univ. Press, 1951, pp. 1-3, 158, 254-258, 273-274; MR0043486 (13,270d).
- [4] M. Hantke, Summen reziproker Eigenwerte , Ph.D. thesis, Martin-Luther-Universit¨t a Halle-Wittenberg, 2006, http://sundoc.bibliothek.uni-halle.de/diss-online/06/06H308/index.htm.
- [5] Ch. Pommerenke, On metric properties of complex polynomials, Michigan Math. J. 8 (1961) 97-115; MR0151580 (27 #1564).
- [6] G. P´ olya and G. Szeg˝, o Problems and Theorems in Analysis. I, Series, Integral Calculus, Theory of Functions , Springer-Verlag, 1998, pp. 129-130, 325-326; MR1492447.
- [7] G. P´ olya and G. Szeg˝, o Problems and Theorems in Analysis. II, Theory of Functions, Zeros, Polynomials, Determinants, Number Theory, Geometry , SpringerVerlag, 1998, pp. 23-24, 192-194; MR1492448.
- [8] Ch. Pommerenke, Univalent Functions , Vandenhoeck & Ruprecht, 1975, pp. 1213; MR0507768 (58 #22526).
- [9] R. Bouffanais, G. D. Weymouth and D. K. P. Yue, Hydrodynamic object recognition using pressure sensing, Proc. Royal Soc. London Ser. A 467 (2011) 19-38; MR2764670.
- [10] G. Cleanthous, Monotonicity theorems for analytic functions centered at infinity, Proc. Amer. Math. Soc. 142 (2014) 3545-3551.
- [11] S. Liesipohja, Numerical Methods for Computing Logarithmic Capacity , M.Sc. thesis, University of Helsinki, 2014, https://helda.helsinki.fi/handle/10138/44698.
- [12] T. Ransford, Potential Theory in the Complex Plane , Cambridge Univ. Press, 1995, pp. 132-137, 152-160; MR1334766 (96e:31001).
- [13] W. N. Bailey, Generalized Hypergeometric Series , Cambridge Univ. Press, 1935, pp. 73-83.
- [14] J. McDougall, L. Schaubroeck and J. Rolf, Exploring Complex Analysis . Ch. 5: Mappings to Polygonal Domains , http://www.jimrolf.com/explorationsInComplexVariables.html.
- [15] Z. Nehari, Conformal Mapping , Dover, 1975, pp. 189-195; MR0377031 (51 #13206).
- [16] T. Ransford, Logarithmic capacity for ˜ , T unpublished note (2014).
- [17] H. R. Haegi, Extremalprobleme und Ungleichungen konformer Gebietsgr¨ssen, o Compositio Math. 8 (1950) 81-111; MR0039811 (12,602b).
- [18] S. Finch, In limbo: Three triangle centers, http://arxiv.org/abs/1406.0836.
- [19] G. P´ olya and G. Szeg˝, Inequalities for the capacity of a condenser, o Amer. J. Math. 67 (1945) 1-32; MR0011871 (6,227e).
- [20] A. Yu. Solynin and V. A. Zalgaller, An isoperimetric inequality for logarithmic capacity of polygons, Annals of Math. 159 (2004) 277-303; MR2052355 (2005a:31002).
- [21] T. A. Driscoll and L. N. Trefethen, Schwarz-Christoffel Mapping , Cambridge Univ. Press, 2002, pp. 9-18; MR1908657 (2003e:30012).
- [22] T. A. Driscoll, Schwarz-Christoffel Toolbox for Matlab, http://www.math.udel.edu/˜driscoll/SC/.
- [23] C. Kimberling, Encyclopedia of Triangle Centers, http://faculty.evansville.edu/ck6/encyclopedia/.
Steven Finch Dept. of Statistics Harvard University Cambridge, MA, USA
steven [email protected]
| null |
[
"Steven R. Finch"
] |
2014-08-05T19:35:36+00:00
|
2014-08-05T19:35:36+00:00
|
[
"math.HO",
"math.MG",
"51M04 (Primary) 30C20, 30C30, 31A15, 33C65, 35J08 (Secondary)"
] |
Appell F1 and Conformal Mapping
|
This is the last of a trilogy of papers on triangle centers. A fairly obscure
"conformal center of gravity" is computed for the class of all isosceles
triangles. This calculation appears to be new. A byproduct is the logarithmic
capacity or transfinite diameter of such, yielding results consistent with
Haegi (1951).
|
1408.1075v2
|
## Searching for dilepton resonances below the Z mass at the LHC
Isaac Hoenig, Gabriel Samach, and David Tucker-Smith Department of Physics, Williams College, Williamstown, MA 01267
(Dated: October 14, 2014)
## Abstract
We consider LHC searches for dilepton resonances in an intermediate mass range, ∼ 10 -80 GeV. We adopt a kinetically mixed Z ′ as an example of weakly coupled new physics that might have evaded detection at previous experiments but which could still be probed by LHC dilepton spectrum measurements in this mass range. Based on Monte Carlo simulations, we estimate that existing data from the 7 and 8 TeV LHC could be used to test values of the kinetic mixing parameter glyph[epsilon1] several times smaller than precision electroweak upper bounds, were an appropriate analysis to be carried out by one of the experimental collaborations.
## I. INTRODUCTION
The appearance of a new dilepton resonance at the Large Hadron Collider (LHC) would mean the discovery of physics beyond the Standard Model (SM). Both ATLAS and CMS have searched for dilepton resonances, focusing mainly on the high-mass region above ∼ 100 GeV. Here we argue that there is something to be gained by searching at lower masses - we focus on the ∼ 10 -80 GeV mass range, above the masses of the heaviest hadronic bound states and below the Z peak. To explore whether LHC searches in this mass range could probe parameter space not already excluded by previous experiments we adopt a particular model for a weakly coupled dilepton resonance: a kinetically mixed Z ′ [1]. A kinetically mixed Z ′ is a simple addition to the SM. It is also a central feature in certain models of dark matter [2-10].
For our purposes, the kinetically mixed Z ′ scenario is parameterized by the Z ′ mass M Z ′ and the kinetic mixing parameter glyph[epsilon1] , which controls the coupling strength of the Z ′ to SM particles. We use Monte Carlo simulations to estimate the potential sensitivity of specialized analyses of existing 7 and 8 TeV data to a kinetically mixed Z ′ , and to estimate the potential sensitivity of 14 TeV data in the longer term. We also investigate whether the CMS analysis of the dilepton invariant-mass spectrum at 7 TeV [11] implies interesting upper bounds on glyph[epsilon1] for various intermediate M Z ′ values, although this exercise is no substitute for an analysis carried out by an experimental collaboration.
Our main results are summarized in Figure 9. We find that CMS results presented in Ref. [11] imply upper bounds on glyph[epsilon1] that lie below precision electroweak upper bounds in the 30 -70 GeV mass range. More specialized analyses of existing data could probe glyph[epsilon1] values as small as a factor of ∼ 5 below precision electroweak upper bounds. The analyses we test start with standard muon p T cuts used in CMS analyses (14 and 9 GeV for the dimuon pair for √ s = 7 TeV, and 20 and 10 GeV for √ s = 8 TeV), but would still be sensitive to glyph[epsilon1] < ∼ 10 -2 for Z ′ masses as small as M Z ′ ∼ 12 GeV, although for M Z ′ below 14 GeV the CMS search for a light pseudoscalar decaying to µ + µ -already implies exclusion down to significantly smaller values of glyph[epsilon1] [12].
We focus on a kinetically mixed Z ′ , but can conclude more generally that, even for dilepton resonances below M Z , the LHC can compete with previous experimental probes. Our results suggest that it would be worthwhile for LHC experimental collaborations to carry
out analyses searching for dilepton resonances in the intermediate mass range considered here. One can interpret our sensitivity estimates in the context of a different model by using the cross sections of Table I and the fact that the production cross section is proportional to glyph[epsilon1] 2 , to convert the glyph[epsilon1] values of Figure 9 to σ pp ( → Z ′ → l + l -) values.
In the following section we review the kinetically mixed Z ′ model and current experimental constraints. In Section III we describe our simulation and analysis methods and present sensitivity estimates for 7 TeV data. In Section IV we present sensitivity estimates for 8 TeV data and, under certain assumptions, for the 14 TeV LHC.
## II. A KINETICALLY MIXED Z ′
The model we consider is defined by the Lagrangian
$$\mathcal { L } = \mathcal { L } _ { S M } - \frac { 1 } { 4 } B ^ { \prime } _ { \mu \nu } B ^ { \prime \mu \nu } - \frac { \epsilon } { 2 } B _ { \mu \nu } B ^ { \prime \mu \nu } + \frac { M _ { B ^ { \prime } } ^ { 2 } } { 2 } B ^ { \prime } _ { \mu } B ^ { \prime \mu } + \cdots,$$
where L SM is the SM Lagrangian, B µν is the hypercharge field strength tensor, and B ′ µν is the field strength tensor for a new U (1) ′ gauge field. The mass for the U (1) ′ gauge field, M B ′ , can be generated by the vacuum expectation value of a SM-singlet scalar field Φ charged under the U (1) , whose Lagrangian terms we do not write explicitly. ′ The phenomenology of Φ can be interesting in its own right, but we do not consider it here. The collider implications of U (1) ′ gauge bosons in this type of setup have previously been considered in Refs. [10, 13-31], for example.
The neutral SU (2), U (1) and U (1) ′ gauge fields W 3 µ , B µ , and B ′ µ are related to masseigenstate gauge fields with diagonal and canonically normalized kinetic terms, A µ , Z µ , and Z ′ µ , by
$$\mu \cdot \cdot \cdot \\ \mu \cdot \\ \mu \cdot \\ B ^ { \prime } _ { \mu } } = \begin{pmatrix} b _ { \mu } \\ c _ { w } & - \left ( s _ { w } c _ { z } + \frac { \epsilon } { \sqrt { 1 - \epsilon ^ { 2 } } } s _ { z } \right ) & \left ( s _ { w } s _ { z } - \frac { \epsilon } { \sqrt { 1 - \epsilon ^ { 2 } } } c _ { z } \right ) \\ s _ { w } & c _ { w } c _ { z } & - c _ { w } s _ { z } \\ 0 & \frac { 1 } { \sqrt { 1 - \epsilon ^ { 2 } } } s _ { z } & \frac { 1 } { \sqrt { 1 - \epsilon ^ { 2 } } } c _ { z } \end{pmatrix} \begin{pmatrix} A _ { \mu } \\ Z _ { \mu } \\ Z ^ { \prime } _ { \mu } \end{pmatrix}, \quad \quad ( 2 ) \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ \mu \cdot \\ },$$
where c w and s w are cosine and sine of the weak mixing angle, and c z and s z are cosine and sine of the angle that parameterizes Z -Z ′ mixing, determined by
$$\alpha \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a u v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v a v e p u m { \ v b u n u n u n u n u n u n } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } { } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } _ } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } \ } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } }, } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } ) } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } \\ } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } ( } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } | } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } }) } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } & \ \) } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
Here M Z is the SM value for the mass of the Z boson, i.e. its value in the glyph[epsilon1] → 0 limit.
One can use the field transformation of Equation (2) and the charges of SM fermions under the SM gauge group to calculate the couplings of SM fermions to the Z ′ . If both
$$\left | \frac { \epsilon s _ { w } \overline { M } _ { Z } ^ { 2 } } { \overline { \overline { \overline { \imath } } ^ { 2 } } \quad \overline { \imath } \, 2 } \right | \ll 1$$
$$\epsilon \ll 1 \quad \text{ and } \quad \left | \frac { \epsilon s _ { w } \overline { M } _ { Z } ^ { 2 } } { \overline { M } _ { Z } ^ { 2 } - M _ { B ^ { \prime } } ^ { 2 } } \right | \ll 1$$
are satisfied, we can work to first order in glyph[epsilon1] , giving
$$\theta _ { z } \simeq \frac { \epsilon s _ { w } \overline { M } _ { Z } ^ { 2 } } { \overline { M } _ { Z } ^ { 2 } - M _ { B ^ { \prime } } ^ { 2 } }.$$
In this regime, the couplings of the Z ′ to SM fermions are
$$g _ { \bar { f } f Z ^ { \prime } } \simeq - \epsilon \left ( \frac { \overline { M } _ { Z } ^ { 2 } c _ { w } e Q _ { f } - M _ { B ^ { \prime } } ^ { 2 } g _ { y } Y _ { f } } { \overline { M } _ { Z } ^ { 2 } - M _ { B ^ { \prime } } ^ { 2 } } \right ),$$
where e and g y are the electromagnetic and hypercharge gauge couplings, Q f and Y f are the electric charge and hypercharge of the fermion, and where we use the notation
$$\mathcal { L } \supset \sum _ { f } g _ { \bar { f } f Z ^ { \prime } } \overline { f } \gamma ^ { \mu } f Z ^ { \prime } _ { \mu },$$
with f = e L , e R , and so on. When either M B ′ glyph[lessmuch] M Z or M B ′ glyph[greatermuch] M Z is satisfied we can further approximate Equation (6) as
$$g _ { \bar { f } f Z ^ { \prime } } \simeq \begin{cases} - \epsilon c _ { w } e Q _ { f } & M _ { B ^ { \prime } } \ll \overline { M } _ { Z } \\ - \epsilon g _ { y } Y _ { f } & M _ { B ^ { \prime } } \gg \overline { M } _ { Z }. \end{cases}$$
In the context of a dark matter model, the Z ′ might have a sizable branching ratios to dark-sector final states. In this work we assume that the Z ′ decays exclusively to SM states. This possibility fits in naturally with viable dark matter models, even for small values of glyph[epsilon1] . For example, if the dark matter is a fermion χ charged under U (1) , ′ Z ′ decays to χ pairs are forbidden for M > M χ Z ′ / 2, and for M > M χ Z ′ the χ relic abundance can be regulated by the annihilation process χ χ ( -) → Z Z ′ ′ for an appropriate value of the U (1) ′ gauge coupling g ′ [32]. For the parameter region relevant to the collider studies considered in this paper ( M ′ Z in the ∼ 10 -80 GeV range and glyph[epsilon1] > 10 -3 ) results from LUX [33] rule out this χ χ ( -) → Z Z ′ ′ scenario for the case where the dark matter is a Dirac fermion, but the Majorana case is viable. A Majorana mass can be generated by Φ -χ interactions for suitably chosen U (1) ′ charges.
Although dark matter offers one motivation to search for weakly coupled dilepton resonances, our results do not depend on the possible connection to dark matter. The key assumption is that the width of the dilepton resonance is dominated by SM final states.
Experimental and observational constraints on the kinetically mixed Z ′ scenario are summarized in Ref. [34]. In the ∼ 10 -80 GeV mass range that interests us, precision electroweak constraints require glyph[epsilon1] < ∼ (2 -3) × 10 -2 [25]. If the Z ′ decays dominantly to SM states, as assumed here, there are also constraints from PEP, PETRA, and TRISTAN measurements of the e + e -→ hadrons cross sections at various center-of-mass energies. As shown in Ref. [25], these constraints are stronger than those from precision electroweak measurements at particular values of M Z ′ close to the experiments' center-of-mass energies, but they do not cover the mass range continuously.
At higher masses, LHC searches for high-mass dilepton resonances can be used to constrain the kinetically mixed Z ′ scenario. These constraints were derived in Ref. [30] using 7 TeV LHC results. Figure 1 shows the upper bounds on glyph[epsilon1] implied by ATLAS [35] and CMS [36] 8 TeV results. To obtain these constraints we calculate the cross section
FIG. 1. 95% CL upper bounds on glyph[epsilon1] from ATLAS [35] (dashed blue line) and CMS [36] (thick green line) searches for high-mass dilepton resonances, along with the precision electroweak constraint taken from Ref. [25] (thin red line).
for pp → Z ′ → l + l -at leading order using Madgraph [37] with CTEQ66 [38, 39] parton distribution functions, and apply a mass-dependent NNLO K-factor determined using the
## ZWPROD code [40, 41].
Given how successful LHC searches for high-mass dilepton resonances have been at probing new parameter space for the kinetically mixed Z ′ scenario, a natural question is whether LHC data can be used to probe new parameter space for M Z ′ < M Z . The results of a CMS search for a light pseudoscalar Higgs in the dimuon channel, which relied on a specialized trigger, imply a limit on glyph[epsilon1] for M Z ′ in the 5 5 . -14 GeV mass range [12] (for an earlier ATLAS analysis, see Ref. [42]). For example, assuming equal acceptances for the Z ′ and the pseudoscalar, we find that the CMS upper limit implies glyph[epsilon1] < 2 × 10 -3 for M Z ′ = 12 GeV. In the following sections we consider whether LHC studies of the dimuon invariant mass distribution could in principle be used to probe the kinetically mixed Z ′ model for a broader range of intermediate Z ′ masses. From simulations with standard cuts taken from existing analyses, we find that current data could be used to probe glyph[epsilon1] values well below the precision electroweak limit for much of the mass range below M Z , although not down to the glyph[epsilon1] ∼ 2 × 10 -3 level probed by the CMS pseudoscalar search for masses below 14 GeV. It is possible that Tevatron data could also be used to probe glyph[epsilon1] values below precision electroweak upper bounds. Our rough estimates suggest that it is unlikely that the sensitivity of Tevatron data would exceed that of LHC data, but we have not investigated the Tevatron sensitivity in detail.
## III. POTENTIAL SENSITIVITY OF 7 TEV LHC DATA
## A. simulation methods
We perform Monte Carlo simulations of dimuon production to estimate the sensitivity of LHC data to the kinetically mixed Z ′ model for M Z ′ < M Z . We implement the model in FeynRules [43] and simulate both the Z ′ signal and the dominant Drell-Yan background using MadGraph5 [37] and Pythia 6.4 [44], with Madgraph's implementation of MLM matching [45] turned on, and up to two jets included at matrix element level. We approximate detector resolution effects by performing a Gaussian smearing of the p T 's of the muons using the
smearing function
$$\frac { \sigma _ { p _ { T } } } { p _ { T } } = \begin{cases} 0. 0 3 & | \eta | < 1. 5 \text{ and } p _ { T } < 2 0 0 \\ 0. 0 4 & | \eta | \geq 1. 5 \text{ and } p _ { T } < 2 0 0 \\ 0. 0 5 & p _ { T } \geq 2 0 0 \end{cases}$$
This smearing function is consistent with the settings in the default CMS card for the Delphes 3.0 fast detector simulator, which have been shown to give reasonable agreement with data [46]. The sizes of our background Monte Carlo samples correspond to effective luminosities roughly a factor of ten times larger than the luminosities used for the associated analyses.
Drell-Yan production in the mass range that interests us has been studied at 7 TeV by ATLAS [47] and CMS [11]. In this section we will frequently refer to the CMS analysis, which uses an integrated luminosity of 4.5 fb -1 . Figure 1 of that paper shows that, for the selection cuts used, Drell Yan production accounts for approximately 90% of dimuon production in the mass range that interests us. The CMS analysis requires the leading and subleading muons to have ( p T ) 1 > 14 GeV and ( p T ) 2 > 9 GeV, respectively, with both muons' pseudorapidities satisfying | η | < 2 4. . For these cuts, Figure 2 compares the dimuon invariant-mass distribution produced by our simulations with that produced by the CMS full simulation of the Drell-Yan process, as reported in Ref. [11]. We rescale our distribution to produce the same number of events as the CMS simulation in the 60 -120 GeV mass range. We find agreement at the 5% level for the bins in the 25 -76 GeV mass range. Larger discrepancies approaching 15% are evident in the two lowest-mass bins, where higher order effects are most important. Larger discrepancies are also evident around the Z mass, where the bin counts depend sensitively on the muon momentum resolution. Our simulations give a Z peak that is slightly too short and broad, suggesting that we are not unreasonably optimistic in our estimation of the resolution.
A light kinetically mixed Z ′ can also be searched for in the dielectron channel, and combining dimuon and dielectron results would likely lead to improved sensitivity. For simplicity we restrict our attention to the dimuon channel, in part because, as shown in Figure 3 of Ref. [11], the efficiency in the dielectric channel is significantly lower in the invariant mass range that interests us.
FIG. 2. Left: with √ s = 7 TeV and an integrated luminosity of 4.5 fb -1 , a comparison between the Drell-Yan dimuon invariant-mass distributions predicted by CMS simulations (red squares) and by our simulations (blue circles). The muons are required to have ( p T ) 1 > 14 GeV, ( p T ) 2 > 9 GeV, and | η | < 2 4. . We rescale our m µµ distribution to give the same total counts as the CMS simulation in the 60 -120 GeV mass range. Right: discrepancies between our simulated bin counts and those of CMS.
## B. A benchmark scenario: M Z ′ = 50 GeV, glyph[epsilon1] = 0 02 .
To establish that a Z ′ signal would be detectable at the LHC for parameters consistent with precision electroweak bounds, we adopt the parameter point ( M Z ′ = 50 GeV, glyph[epsilon1] = 0 02), which lies slightly below the precision electroweak exclusion contour in Figure 1. . We normalize our background invariant-mass distribution to data, using the number of events reported in the 60 -120 GeV mass range in the CMS analysis of Ref. [11]. To normalize the signal we use the matched cross section calculated by MadGraph with CTEQ66 parton distributions [38, 39] and, in addition to selection cuts, we apply an efficiency of 75%, consistent with or slightly below the efficiencies reported in Ref. [11]. Signal cross sections calculated by MadGraph for various Z ′ masses are given in Table I.
The invariant mass distribution for signal plus background for M Z ′ = 50 GeV and glyph[epsilon1] = 0 02, using 1 GeV bins, is shown in Figure 3. For the parameters of this benchmark scenario, . the Z ′ signal would be clearly visible in 7 TeV data, motivating us to investigate further the potential sensitivity of the LHC to the kinetically mixed Z ′ model in the M Z ′ < M Z regime.
FIG. 3. With √ s = 7 TeV and an integrated luminosity of 4.5 fb -1 , simulated signal plus background for the benchmark parameter point M Z ′ = 50 GeV and glyph[epsilon1] = 0 02. . The muons are required to have ( p T ) 1 > 14 GeV, ( p T ) 2 > 9 GeV, and | η | < 2 4. . We normalize background to data and apply a 75% efficiency to signal.
## C. Models for the dimuon invariant mass distributions
To estimate the glyph[epsilon1] reach of 7 TeV LHC data, we adopt models for the background and signal invariant mass distributions. We model the signal using the invariant mass bin counts produced by our Monte Carlo simulations. Throughout we apply a 75% efficiency to the signal. For the background we employ a more flexible model that allows for deviations of actual SM m µµ distributions from those predicted by simulations. The bin counts predicted by our background model are
$$b _ { i } ( \theta, \delta ) = p _ { i } ( \theta ) ( 1 + \delta _ { i } ),$$
where p i represents a fifth-order polynomial with coefficients θ , and the δ i are additional nuisance parameters, one for each bin. When we incorporate a possible signal the predicted bin counts become
$$\nu _ { i } ( \mu, \theta, \delta ) = \mu s _ { i } + b _ { i } ( \theta, \delta ),$$
where s i are the Monte-Carlo derived signal bin counts for particular pair of ( , glyph[epsilon1] M Z ′ ) values, and µ controls the signal strength.
We quantify the compatibility of signal strength µ with data n i by
$$\chi ^ { 2 } ( \mu ) = \min _ { \{ \theta, \, \delta \} } \left \{ \sum _ { i } \left [ \frac { \left [ n _ { i } - \nu _ { i } ( \mu, \theta, \delta ) \right ] ^ { 2 } } { \nu _ { i } ( \mu, \theta, \delta ) } + \left ( \frac { \delta _ { i } } { \kappa } \right ) ^ { 2 } \right ] \right \}.$$
In this expression, the background parameters θ and δ are set to their best-fit values for signal strength µ . That is, they are chosen to minimize the quantity in brackets. The δ parameters are effectively associated with fake measurements of uncertainty κ . In the κ → 0 limit the background model reduces to a fifth-order polynomial. We describe how we choose κ values below.
Taking the signal strength to zero and neglecting the δ parameters (or equivalently, sending κ → 0), we can identify the θ parameters that give the best fifth-order polynomial fit to our Drell-Yan Monte Carlo m µµ distribution. Figure 4 shows this fit for the 35 -75 GeV mass range. At lower masses, matters are complicated by the fact that the muon p T
FIG. 4. For √ s = 7 TeV and an integrated luminosity of 4.5 fb -1 , our simulated Drell-Yan m µµ distributions with and without the ( p T ) µµ cut, along with best-fit fifth-order polynomial curves. The muons are required to have ( p T ) 1 > 14 GeV, ( p T ) 2 > 9 GeV, and | η | < 2 4. .
cuts produce a peak in the invariant mass distribution near m µµ = 30 GeV. Dimuon masses below ∼ 30 GeV require the dimuon system to recoil off of a jet or jets in order to satisfy ( p T ) 1 > 14 GeV and ( p T ) 2 > 9 GeV .
A cut on ( p T ) µµ , the transverse momentum of the dimuon system, can be used to probe Z ′ masses below ∼ 30 GeV. We choose the cut on ( p T ) µµ to be equal to the sum of the two individual muon p T thresholds, 23 GeV for the muon p T cuts used in the CMS analysis of
Ref. [11]. With this choice, an event with very small m µµ that passes the individual muon p T cuts will also pass the cut on ( p T ) µµ . Here and for all of our analyses we require that the separation between the µ + and µ -satisfies ∆ R > 0 3. . Figure 4 shows the best fifth-order polynomial fit for the 12 -40 GeV mass range when this additional cut is applied to to our Drell-Yan Monte Carlo sample.
To calculate a p -value for the background-model fit to a given data set n i , we first use Equation (12) to calculate χ 2 (0) and extract best-fit background parameters θ (0) and δ (0) . We define the p -value for the fit to be
$$p = \int _ { \chi ^ { 2 } ( 0 ) } ^ { \infty } f ( x ; n _ { d } ) \ d x,$$
where f ( x n ; d ) is the probability distribution function for the chi squared distribution with n d degrees of freedom, and we take n d equal to the number of bins minus six, the number of parameters for the fifth-order polynomial [55].
In Section III E we estimate bounds on glyph[epsilon1] for various Z ′ masses based on the CMS results of Ref. [11], and in Section III F we estimate the potential sensitivity of 7 TeV LHC data using two hypothetical analyses with 1-GeV m µµ bins, one with the ( p T ) µµ > 23 GeV cut and the other without. Tables II-IV show the M Z ′ values tested for each analysis and the associated m µµ ranges. For each m µµ range, the third column of the table gives the median p -value for the fifth-order polynomial fit to pseudo data generated by statistically fluctuating Monte Carlo expectations for SM bin counts. We use the Monte Carlo expectations reported by CMS for the analysis based on the CMS results, and we use our own Monte Carlo simulations for the 1-GeV analyses
We choose κ values by imposing two requirements. First, we require that the median p value for background-model fits to SM Monte-Carlo-generated psuedo data is at least 0.5. Second, we require that when a signal corresponding to kinetic mixing parameter glyph[epsilon1] in is injected into these pseudo-data sets, the 95% confidence level upper bound on glyph[epsilon1] , glyph[epsilon1] 95 , is found to be less than glyph[epsilon1] in for fewer than 5% of the pseudo-data sets. Our procedure for calculating glyph[epsilon1] 95 for a particular data set is described in the following section. Because we impose the second requirement independently for each hypothesized value of M Z ′ , different values of M Z ′ can have different values of κ . For each value of M Z ′ , the second requirement is imposed over a range of glyph[epsilon1] in values extending down to zero and up to twice the median of the glyph[epsilon1] 95 values obtained for the SM-only pseudo-data sets. The κ values obtained by imposing
our two requirements fill the fourth column in Tables II-IV.
The κ parameter is meant to take into account imperfections in the polynomial background model. An actual experimental analysis would deal with systematic effects not considered here, and these effects may make the background m µµ distribution less smooth than our Monte Carlo simulations would suggest. Along with the κ values given in Tables II-IV, we present results for κ = 0, κ = 10 -2 , κ = 2 × 10 -2 , and κ = 3 × 10 -2 to give a sense of how the potential sensitivity is affected if systematic effects introduce an additional source of bin-to-bin randomness.
## D. Statistical procedure
Given a hypothesized value of M Z ′ and an observed invariant mass distribution n i - which could either stand for actual data reported by CMS in Ref. [11], or for pseudo data generated from our Monte Carlo simulations - we use the CL s method [48] to determine glyph[epsilon1] 95 , the 95% confidence level upper limit on the kinetic mixing parameter for that value of M Z ′ . We use a test statistic based on the χ 2 defined in Equation (12):
$$q _ { \mu } = \begin{cases} \chi ^ { 2 } ( \mu ) - \chi ^ { 2 } ( \hat { \mu } ) & \mu > \hat { \mu } \\ 0 & \mu \leq \hat { \mu }, \end{cases}$$
where ˆ, is the best-fit signal strength. µ Using the asymptotic formulae of Ref. [49], we can approximate CL s as √
$$C L _ { s } = \frac { 1 - \Phi ( \sqrt { q _ { \mu } } ) } { 1 - \Phi ( \sqrt { q _ { \mu } } - \sqrt { q _ { \mu, A } } ) },$$
where Φ is the cummulative distribution of the standard Gaussian and q µ,A is the value of q µ obtained when the data n i are taken to be equal to ν i (0 , θ (0) , δ (0) ), the expected bin counts for the best-fit background model. For the purpose of calculating q µ,A , we replace δ i is with δ i -δ (0) i in the χ 2 expression of Equation (12). We define glyph[epsilon1] 95 to be the kinetic mixing parameter that gives CL s = 0 05 for . µ = 1.
## E. Estimates of bounds on glyph[epsilon1] from the 7 TeV CMS analysis
Focusing on the 30 -76 GeV bins, the smoothness of the m µµ distribution for the data shown in Figure 1 of the CMS analysis of Ref. [11] implies a constraint on the couplings
of a Z ′ with a mass in this range. We use the background model and statistical procedure outlined in Sections III C and III D to estimate upper bounds on glyph[epsilon1] based on the CMS results. The Z ′ masses tested, associated m µµ fit ranges, and κ values are shown in Table II.
GLYPH<230>
e
'
FIG. 5. Estimated 95% CL upper bounds on glyph[epsilon1] based on Ref. [11], the √ s = 7 TeV CMS study of the dimuon invariant mass distribution. For the thick blue line, κ is set according to the procedure described in Section III C. The dashed lines show results for other κ values. The thin red line is the precision electroweak constraint taken from Ref. [25].
Based on the reported data bin counts from Figure 1 of Ref. [11], we obtain the glyph[epsilon1] 95 estimates presented in Figure 5. The estimated glyph[epsilon1] bounds corresponding to the κ values of Table II lie below precision electroweak constraints by as much as a factor of two. Of the five masses tested, the four lowest lie on the boundary between two of the invariant mass bins used in the CMS analysis, meaning that the signal events are roughly equally distributed between two relatively wide mass bins. This pushes our estimates in the conservative direction, as far as other masses are concerned.
## F. Sensitivity estimates for 7 TeV data
A more specialized analysis of LHC data could probe lower values of glyph[epsilon1] and a broader range of Z ′ masses. To estimate the potential sensitivity of √ s = 7 TeV data we calculate median values of glyph[epsilon1] 95 for fake data generated from our Drell-Yan Monte Carlo simulations. Although it would be more interesting for the LHC to find evidence for a light Z ′ rather
than to rule out additional parameter space, we use potential exclusion sensitivity in the absence of new physics as a diagnostic for whether new parameter space can be probed.
We consider two hypothetical analyses with 1 GeV m µµ bins, both based on the same integrated luminosity as the CMS analysis. For the first, which incorporates the same muon p T and η cuts as the CMS analysis, we use the Z ′ masses, m µµ fit ranges, and κ values from Table III. The second analysis adds the cut ( p T ) µµ > 23 GeV, and has been tested with the Z ′ masses, m µµ fit ranges, and κ values from Table IV. We normalize our Monte Carlo simulations so that, in the absence of the ( p T ) µµ cut, we get the same number of selected events in the 60 -120 m µµ range as the CMS analysis, as determined using Figure 1 of Ref. [11].
FIG. 6. Exclusion sensitivity estimates for 7 TeV LHC data based on an integrated luminosity of 4.5 fb -1 , with the cut ( p T ) µµ > 23 GeV (left), and without a ( p T ) µµ cut (right). For the thick blue line, κ is set according to the procedure described in Section III C. The dashed lines show results for other κ values. The thin red line is the precision electroweak constraint taken from Ref. [25].
Our resulting sensitivity estimates are shown in Figure 6. For the κ values of Tables III and IV, the median values of glyph[epsilon1] 95 go as low as a factor of four below the upper bound on glyph[epsilon1] from precision electroweak constraints.
## IV. POTENTIAL SENSITIVITY OF 8 TEV AND 14 TEV LHC DATA
At √ s = 8 TeV, the advantages of a larger integrated luminosity ∼ 20 fb -1 and slightly higher signal cross sections compete against higher muon p T thresholds required by the higher instantaneous luminosity. Based on Ref. [50], we conclude that it is realistic to
consider hypothetical analyses based on muon p T cuts of 20 GeV and 10 GeV. These cuts push the peak in the m µµ invariant mass distribution to larger values of m µµ , as shown in Figure 7.
FIG. 7. For √ s = 8 TeV and an integrated luminosity of 20 fb -1 , our simulated Drell-Yan m µµ distributions with and without the ( p T ) µµ cut, along with best-fit fifth-order polynomial curves. The muons are required to have ( p T ) 1 > 20 GeV, ( p T ) 2 > 10 GeV, and | η | < 2 4. .
As for the 7 TeV case, we consider two hypothetical analyses with 1-GeV m µµ bins, one with a ( p T ) µµ cut and one without. We take the integrated luminosity to be 20 fb -1 . Fifthorder polynomial fits to our Drell-Yan Monte Carlo m µµ distributions for the 12 -40 and 40 -75 GeV mass ranges are shown in Figure 7. Tables V and VI show the Z ′ masses tested for these analyses, the associated m µµ ranges, and the κ values determined the procedure described in Section III C.
As before we apply a 75% efficiency for the signal. We normalize our Drell-Yan Monte Carlo simulations by requiring that the number of events in the 60 -120 GeV m µµ mass range is equal to the number of events in that mass range for the 7 TeV case, multiplied by by three factors: the ratio of the 8 and 7 TeV integrated luminosities, the ratio of the 8 and 7 TeV acceptances in that m µµ range for the relevant muon p T and η cuts, as determined by our simulations, and the ratio of the 8 and 7 TeV NNLO dimuon production cross sections in that m µµ range. We use σ 8 TeV = 1 12 nb and . σ 7 TeV = 0 97 nb, consistent with the values . shown in Refs. [51] based on calculations using FEWZ [52].
Our sensitivity estimates are shown in Figure 8. These results suggest that, relative to
95
e
0.030
0.025
0.020
0.015
0.010
0.005
0.000
GLYPH<230>
GLYPH<230>
GLYPH<230>
GLYPH<230>
GLYPH<230>
15
GLYPH<230>
GLYPH<230>
GLYPH<230>
GLYPH<230>
GLYPH<230>
20
25
GLYPH<230>
GLYPH<230>
GLYPH<230>
GLYPH<230>
GLYPH<230>
30
35
GLYPH<230>
GLYPH<230>
GLYPH<230>
GLYPH<230>
k
k
k
GLYPH<230>
= 0.03
k
= 0.02
= 0.01
= 0
40
45
MZ
'
H
GeV
L
FIG. 8. Exclusion sensitivity estimates for 8 TeV LHC data based on an integrated luminosity an integrated luminosity of 20 fb -1 , with the cut ( p T ) µµ > 30 GeV (left), and without a ( p T ) µµ cut (right). For the thick blue line, κ is set according to the procedure described in Section III C. The dashed lines show results for other κ values. The thin red line is the precision electroweak constraint taken from Ref. [25].
the 7 TeV case, an analysis based on 8 TeV data would yield improved sensitivity for at least some of the m µµ range below the Z mass. Figure 9 collects our results for estimated bounds on glyph[epsilon1] based on the 7 TeV CMS analysis along with estimates of potential sensitivity for 7 and 8 TeV.
The ability of the 14 TeV LHC to probe relatively light dilepton resonances depends on trigger issues. Refs. [53, 54] give reason to hope that various upgrades should allow trigger thresholds to be close to their 8 TeV values, even at the High Luminosity LHC. We will estimate the 14 TeV reach for the kinetically mixed Z ′ model under the assumption that the dimuon p T cuts we considered for 8 TeV analyses would be feasible at 14 TeV.
For the mass range considered here, our simulations indicate that the the glyph[epsilon1] reach for 14 TeV and an integrated luminosity of 10 fb -1 is approximately the same as the reach for 8 TeV and an integrated luminosity of 20 fb -1 . If we optimistically assume that the sensitivity for higher luminosities depends on S/ √ B , and therefore that the glyph[epsilon1] reach is proportional to (∫ L dt ) -1 4 / , we obtain the 14 TeV sensitivity estimates shown in Figure 9 for an integrated luminosity of 3000 fb -1 . These reach down to glyph[epsilon1] ∼ 10 -3 .
GLYPH<230>
GLYPH<230>
GLYPH<230>
GLYPH<230>
GLYPH<230>
95
e
e
FIG. 9. Estimated upper bounds on glyph[epsilon1] based on the CMS analysis of Ref. [11] (black, solid), and exclusion sensitivity estimates for 7 TeV LHC data (blue, short dashed), for 8 TeV LHC data (green, long dashed), and, under the assumptions stated in the text, for 3000 fb -1 of 14 TeV data (orange, dot-dashed). Results for the low-mass analyses incorporating a ( p T ) µµ cut are indicated by square markers. The κ parameters are set according to the procedure described in Section III C. The thin red line is the precision electroweak constraint taken from Ref. [25].
## V. CONCLUSIONS
LHC measurements of dilepton invariant mass distributions below M Z can be used as sensitives probe of new physics. For the kinetically mixed Z ′ model, analyses of 7 and 8 TeV data would be sensitive to glyph[epsilon1] values several times smaller than the upper bounds from precision electroweak constraints. If trigger thresholds can be held near their 8 TeV values, glyph[epsilon1] values as small as ∼ 10 -3 may be accessible to the High Luminosity LHC.
## ACKNOWLEDGMENTS
We are grateful to Kyle Cranmer, David Curtin, Stefania Gori, and Abi Sofer for helpful conversations. This work was supported by NSF Grant #1216168. DTS thanks the Aspen Center for Physics and NSF Grant #1066293 for hospitality as this paper was being
completed.
- [1] B. Holdom, Phys. Lett. B 166 , 196 (1986).
- [2] C. Boehm and P. Fayet, Nucl. Phys. B 683 , 219 (2004) [hep-ph/0305261].
- [3] M. Pospelov, A. Ritz and M. B. Voloshin, Phys. Lett. B 662 , 53 (2008) [arXiv:0711.4866 [hep-ph]].
- [4] R. Foot, Phys. Rev. D 78 , 043529 (2008) [arXiv:0804.4518 [hep-ph]].
- [5] N. Arkani-Hamed, D. P. Finkbeiner, T. R. Slatyer and N. Weiner, Phys. Rev. D 79 , 015014 (2009) [arXiv:0810.0713 [hep-ph]].
- [6] H. An, S. -L. Chen, R. N. Mohapatra and Y. Zhang, JHEP 1003 , 124 (2010) [arXiv:0911.4463 [hep-ph]].
- [7] E. J. Chun, J. -C. Park and S. Scopel, JHEP 1102 , 100 (2011) [arXiv:1011.3300 [hep-ph]].
- [8] S. Andreas, M. D. Goodsell and A. Ringwald, Phys. Rev. D 87 , 025007 (2013) [arXiv:1109.2869 [hep-ph]].
- [9] H. Davoudiasl and I. M. Lewis, Phys. Rev. D 89 , 055026 (2014) [arXiv:1309.6640 [hep-ph]].
- [10] J. M. Cline, G. Dupuis, Z. Liu and W. Xue, arXiv:1405.7691 [hep-ph].
- [11] S. Chatrchyan et al. [CMS Collaboration], JHEP 1312 , 030 (2013) [arXiv:1310.7291 [hep-ex]].
- [12] S. Chatrchyan et al. [CMS Collaboration], Phys. Rev. Lett. 109 , 121801 (2012) [arXiv:1206.6326 [hep-ex]].
- [13] K. S. Babu, C. F. Kolda and J. March-Russell, Phys. Rev. D 57 , 6788 (1998) [hep-ph/9710441].
- [14] T. G. Rizzo, Phys. Rev. D 59 , 015020 (1998) [hep-ph/9806397].
- [15] M. J. Strassler and K. M. Zurek, Phys. Lett. B 651 , 374 (2007) [hep-ph/0604261].
- [16] J. Kumar and J. D. Wells, Phys. Rev. D 74 , 115017 (2006) [hep-ph/0606183].
- [17] W. -F. Chang, J. N. Ng and J. M. S. Wu, Phys. Rev. D 74 , 095005 (2006) [Erratum-ibid. D 79 , 039902 (2009)] [hep-ph/0608068].
- [18] D. Feldman, Z. Liu and P. Nath, Phys. Rev. D 75 , 115001 (2007) [hep-ph/0702123 [HEP-PH]].
- [19] N. Arkani-Hamed and N. Weiner, JHEP 0812 (2008) 104 [arXiv:0810.0714 [hep-ph]].
- [20] S. Cassel, D. M. Ghilencea and G. G. Ross, Nucl. Phys. B 827 , 256 (2010) [arXiv:0903.1118 [hep-ph]].
- [21] B. Batell, M. Pospelov and A. Ritz, Phys. Rev. D 79 , 115008 (2009) [arXiv:0903.0363 [hepph]].
- [22] R. Essig, P. Schuster and N. Toro, Phys. Rev. D 80 , 015003 (2009) [arXiv:0903.3941 [hep-ph]].
- [23] J. D. Bjorken, R. Essig, P. Schuster and N. Toro, Phys. Rev. D 80 , 075018 (2009) [arXiv:0906.0580 [hep-ph]].
- [24] B. Batell, M. Pospelov and A. Ritz, Phys. Rev. D 80 , 095024 (2009) [arXiv:0906.5614 [hepph]].
- [25] A. Hook, E. Izaguirre and J. G. Wacker, Adv. High Energy Phys. 2011 , 859762 (2011) [arXiv:1006.0973 [hep-ph]].
- [26] M. Williams, C. P. Burgess, A. Maharana and F. Quevedo, JHEP 1108 , 106 (2011) [arXiv:1103.4556 [hep-ph]].
- [27] M. T. Frandsen, F. Kahlhoefer, A. Preston, S. Sarkar and K. Schmidt-Hoberg, JHEP 1207 , 123 (2012) [arXiv:1204.3839 [hep-ph]].
- [28] N. Toro and I. Yavin, Phys. Rev. D 86 , 055005 (2012) [arXiv:1202.6377 [hep-ph]].
- [29] V. Barger, D. Marfatia and A. Peterson, Phys. Rev. D 87 , 015026 (2013) [arXiv:1206.6649 [hep-ph]].
- [30] J. Jaeckel, M. Jankowiak and M. Spannowsky, Phys. Dark Univ. 2 , 111 (2013) [arXiv:1212.3620 [hep-ph]].
- [31] H. Davoudiasl, H. -S. Lee, I. Lewis and W. J. Marciano, Phys. Rev. D 88 , no. 1, 015022 (2013) [arXiv:1304.4935 [hep-ph]].
- [32] D. Hooper, N. Weiner and W. Xue, Phys. Rev. D 86 , 056009 (2012) [arXiv:1206.2929 [hep-ph]].
- [33] D. S. Akerib et al. [LUX Collaboration], Phys. Rev. Lett. 112 , 091303 (2014) [arXiv:1310.8214 [astro-ph.CO]].
- [34] J. Jaeckel, Frascati Phys. Ser. 56 , 172 (2012) [arXiv:1303.1821 [hep-ph]].
- [35] [ATLAS Collaboration], ATLAS-CONF-2013-017.
- [36] [CMS Collaboration], CMS PAS EXO-12-061.
- [37] J. Alwall, M. Herquet, F. Maltoni, O. Mattelaer and T. Stelzer, JHEP 1106 , 128 (2011) [arXiv:1106.0522 [hep-ph]].
- [38] S. Kretzer, H. L. Lai, F. I. Olness and W. K. Tung, Phys. Rev. D 69 , 114005 (2004) [hepph/0307022].
- [39] P. M. Nadolsky, H. -L. Lai, Q. -H. Cao, J. Huston, J. Pumplin, D. Stump, W. -K. Tung and C. -P. Yuan, Phys. Rev. D 78 , 013004 (2008) [arXiv:0802.0007 [hep-ph]].
- [40] R. Hamberg, W. L. van Neerven and T. Matsuura, Nucl. Phys. B 359 , 343 (1991) [Erratumibid. B 644 , 403 (2002)].
- [41] W. L. van Neerven and E. B. Zijlstra, Nucl. Phys. B 382 , 11 (1992) [Erratum-ibid. B 680 , 513 (2004)].
- [42] [ATLAS Collaboration], ATLAS-CONF-2011-020.
- [43] N. D. Christensen and C. Duhr, Comput. Phys. Commun. 180 , 1614 (2009) [arXiv:0806.4194 [hep-ph]].
- [44] T. Sjostrand, S. Mrenna and P. Z. Skands, JHEP 0605 , 026 (2006) [hep-ph/0603175].
- [45] J. Alwall, S. Hoche, F. Krauss, N. Lavesson, L. Lonnblad, F. Maltoni, M. L. Mangano and M. Moretti et al. , Eur. Phys. J. C 53 , 473 (2008) [arXiv:0706.2569 [hep-ph]].
- [46] J. de Favereau et al. [DELPHES 3 Collaboration], JHEP 1402 , 057 (2014) [arXiv:1307.6346 [hep-ex]].
- [47] G. Aad et al. [ATLAS Collaboration], JHEP 1406 , 112 (2014) [arXiv:1404.1212 [hep-ex]].
- [48] A. L. Read, J. Phys. G 28 , 2693 (2002).
- [49] G. Cowan, K. Cranmer, E. Gross and O. Vitells, Eur. Phys. J. C 71 , 1554 (2011) [arXiv:1007.1727 [physics.data-an]].
- [50] S. Chatrchyan et al. [CMS Collaboration], JHEP 1401 , 096 (2014) [arXiv:1312.1129 [hep-ex]].
- [51] S. Chatrchyan et al. [CMS Collaboration], Phys. Rev. Lett. 112 , 191802 (2014) [arXiv:1402.0923 [hep-ex]].
- [52] R. Gavin, Y. Li, F. Petriello and S. Quackenbush, Comput. Phys. Commun. 182 , 2388 (2011) [arXiv:1011.3540 [hep-ph]].
- [53] [ATLAS Collaboration], ATL-PHYS-PUB-2013-004.
- [54] A. Tapper et al. [CMS Collaboration], CERN-LHCC-2013-011.
- [55] Starting with pseudo data sets generated by statistically fluctuating the background-model prediction b i ( θ (0) , δ (0) ), and associated sets of δ fake i generated by fluctuating about δ (0) i with standard deviation κ , the p -value that we have defined is the fraction of those pseudo data/ δ fake i sets for which χ 2 (0) is greater than for the original data set. When calculating χ 2 (0) for the pseudo data sets, one replaces δ i with δ i -δ fake i in Equation (12).
TABLE I. Cross sections for pp → Z ′ → µ + µ -used for our analyses, taking glyph[epsilon1] = 2 × 10 -2 . The cross sections are proportional to glyph[epsilon1] 2 .
| M Z ′ (GeV) | σ 7 TeV (pb) | σ 8 TeV (pb) | σ 14 TeV (pb) |
|---------------|----------------|----------------|-----------------|
| 12 | 274 | 305 | 476 |
| 15 | 163 | 182 | 288 |
| 20 | 80.6 | 90.3 | 147 |
| 30 | 28.5 | 32.4 | 54.6 |
| 40 | 13.3 | 15.3 | 26.3 |
| 50 | 7.43 | 8.56 | 15.1 |
| 60 | 4.76 | 5.51 | 9.96 |
| 70 | 3.6 | 4.21 | 7.7 |
| 80 | 4.24 | 4.94 | 9.28 |
TABLE II. Masses tested for the analysis based on the CMS results of Ref. [11], and associated fit information. The third column gives the median p -value for the polynomial ( κ → 0) fit to pseudo data sets generated by statistically fluctuating the CMS Monte-Carlo-predicted SM bin counts reported in Figure 1 of Ref. [11]. No ( p T ) µµ cut is imposed in this analysis.
| M Z ′ (GeV) | m µµ range (GeV) | median p -value, κ → 0 | κ |
|---------------|--------------------|--------------------------|----------------|
| 30 | 30 - 76 | 0.2 | 6 . 1 × 10 - 3 |
| 40 | 35 - 76 | 0.25 | 7 . 5 × 10 - 3 |
| 50 | 35 - 76 | 0.25 | 4 . 8 × 10 - 3 |
| 60 | 35 - 76 | 0.25 | 4 . 8 × 10 - 3 |
| 70 | 35 - 76 | 0.25 | 4 . 8 × 10 - 3 |
TABLE III. Masses tested for the √ s = 7 TeV analysis based on 1-GeV bins and no ( p T ) µµ cut, and associated fit information. The third column gives the median p -value for the polynomial ( κ → 0) fit to pseudo data sets generated by statistically fluctuating our simulated Drell-Yan m µµ distribution.
| M Z ′ (GeV) | m µµ range (GeV) | median p -value, κ → 0 | κ |
|---------------|--------------------|--------------------------|----------------|
| 30 | 30 - 60 | 0.26 | 9 . 6 × 10 - 3 |
| 40 | 35 - 75 | 0.27 | 1 . 1 × 10 - 2 |
| 50 | 35 - 75 | 0.27 | 4 . 9 × 10 - 3 |
| 60 | 35 - 75 | 0.27 | 4 . 9 × 10 - 3 |
| 70 | 35 - 75 | 0.27 | 7 . 2 × 10 - 3 |
| 80 | 60 - 82 | 0.21 | 1 . 9 × 10 - 2 |
TABLE IV. Masses tested for the √ s = 7 TeV analysis based on 1-GeV bins with a ( p T ) µµ > 23 GeV cut, and associated fit information. The third column gives the median p -value for the polynomial ( κ → 0) fit to pseudo data sets generated by statistically fluctuating our simulated Drell-Yan m µµ distribution.
| M Z ′ (GeV) | m µµ range (GeV) | median p -value, κ → 0 | κ |
|---------------|--------------------|--------------------------|----------------|
| 12 | 12 - 40 | 0.38 | 6 . 0 × 10 - 3 |
| 15 | 12 - 40 | 0.38 | 6 . 0 × 10 - 3 |
| 20 | 12 - 40 | 0.38 | 6 . 0 × 10 - 3 |
| 30 | 12 - 40 | 0.38 | 9 . 2 × 10 - 3 |
TABLE V. Masses tested for the √ s = 8 TeV analysis with no ( p T ) µµ cut, and associated fit information. The third column gives the median p -value for the polynomial ( κ → 0) fit to pseudo data sets generated by statistically fluctuating our simulated Drell-Yan m µµ distribution.
| M Z ′ (GeV) | m µµ range (GeV) | median p -value, κ → 0 | κ |
|---------------|--------------------|--------------------------|----------------|
| 40 | 40 - 75 | 0.28 | 9 . 1 × 10 - 3 |
| 50 | 40 - 75 | 0.28 | 6 . 6 × 10 - 3 |
| 60 | 40 - 75 | 0.28 | 5 . 3 × 10 - 3 |
| 70 | 40 - 75 | 0.28 | 2 . 6 × 10 - 3 |
| 80 | 65 - 82 | 0.38 | 5 . 0 × 10 - 3 |
TABLE VI. Masses tested for the √ s = 8 TeV analysis with a ( p T ) µµ > 30 GeV cut, and associated fit information. The third column gives the median p -value for the polynomial ( κ → 0) fit to pseudo data sets generated by statistically fluctuating our simulated Drell-Yan m µµ distribution.
| M Z ′ (GeV) | m µµ range (GeV) | median p -value, κ → 0 | κ |
|---------------|--------------------|--------------------------|----------------|
| 12 | 12 - 40 | 0.32 | 4 . 5 × 10 - 3 |
| 15 | 12 - 40 | 0.32 | 6 . 2 × 10 - 3 |
| 20 | 12 - 40 | 0.32 | 4 . 5 × 10 - 3 |
| 30 | 12 - 40 | 0.32 | 4 . 5 × 10 - 3 |
| 40 | 12 - 45 | 0.21 | 2 . 2 × 10 - 2 |
|
10.1103/PhysRevD.90.075016
|
[
"Isaac Hoenig",
"Gabriel Samach",
"David Tucker-Smith"
] |
2014-08-05T19:35:50+00:00
|
2014-10-13T13:48:18+00:00
|
[
"hep-ph",
"hep-ex"
] |
Searching for dilepton resonances below the Z mass at the LHC
|
We consider LHC searches for dilepton resonances in an intermediate mass
range, $\sim 10 -80$ GeV. We adopt a kinetically mixed $Z'$ as an example of
weakly coupled new physics that might have evaded detection at previous
experiments but which could still be probed by LHC dilepton spectrum
measurements in this mass range. Based on Monte Carlo simulations, we estimate
that existing data from the 7 and 8 TeV LHC could be used to test values of the
kinetic mixing parameter $\epsilon$ several times smaller than precision
electroweak upper bounds, were an appropriate analysis to be carried out by one
of the experimental collaborations.
|
1408.1076v1
|
## Lime: Data Lineage in the Malicious Environment
Michael Backes, Niklas Grimm, and Aniket Kate
## Saarland University, Germany
{ backes@cs, s9nigrim@stud, aniket@mmci .uni-saarland.de }
## Abstract
Intentional or unintentional leakage of confidential data is undoubtedly one of the most severe security threats that organizations face in the digital era. The threat now extends to our personal lives: a plethora of personal information is available to social networks and smartphone providers and is indirectly transferred to untrustworthy third party and fourth party applications.
In this work, we present a generic data lineage framework LIME for data flow across multiple entities that take two characteristic, principal roles (i.e., owner and consumer). We define the exact security guarantees required by such a data lineage mechanism toward identification of a guilty entity, and identify the simplifying non-repudiation and honesty assumptions. We then develop and analyze a novel accountable data transfer protocol between two entities within a malicious environment by building upon oblivious transfer, robust watermarking, and signature primitives. Finally, we perform an experimental evaluation to demonstrate the practicality of our protocol.
## 1 Introduction
In the digital era, information leakage through unintentional exposures, or intentional sabotage by disgruntled employees and malicious external entities, present one of the most serious threats to organizations. According to an interesting chronology of data breaches maintained by the Privacy Rights Clearinghouse (PRC), in the United States alone, 867 525 654 , , records have been breached from 4 289 , data breaches made public since 2005 [1]. It is not hard to believe that this is just the tip of the iceberg, as most cases of information leakage go unreported due to fear of loss of customer confidence or regulatory penalties: it costs companies on average $ 214 per compromised record [2]. Large amounts of digital data can be copied at almost no cost and can be spread through the internet in very short time. Additionally, the risk of getting caught for data leakage is very low, as there are currently almost no accountability mechanisms. For these reasons, the problem of data leakage has reached a new dimension nowadays.
Not only companies are affected by data leakage, it is also a concern to individuals. The rise of social networks and smartphones has made the situation worse. In these environments, individuals disclose their personal information to various service providers, commonly known as third party applications , in return for some possibly free services. In the absence of proper regulations and accountability mechanisms, many of these applications share individuals' identifying information with dozens of advertising and Internet tracking companies.
Even with access control mechanisms, where access to sensitive data is limited, a malicious authorized user can publish sensitive data as soon as he receives it. Primitives like encryption offer protection only as long as the information of interest is encrypted, but once the recipient decrypts a message, nothing can prevent him from publishing the decrypted content. Thus it seems impossible to prevent data leakage proactively.
Privacy, consumer rights, and advocacy organizations such as PRC [7] and EPIC [8] try to address the problem of information leakages through policies and awareness. However, as seen in the following scenarios the effectiveness of policies is questionable as long as it is not possible to provably associate the guilty parties to the leakages.
Scenario 1: Social Networking. It was reported that third party applications of the widely used online social network Facebook leak sensitive private information about the users or even their friends to advertising companies [3]. In this case, it was possible to determine that several applications were leaking data by analyzing their behaviour and so these applications could be disabled by Facebook. However, it is not possible to make a particular application responsible for leakages that already happened, as many different applications had access to the private data.
Scenario 2: Outsourcing. Up to 108 000 , Florida state employees were informed that their personal information has been compromised due to improper outsourcing [5]. The outsourcing company that was handed sensitive data hired a further subcontractor that hired another subcontractor in India itself. Although the offshore subcontractor is suspected, it is not possible to provably associate one of the three companies to the leakage, as each of them had access to the data and could have possibly leaked it.
Wefind that the above and other data leakage scenarios can be associated to an absence of accountability mechanisms during data transfers: leakers either do not focus on protection, or they intentionally expose confidential data without any concern, as they are convinced that the leaked data cannot be linked to them. In other words, when entities know that they can be held accountable for leakage of some information, they will demonstrate a better commitment towards its required protection.
In some cases, identification of the leaker is made possible by forensic techniques, but these are usually expensive and do not always generate the desired results. Therefore, we point out the need for a general accountability mechanism in data transfers. This accountability can be directly associated with provably detecting a transmission history of data across multiple entities starting from its origin. This is known as data provenance, data lineage or source tracing. The data provenance methodology, in the form of robust watermarking techniques [32] or adding fake data [35], has already been suggested in the literature and employed by some industries. However, most efforts have been ad-hoc in nature and there is no formal model available. Additionally, most of these approaches only allow identification of the leaker in a nonprovable manner, which is not sufficient in many cases.
Our Contributions. In this paper, we formalize this problem of provably associating the guilty party to the leakages, and work on the data lineage methodologies to solve the problem of information leakage in various leakage scenarios.
As our first contribution, we define LIME, a generic data lineage framework for data flow across multiple entities in the malicious environment. We observe that entities in data flows assume one of two roles: owner or consumer. We introduce an additional role in the form of auditor, whose task is to determine a guilty party for any data leak, and define the exact properties for communication between these roles. In the process, we identify an optional non-repudiation assumption made between two owners, and an optional trust (honesty) assumption made by the auditor about the owners.
The key advantage of our model is that it enforces accountability by design ; i.e., it drives the system designer to consider possible data leakages and the corresponding accountability constraints at the design stage. This helps to overcome the existing situation where most lineage mechanisms are applied only after a leakage has happened.
As our second contribution, we present an accountable data transfer protocol to verifiably transfer data between two entities. To deal with an untrusted sender and an untrusted receiver scenario associated with data transfer between two consumers, our protocols employ an interesting combination of the robust watermarking, oblivious transfer, and signature primitives. We also implement our protocol and demonstrate the practicality for real-life data transfer scenarios such as online social networks and outsourcing.
Paper Outline. We organize the remainder of this paper as follows: In Section 2 we introduce our model LIME and give a threat model and design goals. Section 3 describes primitives used in the paper and Section 4 presents and analyzes protocols for accountable data transfer. Section 5 shows results we obtained from a practical implementation and Section 6 gives examples of how our model can be applied to real world settings. We discuss additional features of our approach in Section 7 and related work in Section 8. Finally we present our conclusions in Section 9.
## 2 The Lime Framework
As we want to address a general case of data leakage in data transfer settings, we propose the simplifying model LIME ( L ineage i n the m alicious e nvironment). With LIME we assign a clearly defined role to each involved party and define the inter-relationships between these roles. This allows us to define the exact properties that our transfer protocol has to fulfill in order to allow a provable identification of the guilty party in case of data leakage.
## 2.1 Model
As LIME is a general model and should be applicable to all cases, we abstract the data type and call every data item document . There are three different roles that can be assigned to the involved parties in LIME: data owner , data consumer and auditor . The data owner is responsible for the management of documents and the consumer receives documents and can carry out some task using them. The auditor is not involved in the transfer of documents, he is only invoked when a leakage occurs and then performs all steps that are necessary to identify the leaker. All of the mentioned roles can have multiple instantiations when our model is applied to a concrete setting. We refer to a concrete instantiation of our model as scenario .
In typical scenarios the owner transfers documents to consumers. However, it is also possible that consumers pass on documents to other consumers or that owners exchange documents with each other. In the outsourcing scenario [5] the employees and their employer are owners, while the outsourcing companies are untrusted consumers.
In the following we show relations between the different entities and introduce optional trust assumptions. We only use these trust assumptions because we find that they are realistic in a real world scenario and because it allows us to have a more efficient data transfer in our framework. At the end of this section we explain how our framework can be applied without any trust assumptions.
When documents are transferred from one owner to another one, we can assume that the transfer is governed by a non-repudiation assumption . This means that the sending owner trusts the receiving owner to take responsibility if he should leak the document. As we consider consumers as untrusted participants in our model, a transfer involving a consumer cannot be based on a non-repudiation assumption. Therefore, whenever a document is transferred to a consumer, the sender embeds information that uniquely identifies the recipient. We call this fingerprinting . If the consumer leaks this document, it is possible to identify him with the help of the embedded information.
As presented, LIME relies on a technique for embedding identifiers into documents, as this provides an instrument to identify consumers that are responsible for data leakage. We require that the embedding does not not affect the utility of the document. Furthermore, it should not be possible for a malicious consumer to remove the embedded information without rendering the document useless. A technique that can offer these properties is robust watermarking . We give a definition of watermarking and a detailed description of the desired properties in Section 3.1.
A key position in LIME is taken by the auditor. He is not involved in the transfer, but he takes action once a leakage occurs. He is invoked by an owner and provided with the leaked data. If the leaked data was
Figure 1: The LIME framework
transferred using our model, there is identifying information embedded for each consumer who received it. Using this information the auditor can create an ordered chain of consumers who received the document. We call this chain the lineage of the leaked document. The last consumer in the lineage is the leaker. In the process of creating the lineage each consumer can reveal new embedded information to the auditor to point to the next consumer - and to prove his own innocence. In order to create a complete lineage it is necessary that the auditor receives information from the owner, as only the owner can reveal the information embedded during the first transfer. We assume that the auditor is always invoked by the owner or that he is at least provided with information about the owners identity, so that the auditor can start his investigation with the owner and a complete lineage can be created.
We can assume that the auditor trusts the owner to be honest. Honesty in this case means, that the owner does not leak a document and blame another party. We can make this assumption as the owner is concerned with the document's privacy in the first place. However, the auditor does not trust the consumers. In a real world setting the auditor can be any authority, for example a governmental institution, police, a legal person or even some software. In the outsourcing scenario [5], the employer can invoke the auditor who recreates the lineage and thereby uncovers the identity of the leaker. The employer can use this information to take legal actions against the outsourcing company. We show the flow of documents, the optional nonrepudiation and honesty assumptions the cases in which fingerprinting is used in Figure 1. In Section 6, we present how LIME can be applied to the outsourcing scenario presented in the introduction.
Remark. We choose these honesty and non-repudiation assumptions because they realistically model many real-world scenarios. Due to these assumptions we can reduce the overhead introduced to transfers by LIME: In a transfer between two owners no fingerprinting has to be used and the owner can run a simplified transfer protocol because he is trusted by the auditor. However, these trust assumptions are not necessary for the correctness of LIME. If an owner is untrusted we can easily treat him as a consumer, so we do not have to make any trust assumptions about him.
## 2.2 Threat Model and Design Goals
Although we try to address the problem of data leakage, LIME cannot guarantee that data leakage does not occur in the first place; once a consumer has received a document, nothing can prevent him from publishing it. We offer a method to provably identify the guilty party once a leakage has been detected. By introducing this reactive accountability, we expect that leakage is going to occur less often, since the identification of the guilty party will in most cases lead to negative consequences. As our only goal is to identify guilty parties, the attacks we are concerned about are those that disable the auditor from provably identifying the guilty party.
Therefore, we consider attackers in our model as consumers that take every possible step to publish a document without being held accountable for their actions. As the owner does not trust the consumer, he
uses fingerprinting every time he passes a document to a consumer. However, we assume that the consumer tries to remove this identifying information in order to be able to publish the document safely. As already mentioned previously, consumers might transfer a document to another consumer, so we also have to consider the case of an untrusted sender . This is problematic because a sending consumer who embeds an identifier and sends the marked version to the receiving consumer could keep a copy of this version, publish it and so frame the receiving consumer. Another possibility to frame other consumers is to use fingerprinting on a document without even performing a transfer and publish the resulting document.
A different problem that arises with the possibility of false accusation is denial. If false accusation is possible, then every guilty receiving consumer can claim that he is innocent and was framed by the sending consumer.
The crucial phase in our model is the transfer of a document involving untrusted entities, so we clearly define which properties we require our protocol to fulfill. We call the two parties sender and recipient . We expect a transfer protocol to fulfill the following properties and only tolerate failures with negligible probabilities.
- 1. correctness: When both parties follow the protocol steps correctly and only publish their version of the document, the guilty party can be found.
- 2. no framing: The sender cannot frame recipients for the sender's leakages.
- 3. no denial: If the recipient leaks a document, he can be provably associated with it.
We also require our model to be collusion resistant, i.e. it should be able to tolerate a small number of colluding attackers [38]. We also assume that the communication links between parties are reliable.
Non-Goals. Wedo not aim at proactively stopping data leakage, we only provide means to provably identify the guilty party in case a leak should occur, so that further steps can be taken. We also do not aim for integrity, as at any point an entity can decide to exchange the document to be sent with another one. However, in our settings, the sender wants the receiver to have the correct document, as he expects the recipient to perform a task using the document so that he eventually obtains a meaningful result.
Our approach does not account for derived data (derived data can for example be generated by applying aggregate functions or other statistical operations), as much of the original information can be lost during the creation process of derived data. Nevertheless, we show in Section 7.1 how LIME can operate on composed data. We think of composed data as a form of data created from multiple single documents, so that the original documents can be completely extracted (e.g. concatenation of documents).
We do not consider fairness issues in our accountable transfer protocol; more precisely, we do not consider scenarios in which a sender starts to run the transfer protocol but aborts before a recipient received the document, or when a recipient, despite of receiving the document, falsely claims that he did not receive it. In real-world scenarios, we find fairness not to be an issue as senders and recipients expect some utility from the transfer, and are worried about their reputation and corporate liabilities.
## 3 Primitives
A function f is negligible if for all c > 0 there is a n c so that for all n ≥ n c f ( n ) ≤ 1 n c . In our scheme we make use of digital signatures . More precisely, we use a CMA-secure signature [24], i.e., no polynomialtime adversary is able to forge a signature with non-negligible probability. For a message m that has been signed with party A 's signing key, we write [ m ] sk A . We use a symmetric encryption scheme that offers security under chosen plaintext attacks, writing c = enc m,ek ( ) for encryption of a message m and m = dec c, ek ( ) for decryption of a ciphertext c with symmetric key ek .
## 3.1 Robust Watermarking
We use the definition of watermarking by Adelsbach et al. [10]. To argue about watermarking, we need a so-called similarity function sim D,D ( ′ ) that returns /latticetop if the two documents D and D ′ are considered similar in the used context and ⊥ otherwise. The similarity function is a different one for each data type used and we assume it is given.
Let D be the set of all possible documents, WM⊆{ 0 1 , } + the set of all possible watermarks, K the set of keys and κ the security parameter of the watermarking scheme. A symmetric, detecting watermarking scheme is defined by three polynomial-time algorithms:
- · The probabilistic Key Generation Algorithm GenKey WM (1 κ ) outputs a key k ∈ K for a given security parameter κ .
- · The probabilistic Embedding Algorithm generates a watermarked document D ′ = W ( D,w,k ) on input of the original document D ∈ D , the watermark w ∈ WM and the key k ∈ K .
- · The Detection Algorithm Detect D , w, D, k ( ′ ) outputs /latticetop or ⊥ on input of a (potentially watermarked) document D ′ ∈ D , a watermark w ∈ WM , the original document D ∈ D and a key k ∈ K . /latticetop means that the watermark is detectable; ⊥ means, that it is not.
We require the following properties:
- · imperceptibility: ∀ D ∈ D ∀ , w ∈ WM ∀ , k ∈ K .D ′ ← W ( D,w,k ) ⇒ sim D, D ( ′ ) = /latticetop , i.e., the original document and the watermarked document are similar.
- · effectiveness: ∀ D ∈ D ∀ , w ∈ WM ∀ , k ∈ K .D ′ ←W ( D,w,k ) ⇒ Detect D , w, D, k ( ′ ) = /latticetop , i.e., if a watermark is embedded using a key k , the same watermark should be detectable using the same key k .
- · robustness: For a watermarked document D ′ = W ( D,w,k ,D ) ∈ D , w ∈ WM , k ∈ K there is no polynomial-time adversary that can compute a D ′′ ∈ D given D ′ and w so that sim D ,D ( ′ ′′ ) = /latticetop and Detect D , w, D, k ( ′′ ) = ⊥ with non-negligible probability. This means that no adversary can efficiently remove or change a watermark without rendering the document unusable.
Additionally, we require our watermarking scheme to support multiple re-water- marking , i.e., it should allow for multiple (bounded by the dataflow path length) watermarks to be embedded successively without influencing their individual detectability. This property can also be considered as a special kind of robustness, as it prevents adversaries from making a watermark undetectable simply by adding more watermarks using the same algorithm. More information and some experimental results about this property can be found in [32].
We also expect the watermarking scheme to be collusion resistant [30], i.e., even if an attacker can obtain differently watermarked versions of a document, he should not be able to create a version of the document were none of these watermarks is detectable. Further, for some watermarking schemes the input of the original document is not required for detection of watermarks. We call those watermarking schemes blind . As already stated in [10] this definition of watermarking is very strong and its properties are not yet provided by available schemes. Although we chose this strong definition to prove the correctness of our scheme, there are existing schemes whose properties are arguably sufficient in practice such as the Cox watermarking scheme [17].
Cox Watermarking Scheme [17]. This scheme uses spread-spectrum embedding of Gaussian noise to watermark images. To provide robustness, the watermark is embedded in the most significant part of the picture, so that removing the watermark should not be possible without destroying the underlying picture. The α -factor of the algorithm is a parameter that determines how strong the Gaussian noise is influencing the original image, thus it can be used to trade robustness against imperceptibility. Although our strong robustness requirement is not formally fulfilled, it is shown that the scheme is robust against many common attacks such as scaling, JPEG compression, printing, xeroxing and scanning, multiple re-watermarking and others [17].
Figure 2: Oblivious Transfer protocol by Naor and Pinkas [34]
| Initialization | Initialization |
|-----------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------|
| Sender | Chooser |
| Pick random C,r ∈ Z q and compute g r and C r . = | Pick random C,r ∈ Z q and compute g r and C r . = |
| Send C . | ⇒ |
| Transfer | Transfer |
| Sender holding M 0 , M 1 Compute PK r 0 and PK r 1 = C r /PK r 0 . Encrypt M 0 as E 0 = H ( PK r 0 ) ⊕ M 0 and M 1 as E 1 = | Chooser requesting M σ Pick random 1 ≤ k ≤ q and compute PK σ = g k and PK 1 - σ = C/PK σ . ⇐ = Send PK 0 . |
| H ( PK r 1 ) ⊕ M 1 Send g r , E 0 , E 1 . | = ⇒ Compute H (( g r ) k ) = H ( PK r σ ) and use it to decrypt M σ . |
## 3.2 1-out-of-2 Oblivious Transfer
1-out-of-2 Oblivious Transfer ( OT 2 1 ) involves two parties, the sender and the chooser . The sender offers two items M 0 and M 1 and the chooser chooses a bit σ . The chooser obtains M σ but no information about M 1 -σ and the sender learns nothing regarding σ . In this context, when speaking of learning nothing, we actually mean nothing can be learned with non-negligible probability. There are different concrete instantiations of this primitive.
As an example implementation of OT 2 1 we show a protocol by Naor and Pinkas [34]: The protocol operates over a group Z q of prime order q with a generator g for which the computational Diffie-Hellman assumption holds. The protocol is proven secure in the random oracle model using an ideal hash function H . The construction can be found in Figure 2.
In the initialization phase, the sender picks a random group element C and sends it to the chooser. It is important that the chooser does not know DLog C ( ) . Additionally, the sender chooses another random group element r and computes C r . These values are not used in the initialization phase, but for efficiency reasons the sender can compute these values in this phase, as it can be performed offline.
In the transfer phase, the chooser picks a random 1 ≤ k ≤ q and computes PK σ = g k and PK 1 -σ = C/PK σ , so that he knows DLog PK ( σ ) . Then he sends PK 0 to the sender. The sender computes PK r 0 and without further exponentiation PK r 1 = C /PK r r 0 . Then he encrypts the two messages M 0 and M 1 as E 0 = H PK ( r 0 ) ⊕ M 0 and E 1 = H PK ( r 1 ) ⊕ M 1 and sends g , E r 0 and E 1 to the chooser. The chooser can now compute H (( g r ) k ) = H PK ( r σ ) and so decrypt M σ = E σ ⊕ H PK ( r σ ) .
The proof for the correctness and security of the protocol in the random oracle model and under the Diffie-Hellman assumption can be found in [34]. Additionally, they provide an approach to improve efficiency for multiple executions of OT 2 1 using a bandwidth/computation tradeoff. This is achieved by performing one single instance of OT 2 n 1 offering all 2 n possible combinations instead of n instances of OT 2 1 offering every part on its own. Depending on available computational power and bandwidth, this can improve efficiency.
When we use OT 2 1 in our protocols to send messages, the sender actually encrypts the messages, sends both encryptions to the chooser and performs OT 2 1 just on the decryption keys. This allows us to use the OT 2 1 protocol with a fixed message size while actually sending messages of arbitrary size. Note that this
Figure 3: protocol for trusted senders
could only be a security risk if the chooser was able to break the encryption scheme.
## 4 Accountable Data Transfer
In this section we specify how one party transfers a document to another one, what information is embedded and which steps the auditor performs to find the guilty party in case of data leakage. We assume a public key infrastructure to be present, i.e. both parties know each others signature verification key.
## 4.1 Trusted Sender
In the case of a trusted sender it is sufficient for the sender to embed identifying information, so that the guilty party can be found. As the sender is trusted, there is no need for further security mechanisms. In Figure 3, we present a transfer protocol that fulfills the properties of correctness and no denial as defined in Section 2.2. As the sender is trusted to be honest, we do not need the no framing property.
The sender, who is in possession of some document D , creates a watermarking key k , embeds a triple σ = ( C ,C S R , τ ) consisting of the two parties' identifiers and a timestamp τ into D to create D w = W ( D,σ,k ) . He then sends D w to the recipient, who will be held accountable for this version of the document. As the sender also knows D w , this very simple protocol is only applicable if the sender is completely trusted; otherwise the sender could publish D w and blame the recipient.
## 4.2 Untrusted Sender
In the case of an untrusted sender we have to take additional actions to prevent the sender from cheating, i.e. we have to fulfill the no framing property. To achieve this property, the sender divides the original document into n parts and for each part he creates two differently watermarked versions. He then transfers one of each of these two versions to the recipient via OT 2 1 . The recipient is held accountable only for the document with the parts that he received, but the sender does not know which versions that are. The probability for the sender to cheat is therefore 1 2 n . We show the protocol in Figure 4 and provide an analysis of the protocol properties in Section 4.4.
First, the sender generates two watermarking keys k 1 and k 2 . It is in his own interest that these keys are fresh and distinct . The identifying information that the sender embeds into the document D is a signed statement σ = [ C ,C S R , τ ] sk C R containing the sender's and recipient's identifiers and a timestamp τ , so that every valid watermark is authorized by the recipient. The sender computes the watermarked document D ′ = W ( D,σ,k 1 ) , splits the document D ′ into n parts and creates two different versions D i,j = W ( D , j, k i 2 ) of each part by adding an additional watermark j ∈ { 0 1 , } . For each version of each part D i,j , i ∈ { 1 , . . . , n } , j ∈ { 0 1 , } he creates a signed message m i,j = [ τ, i, j ] sk C S containing the timestamp of the current transfer, the part's index and the content of the version's second watermark. Then he generates an AES key ek i,j , encrypts c i,j = enc ( 〈 D i,j , m i,j 〉 , ek i,j ) and sends c i,j to the recipient. The recipient chooses a bit b i ∈ { 0 1 , } for each i ∈ { 1 . . . n } and receives ek i,b i via oblivious transfer (note that the n executions of OT 1 2 can be parallelized). He then decrypts all c i,b i using ek i,b i and reconstructs the document by joining the parts D 1 ,b 1 . . . D n,b n . The signed statements m 1 ,b 1 . . . m n,b n serve as proof of his
choice. As for each part he chooses a bit b i , the final version is identified by a bitstring b ∈ { 0 1 , } n . As he will only be held accountable for the version watermarked with b , we have a failure probability (i.e., the sender correctly guessing b ) of 1 2 n .
## 4.3 Data Lineage Generation
The auditor is the entity that is used to find the guilty party in case of a leakage. He is invoked by the owner of the document and is provided with the leaked document. In order to find the guilty party, the auditor proceeds in the following way:
- 1. The auditor initially takes the owner as the current suspect.
- 2. The auditor appends the current suspect to the lineage.
- 3. The auditor sends the leaked document to the current suspect and asks him to provide the detection keys k 1 and k 2 for the watermarks in this document as well as the watermark σ . If a non-blind watermarking scheme is used, the auditor additionally requests the unmarked version of the document.
- 4. If, with key k 1 , σ cannot be detected, the auditor continues with 9.
- 5. If the current suspect is trusted, the auditor checks that σ is of the form ( C ,C S R , τ ) where C S is the identifier of the current suspect, takes C R as current suspect and continues with 2.
- 6. The auditor verifies that σ is of the form [ C ,C S R , τ ] sk C R where C S is the identifier of the current suspect. He also verifies the validity of the signature.
- 7. The auditor splits the document into n parts and for each part he tries to detect 0 and 1 with key k 2 . If none of these or both of these are detectable, he continues with 9. Otherwise he sets b ′ i as the detected bit for the i th part. He sets b ′ = b ′ 1 . . . b ′ n .
- 8. The auditor asks C R to prove his choice of b = b 1 · · · b n for the given timestamp τ by presenting the m i,b i = [ τ, i, b i ] sk C S . If C R is not able to give a correct proof (i.e., m i,b i is of the wrong form or the signature is invalid) or if b = b ′ , then the auditor takes C R as current suspect and continues with 2.
- 9. The auditor outputs the lineage. The last entry is responsible for the leakage.
Remark. It would also be possible to include the signed statement σ in every single part of the document, but as the maximum size of a watermark is limited by the document's size, this might be problematic for scenarios where the parts are small. Therefore, we embed σ in the complete original document and only embed single bits to the (possibly small) parts of the document. We use a timestamp τ to uniquely identify a specific transfer between two parties, and thus assume that no two transfers between the same two parties take place at the same time. However, it would be possible to use a counter that is incremented on each transfer to allow multiple transfers at the exact same time.
## 4.4 Analysis of the Protocol
We now show that the protocol presented in Figure 4 fulfills the required properties of correctness, no framing and no denial as presented in Section 2.2 :
- 1. correctness: Assume that both parties follow the protocol steps correctly. We show that for all possible scenarios the guilty party is determined correctly:
```
}
} \
\
```
Figure 4: protocol for untrusted senders
- (a) the sender publishes D or D ′ : The auditor does not detect σ (in the case of D ) or the b i (in the case of D ′ ) and correctly blames the sender.
- (b) the recipient publishes D w : The auditor successfully detects σ and b ′ and verifies that σ is of the correct form. The recipient is able to provide the proof of his choice of b ; the auditor verifies b ′ = b and suspects the recipient. As there are no further watermarks embedded, the auditor correctly blames the recipient.
- 2. no framing: In a first step we show that the sender cannot obtain the version of the document for which the recipient can prove his choice (i.e. the version watermarked with the bitstring b ): The sender knows all the D i,j , that are used for the computation of D w , but he does not know the bitstring b that the recipient chose due to the properties of oblivious transfer. So all he can do is guess a bitstring b ∗ = b ∗ 1 · · · b ∗ n and create D ∗ w = join D ( 1 ,b ∗ 1 , . . . , D n,b ∗ n ) . But b and b ∗ (and therefore D w and D ∗ w ) are the same only with negligible probability of 1 2 n , so the probability for the sender to learn D w is negligible if he follows the protocol correctly.
The sender might try to learn about b by offering the same version twice during the oblivious transfer. Usually the recipient would have no possibility of realizing this, as he cannot detect the watermark, but as the sender additionally has to send the signed statements m i,j = [ τ, i, j ] sk C S the recipient knows that he received what he asked for. This is the case because j has by construction the same value as b i chosen by the recipient. The sender can still send a wrong version D i, 1 -j and so know the bitstring the document is watermarked with, but as the recipient only proves his choice of b , the sender still cannot frame him; he would only lose the ability to prove the recipients' guilt in case the recipient publishes the document.
In a second step we show that a malicious sender cannot create a document that the recipient will be held accountable for without running the transfer protocol:
As the correctness of the signed statement σ is verified in the auditing process and as the sender can only forge the recipient's signature with negligible probability, the only possibility to mount this attack is to reuse a valid signed statement from a past transaction. This implies that the included timestamp τ is the same, too. As the auditor asks the recipient to prove his choice of b for this τ , the recipient is able to provide a correct proof, as a valid transfer with timestamp τ actually happened. Analogous to the previous case, the sender can only chose b ∗ ∈ { 0 1 , } n randomly and therefore he can only succeed with negligible probability.
Figure 5: (a) shows computation times for different numbers of document parts; (b) shows computation times for different image sizes.
From these two steps it follows that the sender is not able to frame a recipient.
- 3. no denial: Wefirst show that the recipient cannot publish a version of the document whose embedded watermarks are different from those embedded in the version D w of the document that he righteously obtained. He also cannot obtain the watermark-free version D : As the recipient only receives the watermarked version, he could only learn D by removing the watermark, which he can only do with negligible probability due to the robustness property of the watermarking scheme. The recipient could also create a watermarked version with a different bitstring embedded, if he is able to get D i, 1 -b i for some i , but this is only possible if he breaks the OT 1 2 scheme or the encryption scheme, which is only possible with negligible probability.
We now show that a recipient cannot cheat during the auditing process, when he proves which version of the document he asked for during the transfer protocol: In order to prove another choice of b he would again have to break the OT 1 2 scheme or the encryption scheme in order to learn the m i, 1 -b i for some i or he would need to forge the sender's signature to create m i, 1 -b i for some i . As all of this is only possible with negligible probability, the overall probability for the recipient to succeed is negligible.
From these two steps it follows that the recipient is not able to publish a document without being caught.
This proves our protocol's security up to fairness issues. However, as discussed in Section 2.2, we do not find fairness issues to be problematic in our application scenarios.
## 5 Implementation and Microbenchmarking
Weimplemented the protocol in Figure 4 as a proof-of-concept and to analyze its performance. For the oblivious transfer subprotocol we implemented the protocol by Naor and Pinkas [34] using the PBC library [6], which itself makes use of the GMP library [4]. For signatures we implemented the BLS scheme [14], also using the PBC library. For symmetric encryption we used an implementation of AES from the Crypto++ [19] library. For watermarking we used an implementation of the Cox algorithm for robust image watermarking [17] from Peter Meerwald's watermarking toolbox [33]. We set the α -factor, which determines the strength of the watermark, to a value of 0.1.
(a)
(b)
Figure 6: (a) shows an image transferred with our algorithm. In (b) we set the α -factor used by the Cox algorithm to 0.5 in order to obtain very strong watermarks for the small parts. As a result differences between adjacent document parts are visible.
We executed the experiment with different parameters to analyze the performance. The sender and recipient part of the protocol are both executed in the same program, i.e., we do not analyze network sending, but only computational performance. The executing machine is a Lenovo ThinkPad model T430 with 8 GB RAMand 4 × 2.6GHz cores, but all executions were performed sequentially. We measured execution times for different phases of the protocol: watermarking, signature creation, encryption, oblivious transfer and detection. We executed each experiment 250 times and determined the average computation time and the standard deviation.
In the first experiment we used an image of size 512 × 512 pixels and changed the number of parts the image was split into. We show the results in Figure 5(a). We can see that the execution time of watermarking, signatures, oblivious transfer and detection is linear in the number of document parts. The execution time of encryption is also increasing slowly, but it is still insignificant compared to the other phases.
In the second experiment we used a fixed number of document parts of 256 and changed the size of the image used. We used 4 different sizes: 256 × 256 px (65KB), 512 × 512 px (257KB), 1024 × 1024 px (1.1 MB) and 2048 × 2048 px (4.1 MB). The results can be seen in Figure 5(b). We can observe that the execution time for watermarking and detection is linear in the image size, but in contrast to the previous result, we already start with a long execution time for small sizes and the growth is rather slow. The execution time for oblivious transfer stays constant, which might be surprising at first sight, but this can be explained easily, as we only transfer AES keys in the oblivious transfer phase and these are of constant size for all image sizes. Of course the encrypted files also have to be sent over the network (so there would be a higher communication overhead for bigger images), but in our implementation we only considered the computational costs. The execution time for the creation of the signatures is also constant as the number and form of the signed statements is the same for all images. Again the execution time of encryption is increasing slowly, but it is of no significance compared to the other phases.
The tests showed that from the resulting files all watermarks (i.e. the identifying watermark σ and all the
Figure 7: Outsourcing Scenario
single bits b i forming the bitstring b ) could be correctly detected, showing the correctness of our protocol. In Figure 6(a) in the appendix we give an example of an image that was transferred using our algorithm.
We find that latencies of a few seconds are acceptable in the scenarios that we considered. Additionally, as we show in our experiments in Figure 5(a), one can easily perform a tradeoff between performance and security. It is also possible to use the OT extension technique presented in [27] to increase the efficiency of oblivious transfer.
Although we use only image files as documents in our experimental implementation, we stress that the same mechanism can be used for all types of data for which robust watermarking schemes exist.
## 6 Scenarios
## 6.1 Outsourcing
The first diagram in Figure 7 shows a typical outsourcing scenario. An organization acts as owner and can outsource tasks to outsourcing companies which act as consumers in our model. It is possible that the outsourcing companies receive sensitive data to work on and as the outsourcing companies are not necessarily trusted by the organization, fingerprinting is used on transferred documents. The outsourcing company itself can outsource tasks to other outsourcing companies and thus relay the documents, again using fingerprinting.
If now at any point one of the involved outsourcing companies leaks a confidential document, the organization can invoke the auditor to find the responsible party. The auditor then examines the fingerprints in the document, creates a lineage and is finally able to name the guilty party. In the example given in the introduction [5], there were three outsourcing companies involved and a data leakage could not be clearly associated with one of these. Using LIME the responsible party can be clearly found.
## 6.2 Online Social Network
The second diagram in Figure 7 shows an online social networking scenario. The users of the network are the owners, as they enter their personal information, post messages, etc. The online social network (OSN) uses all this information as a consumer in this scenario. Third party applications that have access to this information in return for some service act as further consumers in this scenario. The users give their information to the OSN which can relay that information to third party applications using fingerprinting.
In the introduction we gave an example where third party applications leaked private information of Facebook users to advertising companies [3]. Although it was possible to determine which applications leaked data, given some leaked information it is not possible to find the responsible application. Using LIME this link can be found and proven.
## 7 Discussion
## 7.1 Composability
LIME also allows us to create a lineage for a document that is published as part of a composed object. Consider the following setting: Owner A and Owner B each own a set of database entries D A and D B which they both transfer to Consumer 1, who receives marked versions D 1 A and D 1 B . Consumer 1 then creates a composed object D 1 = D 1 A || D 1 B by concatenation and transfers it further to Consumer 2, who receives a marked version D 2 = D 2 A || D 2 B and then leaks it. If Owner 1 notices that his set of database entries was published as part of the composed object D 2 = D 2 A || D 2 B , he invokes the auditor and provides him with D 2 , D 2 A , D A , the watermark and the necessary detection keys. The auditor detects the watermarks in D 2 A and verifies that this version was transferred to Consumer 1, who proves that he transferred the composed data to Consumer 2 by showing the watermark and the detection keys for the composed document. As Consumer 2 cannot disprove his guilt, he can be held accountable for publishing a version of D A even though he published it as part of composed data.
## 7.2 Collusion Resistance
The collusion resistance of our scheme depends on the collusion resistance of the underlying watermarking scheme. Assume several consumers are working together in order to create an untraceable version of a document. Then their approach is to merge the versions they rightfully obtained to create a new version where the watermarks cannot be detected. As the detection of σ is just a detection of a watermark in the complete document, we obviously have the same collusion resistance as the watermarking scheme for this case. The case of the detection of a bit b i in a part D i is again just a detection of a watermark, so the collusion resistance is again the same as for the watermarking scheme. However, we have to know which detected bit belongs to which consumer, so that we can still guarantee that the sender cannot frame the receiving consumers. Linking the detected bits to the responsible consumers is possible, as for each consumer a different embedding key was used. As for each part multiple bits might be detectable, the probability for a sender to successfully frame the receiving consumers is less than or equal to the probability of framing a single recipient successfully, as he still would have to guess all the bits correctly. However, we have to note that in order to successfully mount a collusion attack against our scheme, it is sufficient to mount a collusion attack against 1 of the n + 1 watermarks that are used, where n is the number of parts the document was split into.
Wecan conclude that our scheme tolerates collusions to a certain extent, when it is used with a collusion resistant watermark, without losing its key properties.
## 7.3 Error Tolerance
Depending on the quality of the underlying watermarking scheme, it may be too strong to require that all bits b i are detected correctly. Therefore, it could be a good idea to introduce some error tolerance. However, we have to keep in mind that this will increase the probability of the sender successfully framing an innocent recipient. There are two different kinds of errors that can occur: the first one is that no bit can be detected, and the second one is that a wrong bit is detected. Assume the document is split into n parts. Tolerating a non-detectable bit increases the probability of successful framing by a factor of 2. Instead of guessing a bitstring b ∈ { 0 1 , } n , it is sufficient to guess b ∈ { 0 1 , } n -1 . Tolerating a wrong bit is worse, as it increases this probability by a factor of ( n + 1) . Instead of accepting just the correct bitstring, we also accept all bitstrings that are changed at exactly one position. As there are n positions, we additionally accept n bitstrings; hence the number of accepted bitstrings and thus the probability of guessing one of these is higher by a factor of n + 1 . If we want to allow some error tolerance while keeping the probability of
successful framing to be small, we have to choose a larger n ; e.g., to tolerate 128 non-detectable bits, we choose n = 256 and have the same framing probability as with n = 128 and no tolerance.
## 7.4 Possible Data Distortion
In our experiment, we used a simple splitting algorithm: We split the image into n equally sized squares. However, when we used a strong watermark for the small parts (that is the α -factor used by the Cox algorithm is 0.5), differences between adjacent parts became visible even though the single watermarks are imperceptible. The resulting image can be seen in Figure 6(b) in the appendix. In some cases, this distortion might affect the usability of the document. We stress however, that we were still able to obtain good results with our approach. In Figure 6(a) we used the Cox algorithm with an alpha factor of 0.1 and no distortion is visible.
It might be interesting to investigate if this problem can be circumvented by using more elaborate splitting algorithms. As most watermarking schemes make use of the contiguity of information in the document, this is not a trivial task.
## 8 Related Work
## 8.1 Other Models
The model introduced in [35] intends to help the data distributor to identify the malicious agent which leaked the information. In addition, they argue that current watermarking techniques are not practical, as they may embed extra information which could affect agents' work and their level of robustness may be inadequate. In LIME the relationship of data distributor and agents corresponds to the relationship between data owner and consumer and the model could be used as an alternative method to trace the information given to the consumers.
Controlled data disclosure is a well-studied problem in the security literature, where it is addressed using access control mechanisms. Although these mechanisms can control release of confidential information and also prevent accidental or malicious destruction of information, they do not cover propagation of information by a recipient that is supposed to keep the information private. For example, once an individual allows a third party app to access her information from a social network, she can no longer control how that app may redistribute the information. Therefore, the prevalent access control mechanisms are not adequate to resolve the problem of information leakages. Data usage control enforcement systems [41, 29] employ preventive measures to ensure that data is transferred in distributed systems in a controlled manner preserving the welldefined policies. Techniques have been developed for securely distributing data by forming coalitions among the data owners [45]. In controlled environments, such techniques can be composed with our protocols to improve data privacy.
In [47] the authors present the problem of an insider attack, where the data generator consists of multiple single entities and one of these publishes a version of the document. Usually methods for proof-of-ownership or fingerprinting are only applied after completion of the generating process, so all entities involved in the generation process have access to the original document and could possibly publish it without giving credit to the other authors, or also leak the document without being tracked. As presented in the paper, this problem can be solved by the usage of watermarking and possibly even by using complete fingerprinting protocols during the generating phase of the document.
## 8.2 Other Fingerprinting Protocols
In [39] Poh addresses the problem of accountable data transfer with untrusted senders using the term fair content tracing . He presents a general framework to compare different approaches and splits protocols into four categories depending on their utilization of trusted third parties, i.e., no trusted third parties, offline trusted third parties, online trusted third parties and trusted hardware. Furthermore, he introduces the additional properties of recipient anonymity and fairness in association with payment. All presented schemes use watermarking to trace the guilty party and most presented protocols make use of watermarking in the encrypted domain , where encrypted watermarks are embedded in encrypted documents [37]. A new scheme presented is based on chameleon encryption [12]. In [43] Sadeghi also examines several fingerprinting schemes and presents new constructions for symmetric, asymmetric and anonymous fingerprinting schemes. The asymmetric scheme uses a homomorphic commitment scheme to compute the fingerprinted version of the document.
Domingo-Ferrer presents the first fingerprinting protocol that makes use of oblivious transfer in [21]. In the scheme, documents are split into smaller parts and for each part two different versions are created. Then the recipient receives one version of each part via oblivious transfer and in return sends a commitment on the received part. The recipient can now be identified by the unique combinations of versions he received. The protocol has several flaws, as discussed in [42] and [16]. The main problem is that a malicious sender can offer the same version twice in the oblivious transfer, so that he will know which version the recipient receives.
Sadeghi [42] and Hanaoka et al. [16] propose different solutions; the former lets the sender open some pairs to validate that they are not equal and the latter uses oblivious transfer with a two-lock cryptosystem where the recipient can compare both versions in encrypted form. However, both proposed solutions have some flaws themselves. The problem is that it is possible to create two different versions with the same watermark, so even if the equality test fails, the two offered versions can still have the same watermark and the sender will know which watermark the recipient received. Also, the fix proposed in [16] ruins the negligible probability of failure, as it does not split the document into parts, but creates n different versions and sends them via 1-out-of-n oblivious transfer.
Domingo-Ferrer presents another protocol based on oblivious transfer in [22], but again the sender can cheat during oblivious transfer. [26] presents another protocol using oblivious transfer. The protocol uses an approach similar to the chameleon encryption [12], and using 1-out-of-n oblivious transfer a decryption key is transmitted so that the sender does not know it. The protocol suffers from the same problems as the one presented in [16]; namely, the sender can guess the key used by the recipient with non-negligible probability 1 n and the sender can even cheat in the oblivious transfer by offering the same key n times, so that he will know the key used by the recipient.
We see that all asymmetric fingerprinting protocols based on oblivious transfer that have been proposed so far suffer from the same weakness. We circumvent this problem in our protocol by additionally sending a signed message including the watermark's content, so that the recipient is able to prove what he asked for. In contrast to the watermark, this message can be read by the recipient, so he can notice if the sender cheats.
## 8.3 Broadcasting
Parviainen and Parnes present an approach for distributing data in a multicast system, so that every recipient holds a differently watermarked version [36]. The sender splits the file into blocks and for each block he creates two different versions by watermarking them with different watermarks and encrypting them with different keys. Each recipient is assigned a set of keys, so that he can decrypt exactly one version of each part. The resulting combination of parts can uniquely identify the recipient. In [9] Adelsbach, Huber and Sadeghi show another approach for a broadcasting system that allows identification of recipients by their
received files. With a technique called fingercasting , recipients automatically embed a watermark in files during the decryption process. The process is based on the chameleon cipher [12], which allows one to decrypt an encrypted file with different decryption keys, to introduce some noise that can be used as a means of identification. In [28] Katzenbeisser et al. use the technique of fingercasting together with a randomized fingerprinting code in order to provide better security against colluding attackers. However, in these broadcasting approaches the problem of an untrusted sender is not addressed.
## 8.4 Watermarking
LIME can be used with any type of data for which watermarking schemes exist. Therefore, we briefly describe different watermarking techniques for different data types. Most watermarking schemes are designed for multimedia files such as images [40], videos [20], and audio files [11]. In these multimedia files, watermarks are usually embedded by using a transformed representation (e.g. discrete cosine, wavelet or Fourier transform) and modifying transform domain coefficients.
Watermarking techniques have also been developed for other data types such as relational databases, text files and even Android apps. The first two are especially interesting, as they allow us to apply LIME to user databases or medical records. Watermarking relational databases can be done in different ways. The most common solutions are to embed information in noise-tolerant attributes of the entries or to create fake database entries [25]. For watermarking of texts, there are two main approaches. The first one embeds information by changing the text's appearance (e.g. changing distance between words and lines) in a way that is imperceptible to humans [15]. The second approach is also referred to as language watermarking and works on the semantic level of the text rather than on its appearance [13]. A Mechanism also has been proposed to insert watermarks to Android apps [49]. This mechanism encodes a watermark in a permutation graph and hides the graph as a linked list in the application. Due to the list representation, watermarks are encoded in the execution state of the application rather than in its syntax, which makes it robust against attacks.
Embedding multiple watermarks into a single document has been discussed in literature and there are different techniques available [46, 32, 23]. In [32] they discuss multiple re-watermarking and in [23] the focus is on segmented watermarking. Both papers show in experimental results that multiple watermarking is possible which is very important for our scheme, as it allows us to create a lineage over multiple levels.
It would be desirable not to reveal the private watermarking key to the auditor during the auditor's investigation, so that it can be safely reused, but as discussed in [18, 31, 48] current public key watermarking schemes are not secure and it is doubtful if it is possible to design one that is secure. In [44] Sadeghi presents approaches to zero-knowledge watermark detection. With this technology it is possible to convince another party of the presence of a watermark in a document without giving any information about the detection key or the watermark itself. However, the scheme discussed in [44] also hides the content of the watermark itself and are therefore unfit for our case, as the auditor has to know the watermark to identify the guilty person. Furthermore, using a technology like this would come with additional constraints for the chosen watermarking scheme.
## 9 Conclusion and Future Directions
We present LIME, a model for accountable data transfer across multiple entities. We define participating parties, their inter-relationships and give a concrete instantiation for a data transfer protocol using a novel combination of oblivious transfer, robust watermarking and digital signatures. We prove its correctness and show that it is realizable by giving microbenchmarking results. By presenting a general applicable framework, we introduce accountability as early as in the design phase of a data transfer infrastructure.
Although LIME does not actively prevent data leakage, it introduces reactive accountability. Thus, it will deter malicious parties from leaking private documents and will encourage honest (but careless) parties to provide the required protection for sensitive data. LIME is flexible as we differentiate between trusted senders (usually owners) and untrusted senders (usually consumers). In the case of the trusted sender, a very simple protocol with little overhead is possible. The untrusted sender requires a more complicated protocol, but the results are not based on trust assumptions and therefore they should be able to convince a neutral entity (e.g. a judge).
Our work also motivates further research on data leakage detection techniques for various document types and scenarios. For example, it will be an interesting future research direction to design a verifiable lineage protocol for derived data.
Acknowledgements. We thank Deepak Garg for encouraging discussions and Amit Roy, Emre Goynugur and Isaac Alpizar Chacon for their work towards an initial draft. We are also grateful to our anonymous reviewers for their comments.
## References
- [1] Chronology of data breaches. http://www.privacyrights.org/data-breach .
[2]
Data breach cost.
http://www.symantec.com/about/news/release/article.jsp?prid=20110308\_01
.
- [3] Facebook in Privacy Breach. http://online.wsj.com/article/SB10001424052702304772804575558484075236968.html .
- [4] GNU Multiple Precision Arithmetic Library (GMP). http://gmplib.org/ .
[5]
Offshore outsourcing.
http://www.computerworld.com/s/article/109938/Offshore\_outsourcing\_cited\_in\_Florida\_data\_leak
.
- [6] Pairing-Based Cryptography Library (PBC). http://crypto.stanford.edu/pbc .
- [7] Privacy rights clearinghouse. http://www.privacyrights.org .
- [8] Electronic Privacy Information Center (EPIC). http://epic.org , 1994.
- [9] A. Adelsbach, U. Huber, and A.-R. Sadeghi. Fingercasting - Joint Fingerprinting and Decryption of Broadcast Messages. T. Data Hiding and Multimedia Security , 2:1-34, 2007.
- [10] A. Adelsbach, S. Katzenbeisser, and A.-R. Sadeghi. A computational model for watermark robustness. In Information Hiding , pages 145-160. Springer, 2007.
- [11] M. A. Alsalami and M. M. Al-Akaidi. Digital audio watermarking: survey. School of Engineering and Technology, De Montfort University, UK , 2003.
- [12] R. Anderson and C. Manifavas. Chameleon - A new kind of stream cipher. In Fast Software Encryption , pages 107-113. Springer, 1997.
- [13] M. J. Atallah, V . Raskin, C. Hempelmann, M. Karahan, R. Sion, U. Topkara, and K. E. Triezenberg. Natural Language Watermarking and Tamperproofing. In Information Hiding , pages 196-212, 2002.
- [14] D. Boneh, B. Lynn, and H. Shacham. Short signatures from the Weil pairing. In Advances in Cryptology-ASIACRYPT 2001 , pages 514-532. Springer, 2001.
- [15] J. T. Brassil, S. Low, and N. F. Maxemchuk. Copyright protection for the electronic distribution of text documents. Proceedings of the IEEE , 87(7):1181-1196, 1999.
- [16] J.-G. Choi, G. Hanaoka, K. H. Rhee, and H. Imai. How to break COT-based fingerprinting schemes and design new one. IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences , 88(10):2800-2807, 2005.
[17] I. J. Cox, J. Kilian, F. T. Leighton, and T. Shamoon. Secure spread spectrum watermarking for multimedia. Image Processing, IEEE Transactions on , 6(12):1673-1687, 1997.
- [18] I. J. Cox and J.-P. M. Linnartz. Public watermarks and resistance to tampering. In International Conference on Image Processing (ICIP'97) , pages 26-29, 1997.
[19] W. Dai. Crypto++ Library. http://cryptopp.com .
| [20] | G. Do¨rr e and J.-L. Dugelay. A guide tour of video waterm arking. Signal processing: Image communication , 18(4):263-282, 2003. |
|--------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [21] | J. Domingo-Ferrer. Anonymous fingerprinting based on committed oblivious transfer. In Public Key Cryptography , pages 43-52. Springer, 1999. |
| [22] | J. Domingo-Ferrer and J. Herrera-Joancomart´. ı Efficient smart-card based anonymous fingerprinting. In Smart Card Research and Applications , pages 221-228. Springer, 2000. |
| [23] | K. Frikken and M. Atallah. Cropping-resilient segmented multiple watermarking. In Algorithms and Data Structures , pages 231-242. Springer, 2003. |
| [24] | S. Goldwasser, S. Micali, and R. L. Rivest. A digital signature scheme secure against adaptive chosen-message attacks. SIAM J. Comput. , 17(2):281-308, 1988. |
| [25] | R. Halder, S. Pal, and A. Cortesi. Watermarking techniques for relational databases: Survey, classification and comparison. Journal of Universal Computer Science , 16(21):3164-3190, 2010. |
| [26] | D. Hu and Q. Li. Asymmetric fingerprinting based on 1-out-of-n oblivious transfer. Communications Letters, IEEE , 14(5):453-455, 2010. |
| [27] | Y. Ishai, J. Kilian, K. Nissim, and E. Petrank. Extending oblivious transfers efficiently. In Advances in Cryptology-CRYPTO 2003 , pages 145-161. Springer, 2003. |
| [28] | S. Katzenbeisser, B. Skoric, M. U. Celik, and A.-R. Sadeghi. Combining Tardos Fingerprinting Codes and Fingercasting. In Information Hiding , pages 294-310, 2007. |
| [29] | F. Kelbert and A. Pretschner. Data usage control enforcement in distributed systems. In CODASPY , pages 71-82, 2013. |
| [30] | J. Kilian, F. T. Leighton, L. R. Matheson, T. G. Shamoon, R. E. Tarjan, and F. Zane. Resistance of digital watermarks to collusive attacks. In IEEE International Symposium on Information Theory , pages 271-271, 1998. |
| [31] | J.-P. M. Linnartz and M. Van Dijk. Analysis of the sensitivity attack against electronic watermarks in images. In Information Hiding , pages 258-272. Springer, 1998. |
| [32] | A. Mascher-Kampfer, H. St¨gner, o and A. Uhl. Multiple r e-watermarking scenarios. In Proceedings of the 13th International Conference on Systems, Signals, and Image Processing (IWSSIP 2006) , pages 53-56. Citeseer, 2006. |
| [33] | P. Meerwald. Watermarking toolbox. http://www.cosy.sbg.ac.at/˜pmeerw/Watermarking/source . |
| [34] | M. Naor and B. Pinkas. Efficient oblivious transfer protocols. In Proceedings of the Twelfth Annual ACM-SIAM Symposium on Discrete Algorithms , pages 448-457, 2001. |
| [35] | P. Papadimitriou and H. Garcia-Molina. Data leakage detection. Knowledge and Data Engineering, IEEE Transactions on , 23(1):51-63, 2011. |
| [36] | R. Parviainen and P. Parnes. Large scale distributed watermarking of multicast media through encryption. In Proceedings of the IFIP TC6/TC11 International Conference on Communications and Multimedia Security Issues of the New Century , volume 192, pages 149-158, 2001. |
| [37] | R. Petrovic and B. Tehranchi. Watermarking in an encrypted domain, July 7 2006. US Patent App. 11/482,519. |
| [38] | B. Pfitzmann and M. Waidner. Asymmetric fingerprinting for larger collusions. In Proceedings of the 4th ACM conference on Computer and communications security , CCS '97, pages 151-160, 1997. |
| [39] | G. S. Poh. Design and Analysis of Fair Content Tracing Protocols . PhD thesis, 2009. |
| [40] | V. M. Potdar, S. Han, and E. Chang. A survey of digital image watermarking techniques. In International Conference on Industrial Informatics, 2005. INDIN'05. 2005 3rd IEEE , pages 709-716. IEEE, 2005. |
| [41] | A. Pretschner, M. Hilty, F. Sch¨tz, u C. Schaefer, and T.Walter. Usage Control Enforcement: Present and Future. IEEE Security &Privacy , 6(4):44-53, 2008. |
| [42] | A.-R. Sadeghi. How to break a semi-anonymous fingerprinting scheme. In Information Hiding , pages 384-394. Springer, 2001. |
| [43] | A.-R. Sadeghi. Secure Fingerprinting on Sound Foundations . PhD thesis, 2004. |
| [44] | A.-R. Sadeghi. The Marriage of Cryptography and Watermarking - Beneficial and Challenging for Secure Watermarking and Detection. In Proceedings of the 6th International Workshop on Digital Watermarking , IWDW'07, pages 2-18, 2008. |
| [45] | F. Salim, N. P. Sheppard, and R. Safavi-Naini. A Rights Management Approach to Securing Data Distribution in Coalitions. In NSS , pages 560-567, 2010. |
- [46] N. P. Sheppard, R. Safavi-Naini, and P. Ogunbona. On multiple watermarking. In MM&Sec , pages 3-6, 2001.
- [47] N. P. Sheppard, R. Safavi-Naini, and P. Ogunbona. Secure multimedia authoring with dishonest collaborators. EURASIP J. Appl. Signal Process. , 2004:2214-2223, 2004.
- [48] Y. Wu and R. H. Deng. A Pollution Attack to Public-key Watermarking Schemes. In Multimedia and Expo (ICME), 2012 IEEE International Conference on , pages 230-235. IEEE, 2012.
- [49] W. Zhou, X. Zhang, and X. Jiang. Appink: watermarking android apps for repackaging deterrence. In Proceedings of the 8th ACM SIGSAC symposium on Information, computer and communications security , pages 1-12. ACM, 2013.
| null |
[
"Michael Backes",
"Niklas Grimm",
"Aniket Kate"
] |
2014-08-05T19:45:38+00:00
|
2014-08-05T19:45:38+00:00
|
[
"cs.CR"
] |
Lime: Data Lineage in the Malicious Environment
|
Intentional or unintentional leakage of confidential data is undoubtedly one
of the most severe security threats that organizations face in the digital era.
The threat now extends to our personal lives: a plethora of personal
information is available to social networks and smartphone providers and is
indirectly transferred to untrustworthy third party and fourth party
applications.
In this work, we present a generic data lineage framework LIME for data flow
across multiple entities that take two characteristic, principal roles (i.e.,
owner and consumer). We define the exact security guarantees required by such a
data lineage mechanism toward identification of a guilty entity, and identify
the simplifying non repudiation and honesty assumptions. We then develop and
analyze a novel accountable data transfer protocol between two entities within
a malicious environment by building upon oblivious transfer, robust
watermarking, and signature primitives. Finally, we perform an experimental
evaluation to demonstrate the practicality of our protocol.
|
1408.1077v2
|
Mar´ ıa G´mez-Rocha o · Thomas Hilger · Andreas
Krassnigg
## First look at heavy-light mesons with a dressed quark-gluon vertex
Received: December 4, 2014
Abstract Following up on earlier work, we investigate possible effects of a dressed quark-gluon vertex in heavy-light mesons. In particular, we study corrections to the popular rainbow-ladder truncation of the Dyson-Schwinger-Bethe-Salpeter equation system. We adopt a simple interaction kernel which reduces the resulting set of coupled integral equations to a set of coupled algebraic equations, which are solved numerically. In this way, we extend previous studies to quark-antiquark systems with unequal current-quark masses, at first for the pseudoscalar case, and investigate the resulting set of problems and solutions. We attempt to find patterns in - as well as to quantify corrections to - the rainbowladder truncation. In addition, we open this approach to phenomenological predictions of the heavy quark symmetry.
Keywords Heavy-light mesons · Dyson-Schwinger Equations · Bethe-Salpeter Equation
## 1 Motivation
In modern studies of Quantum Chromodynamics (QCD) meson states made of one light and one heavy quark have attracted particular attention for a number of years. The reasons are manifold, e. g., the possibility to construct a theoretical bridge between effective field theories valid in opposite regimes like chiral perturbation theory as well as perturbative nonrelativistic QCD and heavy-quark effective theory, respectively. Another challenging reason is the need to understand a system determined by different scales of the order of Λ QCD on one hand and the bottom-quark mass on the other.
Among the most promising approaches that are able to deal with all of these requirements as well as able to provide a solid and modern theoretical framework for bound-state studies are lattice-regularized QCD studies and also methods of continuum quantum field theories. Our method of choice herein is the Dyson-Schwinger-Bethe-Salpeter-equation (DSBSE) approach, which has been applied to various problems of theoretical hadron physics in the past decades with increasing success and comprehension, see, e. g., [1, 2] and references therein.
Typically, sophisticated hadron studies in the DSBSE approach are numerical in nature and make use of a truncation of the infinite tower of coupled integral equations involved; see, e. g., [3-5] for the details of such a setup. A popular variant is the so-called rainbow-ladder (RL) truncation, where the dressed quark-gluon vertex [6] assumes a strongly simplified structure; it models the gluon-quark interaction and reduces the effort of solving relevant equations to the Dyson-Schwinger equation (DSE) of the quark propagator, and subsequently the quark-antiquark Bethe-Salpeter equation (BSE) with a suitable interaction kernel.
Presented by Mar´ ıa G´mez-Rocha at Light-Cone 2014, NC State University, Raleigh, USA, 26 - 30 May, 2014 o
M. G´ omez-Rocha et al.
University of Graz, Institute of Physics, NAWI Graz, A-8010 Graz, Austria E-mail: [email protected]
Despite its success, in particular for pseudoscalar mesons and their various properties [7-11], vector mesons [12, 13], the heavy-quark domain [14, 15], and an extension to a quark-diquark and later threequark formulation for baryons [16-20], always see also references therein, this truncation has appeared limited in certain respects. Typically such limitations appear where the bare covariant (vector) structure left in the quark-gluon vertex in this truncation is expected to be insufficient for the treatment of certain states, such as those identified as orbital-angular-momentum excitations [21, 22]. A situation similar to the latter case is presented by a heavy-light meson state, since one cannot expect the same simplicity as apparent in a pseudoscalar or vector meson ground state with equal-mass quarks. While heavy-light meson states have been investigated in the DSBSE approach, cf. [23] and references therein, the assumptions made there about the connection between the DSE and BSE kernels are different from the truncation scheme considered here and cannot easily be related to our present setup.
Herein, we investigate dressing effects in the quark-gluon vertex in a scheme developed earlier in the equal-mass-constituent case [24, 25] for a simple interaction model. This goes beyond an introduction of trivial dressing factors accompanying the bare structure in the quark-gluon vertex. For the sake of brevity, we merely sketch the formalism and rely on illustrations to highlight the problems and possibilities given by our study. We quantify dressing effects for select pseudoscalar states and present further directions of our work. The calculations are performed in Euclidean space.
## 2 Setup and interaction model
Meson studies in the DSBSE approach have a long-standing successful record. In the context of the present setup it is always instructive to note that first simplified attempts at meson spectroscopy analogous to the ones undertaken today were performed already several decades ago [26]. In fact, the interaction model defined there is the basis for the investigation presented here.
In QCD the DSE for the quark propagator reads
$$S ^ { - 1 } ( p ) = i \mathfrak { p } + \mathbf 1 m _ { q } + \int \frac { d ^ { 4 } q } { ( 2 \pi ) ^ { 4 } } g ^ { 2 } D _ { \mu \nu } ( p - q ) \frac { \lambda ^ { a } } { 2 } \gamma _ { \mu } S ( q ) \Gamma _ { \nu } ^ { a } ( q ; p )$$
where p and m q are the quark momentum and current mass; S , D µν , and Γ a ν are the renormalized dressed quark propagator, gluon propagator, and quark-gluon vertex, respectively. The color structure of the interaction is encoded in the equation via the superscript a , g is the strong coupling constant, and S has the general Dirac structure of a fermion propagator.
In order to study mesons, one needs to solve the BSE that contains two dressed quark propagators, obtained by solving Eq. (1), together with the quark-antiquark scattering kernel K in the form
$$\Gamma ( p ; P ) = \int _ { q } ^ { \Lambda } K ( p ; q ; P ) \, S ( q _ { + } ) \Gamma ( q ; P ) S ( q _ { - } ) \,,$$
where Γ is the Bethe-Salpeter amplitude (BSA) and Λ a regularization scale. The (anti-)quark momenta are given by q + = q + ηP and q -= q -(1 -η P ) ; the momentum partitioning parameter η ∈ [0 , 1] is an in principle free parameter, which accounts for the arbitrariness in the definition of the relative quark-antiquark momentum q . In the case of equal-mass quarks, the obvious (and most convenient) choice is η = 1 2, but in the present work we have to pay some attention to this particular parameter / and its effects on the results, explained in detail below.
With S precomputed, a choice for K is needed to proceed with the solution of the BSE. Such a choice is made possible via guiding restrictions by the axial-vector Ward-Takahashi identity (AVWTI), which guards the effects of chiral symmetry and its dynamical breaking in any model description of hadrons that implements a consistent set of interactions. While the simplest truncation of the DSBSE system to satisfy this constraint, the RL truncation, is well-known, steps beyond it have been sought and developed with the goal of a nonperturbative systematic scheme [24, 25] that would enable numerical studies of increasing sophistication, such as [21, 22, 27-32]. On a more general basis the AVWTI has been investigated to find construction principles for the meson-BSE kernel for a given type of quark-gluon vertex [33-35].
The setup employed here is a straight-forward adaptation of the one presented in [25], and we refer the reader to this reference for more details regarding the formalism and a comprehensive set of results
Π
C
/Equal
0.51
Fig. 1 Left panel : Pion mass as a function of the number of recursive steps used in the construction of the relevant quark-gluon vertex. For each n , several lines are plotted for a set of representative values of η spanning the entire range of possible values. Right panel : Pion mass plotted as a function of η ; different lines correspond to different n . The symmetry of the graph with respect to η ↔ (1 -η ) characterizes the equal-mass-constituent case. In both panels, the dashed line marks the experimental mass value for comparison.
for both the quark-propagator dressing functions as well as results for pseudoscalar and vector mesons with equal-mass quarks. Here we sketch only the most important ingredients of our study; more results and a detailed account of the involvement due to the unequal masses of the constituents in the BSE will be published elsewhere [36].
In order to understand and discuss the results presented herein, two more ingredients are necessary, namely the interaction model and the construction principle for the quark-gluon vertex. To define the interaction in this model, we follow [25, 26] in replacing the gluon propagator by
$$g ^ { 2 } \, D _ { \mu \nu } ( k ) \coloneqq \left ( \delta _ { \mu \nu } - \frac { k _ { \mu } k _ { \nu } } { k ^ { 2 } } \right ) ( 2 \pi ) ^ { 4 } \, \mathcal { G } ^ { 2 } \, \delta ^ { 4 } ( k ) \,.$$
The parameter G defines the (constant) interaction strength, sets a scale in the computations, and is determined a priori. All integrals collapse via the δ 4 -term and one is left with a coupled system of algebraic equations which are then solved numerically. Apart from this simplifying effect, one has to note as well that also in the BSA several structures (more precisely those involving factors of the relative momentum) become trivial, i. e., they vanish. As a result the manifest covariance of the BSA in this model is destroyed, which immediately impacts on the role of the momentum partitioning parameter η : meson masses and other results are no longer η -independent like in the case with all covariants taken into account.
In practice this means that the model assumptions oversimplify meson structure to the extent of partial loss of covariance and seem to have even more but indirect consequences such as: the lack of solutions for scalar and axial-vector meson states; the trivialization of the entire BSA for mesons with spin J > 1 [5, 37]; the lack of solutions for radially excited meson states; an unnatural connection of interaction strength and the resulting amount of dynamical chiral symmetry breaking, in particular at finite temperatures [38]. Nonetheless, the benefits outweigh these shortcomings in that one can probe dressing effects up to infinite order in any given scheme as well as more complicated approaches and new ideas with unparalleled ease as compared to studies using more realistic interactions.
As a first illustration we gauge the dependence on η for the already-known case of equal-mass quarks in the case of the pion, as presented in Fig. 1; this serves as a guide for the heavy-light case in terms of a systematic error due to one particular choice for η . What we find is only a light variation of the order of a few percent of the pion's mass for each stage in the recursive approach compared to the fully dressed quark-gluon vertex in our scheme (see below). The same needs to be done for the pseudoscalar quarkonia, presented in Fig. 2: there, while the effects look larger in the figure, their relative order of magnitude is still at the few-percent level, comparable with the pion case. In the figures, we also show a dashed horizontal line to indicate the respective experimental mass value for easy reference. Note also that not for all cases of η a numerical solution can easily be found, in which cases the corresponding lines in the figure are missing. More precisely, the correction to RL ∼ 1.6% for η = 0 5 in the charmonium case. The systematic error due to the model-artificial . η -dependence can be quantified to be smaller than 6%. For pseudoscalar bottomonium, we obtain a correction to the RL result of ∼ 0.2% for η = 0 5, where the systematic error due to the . η -dependence is smaller than 1.5%.
c
Η
Η
c
b
C
/Equal
0.51
/Equal
0.51
Fig. 2 Same as the left panel of Fig. 1 for the charmonium (left panel) and bottomonium (right panel) pseudoscalar ground states. Note that no lines are plotted where no solution for the BSE was found at a particular order n or for a particular value of η . Note also that as a general pattern no solutions are found for those situations considered here with an odd value of n . This qualitative difference between odd and even n is not new; it has been explored and detailed in [25], and also will be further investigated in future works [36].
The construction of the quark-gluon vertex follows a recursive pattern based on its DSE. Two correction terms to the bare vertex, the so-called abelian and non-abelian gluon corrections, are considered and combined effectively to a single term with unified form and variable dependence. The combination is weighted with an overall factor C , which, in turn can be used to serve as a model parameter and adjusted to data; in the case of [25] this was fit to data from lattice-regularized QCD on the propagator level. The effective equation to obtain the quark-gluon vertex reads
$$\Gamma _ { \mu } ^ { \mathcal { C } } ( p ) = \gamma _ { \mu } - \mathcal { C } \, \gamma _ { \rho } \, S ( p ) \, \Gamma _ { \mu } ^ { \mathcal { C } } ( p ) \, S ( p ) \, \gamma _ { \rho }$$
which illustrates the dependence on C . At the same time, it is clear that this equation can be approached with a recursive procedure and that the bare quark-gluon vertex serves as a starting value, Γ C µ, 0 ( p ) = γ µ . The recursion relation is
$$\Gamma ^ { \mathcal { C } } _ { \mu, n } ( p ) = - \mathcal { C } \, \gamma _ { \rho } \, S ( p ) \, \Gamma ^ { \mathcal { C } } _ { \mu, n - 1 } ( p ) \, S ( p ) \, \gamma _ { \rho }$$
∞
and the final result for the quark-gluon vertex is obtained by Γ C µ ( p ) = ∑ n =0 Γ C µ,n ( p ). In view of these relations, additional information for the interpretation of Figs. 1 and 2 is at hand: The value of C = 0 51 was chosen in [25] and is adopted here without change together with the values . G = 0 69 GeV, . m u,d = 0 01 GeV, . m s = 0 166 GeV, . m c = 1 33 GeV, and . m b = 4 62 GeV; this enables a . reasonable phenomenological description of the equal-mass pseudoscalar and vector states and is thus an ideal choice for our study, where a quantification of systematics beyond RL truncation is sought. Also, we can now interpret n as the number of times the recursion in Eq. (5) has been applied to the bare-vertex RL case in the quark DSE and the consistent BSE: n = 0 corresponds to the RL truncation itself and is thus completely insensitive to the parameter C . n = ∞ is obtained by an exact summation of all terms in the series construction of Γ C µ ( p ) and offers a unique perspective that reaches beyond a set of results obtained using any finite number of terms.
## 3 Results and Discussion
After the necessary details have been laid out and checks for the equal-mass cases have been presented, we now focus on the goal of our study, namely, to obtain information about the importance of dressings of the quark-gluon vertex such as the one sketched above in heavy-light mesons. This is particularly interesting, since an immediate and comprehensive phenomenologically successful sophisticated RLstudy of heavy-light mesons is still missing in the repertoire of the DSBSE approach; see [39] for a recent account of efforts and status of current research in this direction.
As a first proving ground for our approach we investigated the kaon which is the prototype of the unequal-mass constituent meson in every DSBSE study. The results shown in Fig. 3 are completely analogous to the ones shown for the pion in Fig. 1. In our setup we find a correction to RL of ∼ 7% for η = 0 5. The systematic error due to the . η -dependence can be quantified to be smaller than 10%. The symmetry with regard to η ↔ (1 -η ) apparent for the pion in the right panel of Fig. 1 is clearly
K C
/Equal
0.51
Fig. 3 Same as Fig. 1 for the kaon mass. The experimental value is above results and not shown in right panel.
broken in the kaon case due to the unequal masses of the quarks. We also note here that for the kaon there are no intersections of curves in the right panel of Fig. 3: for each value of η the convergence pattern from n = 0 towards n = ∞ is the same. While not shown here for the sake of brevity, we note that this behavior is not guaranteed by any particular aspect of our setup and that the curves for the different n can indeed intersect in other cases.
The two heavy-light meson masses investigated here are those of the D and B mesons, for which we show our results in the left and right panels of Fig. 4, respectively. In terms of quantitative information we extract the following numbers: For the D meson we find a correction to RL truncation of ∼ 5% for a choice of η = 0 75. The corresponding systematic error due to the model-artificial . η -dependence can only be weakly constrained to be less than 15%, which somewhat blurs possible conclusions from the differences among results for different n , but is still a valuable guide for sophisticated studies analogous to the present. The result for the D meson mass in the fully dressed case in our scheme underestimates the experimental value by about 10%, which is still reasonable, given the considerations above.
For the B meson we obtain a correction to RL truncation of ∼ 2% for η = 0 95. The error due to the . η -dependence is somewhat smaller than for the D , namely below 12%. Our result for the B mass in the fully-dressed case again underestimates the experimental value, by about 4%, but yet again lies within a reasonable range given the η -variations quoted above. Regarding the values chosen for η we note that they are typical values chosen in a DSBSE meson calculation. For example, in the original treatment of the present interaction model in its RL setup in [26], the values chosen for η in the D and B meson cases were 0 8 and 0 92, respectively. We also note that the authors in [26] find mass-variations with . . η of similar size to our own; they quote that such variations are smaller than 15 %.
## 4 Conclusions
Based on previous works we investigated the K D , , and B meson masses in a systematic truncation scheme of the DSBSE system with a simplifying but far-reaching model-interaction setup. After an estimate of systematic model-inherent errors we proceed to quantify corrections to the popular RL truncation. In particular, a comparison of the dressing effects for the equal-mass and heavy-light cases
D C
B
/Equal
0.51
C
/Equal
0.51
Fig. 4 Same as the left panel in Fig. 1 for the D (left panel) and B (right panel) meson masses.
could suggest that a simultaneous phenomenological description of both kinds of pseudoscalars in an RL treatment can in fact be successful, which is an important piece of outlook for future sophisticated hadron studies in this approach [40]. Next steps for the work presented here are the inclusion of vector mesons and a direct check of heavy-quark symmetry predictions like they have been performed recently in a similar setup in relativistic Hamiltonian dynamics [41].
Acknowledgements We acknowledge helpful conversations with C. Popovici, H. Sanchis-Alepuz, P. C. Tandy, and R. Williams. This work was supported by the Austrian Science Fund (FWF) under project no. P25121-N27.
## References
- 1. A. Bashir, L. Chang, I. C. Cloet, B. El-Bennich, et al. Commun.Theor.Phys. , 58:79-134, 2012.
- 2. Ian C. Cloet and Craig D. Roberts. Prog.Part.Nucl.Phys. , 77:1-69, 2014.
- 3. M. Blank and A. Krassnigg. Comput. Phys. Commun. , 182:1391, 2011.
- 4. A. Krassnigg. Phys. Rev. D , 80:114010, 2009.
- 5. C. S. Fischer, S. Kubrak, and R. Williams arXiv:1406.4370.
- 6. R. Alkofer, C. S. Fischer, F. Llanes-Estrada, and K. Schwenzer Ann. Phys. , 324:106, 2009.
- 7. Pieter Maris and Craig D. Roberts. Phys. Rev. C , 56:3369-3383, 1997.
- 8. Pieter Maris and Peter C. Tandy. Phys. Rev. C , 62:055204, 2000.
- 9. A. Holl, A. Krassnigg, and C. D. Roberts. Phys. Rev. C , 70:042203(R), 2004.
- 10. P. Maris and P. C. Tandy. Nucl. Phys. Proc. Suppl. , 161:136-152, 2006.
- 11. A. Holl, A. Krassnigg, P. Maris, C. D. Roberts, and S. V. Wright. Phys. Rev. C , 71:065204, 2005.
- 12. Pieter Maris and Peter C. Tandy. Phys. Rev. C , 60:055214, 1999.
- 13. M. S. Bhagwat and P. Maris. Phys. Rev. C , 77:025203, 2008.
- 14. Pieter Maris. AIP Conf. Proc. , 892:65-71, 2007.
- 15. M. Blank and A. Krassnigg. Phys. Rev. D , 84:096014, 2011.
- 16. G. Eichmann, A. Krassnigg, M. Schwinzerl, and R. Alkofer. Annals Phys. , 323:2505-2553, 2008.
- 17. G. Eichmann, R. Alkofer, A. Krassnigg, and D. Nicmorus. Phys. Rev. Lett. , 104:201601, 2010.
- 18. H. Sanchis-Alepuz, G. Eichmann, S. Villalba-Chavez, and R. Alkofer. Phys.Rev. , D84:096003, 2011.
- 19. Gernot Eichmann and Christian S. Fischer. Phys.Rev. , D87:036006, 2013.
- 20. H. Sanchis-Alepuz, R. Williams, and R. Alkofer. Phys.Rev. , D87:096015, 2013.
- 21. Christian S. Fischer and Richard Williams. Phys. Rev. D , 78:074006, 2008.
- 22. Christian S. Fischer and Richard Williams. Phys. Rev. Lett. , 103:122001, 2009.
- 23. Mikhail A. Ivanov, Yu. L. Kalinovsky, and Craig D. Roberts. Phys. Rev. D , 60:034018, 1999.
- 24. A. Bender, W. Detmold, C. D. Roberts, and A. W. Thomas. Phys. Rev. C , 65:065203, 2002.
- 25. M. S. Bhagwat, A. Holl, A. Krassnigg, C. D. Roberts, and P. C. Tandy. Phys. Rev. C , 70:035205, 2004.
- 26. H. J. Munczek and A. M. Nemirovsky. Phys. Rev. D , 28:181, 1983.
- 27. Peter Watson and Wolfgang Cassing. Few-Body Syst. , 35:99-115, 2004.
- 28. P. Watson, W. Cassing, and P. C. Tandy. Few-Body Syst. , 35:129-153, 2004.
- 29. C. S. Fischer, P. Watson, and W. Cassing. Phys. Rev. D , 72:094025, 2005.
- 30. H. H. Matevosyan, A. W. Thomas, and P. C. Tandy. Phys. Rev. C , 75:045201, 2007.
- 31. H. H. Matevosyan, A. W. Thomas, and P. C. Tandy. J. Phys. G , 34:2153-2164, 2007.
- 32. R. Williams. arXiv:1404.2545.
- 33. H. J. Munczek. Phys. Rev. D , 52:4736-4740, 1995.
- 34. Lei Chang and Craig D. Roberts. Phys. Rev. Lett. , 103:081601, 2009.
- 35. Walter Heupel, Tobias Goecke, and Christian S. Fischer. Eur.Phys.J. , A50:85, 2014.
- 36. M. Gomez-Rocha, T. Hilger, and A. Krassnigg. in preparation .
- 37. A. Krassnigg and M. Blank. Phys. Rev. D , 83:096006, 2011.
- 38. M. Blank and A. Krassnigg. Phys. Rev. D , 82:034006, 2010.
- 39. E. Rojas, B. El-Bennich, and J.P.B.C. de Melo. arXiv:1407.3598.
- 40. C. Popovici, T. Hilger, M. Gomez-Rocha, and A. Krassnigg. arXiv:1407.7970.
- 41. M. Gomez-Rocha and W. Schweiger. Phys.Rev. , D86:053010, 2012.
|
10.1007/s00601-014-0938-8
|
[
"Maria Gomez-Rocha",
"Thomas Hilger",
"Andreas Krassnigg"
] |
2014-08-05T19:46:54+00:00
|
2014-12-03T20:36:33+00:00
|
[
"hep-ph"
] |
First look at heavy-light mesons with a dressed quark-gluon vertex
|
Following up on earlier work, we investigate possible effects of a dressed
quark-gluon vertex in heavy-light mesons. In particular, we study corrections
to the popular rainbow-ladder truncation of the Dyson-Schwinger-Bethe-Salpeter
equation system. We adopt a simple interaction kernel which reduces the
resulting set of coupled integral equations to a set of coupled algebraic
equations, which are solved numerically. In this way, we extend previous
studies to quark-antiquark systems with unequal current-quark masses, at first
for the pseudoscalar case, and investigate the resulting set of problems and
solutions. We attempt to find patterns in - as well as to quantify corrections
to - the rainbow-ladder truncation. In addition, we open this approach to
phenomenological predictions of the heavy quark symmetry.
|
1408.1078v2
|
## Einstein Metrics, Harmonic Forms, and Symplectic Four-Manifolds
Claude LeBrun Stony Brook University
## Abstract
If M is the underlying smooth oriented 4-manifold of a Del Pezzo surface, we consider the set of Riemannian metrics h on M such that W + ( ω, ω ) > 0, where W + is the self-dual Weyl curvature of h , and ω is a non-trivial self-dual harmonic 2-form on ( M,h ). While this open region in the space of Riemannian metrics contains all the known Einstein metrics on M , we show that it contains no others. Consequently, it contributes exactly one connected component to the moduli space of Einstein metrics on M .
## 1 Introduction
Given a smooth compact 4-manifold M , one would like to completely understand the moduli space E ( M ) of the Einstein metrics it carries. Recall that an Einstein metric [4] means a Riemannian metric h which has constant Ricci curvature, in the sense that it solves the Einstein equation r = λh
where r is the Ricci tensor of h and λ is a real number, called the Einstein constant of h . The moduli space E ( M ) of Einstein metrics on M is by definition the quotient of the set of Einstein metrics by the action of the group Diff ( M ) × R × of self-diffeomorphisms and constant rescalings. For simplicity, we may give E ( M ) the quotient topology induced by the C ∞ -topology on the space of smooth metric tensors; however, it is worth noting that, for reasons of elliptic regularity [13], this coincides [1] with the metric topology induced by the Gromov-Hausdorff distance between unit-volume Einstein metrics.
While our understanding of this problem is rather limited for general 4-manifolds, there are specific cases where our knowledge is quite complete.
In particular, if M is the 4-torus, or K 3, or a compact real or complexhyperbolic 4-manifold, the Einstein moduli space E ( M ) is known to be connected [4, 5, 20]. This should be contrasted with the pattern that predominates in higher dimensions, where Einstein moduli spaces are typically disconnected, and indeed often have infinitely many connected components [7, 8, 34].
This article will explore related uniqueness questions for Einstein metrics on the small but important class of smooth compact 4-manifolds that arise as Del Pezzo surfaces . These 4-manifolds are characterized [10] by two properties: they admit Einstein metrics with λ > 0, and they also admit symplectic structures. Up to diffeomorphism, there are exactly ten such manifolds, namely S 2 × S 2 and the nine connected sums CP 2 # m CP 2 , m = 0 1 , , . . . , 8. The known Einstein metrics on these spaces all have λ > 0, and our objective here is to completely characterize these known Einstein metrics by a curvature condition. To this end, notice that, with their standard orientations, each of these 4-manifolds M has b + ( M ) = 1. This is equivalent to saying that, for any Riemann metric h on any of these compact oriented 4-manifolds, there is, up to an overall multiplicative constant, a unique non-trivial self-dual harmonic 2-form ω . We can therefore consider those Riemannian metrics h on M which satisfy the curvature inequality
$$W ^ { + } ( \omega, \omega ) > 0$$
at every point of M , where W + is the self-dual Weyl tensor of h . Note that, whenever b + ( M ) = 1, this condition only depends on the metric h , since the harmonic form ω is then uniquely determined up to a non-zero multiplicative constant. Our first main result is the following:
Theorem A. Let ( M,h ) be a compact oriented Einstein 4 -manifold with b + = 1 , and suppose that condition (1) holds at every point of M . Then M is diffeomorphic to a Del Pezzo surface, in such a way that h becomes
- · a K¨ ahler-Einstein metric with λ > 0 ; or
- · a constant multiple of the Page metric on CP 2 # CP 2 ; or
- · a constant multiple of the Chen-LeBrun-Weber metric on CP 2 #2 CP 2 .
Conversely, every metric on this list satisfies (1) at every point.
The proof of this result is given in § 2 below, and proceeds by proving that the given Einstein metric must be conformally K¨hler. a Our method makes
strong use of the fact that the second Bianchi identity implies that the selfdual Weyl curvature W + of an oriented 4-dimensional Einstein manifold is harmonic as a bundle-valued 2-form. In fact, the proof does not really require the assumption that b + ( M ) = 1; it suffices to assume that there is a harmonic self-dual 2-form ω on ( M,h ) such that (1) holds at every point.
While the inequality (1) may have a somewhat unfamiliar flavor, it is interestingly related to the positivity of scalar curvature. Indeed, any harmonic self-dual 2-form ω satisfies the Weitzenb¨ck formula o
$$\frac { 1 } { 2 } \Delta | \omega | ^ { 2 } + | \nabla \omega | ^ { 2 } + \frac { s } { 3 } | \omega | ^ { 2 } = 2 W ^ { + } ( \omega, \omega ),$$
and it therefore follows that any metric of scalar curvature s > 0 must at least satisfy (1) at some points of M . Thus, while Theorem A does not provide a complete classification of λ > 0 Einstein metrics on Del Pezzo surfaces, it does represent a step in that direction.
/negationslash
/negationslash
However, our method requires (1) to hold everywhere , rather than just at certain points. This is a strong condition, because it guarantees that the closed self-dual 2-form ω is nowhere zero, and therefore implies that ( M,ω ) is a symplectic manifold. If h is a Riemannian metric on a smooth compact oriented 4-manifold M with b + ( M ) = 0, we will thus say that h is of symplectic type if there is a self-dual harmonic 2-form on ( M,h ) such that ω = 0 at every point of M . This is actually a conformally invariant condition; if h is of symplectic type, and if u is a smooth positive function, then uh is also of symplectic type. For this reason, it is also natural to say that the conformal class [ h ] is of symplectic type if there is a self-dual harmonic 2-form on ( M, h [ ]) which is everywhere non-zero. This is an open condition on [ h ], in the sense that the set of conformal classes of symplectic type is automatically open [22] in the C 2 topology.
/negationslash
Condition (1) is also conformally invariant. Namely, if we replace h with uh for some positive function u , then W + ( ω, ω ) is replaced with u -3 W + ( ω, ω ), thereby leaving the the sign of W + ( ω, ω ) unaltered at any given point. We will henceforth say that the conformal class [ h ] is of positive symplectic type if, for some choice of h -compatible self-dual harmonic 2-form ω , condition (1) holds everywhere on M . This obviously implies that ω = 0 everywhere, so positive symplectic type implies symplectic type. The condition of positive symplectic type is once again open in the C 2 topology.
With these concepts in place, we are now ready to formulate our other main result, which is a direct consequence of Theorem A:
Theorem B. Let M be the underlying smooth compact 4 -manifold of a Del Pezzo surface. Let E ( M ) denote the the moduli space of Einstein metrics
h on M , and let E + ω ( M ) ⊂ E ( M ) be the open subset arising from Einstein metrics h for which the corresponding conformal classes [ h ] are of positive symplectic type. Then E + ω ( M ) is connected. Moreover, if b 2 ( M ) ≤ 5 , then E + ω ( M ) exactly consists of a single point.
Proof. A Del Pezzo surface is by definition a compact complex surface ( M ,J 4 ) whose first Chern class is a K¨hler class. a As complex manifolds, the Del Pezzo surfaces are exactly CP 1 × CP 1 and the blow-ups of CP 2 at m distinct points, 0 ≤ m ≤ 8, such that no three points are on a line, no six are on a conic, and no eight are on a nodal cubic with one of the given points at the node [11, 24]. When b 2 ≤ 5, there is consequently, up to biholomorphism, only one Del Pezzo complex structure for each diffeotype, since we can simultaneously move up to four generically located points in the projective plane to standard positions via a suitable projective linear transformation. For larger values of b 2 , the choice of complex structure instead essentially depends on 2 b 2 -10 complex parameters; however, the various possibilities still form a single connected family, since the set of prohibited configurations of m = b 2 -1 points in CP 2 is a finite union of complex hypersurfaces, and so has real codimension 2.
Given a Del Pezzo surface ( M,J ) with fixed complex structure, there is always a λ > 0 Einstein metric h which can be written as h = s -2 g for a J -compatible extremal K¨hler metric a g with scalar curvature s > 0. In most cases, one can simply take h = g , so that h is a K¨ ahler-Einstein metric. By a result of Tian [26, 32], a Del Pezzo surface ( M,J ) admits a J -compatible K¨hler-Einstein metric iff it has reductive automorphism group. a This excludes only two Del Pezzo surfaces, namely the ones diffeomorphic to CP 2 # CP 2 and CP 2 #2 CP 2 ; these two do not admit K¨hler-Einstein metrics, a but they nonetheless admit conformally K¨hler, a λ > 0 Einstein metrics, known as the Page and Chen-LeBrun-Weber metrics [10, 12, 27], respectively. In [23], it was then shown that any conformally K¨hler, Einstein a metric on a compact complex surface is either K¨hler-Einstein, a or else is isometric to a constant multiple of one of these two special metrics. In particular, up to complex automorphisms and constant rescalings, there is exactly one conformally K¨hler, Einstein metric for each Del Pezzo complex a structure. Theorem A therefore tells us that the moduli space E + ω ( M ) can be identified with the moduli space of Del Pezzo complex structures. Since we have seen that the latter is always pathwise connected, and moreover consists of a single point when b 2 ( M ) ≤ 5, the claim follows.
We now conclude this introduction with a consequence of Theorem B:
Corollary 1. For any Del Pezzo surface M , E + ω ( M ) is exactly a connected component of E ( M ) .
Indeed, it suffices to prove that the path-connected space E + ω ( M ) is both open and closed in E ( M ). Since E ( M ) has the quotient topology, the fact that it is open follows from the fact that the set of metrics with positive symplectic conformal class is open and invariant under the action of Diff ( M ) × R × . On the other hand, it is also closed, because, except in cases where E + ω ( M ) is now known to be a single point, the Einstein metrics in question are all K¨hler, and requiring that a Riemannian metric carry a a parallel almost-complex structure is a closed condition.
## 2 Harmonic Self-Dual Weyl Curvature
/negationslash
Recall that we say that a conformal class [ h ] on an a compact oriented 4manifold M is of symplectic type if there is a harmonic self-dual 2-form ω on ( M,h ) such that ω = 0 everywhere on M . This is indeed a conformally invariant condition, because the Hodge star operator is conformally invariant; moreover, it is an open condition [22] with respect to the C 2 topology. Since any self-dual 2-form ω satisfies
$$\omega \wedge \omega = \omega \wedge \ast \omega = | \omega | _ { h } ^ { 2 } \ d \mu _ { h },$$
it follows that an appropriate ω is actually a symplectic form on M if [ h ] is of symplectic type. Assuming this, the conformally related metric g ∈ [ h ] given by g = 2 -1 2 / | ω | h h is then an almost-K¨hler metric a , in the sense that g is related to the symplectic form ω via g = ω ( · , J · ) for a unique almostcomplex structure J on M . For our purposes, the important point is that, in dimension 4, the almost-K¨hler condition is equivalent to saying that a ω is harmonic and self-dual with respect to g , and that | ω | 2 g ≡ 2.
While our primary aim here is to learn something about Einstein metrics, we will more generally focus on oriented Riemannian 4-manifolds ( M,h ) with harmonic self-dual Weyl curvature , in the sense that δW + := -∇· W + = 0. When h is Einstein, this property holds, as a consequence of the second Bianchi identity. However, we will see in due course that δW + = 0 is in general much weaker than the Einstein condition.
When [ h ] is of symplectic type, it will prove profitable to study this equation from the point of view of the conformally related almost-K¨hler a metric g . This is quite tractable, because the divergence-free condition on a section of /circledot 2 0 Λ + is conformally invariant, albeit [28] with an unexpected
conformal weight . In practice, this means that if h = f 2 g has the property that δW + = 0, then g will instead have the property that δ fW ( + ) = 0. For us, the important point is that this then implies a Weitzenb¨ck formula o
$$0 = \nabla ^ { * } \nabla ( f W ^ { + } ) + \frac { s } { 2 } f W ^ { + } - 6 f W ^ { + } \circ W ^ { + } + 2 f | W ^ { + } | ^ { 2 } I \quad \ \ ( 2 )$$
for fW + , considered as a section of End(Λ + ); cf. [12, 15, 28].
To exploit this effectively, we will need the following identity:
Lemma 1. Any 4 -dimensional almost-K¨hler manifold satisfies a
$$\langle W ^ { + }, \nabla ^ { * } \nabla ( \omega \otimes \omega ) \rangle = [ W ^ { + } ( \omega, \omega ) ] ^ { 2 } + 4 | W ^ { + } ( \omega ) | ^ { 2 } - s W ^ { + } ( \omega, \omega )$$
at every point.
Proof. First notice that the oriented Riemannian 4-manifold ( M,g ) satisfies
$$\Lambda ^ { + } \otimes \mathbb { C } = \mathbb { C } \omega \oplus K \oplus \overline { K },$$
where K = Λ 2 0 , J is the canonical line bundle of the almost-complex manifold ( M,J ). Locally choosing a unit section ϕ of K , we thus have
$$\nabla \omega = \alpha \otimes \varphi + \bar { \alpha } \otimes \bar { \varphi }$$
for a unique 1-form α ∈ Λ 1 0 , J , since ∇ [ a ω bc ] = 0 and ω bc ∇ a ω bc = 0. If
$$\circledast \colon \Lambda ^ { + } \times \Lambda ^ { + } \rightarrow \odot _ { 0 } ^ { 2 } \Lambda ^ { + }$$
denotes the symmetric trace-free product, we therefore have
$$\left ( \nabla _ { e } \omega \right ) \circledast \left ( \nabla ^ { e } \omega \right ) = 2 | \alpha | ^ { 2 } \varphi \circledast \bar { \varphi } = - \frac { 1 } { 4 } | \nabla \omega | ^ { 2 } \omega \circledast \omega$$
and we thus deduce that
$$\text{and we thus deduce that} \\ \langle W ^ { + }, \nabla ^ { * } \nabla ( \omega \otimes \omega ) \rangle & \ = \ 2 W ^ { + } ( \omega, \nabla ^ { * } \nabla \omega ) - 2 W ^ { + } ( \nabla _ { e } \omega, \nabla ^ { e } \omega ) \\ & \ = \ 2 W ^ { + } ( \omega, \nabla ^ { * } \nabla \omega ) + \frac { 1 } { 2 } | \nabla \omega | ^ { 2 } W ^ { + } ( \omega, \omega ) \\ & \ = \ 2 W ^ { + } ( \omega, 2 W ^ { + } ( \omega ) - \frac { 8 } { 3 } \omega ) + \left [ W ^ { + } ( \omega, \omega ) - \frac { 8 } { 3 } \right ] W ^ { + } ( \omega, \omega ) \\ & \ = \ - \frac { 2 } { 3 } s W ^ { + } ( \omega, \omega ) + 4 | W ^ { + } ( \omega ) | ^ { 2 } + \left [ W ^ { + } ( \omega, \omega ) - \frac { 8 } { 3 } \right ] W ^ { + } ( \omega, \omega ) \\ & \ = \ [ W ^ { + } ( \omega, \omega ) ] ^ { 2 } + 4 | W ^ { + } ( \omega ) | ^ { 2 } - s W ^ { + } ( \omega, \omega ) \\ & \ e$$
where we have used the Weitzenb¨ck formula o
$$0 = \nabla ^ { * } \nabla \omega - 2 W ^ { + } ( \omega ) + \frac { s } { 3 } \omega$$
for the harmonic self-dual 2-form ω , as well as the associated key identity
$$\frac { 1 } { 2 } | \nabla \omega | ^ { 2 } = W ^ { + } ( \omega, \omega ) - \frac { s } { 3 }$$
resulting from the fact that | ω | 2 ≡ 2.
Plugging this into our Weitzenb¨ck formula (2) and integrating by parts, o we thus see that whenever a compact almost-K¨hler 4-manifold ( a M,g,ω ) satisfies δ fW ( + ) = 0, we then automatically have
$$\text{satisfies} \, \delta ( f W ^ { + } ) = 0, \, \text{we then automatically have} \\ 0 \, & = \, \int _ { M } \left \langle \left ( \nabla ^ { * } \nabla f W ^ { + } + \frac { s } { 2 } f W ^ { + } - 6 f W ^ { + } \circ W ^ { + } + 2 f | W ^ { + } | ^ { 2 } I \right ), \omega \otimes \omega \right \rangle d \mu \\ \, & = \, \int _ { M } \left [ \langle W ^ { + }, \nabla ^ { * } \nabla ( \omega \otimes \omega ) \rangle + \frac { s } { 2 } W ^ { + } ( \omega, \omega ) - 6 | W ^ { + } ( \omega ) | ^ { 2 } + 2 | W ^ { + } | ^ { 2 } | \omega | ^ { 2 } \right ] f \, d \mu \\ \, & = \, \int _ { M } \left [ \left ( [ W ^ { + } ( \omega, \omega ) ] ^ { 2 } + 4 | W ^ { + } ( \omega ) | ^ { 2 } - s W ^ { + } ( \omega, \omega ) \right ) \\ & \quad + \frac { s } { 2 } W ^ { + } ( \omega, \omega ) - 6 | W ^ { + } ( \omega ) | ^ { 2 } + 4 | W ^ { + } | ^ { 2 } \right ] f \, d \mu \\ \, & = \, \int _ { M } \left [ [ W ^ { + } ( \omega, \omega ) ] ^ { 2 } - \frac { s } { 2 } W ^ { + } ( \omega, \omega ) - 2 | W ^ { + } ( \omega ) | ^ { 2 } + 4 | W ^ { + } | ^ { 2 } \right ] f \, d \mu \,. \\ \text{In other words, letting} \, W ^ { + } ( \omega ) ^ { + } \text{ denote the component of} \, W ^ { + } ( \omega ) \text{ perpendicular} \,. \\. \, \quad \, \, \,$$
In other words, letting W + ( ω ) ⊥ denote the component of W + ( ω ) perpendicular to ω , any compact almost-K¨hler manifold ( a M,g,ω ) with δ fW ( + ) = 0 satisfies the identity
$$\int _ { M } s W ^ { + } ( \omega, \omega ) f \ d \mu = 8 \int _ { M } \left ( | W ^ { + } | ^ { 2 } - \frac { 1 } { 2 } | W ^ { + } ( \omega ) ^ { \perp } | ^ { 2 } \right ) f \ d \mu \. \quad ( 3 )$$
To proceed further, we will now need another algebraic observation:
Lemma 2. Any 4 -dimensional almost-K¨hler manifold satisfies a
$$| W ^ { + } | ^ { 2 } - \frac { 1 } { 2 } | W ^ { + } ( \omega ) ^ { \perp } | ^ { 2 } & \geq \frac { 3 } { 8 } \left [ W ^ { + } ( \omega, \omega ) \right ] ^ { 2 } \\ \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \colon$$
at every point, and equality can only hold at points where W + ( ω ) ⊥ = 0 .
Proof. If A = [ A jk ] is any symmetric trace-free 3 × 3 matrix, the fact that A 33 = -( A 11 + A 22 ) implies that
$$\sum _ { j k } A _ { j k } ^ { 2 } \geq 2 A _ { 2 1 } ^ { 2 } + A _ { 1 1 } ^ { 2 } + A _ { 2 2 } ^ { 2 } + A _ { 3 3 } ^ { 2 } = 2 A _ { 2 1 } ^ { 2 } + \frac { 3 } { 2 } A _ { 1 1 } ^ { 2 } + 2 ( \frac { A _ { 1 1 } } { 2 } + A _ { 2 2 } ) ^ { 2 }$$
and we therefore conclude that
$$| A | ^ { 2 } \geq 2 A _ { 2 1 } ^ { 2 } + \frac { 3 } { 2 } A _ { 1 1 } ^ { 2 }.$$
If we now let A represent W + : Λ + → Λ + with respect to an orthogonal basis ε , ε 1 2 , ε 3 for Λ + such that ω = √ 2 ε 1 and W + ( ω ) ⊥ ∝ ε 2 , this inequality becomes which not only proves the desired inequality, but shows that it is actually strict whenever ω is not an eigenvector of W + .
$$| W ^ { + } | ^ { 2 } & \geq | W ^ { + } ( \omega ) ^ { \perp } | ^ { 2 } + \frac { 3 } { 8 } \left [ W ^ { + } ( \omega, \omega ) \right ] ^ { 2 } \\ \text{nly proves the desired inequality, but shows that i}$$
Combining (3) with Lemma 2 now yields the global inequality
$$\int _ { M } s W ^ { + } ( \omega, \omega ) f \ d \mu \geq 3 \int _ { M } \left [ W ^ { + } ( \omega, \omega ) \right ] ^ { 2 } f \ d \mu, \quad \quad ( 4 ) \\ \dots \quad \cdot \dots \dots \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \end{array}$$
with equality only if W + ( ω ) ⊥ ≡ 0. It thus follows that
$$0 \geq \int _ { M } W ^ { + } ( \omega, \omega ) \left ( W ^ { + } ( \omega, \omega ) - \frac { s } { 3 } \right ) \ f \ d \mu.$$
However, since 1 2 |∇ ω | 2 = W + ( ω, ω ) -s 3 for any almost-K¨hler 4-manifold, a this proves the following:
Proposition 1. Let ( M ,g,ω 4 ) be a compact almost-K¨hler manifold, and a suppose that, for some positive function f , the conformally related metric h = f 2 g has harmonic self-dual Weyl curvature. Then ( M,g,ω ) satisfies the inequality
$$0 \geq \int _ { M } W ^ { + } ( \omega, \omega ) | \nabla \omega | ^ { 2 } f \ d \mu.$$
This has an interesting immediate consequence:
Proposition 2. Let ( M ,g,ω 4 ) be a compact connected almost-K¨hler mana ifold with W + ( ω, ω ) ≥ 0 , and suppose that the conformally related metric h = f 2 g satisfies δW + = 0 . Then either g is a K¨ ahler metric with scalar curvature s = c/f for some constant c > 0 , or else g satisfies W + ≡ 0 , and so is an anti-self-dual metric.
Proof. Recall that f > 0 by convention, and that W + ( ω, ω ) ≥ 0 by assumption. Thus (5) implies that
$$\int _ { M } W ^ { + } ( \omega, \omega ) | \nabla \omega | ^ { 2 } f \ d \mu = 0,$$
/negationslash so that ∇ ω = 0 wherever W + ( ω, ω ) = 0. If U ⊂ M is the open subset where W + ( ω, ω ) = 0, the restriction of g to U is therefore K¨hler. a On the other hand, by hypothesis, g satisfies δ fW ( + ) = 0. However, for any K¨hler mana ifold of real dimension 4, W + is the trace-free part of ( s/ 4) ω ⊗ ω , where the scalar curvature s satisfies s = 3 W + ( ω, ω ). It follows that d fW [ + ( ω, ω )] = 0 on U . By continuity, we therefore have d fW [ + ( ω, ω )] = 0 on the closure U of U , too. On the other hand, fW + ( ω, ω ) ≡ 0 on M -U , so we also have d fW [ + ( ω, ω )] = 0 on the open set M -U . Hence d fW [ + ( ω, ω )] = 0 on all of M . Since M is connected, it follows that fW + ( ω, ω ) = c/ 3 for some non-negative constant c ≥ 0. If c > 0, M = U , and ( M,g ) is a K¨hler a manifold, with s = 3 W + ( ω, ω ) = c/f . On the other hand, if c = 0, we have W + ( ω, ω ) ≡ 0, and therefore have equality in (4). However, this implies that W + ( ω ) ⊥ ≡ 0, and (3) therefore implies that W + ≡ 0, as claimed.
/negationslash
As a special case, we therefore obtain the following key result:
Theorem 1. Let ( M,h ) be a compact oriented Riemannian 4 -manifold with δW + = 0 . If the conformal class [ h ] is of positive symplectic type, then h = s -2 g for a unique K¨hler metric a g of scalar curvature s > 0 . Conversely, if g is any K¨hler metric of positive scalar curvature, the conformally related a metric h = s -2 g satisfies δW + = 0 .
Proof. To say that [ h ] is of positive symplectic type means that there is a self-dual harmonic 2-form ω on ( M,h ) such that W + ( ω, ω ) > 0 at every point of M . Rescaling h to make ω have constant norm √ 2 results in an almost-K¨hler metric ˆ such that a g h = ˆ ˆ for some positive function f 2 g ˆ . f If h satisfies δW + = 0, the almost-K¨hler metric ˆ then satisfies a g W + ( ω, ω ) > 0 and δ fW ( ˆ + ) = 0, so Proposition 2 then tells us that ˆ is actually K¨hler, g a with scalar curvature ˆ = s c/f ˆ for some positive constant c . In particular, M admits a K¨hler metric with positive scalar curvature, and Yau's vanishing a theorem [35] for the geometric genus therefore implies that b + ( M ) = 1. Thus the choice of ω is in fact unique up to an overall multiplicative constant, and the choice of ˆ is therefore determined up to constant rescalings. g But if, for a positive constant a , we replace ˆ with g g = a g 2 ˆ, we must also replace ˆ f with f = a -1 ˆ ; f and note that the scalar curvature of g is then s = a -2 ˆ. s Since ˆ = f cs ˆ -1 , we then have h = ˆ ˆ = f 2 g c 2 ˆ s -2 ˆ = g c 2 ( a s 2 ) -2 ( a -2 g ) =
( ca -3 ) 2 s -2 g . This shows that setting a = 3 √ c results in a K¨hler metric a g such that h = s -2 g , and moreover shows that this choice yields the only K¨ ahler metric with this property.
Conversely [12], if g is a K¨ ahler metric with s > 0, s -1 W + is parallel, so that, in particular, we have δ s ( -1 W + ) = 0. Thus h = s -2 g satisfies δW + = 0, as promised.
Theorem A is now a straightforward consequence. Indeed, since the second Bianchi identity implies that any Einstein metric on an oriented 4manifold satisfies δW + = 0, Theorem 2 tells us that every Einstein metric h of positive symplectic type must be conformally K¨hler. a Moreover, since the conformal class [ h ] contains a representative g with s > 0, the constant scalar curvature 4 λ of h must [33] be positive, too. Theorem A therefore follows from the known classification [23] of conformally K¨hler, Einstein a metrics on compact 4-manifolds.
## 3 Almost-K¨ ahler Manifolds Revisited
As an added bonus, the results of § 2 also have interesting consequences in the narrower context of almost-K¨hler geometry; for related work, see [6]. a Our main such application is the following:
Theorem 2. Let ( M,g,ω ) be a compact almost-K¨hler a 4 -manifold with nonnegative scalar curvature and harmonic self-dual Weyl tensor:
$$s \geq 0, \quad \delta W ^ { + } = 0.$$
Then ( M,g,ω ) is a constant-scalar-curvature K¨hler manifold. a
Proof. For any almost-K¨hler manifold, a
$$W ^ { + } ( \omega, \omega ) = \frac { s } { 3 } + \frac { 1 } { 2 } | \nabla \omega | ^ { 2 }$$
so that the hypothesis s ≥ 0 implies W + ( ω, ω ) ≥ 0. Proposition 2, with f = 1, therefore tells us that ( M,g ) is K¨ ahler, with scalar curvature s = c/f = c for some positive constant c , or else that W + ≡ 0. In the latter case, we then have 0 = 3 W + ( ω, ω ) ≥ s ≥ 0, so s ≡ 0, and hence |∇ ω | 2 = 2 W + ( ω, ω ) -2 s/ 3 = 0. Thus ( M,g ) is constant-scalar-curvature K¨ ahler, even in the exceptional case.
Conversely, any constant-scalar-curvature K¨ ahler manifold of real dimension 4 satisfies δW + = 0, independent of the sign of s . While the study of 'cscK' (constant-scalar-curvature K¨hler) metrics on compact complex a surfaces is an active area of ongoing research, many existence results are already available [3, 14, 17, 19, 29, 31]. However, we should emphasize that the non-negativity of the scalar curvature plays a crucial role in Theorem 2. For example, there exist many compact almost-K¨hler manifolds with a W + ≡ 0 which are not K¨hler. a Indeed, such examples can be obtained [16] by deforming scalar-flat K¨hler metrics through anti-self-dual confora mal classes, and then conformally rescaling to make | ω | ≡ √ 2. Examples of this type automatically have s ≤ 0, with s < 0 on an open dense subset.
Since any Einstein 4-manifold satisfies δW + = 0, Theorem 2 provides a new proof of Sekigawa's breakthrough result [30] on the Goldberg conjecture:
Corollary 2 (Sekigawa) . Every compact almost-K¨hler Einstein a 4 -manifold with non-negative Einstein constant is K¨hler-Einstein. a
This fact provided a useful guidepost en route to the present results.
The proof of Theorem 2 in fact still works if we merely impose the ostensibly weaker hypothesis that s + tW + ( ω, ω ) ≥ 0 for some constant t ≥ 0, since any such hypothesis will imply that W + ( ω, ω ) ≥ 0, with s = 0 if equality holds. In particular, one reaches exactly the same conclusion if we merely assume that the so-called star-scalar curvature
$$s ^ { * } = s + | \nabla \omega | ^ { 2 } = \frac { s } { 3 } + 2 W ^ { + } ( \omega, \omega )$$
is non-negative:
Proposition 3. Let ( M,g,ω ) be a compact almost-K¨hler a 4 -manifold with non-negative star-scalar curvature and harmonic self-dual Weyl tensor:
$$s ^ { * } \geq 0, \quad \delta W ^ { + } = 0.$$
Then ( M,g,ω ) is a constant-scalar-curvature K¨hler manifold. a
Kirchberg [18] has elsewhere investigated almost-K¨hler 4-manifolds with a harmonic Weyl tensor and positive star-scalar curvature. Since the hypothesis δW = 0 is equivalent to δW + = δW -= 0 , and is therefore stronger than the hypothesis δW + = 0 of Proposition 3, we can recover many of Kirchberg's results from our own. For example, we can deduce the following clarification of [18, Corollary 3.13]:
Corollary 3. Let ( M,g,ω ) be a compact almost-K¨hler a 4 -manifold with non-negative scalar curvature and harmonic Weyl tensor:
$$s \geq 0, \quad \delta W ^ { + } = \delta W ^ { - } = 0.$$
Then ( M ,g,J 4 ) is either a K¨hler-Einstein manifold with a λ ≥ 0 , or else is locally symmetric, with universal cover ( ˜ M,g ˜) isometric to the Riemannian product of two constant-curvature surfaces, where one factor is a 2 -sphere.
Proof. Theorem 2 tells us that ( M,g,J ) is K¨ ahler and has constant scalar curvature. But since the entire Weyl tensor is actually assumed to be harmonic, the second Bianchi identity also tells us that
$$\nabla _ { [ c } r _ { d ] b } = \nabla _ { a } W ^ { a } _ { \, b c d } + \frac { 1 } { 6 } g _ { b [ c } \nabla _ { d ] } s = 0,$$
so that the covariant derivative ∇ r of the Ricci tensor must therefore be completely symmetric. Decomposing ⊗ 3 Λ 1 C into ⊗ 3 (Λ 1 0 , ⊕ Λ 0 1 , ), we thus have
$$\nabla _ { \kappa } r _ { \mu \bar { \nu } } = \nabla _ { \bar { \nu } } r _ { \mu \kappa } = 0 \quad \text{ and } \quad \nabla _ { \bar { \kappa } } r _ { \mu \bar { \nu } } = \nabla _ { \mu } r _ { \bar { \kappa } \bar { \nu } } = 0. \\ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \text{...} \ \end{cases}$$
The Ricci tensor of our K¨hler manifold is therefore parallel, and the prima itive part ˚ ρ ∈ Λ -of its Ricci form must therefore be parallel, too. If ˚= 0, ρ ( M,g,J ) is K¨ ahler-Einstein. Otherwise, the holonomy of ( M,g ) fixes both ω and ˚ , and so must be contained in ρ U (1) × U (1) ⊂ U (2) ⊂ SO (4). In the latter case, the de Rham splitting theorem [4] then implies that the universal cover ( ˜ M,g ˜) of ( M,g ) is a Riemannian product ( M ,g 1 1 ) × ( M ,g 2 2 ) of two complete, constant-curvature Riemann surfaces; and if g is not Einstein, and therefore not flat, the assumption that s ≥ 0 then forces at least one factor ( M ,g j j ) to have positive Gauss curvature.
The Narasimhan-Seshadri theorem [25] provides a complete existence theory for the non-Einstein metrics of Corollary 3. Indeed, a compact complex manifold ( M ,J 4 ) admits such a locally-product K¨hler metric iff it is a a geometrically ruled complex surfaces that arises as the projectivization of a polystable rank-2 holomorphic vector bundles over a compact complex curve. For related results, see [2, 9, 21].
Acknowledgments. The author would like to thank Tedi Draghici for subsequently pointing out some of his own related work, and the anonymous referee for suggesting ways to streamline and clarify the exposition. This work was supported in part by NSF grant DMS-1205953.
## References
- [1] M. T. Anderson , The L 2 structure of moduli spaces of Einstein metrics on 4 -manifolds , Geom. Funct. Anal., 2 (1992), pp. 29-89.
- [2] V. Apostolov and C. Tønnesen-Friedman , A remark on K¨ ahler metrics of constant scalar curvature on ruled complex surfaces , Bull. London Math. Soc., 38 (2006), pp. 494-500.
- [3] C. Arezzo and F. Pacard , Blowing up and desingularizing constant scalar curvature K¨hler manifolds a , Acta Math., 196 (2006), pp. 179228.
- [4] A. L. Besse , Einstein manifolds , vol. 10 of Ergebnisse der Mathematik und ihrer Grenzgebiete (3), Springer-Verlag, Berlin, 1987.
- [5] G. Besson, G. Courtois, and S. Gallot , Entropies et rigidit´s e des espaces localement sym´ etriques de courbure strictement n´ egative , Geom. and Func. An., 5 (1995), pp. 731-799.
- [6] D. E. Blair and T. Draghici , Remarks on Weyl curvature in almost K¨ ahler geometry , Bull. Math. Soc. Sci. Math. Roumanie (N.S.), 52(100) (2009), pp. 241-249.
- [7] C. B¨hm o , Inhomogeneous Einstein metrics on low-dimensional spheres and other low-dimensional spaces , Invent. Math., 134 (1998), pp. 145176.
- [8] C. P. Boyer, K. Galicki, and J. Koll´r a , Einstein metrics on spheres , Ann. of Math. (2), 162 (2005), pp. 557-580.
- [9] D. Burns and P. De Bartolomeis , Stability of vector bundles and extremal metrics , Invent. Math., 92 (1988), pp. 403-407.
- [10] X. X. Chen, C. LeBrun, and B. Weber , On conformally K¨hler, a Einstein manifolds , J. Amer. Math. Soc., 21 (2008), pp. 1137-1168.
- [11] M. Demazure , Surfaces de del Pezzo, II, III, IV, V , in S´minaire sur e les Singularit´s des Surfaces, vol. 777 of Lecture Notes in Mathematics, e Berlin, 1980, Springer, pp. 21-69.
- [12] A. Derdzi´ski n , Self-dual K¨hler manifolds and Einstein manifolds of a dimension four , Compositio Math., 49 (1983), pp. 405-433.
| [13] | D. M. DeTurck and J. L. Kazdan , Some regularity theorems in Rie- mannian geometry , Ann. Sci. ´ Ecole Norm. Sup. (4), 14 (1981), pp. 249- 260. |
|--------|----------------------------------------------------------------------------------------------------------------------------------------------------------|
| [14] | S. K. Donaldson , Scalar curvature and stability of toric varieties , J. Differential Geom., 62 (2002), pp. 289-349. |
| [15] | M. J. Gursky , Four-manifolds with δW + = 0 and Einstein constants of the sphere , Math. Ann., 318 (2000), pp. 417-431. |
| [16] | I. Kim , Almost-K¨hler a Anti-Self-Dual Metrics , PhD thesis, State Uni- versity of New York at Stony Brook, 2014. |
| [17] | J. Kim, C. LeBrun, and M. Pontecorvo , Scalar-flat K¨hler a sur- faces of all genera , J. Reine Angew. Math., 486 (1997), pp. 69-95. |
| [18] | K.-D. Kirchberg , Some integrability conditions for almost K¨hler a manifolds , J. Geom. Phys., 49 (2004), pp. 101-115. |
| [19] | C. LeBrun , Scalar-flat K¨hler a metrics on blown-up ruled surfaces , J. Reine Angew. Math., 420 (1991), pp. 161-177. |
| [20] | , Einstein metrics and Mostow rigidity , Math. Res. Lett., 2 (1995), pp. 1-8. |
| [21] | , Polarized 4-manifolds, extremal K¨hler a metrics, and Seiberg- Witten theory , Math. Res. Lett., 2 (1995), pp. 653-662. |
| [22] | , Yamabe constants and the perturbed Seiberg-Witten equations , Comm. An. Geom., 5 (1997), pp. 535-553. |
| [23] | , On Einstein, Hermitian 4-manifolds , J. Differential Geom., 90 (2012), pp. 277-302. |
| [24] | Y. I. Manin , Cubic Forms: Algebra, Geometry, Arithmetic , North- Holland Publishing Co., Amsterdam, 1974. Translated from the Russian by M. Hazewinkel. |
| [25] | M. S. Narasimhan and C. S. Seshadri , Stable and unitary vector bundles on a compact Riemann surface , Ann. of Math. (2), 82 (1965), pp. 540-567. |
| [26] | Y. Odaka, C. Spotti, and S. Sun , Compact moduli spaces of Del Pezzo surfaces and K¨hler-Einstein a metrics . e-print arXiv:1210.0858 [math.DG], 2012. |
- [27] D. Page , A compact rotating gravitational instanton , Phys. Lett., 79B (1979), pp. 235-238.
- [28] R. Penrose and W. Rindler , Spinors and space-time. Vol. 2 , Cambridge University Press, Cambridge, 1986. Spinor and twistor methods in space-time geometry.
- [29] Y. Rollin and M. Singer , Construction of K¨hler surfaces with cona stant scalar curvature , J. Eur. Math. Soc. (JEMS), 11 (2009), pp. 979997.
- [30] K. Sekigawa , On some 4 -dimensional compact Einstein almost K¨hler a manifolds , Math. Ann., 271 (1985), pp. 333-337.
- [31] Y. Shu , Compact complex surfaces and constant scalar curvature K¨ ahler metrics , Geom. Dedicata, 138 (2009), pp. 151-172.
- [32] G. Tian , On Calabi's conjecture for complex surfaces with positive first Chern class , Invent. Math., 101 (1990), pp. 101-172.
- [33] N. Trudinger , Remarks concerning the conformal deformation of metrics to constant scalar curvature , Ann. Scuola Norm. Sup. Pisa, 22 (1968), pp. 265-274.
- [34] C. van Coevering , Sasaki-Einstein 5-manifolds associated to toric 3-Sasaki manifolds , New York J. Math., 18 (2012), pp. 555-608.
- [35] S. T. Yau , On the curvature of compact Hermitian manifolds , Invent. Math., 25 (1974), pp. 213-239.
Department of Mathematics, State University of New York, Stony Brook, NY 11794-3651 USA
e-mail: [email protected]
Keywords: Einstein metric, Del Pezzo surface, Weyl curvature, moduli space, harmonic form, K¨hler, almost-K¨hler, symplectic. a a
MSC classification: 53C25 (Primary), 14J26, 32J15, 53C55, 53D05.
|
10.1007/s10455-015-9458-0
|
[
"Claude LeBrun"
] |
2014-08-05T19:53:28+00:00
|
2015-04-28T10:01:54+00:00
|
[
"math.DG",
"math.SG",
"53C25 (primary), 14J45, 53A30, 53C57 (secondary)"
] |
Einstein Metrics, Harmonic Forms, and Symplectic Four-Manifolds
|
If $M$ is the underlying smooth oriented $4$-manifold of a Del Pezzo surface,
we consider the set of Riemannian metrics $h$ on $M$ such that $W^+(\omega ,
\omega )> 0$, where $W^+$ is the self-dual Weyl curvature of $h$, and $\omega$
is a non-trivial self-dual harmonic $2$-form on $(M,h)$. While this open region
in the space of Riemannian metrics contains all the known Einstein metrics on
$M$, we show that it contains no others. Consequently, it contributes exactly
one connected component to the moduli space of Einstein metrics on $M$.
|
1408.1079v1
|
## COLOURING ISONEMAL FABRICS WITH MORE THAN TWO COLOURS BY THIN STRIPING
## ROBERT S. D. THOMAS
Abstract. Perfect colouring of isonemal fabrics by thin striping of warp and weft with more than two colours is examined. Examples of thin striping in all possible species with no redundancy and with redundant cells arranged as twills are given. Colouring woven flat tori is discussed.
## 1. Background
This paper takes up the subject of colouring fabrics perfectly with more than two colours introduced in [12] and extends it to thin striping, that is, to colouring the strands in both directions with a finite number of colours cyclically. The first three sections of [12] and the standard references [1, 4] are relevant and will not be repeated here.
At a referee's request three paragraphs will give some idea of terms and notations to be used. A fabric is a two-fold weave that does not fall apart having warps (vertical and dark in colour) crossing wefts (horizontal and pale) in square cells , in each of which one strand (warp or weft) is uppermost or visible in an illustration of the obverse side of the weave. The other side, the reverse , is illustrated (if at all) as reflected in a mirror set up behind/below the plane of the fabric to preserve symmetries. Isonemal fabrics have symmetry group G 1 transitive on the strands. The pattern of over and under is periodic along a strand, this period being called order , distinct from the two-dimensional period , the area of the smallest period parallelograms. A group's period parallelogram with corners that are translates of one another is called a lattice unit . Two catalogues of fabrics have been published [2, 3] giving catalogue numbers for isonemal fabrics of small order in the form a b c --, where a is order, b is a unique representation of the over-and-under pattern, and c is a serial number.
Most isonemal fabrics were divided into five genera by Gr¨ unbaum and Shephard [2] and 39 species in [7, 8, 9], based on Roth's [5] description of 39 types of symmetry-group pairs 〈 G ,H 1 1 〉 , where H 1 is the subgroup of G 1 consisting of elements not including reflection in the plane of the fabric,
2000 Mathematics Subject Classification. primary 52C20; secondary 05B45, 51M20. Key words and phrases. colour, fabric, isonemal, striping, weaving.
Work on this material was done at home and at Wolfson College, Oxford (2005-2006). I am grateful for Oxford hospitality and Winnipeg patience.
Figure 1. Designs of fabrics that could be patterns of redundancy with symmetries marked. a. The 2/1 twill. b. The 4/1 twill. c. 4-1-1.
represented τ . (The few fabrics outside a genus or species are called exceptional .) G 1 consists of elements 〈 s, t 〉 , where s is a two-dimensional transformation (translation, reflection, glide-reflection, half-turn ( ♦ ), or quarter-turn ( /square )) and t is either the identity e or the reflection τ . H 1 is the subgroup consisting of elements 〈 s, e 〉 , and G 2 is the group that is the projection of G 1 defined by 〈 s, τ 〉 ↦→ 〈 s, e 〉 and 〈 s, e 〉 ↦→ 〈 s, e 〉 to avoid the issue of reflections in the plane of the fabric. Axes of reflection pass diagonally across cells (Fig. 1); axes of glide-reflection are not so restricted (Fig. 1, Fig. 2a). When they do so (Fig. 3a), they are said to be in mirror position . Figure 1ab illustrates the simplest isonemal fabrics, twills , in which each strand's pattern is the adjacent strand's shifted by one; a twill's over-and-under pattern is represented a /a /b /b / . . . 1 2 1 2 until its order, the sum of a 1 + a 2 + . . . , is used up.
The standard colouring (warp dark, weft pale) is a perfect colouring in the sense that each symmetry in G 1 either preserves or reverses the colour of all cells. A non-standard colouring is called perfect if each symmetry permutes the colours of all cells of each colour. When the two strands passing through a cell are of the same colour, the cell is called redundant in that colouring because which strand is uppermost has no effect on the cell's appearance.
Thin stripes overlap in individual cells-all cells-of the fabric. Thin stripes of the same colour overlap in redundant cells, in this paper arranged as twills. The natural unit of measurement in the fabric is the side of a cell, in which the diagonal of the square cells have the length √ 2, which will be abbreviated δ . It is convenient to represent δ/ 2 by β . In [12] the following results about thin striping were shown.
Theorem 1. Some non-exceptional fabrics in all species can be perfectly coloured by the assignment of different colours to all strands.
Theorem 2. If no colour is common to warps and wefts in the assignment of a finite number of colours to the striping of an isonemal fabric, then the necessary conditions: the same number of colours are assigned to warps as to
Figure 2. a. Order-30 species-1 o example of Figure 10b of [7]. b. Three-colouring by thin striping with redundant cells strictly between the cells through which the axes run.
Figure 3. a. Design of Roth's [5] example 12-183-1 of species 1 m . b. Three-colouring by thin striping with redundant cells along the axes.
wefts, and the vertical and horizontal sequences of the colours of the stripes are periodic with the same period, are sufficient constraints on the colours to allow perfect colouring by thin striping.
## 2. No Redundancy
Sections 5 and 6 of [12] were written on the erroneous assumption, conflicting with earlier sections of that paper, that there will always be redundant cells in a colouring. 1 That is of course only the case when there are perpendicular strands of the same colour. This section, which should have been in [12],
1 This apparently absurd event is an inexcusable result of sections 5 and 6 having been written when two-colourings were generalized to more than two colours with redundancy before the further generalization of earlier sections of the paper had been thought of.
Figure 4. Fabric of species 11 e in Fig. 15b thickly 4coloured two ways with no redundant cells.
.
Figure 5. Fabric 12-315-4 of species 30 in Fig. 12a thickly 4-coloured two ways with no redundant cells.
.
takes up the interesting possibility where there are no such strands and so no redundancy, there being stripes of two colours in each direction, a total of four colours - not too many to make an attractive pattern.
Lemma 3.1 of [12] gives the necessary and sufficient condition for an assignment of all different colours to a thick striping of a fabric to be perfect, namely preservation of the stripes by the symmetries, followed by the conditions on the symmetries to allow this. No mention is made of conditions on the striping, but striping almost always has to be done with some care, as was noted by Roth [6]. In particular, the striping of wefts dictates how the warps can be striped, as is illustrated with four colours in Fig. 4a and b. These diagrams show two ways in which the fabric of Fig. 15b can be thickly striped perfectly with two colours in each direction. In each case, the striping in one direction dictates the boundaries of the stripes in the
Figure 6. a. Fabric 10-55-2 of species 33 3 (Fig. 5a of [12]). b, c. Thickly 4-coloured two ways with no redundant cells.
.
other direction. What one cannot do is to stripe the fabric with its horizontal stripes as in Fig. 4a and its vertical stripes as in Fig. 4b. Not that this cannot be woven, but warp-weft-interchanging symmetries cannot take stripes to stripes. The same condition of Lemma 3.1 suffices for the same finite number of colours in each direction repeating cyclically with no redundancy. While this condition is a slight relaxation of the conditions for striping with two colours - and therefore unavoidably with redundancy, it still bans twills and twillins as in Section 3 of [11]. That eliminates species 1-10, 12, 14, 16, 18, 20, 23, 24, 26, 28, 31, 32, 34, 36, and 39 leaving 11, 13, 15, 17, 19, 21, 22, 25, 27 , 29, 30, 33, 35, 37, and 38. e Any fabric of these species should be perfectly thickly colorable with any number of colours greater than 1 in each direction. Only two colours will be explored briefly because they are likely to be more attractive than more colours. Examples are in Figures 5 and 6 as well as 4. These are all from species that cannot be thickly 4-coloured with twilly redundancy. That they cannot be thickly 4-coloured with twilly redundancy led to the erroneous remark in [12] (which assumed, as explained in footnote 1, that redundancy was inevitable) that fabrics of species 38 (p. 50) and 33-39 (p. 54) cannot be thickly coloured with four colours at all. Theorems 3 and 4 clarify this for thin striping.
Turning to thin striping, one sees that none of the thick-striping restrictions on species is relevant. And any striping in both directions with no redundancy will do; what can go wrong with thick stripes mentioned in the previous paragraph cannot occur with thin stripes. Unfortunately, since what can be coloured that way is just longer or shorter strips, no interesting motif of a single colour can emerge, and two-coloured motifs seem not to be of much interest. 2 As examples I give a thinly striped version of the
2 Not of sufficient interest to be noticed? This may be a matter of psychology or eyesight. When I look at Figure 39b I recognize immediately a red and blue braid motif but must search for the green and yellow one that perfect colouring ensures. A possible
Figure 7. Thin 4-colourings with no redundant cells. a. Fabric of species 30 thickly coloured in Fig. 5. b. 4-1-1 reverse.
.
Figure 8. Fabrics of species 11 e and 33 , thickly coloured 3 in Figg. 4 and 6, thinly 4-coloured with no redundant cells.
.
three fabrics used to illustrate thick striping (there not being two essentially different thin stripings) in Figures 7 and 8.
## 3. Preliminaries
The remainder of this paper determines the constraints on the fabric designs to allow perfect colourings by thin striping with twilly redundancy and how to arrange them. In [12] the following result is Lemma 4.4.
exception to this claim is the smallest possible two-colour motif, just two cells, illustrated in Figure 7b.
Figure 9. a. Roth's [5] example of species 34, 10-107-1 with lattice unit and centres of G 1 marked (Fig. 8a of [9]). b. Five-colouring by thin striping.
Lemma 1. The assignment of a finite number of colours to stripes of an isonemal fabric can be a perfect colouring only if the colours of the warps are the colours of the wefts or no colour is shared.
The order of the colours does not need to be the same vertically as horizontally. For example, Figure 9 shows a five-colouring otherwise of Roth's [5] example of species 34, catalogue number 10-107-1. This topic will not be pursued here. The centres of the cross motifs are redundant cells at which strands of the same colour cross. In order for the colouring to be a perfect colouring, Roth's Colouring Theorem [6] shows that the symmetry group G 1 of the fabric must leave the redundant cells/blocks invariant as a whole and each element of the group must permute them, if at all, in a consistent way. The group G 2 of the fabric must be a subgroup of the group G 2 of the pattern of redundant cells, as in Figures 1a and b. For a specific colouring of the wefts, say, as illustrated in Figure 9b, the positions of the redundant cells must be chosen in accordance with the consistency requirement. This choice determines the colouring of the warps.
From now on, warps and wefts will be coloured with the same finite set of c colours. For thin striping, the redundant cells may be arranged as ( c -1) / 1 twills, and for c ≥ 4 as various other isonemal fabrics, 6-1-1, 8-1-1, 8-1-2, and including in particular the square satins beginning with 5-1-1. Fabrics will have order five or more, and redundant cells will be arranged as twills (Figg. 1a and b).
There is reason to consider mostly isomenal fabrics with order more than four because those of order four and less are few and mostly exceptional, but in order to consider colourings with three colours, one needs - there being no choice - to begin with the 2/1 twill (Figure 1a) as a redundancy pattern. Mainly because of their aesthetic appeal, but also for practical reasons, I intend to use three colours to represent what can be done with a twilly pattern of redundant cells for any odd number of colours, four colours for a twilly redundancy pattern for any even number of colours congruent to 0 mod 4, and six colours for a twilly redundancy pattern for even numbers
of colours congruent to 2 mod 4. Also, because side-reversing symmetries (mirrors and side-reversing glide-reflections, half-turns, and quarter-turns) of a fabric relate its opposite sides and one sees fabrics one side at a time, side-reversing symmetries have less appeal in striped fabrics. Emphasis will be on fabrics with side-preserving symmetries just because their symmetries are visible when the fabric is striped. 3
Theorem 3. For three colours, perfect colouring of non-exceptional fabrics can be done only with a twilly arrangement of redundant cells.
Proof. The same colours must be used for warp and weft because three is odd. The only isonemal design to serve as a redundancy pattern with order three and a single dark cell per order length is the 2/1 twill. /square
Here three strands are not standing proxy, as they will later, for larger odd numbers. Similarly, four strands in the next theorem are not standing proxy for more. These two numbers of colours, like two, are exceptional.
Theorem 4. For four colours, perfect colouring of non-exceptional fabrics can be done either with no redundancy or with a twilly arrangement of redundant cells.
Proof. Two colours in each direction will colour with no redundancy. With redundancy, there is the 3/1 twill and the apparent exception, 4-1-1 (Figure 1c), which must be shown to be useless. This design is exceptional both in being of order four and in being a member of no genus.
It is impossible for the group G 2 of an isonemal fabric to be a subgroup of the group G 2 of 4-1-1 for several reasons. The glide-reflection axes of 4-1-1 are horizontal and vertical, and those of other isonemal fabrics are all at π/ 4 to the horizontal. The mirrors of 4-1-1 are at the right angle to be potentially useful, but they are 2 δ apart. Any selection from them (for genus I or II) will be an even number of δ s apart, and the isonemality constraint of [7] prevents such a selection from combining with a diagonal period of 4 δ (or multiple of 4 δ ). Finally, the centres of rotation, as the diagram indicates, fall only on alternate strand boundaries, preventing an isonemal fabric from falling into genus III, IV, or V. /square
This does not say that 4-1-1 cannot be the pattern of redundant cells for a thin striping, only that the only isonemal fabric for which it is useful is itself. Unfortunately, when a design that is suitable as a redundancy pattern is itself coloured, the result is just stripes. Compare Figure 7b, 4-1-1 4-coloured with no redundancy. All that changes with a different redundancy pattern is the order of the stripes. With 4-1-1 the order of the colours of the stripes is different in the two directions, but with a twill as redundancy pattern the order is the same in the two directions. It is
3 While the presence of side-reversing symmetry makes it uninteresting and therefore unnecessary to look at both sides of such coloured fabrics, the absence of all such symmetry makes the two sides of a fabric that lacks it patterns of independent interest.
obvious in twilly cases that no fabric with quarter-turn symmetry can have G 2 a subgroup of the 2/1 or 3/1 twill-or of any other ( c -1) / 1 twill doubled or not, since these twills fall into species with no quarter-turns in their G 2 s. (2/1 is of 28 o , 3/1 is of 26 , 6-3-1 is of 27 e o , and 8-3-1 is of subspecies 25 .) e Accordingly, no fabric of species 33-39 can be perfectly coloured thickly or thinly with the same three or four colours vertically and horizontally - or, with twilly redundancy, any number of colours greater than two. Call this the quarter-turn ban. On the other hand, some fabrics from all other species are perfectly 3-colourable by thin striping.
## 4. Three colours
The G 1 of the redundant twill 2/1 for thin striping with three colours, of subspecies 28 o and illustrated in Figure 1a, contains perpendicular glidereflections with axes not in mirror position and mirrors, allowing the possibility of fitting axes of species 1-32 among them. As the characteristic feature of these twills is diagonal lines of, say, dark cells, axes of the fabric to be coloured can be specified as through, between, or perpendicular to the 'dark lines' of the redundancy twill. As Figure 1 indicates, there is no shortage of mirrors and axes of glide-reflections perpendicular to the dark lines; they occur at every opportunity, a mirror and two axes not in mirror position through every cell. No ingenuity is required to fit axes in this direction, and little in the direction of the dark lines for various clumps of species.
For subspecies 1 , 2 o o , and 4 , successive glide-reflection axes not in mirror o position an odd multiple m ≥ 1 of 3 β apart can be placed along the glidereflection axes between the dark lines that have that spacing. An example of this treatment is Figure 2, where a species-1 o fabric is 3-coloured this way; the redundant cells run half-way between the fabric's axes of side-preserving glide-reflection, which are not in mirror position. Since the length of the lattice unit in δ units is 5 and the number of colours 3, the length of the lattice unit for fixed-colour translational symmetry is 15. Other ways of arranging the colouring of these and other subspecies (1 , e 2 , e and 4 , for e example) are possible; all that is being indicated here is that there are some fabrics of each species 1-32 that can be so coloured.
For species 1 m , 2 m , 3, 5 o , and 7 , alternate glide-reflection axes in mirror o position (1 m , 2 m , 3) or alternate mirrors (5 , 7 ) an odd multiple o o m ≥ 1 of 3 δ apart can be placed along the mirrors in the dark lines that have that spacing. Then the other axes will fall on intervening mirrors. Glidereflection axes can lie on these mirrors regardless of parity of glide. An example of this is Figure 3, where Roth's [5] example of subspecies 1 m is 3-coloured this way.
For species 5 e and 6, successive mirrors that can be an odd multiple m ≥ 1 of 3 δ apart can lie along mirrors in the dark lines that have that spacing. An example is shown thinly striped with 3 colours in Figure 10.
Figure 10. a. Design of 12-139-1 of species 6. b. Threecolouring by thin striping with redundant cells along the axes.
Figure 11. a. Design of order 30 and species 32 from Figure 17a of [8]. b. Three-colouring by thin striping with redundant cells along the mirrors of positive slope.
For species 8 o and 10, 28 o and 32, successive mirrors that can be an odd multiple m ≥ 1 of 3 β apart can be put on dark lines' mirrors. Intervening glide-reflection axes will lie on intervening axes parallel to the dark lines, and the fabric's perpendicular mirrors and glide-reflection axes, if any (28 o and 32), can be put on the perpendicular mirrors and axes respectively, which are close enough together to make this possible. An example of species 32 (with the spacing of 28 ) is shown in Figure 11. o The disappearance of the mirror symmetry and half of the half-turns makes such a pattern less attractive than one hopes for unless there is extraneous symmetry introduced, and there is hardly room here.
For species 8 , 9, 27 e o , 28 , 28 e n , 30, and 31, successive mirrors that can be an odd multiple m ≥ 1 of 3 δ apart can be put on dark lines' mirrors that are
Figure 12. a. Roth's [5] example 12-315-4 of species 30. b. Three-colouring by thin striping with redundant cells along the mirrors of negative slope.
so spaced. Intervening glide-reflection axes will then lie on intervening parallel mirrors, and then the fabric's perpendicular mirrors and glide-reflection axes (27 , o 28 , e 28 n , 30, 31) can be put on the perpendicular mirrors and axes respectively. An example with the same motif as 12-183-1 (Figure 3b) arranged in a more complex way is 12-315-4 of species 30 in Figure 12b. Another example is the species-31 fabric of Figure 35a, which produces a slightly complicated zig-zag when thinly striped with three colours, visible in Figure 35b if one considers the dark and pale versions of the three colours green, red, and blue to be merged to just the three colours green, red, and blue. This example will be referred to later.
The previous paragraph can be repeated with 6 δ in place of 3 δ for species 27 e and 29. An example is shown in Figures 13 and 14.
For species 11, 13, 15, 17, 19, 21, 22, and 25, successive axes in mirror position or mirrors that can be an odd multiple m ≥ 1 of 3 δ apart can be put on some dark lines' mirrors. Then the fabric's perpendicular axes in mirror position and/or mirrors can be put on the mirrors perpendicular to the dark lines. Examples are Roth's [5] example of species 11 o shown in Figure 15a, the example of species 11 e shown in Figure 15b, an example of species 21 in Figure 17, the example of species 22 from Figure 11a of [8] in Figure 19a, and an example of subspecies 25 o in Figure 20a, coloured respectively in Figures 16a, 16b, 18, 19b, and 20b.
For species 12 , 14 , 16, 18, 20, 23, 24, and 26, successive glide-reflection e e axes in mirror position that can be an odd multiple m ≥ 1 of 3 β apart can be put on the dark lines' mirrors. The fabric's perpendicular axes not in mirror position and perpendicular mirrors can be put on the axes and mirrors perpendicular to the dark lines. Examples are Roth's [5] example of species 12 in Figure 21a and coloured in Figure 21b (having mirrors in the
Figure 13. Order-48 fabric of species 29; the fabric of Figure 35a doubled.
Figure 14. Three-colouring of the design of Figure 13 by thin striping with redundant cells along the mirrors of negative slope.
Figure 15. Fabrics coloured in Figure 16. a. Roth's [5] example 12-111-2 of species 11 . o b. Example of species 11 e from Figure 5a of [8].
Figure 16. Three-colourings by thin striping with some axes of fabric marked. a. Example of Figure 15a. b. Example of Figure 15b.
pattern along what are axes in the design); Roth's [5] example of species 18 in Figure 22a and coloured in Figure 22b (symmetry except for halfturns, of which there are a lot, and translations disappears); an example of a subspecies-23 o fabric in Figure 23a and coloured in Figure 23b (symmetry gained as the side-preserving glide-reflections become mirrors too at the same time as the mirrors become visible glide-reflections); Roth's [5] example of species 24 in Figure 24; and Roth's [5] example of species 26 in Figure 25.
In the above paragraphs, the dimensions 3 β , 3 , and 6 δ δ for three colours need only be changed to pβ , pδ , and 2 pδ for any odd number of colours p to specify how thin striping can be done in each possible species.
Theorem 5. In each of the species 1 -32 , that is, all those allowed by the quarter-turn ban, there are fabrics that can be perfectly coloured by thin
Figure 17. Design of a fabric of species 21 and order 96.
Figure 18. Three-colouring by thin striping of the fabric of Figure 17 with redundant cells along the axes of positive slope. Much of the symmetry disappears but a half-turn centre remains in the exact centre of the diagram and along a diagonal of positive slope.
striping with an odd number 2 q +1 of colours and redundant cells arranged as a 2 q/ 1 twill for q = 1 2 , , . . . .
Figure 19. a. Species-22 example from Figure 11a of [8]. b. Three-colouring by thin striping with redundant cells along the mirrors of negative slope, which appear as axes of glidereflection.
Figure 20. a. Subspecies-25 e example; the fabric 12-619-1 of Figure 25 doubled. b. Three-colouring by thin striping with redundant cells along the mirrors of negative slope.
## 5. Six colours
When the number of colours is 4 p , the redundant twill is of the form (4 p -1) / 1 of subspecies 26 e (Figure 26a). When the number of colours
Figure 21. a. 12-231-1 (Fig. 6a of [8]) of subspecies 12 . e b. Three-colouring by thin striping with redundant cells along the axes in mirror position (negative slope).
Figure 22. a. 12-135-1 (Fig. 9b of [8]) of subspecies 18 . s b. Three-colouring by thin striping with redundant cells along the glide-reflection axes.
Figure 23. a. 12-31-1 (Fig. 11b of [8]) of subspecies 23 o . b. Three-colouring by thin striping with redundant cells along the glide-reflection axes of negative slope.
Figure 24. a. 12-189-1 of subspecies 24 . e b. Threecolouring by thin striping with redundant cells along the mirrors.
Figure 25. a. 12-619-1 of subspecies 26 e (Fig. 4b of [8]). b. Three-colouring by thin striping with redundant cells along the mirrors of negative slope.
is congruent to 2 mod 4, which is what we pursue in this section, the redundant twill is of the form (4 p + 1) / 1 of subspecies 26 o (Figure 26b). In the former case the mirrors between the dark lines have a cell-centre ♦ halfway between cell-centre ♦ s in redundant cells and a cell-corner ♦ half-way between cell-corner ♦ s where redundant cells touch. In the latter case, the cell-centre ♦ in the middle lies between cell-corner ♦ s in the dark lines and vice versa. In both cases, there are only mirrors every β perpendicular to the dark lines. Therefore for all even numbers of colours no glide-reflection axis not in mirror position can be accommodated. This rules out subspecies 1 , 1 e o , 2 , 2 e o , 8 , 18 o o , 18 , and 28 e o and whole species 4, 10, 12, 14, 16, 20, 24, and 32 as for two colours. Call this the mirror-position requirement. Subspecies left from partly affected species are 1 m , 2 m , 8 , e 18 , s 28 , e and 28 n . The quarter-turn ban eliminates species 33-39. Left are 1 m , 2 m , 3, 5-7, 8 , 9, 11, 13, 15, 17, 18 e s , 19, 21-23, 25, 26, 27 e and 29, 27 o and 30, 28 e and 31, and 28 n .
Because the redundant twill 5/1 for thin striping with six colours is of subspecies 26 , no further subspecies are ruled out than those just ruled out o
Figure 26. Twills for colouring with an even number of colours, redundant cells for thin striping dark. a. 5/1 twill for six-colouring. b. 3/1 twill for four-colouring.
Figure 27. a. Design of fabric of species 1 m with glide 3. b. Six-colouring of fabric of Figure 27a with lines of redundant cells across the axes of symmetry.
for all even numbers of colour. Since the aim here is to show which species contain some perfectly colorable designs, 28 n need not be separately considered. All of those listed do contain 6-colorable designs, i.e., all but those eliminated by the mirror-position requirement and the quarter-turn ban. Since the symmetry group for the 5/1 twill is a subgroup of the symmetry group of both the 2/1 twill and plain weave, 6-colorability, which I shall use as a shorthand for perfect colorability by thin striping with 6 colours, implies both 3-colorability and 2-colorability by thin striping. Three-colorability of a fabric does not ensure 6-colorability because of the mirror-position requirement. Two-colorability of a fabric does not ensure 6-colorability because the symmetry group of plain weave is very much larger (both infinite of course) than that of the 5/1 twill, but all of the species that contain 2-colorable fabrics do also contain 6-colorable fabrics.
Figure 28. a. Design of 12-23-1, Roth's [5] example of species 6 (Fig. 6b of [7]). b. Six-colouring by thin striping with lines of redundant cells across axes.
When glide-reflection axes are in mirror position as they must be here, the length of a lattice unit, where there are only parallel axes, must be even (twice the glide). For species 1 m , 2 m , 3, 6, and 7 (where glide-reflection axes are also mirrors), axes of glide-reflection with glide a multiple m ≥ 1 of 3 δ can be placed on mirrors perpendicular to the dark lines. For example, a fabric of species 1 m is illustrated in Figure 27 and a Roth [5] example of species 6 is illustrated in Figure 28. For subspecies 5 , e provided that the length of the lattice unit is a multiple m ≥ 1 of 3 δ (there is no glidereflection), the mirrors can also be placed on mirrors perpendicular to the dark lines. Because the only symmetry axes are mirrors, which involve reflection in the plane of the fabric ( τ ), the only symmetry visible on a single side of such a fabric-unless an extraneous symmetry is introducedis translational. To check the perfection of the colouring it is necessary to see the reverse. Despite the involvement of τ , the reverse is not the same coloured pattern as the obverse reflected because it is coloured differently. We shall however continue the policy of not illustrating reverses. Species 8 e and 9 with glides an odd multiple m ≥ 1 of 3 δ can be treated like the previous subspecies. An example from species 9 (so as to have side-preserving glidereflections) is illustrated in Figure 29. These arrangements have been chosen for simplicity of exposition; there are other possibilities.
For species 11, 13, 15, 17,, 19, 21, 22, and 25, successive axes or mirrors that can be an odd multiple m ≥ 1 of 3 δ apart can be placed on dark lines' mirrors as in 3-colouring. Then the fabric's perpendicular axes, which are in mirror position (the only constraint on them), can be put on mirrors perpendicular to the dark lines. Examples are illustrated for species 11 e (Figg. 15b and 30), 21 (Figg. 17 and 31), 22 (Figg. 19a and 32), and 25 (Figg. 20a and 33a).
Figure 29. a. Design of fabric of species 9. b. Sixcolouring of fabric of Figure 29a with lines of redundant cells across axes.
Figure 30. Six-colouring of fabric of Figure 15b with lines of redundant cells along axes with negative slope.
For species 18 , 23 s o , and 26 e with cell-centre ♦ , successive axes, axes, and mirrors respectively that can be an odd multiple m ≥ 1 of 3 β apart can be placed on the mirrors along and between the dark lines. The fabric's mirrors perpendicular to these will fall on mirrors perpendicular to the dark lines. Examples of 6-colourings are of the species-18 s fabric of Figure 22a in Figure 33b, the species-23 o fabric of Figure 23a in Figure 34a, and the species-26 e fabric of Figure 25a, in Figure 34b, where only the half-turns and translational symmetry remain in the pattern.
For species 27-31, central rectangles of 27 o and 30, 28 e and 31 (and 28 n ), can have one dimension an odd multiple m ≥ 1 of 3 δ and 27 e and 29 can have one dimension an odd multiple m ≥ 1 of 6 . δ The former can have
Figure 31. Six-colouring of the species-21 design of Figure 17 with lines of redundant cells along axes with positive slope.
Figure 32. Six-colouring of the species-22 design of Figure 19a with lines of redundant cells along mirrors/axes with negative slope.
these mirrors placed on dark lines' mirrors, axes falling between them, and the latter can have mirrors and axes placed on dark lines' mirrors. Fabrics' perpendicular mirrors and axes fall on mirrors and axes perpendicular to the dark lines. An example is a species-31 fabric (Figure 35a) 6-coloured with redundant cells along the mirrors of negative slope in Figure 35b. The mirror symmetry and side-reversing half-turns are missing from the pattern, but the ordinary half-turns are there in the middle of what can be seen as
Figure 33. a. Six-colouring of the species-25 e design of Figure 20a with lines of redundant cells along mirrors with negative slope. b. Six-colouring of species-18 s fabric of Figure 22a with redundant cells along axes of negative slope between the zig-zags.
Figure 34. a. Six-colouring by thin striping of species23 o fabric of Figure 23a with lines of redundant cells along the axes of side-preserving glide-reflection that are not also mirrors. b. Six-colouring of the species-26 e design of Figure 25a with lines of redundant cells along mirrors with negative slope.
tri-colour flags and the centre of what can be seen as flag poles. (Each flag and pole has the other motif at both ends-in the same colour or coloursrather than just the normal one on account of the half-turn symmetries.)
Figure 35. a. Species-31 fabric (Fig. 15b of [8]). b. Sixcolouring by thin striping with lines of redundant cells along mirrors of negative slope.
Theorem 6. In each of the species 1 - , 3 5 - , 9 11 , 13 , 15 , 17 -19 , 21 -23 , and 25 -31 , that is, all those allowed by the mirror-position requirement on glide-reflection axes and the quarter-turn ban, there are fabrics that can be perfectly coloured by thin striping with 4 p + 2 colours and redundant cells arranged as a (4 p +1) 1 / twill for p = 1 2 , , . . . .
## 6. Four colours
For four colours there is a further constraint because the species-26 e lattice units of the redundant 3/1 twill (Fig. 26b) are odd by even in δ . Species 27 o and 30 (with 27 o spacing) have glide-reflection axes an odd distance apart in δ in both perpendicular directions; they can therefore not be accommodated on mirrors of the redundant twill because the glides in both perpendicular directions must be odd. In these two species, however, 27 e remains; only 30 is eliminated altogether. However, 1 m , 2 m , 3, 5-7, 8 , 9, 11, 13, 15, 17, e 18 , 19, 21-23, 25, 26, 27 , 28, 29, and 31 can be perfectly coloured by thin s e striping with four colours, that is, fabrics in all the species but 30 and those eliminated by the mirror-position requirement and the quarter-turn ban. Since the redundant cells and symmetry axes for four colours are related as they are to those for two colours, the symmetry group of the redundant twill is a subgroup of that for two colours. Accordingly, if a fabric is fourcolorable by thin striping then it is two-colorable by thin striping but not the converse. The biggest actual difference is the two-colorability of species 30 and 36 . s
When glide-reflection axes are in mirror position as they must be here, the length of a lattice unit, where there are only parallel axes, must be even
Figure 36. a. Species-1 m fabric. b. Four-colouring by thin striping with lines of redundant cells perpendicular to the axes.
(twice the glide); as a result, the width must be odd [7] for species 1 m , 2 m , 3, 5 , e and 7 . e For species 1 m , 2 m , 3, and 7 , axes of glide-reflection e with glide a multiple m ≥ 1 of 2 δ must therefore be placed on mirrors perpendicular to the dark lines. For example, Figure 36a shows a species-1 m fabric with glide 2 δ , and Figure 36b shows it perfectly 4-coloured by thin striping with redundant cells perpendicular to the axes. For subspecies 5 , provided e that the length of the lattice unit is a multiple m ≥ 1 of 2 δ , the mirrors can be placed on mirrors perpendicular to the dark lines exactly like the glide-reflection axes two sentences back. An example is Roth's [5] example of subspecies 5 , e illustrated in Figure 37a, 4-coloured in Figure 37b with mirrors across the dark lines as they must be. Because the only symmetry axes are mirrors, which involve τ , the only symmetry visible on a single side of such a fabric-unless an extraneous symmetry is introduced-is translational. To check the perfection of the colouring it is necessary to see the reverse. Despite the involvement of τ , the reverse is not the same coloured pattern as the obverse reflected because it is coloured differently. We shall however continue the policy of not illustrating reverses. An example of subspecies 7 e is illustrated in Figure 38, where the reflective symmetry with mirrors coincident with the axes of the visible glide-reflections is not visible, but extraneous reflective symmetry introduced by the colouring has mirrors perpendicular to the axes through the lines of redundant cells 2 δ apart. This is side-preserving reflective symmetry, which cannot happen in a design.
Species-6 lattice units have width a multiple m ≥ 1 of 2 δ and so can have their bounding mirrors placed along or between dark lines' mirrors, their intervening mirrors falling on mirrors between or on dark lines. They cannot be arranged perpendicular to the design's axes because the glides are odd. An example is Roth's [5] example of species 6 in Figure 39a, coloured two different ways in Figures 39b and c.
Figure 37. a. The species-5 e fabric 12-35-1 (Fig. 2a of [7]). b. Four-colouring by thin striping with lines of redundant cells perpendicular to the axes.
Figure 38. a. A species-7 e fabric (Fig. 16a of [7]). b. Fourcolouring by thin striping with lines of redundant cells perpendicular to the axes.
For species 8 e and 9, successive mirrors a multiple m ≥ 1 of 2 δ apart can be placed along the dark lines' mirrors or all between them if the glides are even. The axes of glide-reflection will then fall on mirrors in or between the dark lines. Odd glides too can be accommodated by making sure that the mirrors fall all on dark lines or all between them, since that makes room for the axes on or between the lines. An example with odd glides and the redundant cells along the mirrors is Roth's [5] example of species 9 (for the sake of having side-preserving glide-reflections) in Figure 40.
Figure 39. a. 8-11-1 of species 6 (Fig. 6a of [7]). b. Fourcolouring by thin striping with lines of redundant cells along the mirrors. c. Second four-colouring with lines of redundant cells between the mirrors. Extraneous mirrors are introduced along the lines of redundant cells.
Figure 40. a. Species-9 fabric 8-19-2 (Fig. 7b of [7]). b. Four-colouring by thin striping, which looks thick, with lines of redundant cells along the mirrors. Extraneous sidepreserving glide-reflections are along the mirrors.
For species 11 , 13 , 15 , 17 , 18 , 19 , 21, 22, 23 , 25 , and 26 , closeste e e e s e e e e together axes or mirrors (17 , 18 , 21, 23 e s e mirrors; 19 , 22 axes) with sepe aration a multiple m ≥ 1 of 2 δ can be placed along dark lines' mirrors. The fabric's perpendicular axes or mirrors will fall on mirrors perpendicular to the dark lines. Examples are the fabric of Figure 15b of species 11 e coloured in Figure 41a, Roth's [5] species-17 e example in Figure 42, a fabric of species 18 s in Figure 43, Roth's [5] species-19 e example in Figure 44, Roth's [5] species-21 example in Figure 45, the fabric of species 22 in Figure 19a coloured in Figure 46a, Roth's [5] species-23 e example in Figure 47,
Figure 41. a. Species-11 e fabric of Figure 15b (Fig. 5a of [8]) 4-coloured by thin striping with redundant cells along the axes of positive slope. b. Four-colouring of fabric of Figure 20a of species 25 e with redundant cells along mirrors of positive slope.
Figure 42. a. 8-19-7 of species 17 . e b. Four-colouring by thin striping with redundant cells along mirrors.
the fabric of species 25 e in Figure 20a coloured in Figure 41b, and the fabric of species 26 e in Figure 25a coloured in Figure 46b.
Species 27 , 28 , 28 e e n , 29 with the spacing of 27 , and 31 with the spacing e of 28 , can have their mirrors with separation a multiple e m ≥ 1 of 2 δ placed along the dark lines' mirrors. Intervening fabric axes will then lie along intervening mirrors, and perpendicular fabric axes and mirrors will lie along perpendicular mirrors. Examples are fabric 8-5-1 of species 28 n in Figure 48 (compare Fig. 47b), Roth's [5] example of species 29 in Figure 49 coloured in Figure 50a, and the species-31 fabric of Figure 35a coloured in Figure 50b.
The above discussion for four colours contains the expression 'multiple of 2 ' repeatedly. δ For 4 p colours, 2 δ needs to be replaced by 2 pδ .
Figure 43. a. A fabric of species 18 . s b. Four-colouring by thin striping with redundant cells along mirrors.
Figure 44. a. 8-19-4 of species 19 e (Fig. 9c of [8]). b. Fourcolouring by thin striping with redundant cells along axes.
Figure 45. a. 8-7-2 of species 21 (Fig. 10a of [8]). b. Fourcolouring by thin striping with redundant cells along mirrors.
Figure 46. a. Species-22 fabric of Figure 19a (Fig. 11a of [8]) 4-coloured by thin striping with redundant cells along the axes that are not mirrors (positive slope). b. Four-colouring of fabric of Figure 25a of species 26 e with redundant cells along mirrors of positive slope.
Figure 47. a. Roth's [5] species-23 e example 8-27-1. b. Four-colouring with redundant cells along mirrors.
Theorem 7. In each of the species 1 - , 3 5 - , 9 11 , 13 , 15 , 17 , 18 , 19 , 21 -23 , 25 -29 , and 31 , that is, all those allowed by the mirror-position requirement on glide-reflection axes and the quarter-turn ban except 30 , there are fabrics that can be perfectly coloured by thin striping with 4 p colours and redundant cells arranged as a (4 p -1) / 1 twill for p = 1 2 , , . . . .
## 7. Colouring woven flat tori
The weaving and colouring by thick striping of flat tori was discussed in §§ 2 and 7 of [12]. The possibility of weaving Klein bottles is mentioned
Figure 48. a. Fabric 8-5-1 of species 28 n (Fig. 13b of [8]). b. Four-colouring with redundant cells along mirrors of positive slope
.
Figure 49. Roth's [5] example 16-2499 of species 29 (Fig. 15a of [8]).
there but not discussed. The opposite sides of a period parallelogram of a design or pattern can be identified to weave or colour a woven torus, e.g., a quarter of Figure 1a. The rectangular lattice unit in the same figure or a square multiple of the order-by-order square, like the whole of Figure 1a, can also be used. In the latter case, there are mn strands in each direction if there are m by m n × n squares, but in the former case there are usually rather fewer, in this case one strand in each direction. Too few strands are obviously an obstacle to interesting colouring.
As redundancy of cells plays no role in choosing period parallelograms in the plane to map to tori, the patterns with no redundancy discussed in section 2 behave, if their stripes are thick, like those with thick stripes discussed in § 7 of [12] - another reason why section 2 should have been in
Figure 50. a. Four-colouring of the fabric of Figure 49 with redundant cells along mirrors of negative slope. b. Fourcolouring of the fabric of Figure 35 with redundant cells along mirrors of positive slope.
[12]. For the same reason, if their stripes are thin, they behave like those with thin stripes and twilly redundancy, to which we now turn.
The designs of the fabrics coloured in sections 2 and 4-6 can all be used to weave flat tori. For colouring the torus, the topological constraint 'that the rectangle to be mapped to the torus must be a period parallelogram' has in addition 'that the rectangle must also be a period parallelogram of the colouring'.
This topic was ignored altogether in the papers [10] and [11] on striping with two colours. In that case the parities of the dimensions of lattice units or of the smallest oblique rectangle (at π/ 4) enclosing a rhombic lattice unit (e.g., Figure 13) indicate whether they are suitable for two-colouring a torus with thin striping. The species at issue, i.e., those that can be twocoloured in the plane are, according to Theorems 2.1 and 2.2 of [10], 1 m , 2 m , 3 (Fig. 5a of [10]), 5 o , 5 , 6 (Fig. 5b of [10]), 7, 8 e e (Fig. 5c of [10]), 9 (Fig. 5d of [10]), 11 (Fig. 9ab of [10]), 13, 15, 17, 18 , 19, 21, 22 (Fig. 9cd s of [10]), 23, 25-27, 28 e , 28 n , 29, 30 (Fig. 9ef of [10]), 31, and 36 s . Of these, most allow the lattice unit of the striped pattern to be used for a torus. There are two kinds of exception. Patterns of species 1 m , 2 m (Fig. 4ab of [10]), 5 e (Fig. 4ef of [10]), 7 (Fig. 6 of [10]), 18 , and 26 s e need to have one of the odd lattice-unit dimensions doubled, and patterns of species 5 o (Fig. 4cd of [10]) and 26 o must have both doubled. As well, the lattice unit for species 36 s is interesting because, while its square lattice unit can simply be doubled in both directions, taking it from level 2 to level 4, the square at
the intermediate level 3 is sufficient to make the weaving work for a torus, as can be seen in Figures 7 and 8 of [10].
Order-by-order squares are suitable in all cases because all perfectly 2colourable fabrics are of even order.
When we turn to more than two colours, we find that everything is as for thick striping but simpler. Given a perfect c -colouring by thin striping of an ordern isonemal design, one needs to choose a suitable period parallelogram. Since the design is given, with its various period parallelograms, both n -byn and oblique lattice units of the side-preserving subgroup H 1 , it is just a question of how many copies are needed. In the n -byn case, the smallest mn -bymn choice must have mn be the lowest common multiple of n and c , since each stripe of the colouring is one cell wide. The size of a lattice unit must be inflated (if at all) similarly. The smallest period parallelograms of any colouring can of course be assembled to make others as large as one pleases.
Another approach would take the number of colours c and the species as given and select the lattice unit appropriate to the species, if any, so that fabrics based on it are perfectly colourable.
Consider the example of Figure 2, a 3-coloured species-1 o fabric of order n = 30. The lattice unit shown in Figure 2a is too small to become a torus with the colouring of Fig. 2b because it is 3 δ × 5 . δ While the pattern repeats with the same colour with a translation of 3 δ in the direction perpendicular to the axes, the illustrated lattice unit has to be repeated three times along the axes to bring the colours as well as the pattern into phase, making the torus 3 δ × 15 . δ In this case, the mn -bymn square of the first paragraph of the section would be 30 by 30 with area 900 since m can be 1, as against 3 δ × 15 δ = 90 for the oblique rectangle. More interestingly, the 30 × 30 square/torus has 10 thin stripes of each colour in each direction, whereas the rectangle/torus has only a single strand of each colour in each direction, each crossing the torus six times on account of the way the strand's ends in the rectangular lattice unit match up in the torus.
Because a woven flat torus is a link in the sense of knot theory, it is of interest how many strands there are - the same number of each colour in each direction. As an example of what happens, let us pursue the example of the previous paragraph. When the 3 δ × 5 δ H 1 lattice unit is replicated to make the 3 δ × 15 δ torus, a translational symmetry subgroup is created, for the successive images of the 3 δ × 5 δ lattice unit, while woven identically, must differ by a cyclic permutation of the colours. The behaviour now being discussed is independent of how the torus is woven but depends solely on the dimensions of the H 1 lattice unit. Considering vertical strands and starting with the strand next to the bottom corner of the lattice unit marked in Figure 2a, they are yellow, red, and blue cyclically left to right. When these three strands, which we can call a tricoloured band reaches the upper boundary of the lattice unit, they are identified with the next band to the right by the identification of boundaries, and then the next, and so on until
eventually they emerge at the upper right boundary of the third copy of the lattice unit to be identified with the band entering the lower left boundary upward. This is identified with the band with which we began. There is a single band in each direction and so a single strand of each colour.
The multiple-of-three dimensions required for torus formation with colours in phase ensure that the bands of the previous paragraph can be chosen as units that will not be split up as they pass across the torus. The twilly redundancy of the colouring ensures that the colouring of the bands in one direction is a cyclic permutation of the colouring of the strands of the bands in the other direction, but there is no reason for them to be the same.
The 3 δ × 15 δ torus is by no means the only possibility with this lattice unit. The short dimension can be multiplied by any positive integer. Five gives an interesting result for the 15 δ × 15 δ square torus. The five bands that enter the lower right side of the period parallelogram (= the torus) reach its upper right side together, are identified with the five bands entering the lower left side, and rise to the upper left side to be identified with themselves respectively. As happens with any square period parallelogram at π/ 4 to the horizontal, the bands - and so all the strands - remain separate. There are five bands of each colour in each direction. Any multiple of five gives the same result.
What about non-multiples of five? If the 3 δ × 15 δ period parallelogram is multiplied by 7 to be 21 δ × 15 , then as the five bands at the lower right δ boundary rise up through copies of the lattice unit, we know that, as they cross the boundary between the fifth and sixth copies of the lattice unit, they fill it as they did the lower right side - as they did in the previous paragraph. The multiples like 7 can be taken modulo 5 - or whatever makes a square. Seven gives the same result as two. Crossing only two of the 3 δ × 15 δ rectangles brings the fourth and fifth bands to the left end of the upper left side, with the first, second, and third bands following on. As the permutation ( 1 2 3 4 5 4 5 1 2 3 ) has period 5 like ( 1 2 3 4 5 5 1 2 3 4 ) for just one 3 δ × 15 , δ each supposed band passes through each position on its successive crossings of the torus, and so there is actually only one band. The same happens for multiples congruent to 3 and 4 modulo 5, making congruence to zero quite special. The primality of 5, the apparent number of bands entering the lower right side, makes one band be the almost uniform answer.
When the apparent b bands entering one side of the smallest period parallelogram, which is an H 1 lattice unit perhaps replicated to put the colours in phase, reach the opposite side of the period parallelogram they are permuted by running into a perpendicular side. Call that permutation p . A square period parallelogram has actually b distinct bands, the index of the identity in the cyclic subgroup of permutations, in the example ( 1 2 3 4 5 5 1 2 3 4 ) 5 . When b is prime, we have seen that, because the index of the powers of p other than p b is one, there is only one band. When b is composite there is room for different indices and therefore numbers of bands different from 1 and b .
## References
- 1. Gr¨ unbaum, Branko, and Geoffrey C. Shephard, Satins and twills: An introduction to the geometry of fabrics , Mathematics Magazine 53 (1980), 139-161.
- 2. Gr¨ unbaum, Branko, and Geoffrey C. Shephard, A catalogue of isonemal fabrics , in Discrete Geometry and Convexity , Jacob E. Goodman et al., eds. Annals of the New York Academy of Sciences 440 (1985), 279-298.
- 3. Gr¨ unbaum, Branko, and Geoffrey C. Shephard, An extension to the catalogue of isonemal fabrics , Discrete Mathematics 60 (1986), 155-192.
- 4. Gr¨ unbaum, Branko, and Geoffrey C. Shephard, Isonemal fabrics , American Mathematical Monthly 95 (1988), 5-30.
- 5. Roth, Richard L., The symmetry groups of periodic isonemal fabrics , Geometriae Dedicata 48 (1993), 191-210.
- 6. Roth, Richard L., Perfect colorings of isonemal fabrics using two colors , Geometriae Dedicata 56 (1995), 307-326.
- 7. Thomas, R.S.D., Isonemal prefabrics with only parallel axes of symmetry , Discrete Mathematics 309 (2009), 2696-2711. doi:10.1016/j.disc.2008.06.028. Online: http://arxiv.org/abs/math/0612808v2 and on the journal site.
- 8. Thomas, R.S.D., Isonemal prefabrics with perpendicular axes of symmetry , Utilitas Mathematica 82 (2010), 33-70. Online: http://arxiv.org/abs/0805.3791v1 .
- 9. Thomas, R.S.D., Isonemal prefabrics with no axes of symmetry , Discrete Mathematics 310 (2010), 1307-1324. doi:10.1016/j.disc.2009.12.015. Online: http://arxiv.org/abs/0911.1467v2 and on the journal site.
- 10. Thomas, R.S.D., Perfect colourings of isonemal fabrics by thin striping , Bull. Australian Math. Soc. 83 (2011), 63-86. Online: http://arxiv.org/abs/1006.5653 and on the journal site, doi:10.1017/S0004972710001632.
- 11. Thomas, R.S.D., Perfect colourings of isonemal fabrics by thick striping , Bull. Australian Math. Soc. 85 (2012), 325-349. Online: http://arxiv.org/abs/1109.2254 and on the journal site, doi:10.1017/S0004972711002899.
- 12. Thomas, R.S.D., Colouring isonemal fabrics with more than two colours by thick striping , Contributions to Discrete Mathematics 8 (2013), 38-65. Online: http://cdm.math.ucalgary.ca/cdm/index.php/cdm/article/view/365/165 .
St John's College and Department of Mathematics, University of Manitoba, Winnipeg, Manitoba R3T 2N2 Canada
E-mail address : [email protected]
| null |
[
"Robert S. D. Thomas"
] |
2014-08-05T19:53:46+00:00
|
2014-08-05T19:53:46+00:00
|
[
"math.CO",
"primary 52C20, secondary 05B45, 51M20"
] |
Colouring Isonemal Fabrics with more than two Colours by Thin Striping
|
Perfect colouring of isonemal fabrics by thin striping of warp and weft with
more than two colours is examined. Examples of thin striping in all possible
species with no redundancy and with redundant cells arranged as twills are
given. Colouring woven flat tori is discussed.
|
1408.1080v1
|
## Screening of an electrically charged particle in a two-dimensional two-component plasma at Γ = 2
Alejandro Ferrero 1 and Gabriel T´ ellez 2 Universidad de los Andes
Departamento de F´ ısica Bogot´ a, Colombia
## Abstract
We consider the thermodynamic effects of an electrically charged impurity immersed in a two-dimensional two-component plasma, composed by particles with charges ± e , at temperature T , at coupling Γ = e / k 2 ( B T ) = 2, confined in a large disk of radius R . Particularly, we focus on the analysis of the charge density, the correlation functions, and the grand potential. Our analytical results show how the charges are redistributed in the circular geometry considered here. When we consider a positively charged impurity, the negative ions accumulate close to the impurity leaving an excess of positive charge that accumulates at the boundary of the disk. Due to the symmetry under charge exchange, the opposite effect takes place when we place a negative impurity. Both the cases in which the impurity charge is an integer multiple of the particle charges in the plasma, ± e , and a fraction of them are considered; both situations require a slightly different mathematical treatments, showing the effect of the quantization of plasma charges. The bulk and tension effects in the plasma described by the grand potential are not modified by the introduction of the charged particle. Besides the effects due to the collapse coming from the attraction between oppositely charged ions, an additional topological term appears in the grand potential, proportional to -n 2 ln( mR ), with n the dimensionless charge of the particle. This term modifies the central charge of the system, from c = 1 to c = 1 -6 n 2 , when considered in the context of conformal field theories.
## 1 Introduction
A two-dimensional two-component plasma (TCP) is a gas composed by two kinds of electrically charged particles (with charges ± e ) interacting through an electrostatic potential. In two dimensions, the Coulomb interaction between two unit charges separated by a distance r is -ln( r/L ) with L an irrelevant length. The value of the charges, together with the inverse temperature factor β = ( k B T ) -1 determine a coupling constant given by Γ = βe 2 .
The value of Γ is important to determine the behavior and stability of the system. For low temperatures (Γ ≥ 2), the thermal energy is not enough to avoid the collapse of oppositely charged ions and the point particles must be replaced by hard disks of radius a . The point particle view, however, can still be kept when Γ < 2 because the thermal fluctuations allow to avoid such collapse. For the value Γ = 2, which we will assume in this analysis, the classical Coulomb gas is mathematically isomorph to a quantum free Fermi field at zero temperature [1]. Moreover, analytical solutions for the grand potential, densities and correlation functions can be found for this particular value of the coupling constant.
The analogy between the two-component plasma and the sine-Gordon quantum field theory has been exploited to obtain results for the thermodynamic properties of the Coulomb gas under some conditions when Γ < 2 [2, 3, 4]. When Γ > 2, an infinite-order transition that take place when Γ = 4 at low density (the Kosterlitz-Thouless transition) has been studied [5]; additionally, exact results for the plasma for values of Γ larger than 2, but close enough to 2, have also been found [6, 7].
In the particular case Γ = 2 several studies have been done in the past: the onecomponent plasma (a system composed by only one kind of particles immersed in an oppositely charged background) with adsorbing impurities [8, 9], and the two component plasma with adsorbing boundaries [10, 11] and adsorbing impurities [12].
In this work, we study the introduction of an electrically charged impurity into the plasma. The plasma occupies a disk region of radius R . The impurity is an impenetrable disk of radius r 0 and charge q = ne located at the center of the disk domain that contains the plasma. The electrostatic potential generated by the impurity, at a distance r from it, can be written as V ( r ) = -ne ln( r/L ), where n is the dimensionless charge of the impurity. From this point on, we will understand that the charge q of the impurity is 'integer' if q is an integer multiple of e , i.e., q = ne , with n ∈ Z . On the other hand, a 'non-integer' charge means that n / ∈ Z and so there is a fractional part in the charge of the impurity that cannot be compensated by the other particles in the gas. Using the electrostatic potential just described, we can find the fugacities for positive and negative particles; they are
$$\cdot \\ m _ { s } ( \mathbf r ) = m ( \mathbf r ) e ^ { - s \beta e V ( r ) } = m ( \mathbf r ) \left ( \frac { r } { L } \right ) ^ { 2 s n }, \\ \pm 1, \, \text{and} \, m ( \mathbf r ) = 0 \text{ for } | \mathbf r | = r < r _ { 0 } \text{ and } m ( \mathbf r ) = m \text{ for } r > r _ { 0 }.$$
where s = ± 1, and m ( r ) = 0 for | r | = r < r 0 and m ( r ) = m for r > r 0 .
This model can be applied to some physical systems. The present concentrical circular geometry can be seen as a transversal cut of a cylindrical geometry. The location of the impurity at the center of the system generates the redistribution of charge that induces the accumulation of positive or negative ions around the impurity and the boundary of the plasma. Some proteins and other components in biological systems could be modeled as large cylinders with hard core effects and electrical charges spread along their length. To be able to describe the effective interactions between these charged entities immersed in an electrolyte is an important task towards understanding its physical properties.
Although the screening of charges in an electrolyte is fairly well understood in the mean field regime, Γ → 0, described by the Poisson-Boltzmann equation [13, 14, 15], the intermediate and strong coupling regimes remain elusive. There has been several efforts towards the understanding of the strong coupling regime [16, 17, 18, 19, 20, 21]. However, most of the strong-coupling techniques, such as the Wigner strong coupling approach [19, 20], are only targeted at the one-component plasma, and cannot directly be applied to a multicomponent plasma such as the one studied here. Thus, the present work tries to contribute to the understanding of the screening beyond the mean field by studying the intermediate coupling Γ = 2.
/negationslash
There has been previous studies of charged impurities in the two-dimensional twocomponent plasma beyond mean field, for 0 < Γ < 2, but limited to point impurities in the bulk. In these works, the use of the form factors of the sine-Gordon field theory allows to obtain the behavior of the density profiles at large distances from the impurity [22] and using the product operator expansion, one can obtain the short distance behavior of the densities [23]. However, the complete form of the density profiles for the whole range of distances is unknown. At Γ = 2, the equivalent field theory to the two-component plasma is a free fermion one, therefore it will be possible to obtain more explicit results in the present case, and also consider the effets of the nonzero size of the impurity ( r 0 = 0), and the confinement of the plasma ( R < ∞ ).
This paper is organized as follows: in section 2 we briefly discuss the theoretical background for a two-dimensional two-component plasma at Γ = 2. Section 3 discusses the Green's functions that generate the density and correlations of the plasma. The density of the gas and the total charge are discussed in sections 4 and 5 respectively. Chapter 6 explains the calculation of the grand potential. We finally summarize our conclusions and discuss the results in section 7.
## 2 Formalism
In 1989, Cornu and Jancovici developed a general formalism to study a two-dimensional two-component gas at Γ = e β 2 = 2 [1]. Within this model, the two-dimensional position of a particle r = ( x, y ) can be written in complex coordinates as z = x + iy . As explained before, when Γ = 2 the point particles must be replaced by hard spheres of radius a ; its
inverse a -1 could also be viewed as an ultraviolet cutoff in momentum space. In the limit a → 0, the logarithm of the grand partition function Ξ for this system takes the form [1]
/negationslash
$$\ln \Xi = \text{Tr} \left [ \ln \left ( \, \emptyset + m _ { + } ( \mathbf r ) \frac { 1 + \sigma _ { z } } { 2 } + m _ { - } ( \mathbf r ) \frac { 1 - \sigma _ { z } } { 2 } \right ) - \ln \, \emptyset \, \right ]. \quad \, \quad \, ( 2 ) \\ \mathbf r _ { \infty } \sim \text{T_S_S_S_S_S_S_S_S_S_S_S_$$
/negationslash
In a two-dimensional system, / = ∂ σ ∂ x x + σ ∂ y y , where σ x , σ y and σ z are the Pauli matrices. The terms m ± ( r ) = m ( r ) e -βU ± ( r ) are fugacities that take into account the effects of an electric external potential that generates electrostatic energies given by U ± ( r ) for positive and negative particles. The term m ( r ) includes hard-core effects and other non-electrical contributions. The calculation of the grand potential, Ω = -k B T ln Ξ, is performed by finding the eigenvalues { λ k } of the system [1]
/negationslash
$$\left [ \emptyset - \frac { 1 } { \lambda } \left ( \begin{array} { c c } m _ { + } ( { \mathbf r } ) & 0 \\ 0 & m _ { - } ( { \mathbf r } ) \end{array} \right ) \right ] \psi ( { \mathbf r } ) = 0, \text{ with } \psi ( { \mathbf r } ) = \left ( \begin{array} { c c } g ( { \mathbf r } ) \\ f ( { \mathbf r } ) \end{array} \right ) \quad \quad \quad$$
a two-component spinor. Then the grand potential is given by
$$\beta \Omega & = - \sum _ { k } \ln ( 1 + \lambda _ { k } ). & ( 4 ) \\ \intertext { o n s c h o w t h e c u n i v a l o n e } \, \text{the vector this vector and a free Form} \, \sigma$$
These two last equations show the equivalence between this system and a free Fermi gas [1]. The pressure p , in the thermodynamic limit, is given by
$$\beta p = \frac { \partial \ln \Xi } { \partial A } = - \frac { 1 } { A } \beta \Omega,$$
where A is the area of the gas. The truncated p -body densities are given by:
$$\rangle & = ( - 1 ) ^ { p + 1 } m _ { s _ { 1 } } ( \mathbf r _ { 1 } ) m _ { s _ { 2 } } ( \mathbf r _ { 2 } ) \dots m _ { s _ { p } } ( \mathbf r _ { p } ) \\ & \quad \times \sum _ { i _ { 1 }, i _ { 2 }, \dots, i _ { p } } G _ { s _ { i _ { 1 } }, s _ { i _ { 2 } } } ( \mathbf r _ { i _ { 1 } }, \mathbf r _ { i _ { 2 } } ) \dots G _ { s _ { i _ { p } }, s _ { i _ { 1 } } } ( \mathbf r _ { i _ { p } }, \mathbf r _ { i _ { 1 } } ), \quad \quad ( 6 ) \\ \text{where the summations run over cycles } \{ i, \, i \, \partial \, \quad i \, \} \, \ G _ { \lambda } ( \mathbf r, \, \mathbf r \} \, \text{ are matrix elements}$$
where the summations run over cycles { i 1 , i 2 . . . , i n } . G s 1 s 2 ( r 1 , r 2 ) are matrix elements that satisfy the set of differential equations
$$\left ( \begin{array} { c c } m _ { + } ( \mathbf r _ { 1 } ) & 2 \partial _ { z _ { 1 } } \\ 2 \partial _ { z _ { 1 } } & m _ { - } ( \mathbf r _ { 1 } ) \end{array} \right ) \left ( \begin{array} { c c } G _ { + + } ( \mathbf r _ { 1 }, \mathbf r _ { 2 } ) & G _ { + - } ( \mathbf r _ { 1 }, \mathbf r _ { 2 } ) \\ G _ { - + } ( \mathbf r _ { 1 }, \mathbf r _ { 2 } ) & G _ { - - } ( \mathbf r _ { 1 }, \mathbf r _ { 2 } ) \end{array} \right ) = \delta ( \mathbf r _ { 1 } - \mathbf r _ { 2 } ). \quad \, \quad ( 7 ) \\ \mathbb { m } = \mathbf w \, \text{the main case} \, = \, 1 \, \text{and} \, = \, 2 \, \text{me how} \,.$$
For the particular cases n = 1 and n = 2 we have
$$\rho _ { s } ^ { ( 1 ) } ( \mathbf r ) & = m _ { s } ( \mathbf r ) G _ { s s } ( \mathbf r, \mathbf r ), \\ \rho _ { s _ { 1 } s _ { 2 } } ^ { ( 2 ) T } ( \mathbf r _ { 1 }, \mathbf r _ { 2 } ) & = - m _ { s _ { 1 } } ( \mathbf r _ { 1 } ) m _ { s _ { 2 } } ( \mathbf r _ { 2 } ) G _ { s _ { 1 } s _ { 2 } } ( \mathbf r _ { 1 }, \mathbf r _ { 2 } ) G _ { s _ { 2 } s _ { 1 } } ( \mathbf r _ { 2 }, \mathbf r _ { 1 } ).$$
Since the gas is not stable under the collapse of particles of opposite signs, the results will depend on the cutoff distance given by a . While the limit a → 0 can be taken for some quantities such as the correlation functions, the charge density and other quantities will depend on such distance.
## 3 The Green functions
We will now study the effects of a charged particle of radius r 0 located at the center of the plasma with a potential given by U r ( ) = -ne 2 ln( r/L ). This particle is thus immersed in a large electrical system confined in a large disk of radius R . Defining g s 1 s 2 = e -is 1 U r ( 1 ) G s 1 s 2 e -is 2 U r ( 2 ) and using Eq. (7), we obtain the set of differential equations (for r 0 < r < R )
$$\left \{ m ^ { 2 } - \nabla _ { 1 } ^ { 2 } - \frac { 2 i n } { r _ { 1 } ^ { 2 } } \frac { \partial } { \partial \theta _ { 1 } } + \frac { n ^ { 2 } } { r _ { 1 } ^ { 2 } } \right \} _ { \Gamma _ { 1 } ^ { 2 } } & g _ { \pm \pm } ( \mathbf r _ { 1 }, \mathbf r _ { 2 } ) = m \delta ( \mathbf r _ { 1 } - \mathbf r _ { 2 } ),$$
$$g _ { - + } ( \mathbf r _ { 1 }, \mathbf r _ { 2 } ) = - m ^ { - 1 } \overset { \prime } { \left [ 2 \partial _ { \bar { z } _ { 1 } } - n \frac { e ^ { i \theta _ { 1 } } } { r _ { 1 } } \right ] } g _ { + + } ( \mathbf r _ { 1 }, \mathbf r _ { 2 } ), \quad \ \ ( 1 0 ) \\ \cdot \coprod \ \ e ^ { - i \theta _ { 1 } } \coprod \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cd.$$
$$g _ { + - } ( \mathbf r _ { 1 }, \mathbf r _ { 2 } ) & = - m ^ { - 1 } \left [ \stackrel { \mathbf r } { 2 } \partial _ { z _ { 1 } } + n \frac { e ^ { - i \theta _ { 1 } } } { r _ { 1 } } \right ] g _ { - - } ( \mathbf r _ { 1 }, \mathbf r _ { 2 } ). & & ( 1 1 ) \\ \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot$$
We now propose a particular solution in terms of a Fourier expansion with l -modes g l ss ′ ( r 1 , r 2 ) = e ilθ 1 g l ss ′ (˜ r , 1 ˜ ), r 2 where we defined the dimensionless variables ˜ r 1 2 , = mr 1 2 , . In terms of these variables and using the stated Fourier decomposition, we see that the modes g l ss ′ (˜ r , 1 ˜ ) obey the differential equations r 2
$$\left \{ \tilde { r } _ { 1 } ^ { 2 } \frac { d } { d \tilde { r } _ { 1 } ^ { 2 } } + \tilde { r } _ { 1 } \frac { d } { d \tilde { r } _ { 1 } } - \left [ ( l + n ) ^ { 2 } + \tilde { r } _ { 1 } ^ { 2 } \right ] \right \} & g _ { \pm \pm } ^ { l } ( \tilde { r } _ { 1 }, \tilde { r } _ { 2 } ) = m \delta ( \tilde { r } _ { 1 } - \tilde { r } _ { 2 } ) e ^ { - i l \theta _ { 2 } }, \\ & \quad \quad \quad$$
$$g _ { + = } ^ { l } ( \tilde { r } _ { 1 }, \tilde { \mathbf r } _ { 2 } ) = - e ^ { - i \theta _ { 1 } } \overset { \ell } { \overset { \ell } { \left [ \frac { \dot { d } } { d \tilde { r } _ { 1 } } + \frac { \dot { 1 } } { \tilde { r } _ { 1 } } ( l + n ) \right ] } } g _ { - - } ( \tilde { r } _ { 1 }, \tilde { \mathbf r } _ { 2 } ). \quad \quad ( 1 4 )$$
$$g _ { - + } ^ { l } ( \tilde { r } _ { 1 }, \tilde { \mathbf r } _ { 2 } ) = - & e ^ { i \theta _ { 1 } } \left [ \frac { d } { d \tilde { r } _ { 1 } } - \frac { 1 } { \tilde { r } _ { 1 } } ( l + n ) \right ] g _ { + + } ( \tilde { r } _ { 1 }, \tilde { \mathbf r } _ { 2 } ), \\ \Gamma _ { \ } \dots \dots \dots \Delta _ { \ } \Delta _ { \ } d \quad 1 \dots \dots \dots$$
The Fourier modes can thus be written in the form g l ±± (˜ r , 1 ˜ ) r 2 = A l (˜ ) r 2 K l + n (˜ ) + r 1 B l (˜ ) r 2 I l + n (˜ ), where r 1 I l ( x ) and K x l ( ) are the modified Bessel functions of first and second kind and A l (˜ ) and r 2 B l (˜ ) are coefficients to be determined by the boundary conditions. r 2 We thus have (for r 0 < r 1 < R and r 1 < r 2 )
$$g _ { \pm \pm } & = \sum _ { l } e ^ { i l \theta _ { 1 } } \left [ A _ { l } ^ { ( \pm \pm ) } K _ { l + n } ( \tilde { r } _ { 1 } ) + B _ { l } ^ { ( \pm \pm ) } I _ { l + n } ( \tilde { r } _ { 1 } ) \right ], \\ a \quad. & = \sum _ { \varrho } e ^ { i ( l \pm 1 ) \theta _ { 1 } } \left [ \, \Delta ^ { ( \pm \pm ) } K. \quad. \, d ( \tilde { r }. \, \cdot \, = \, R ^ { ( \pm \pm ) } I.$$
$$g _ { \mp \pm } = & \sum _ { l } ^ { \cdot } e ^ { i ( l \pm 1 ) \theta _ { 1 } } \left [ A _ { l } ^ { ( \pm \pm ) } K _ { l + n \pm 1 } ( \tilde { r } _ { 1 } ) - B _ { l } ^ { ( \pm \pm ) } I _ { l + n \pm 1 } ( \tilde { r } _ { 1 } ) \right ], \\.. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &.. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &.. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &. &.$$
and something similar for r 2 < r 1 where the coefficients for that region will be called A ′ ( ±± ) l and B ′ ( ±± ) l .
In the regions r < r 0 and r > R we have that the fugacity vanishes so m = 0. Therefore, the differential equations for G s 1 s 2 imply that
$$G _ { + + } ^ { l } = u _ { l } ^ { ( + + ) } e ^ { i l \theta _ { 1 } } r _ { 1 } ^ { l }, \quad G _ { + - } ^ { l } = u _ { l } ^ { ( + - ) } e ^ { i ( l - 1 ) \theta _ { 1 } } r _ { 1 } ^ { l - 1 }, \quad \ \ ( 1 7 )$$
$$G _ { - - } ^ { l } = u _ { l } ^ { ( - - ) } e ^ { i l \theta _ { 1 } } r _ { 1 } ^ { - l }, \ \ G _ { - + } ^ { l } = u _ { l } ^ { ( - + ) } e ^ { i ( l + 1 ) \theta _ { 1 } } r _ { 1 } ^ { - ( l + 1 ) },$$
where the coefficients u ( s 1 s 2 ) l have to be determined for each one of the two regions. Eqs. (12) -(14), (17) and (18) imply the following boundary conditions:
- 1. g ±± must be continuous at r 1 = r 2 .
- 2. g ±∓ must be discontinuous at r 1 = r 2 .
- 3. G s 1 s 2 must be finite at r = 0.
- 4. G s 1 s 2 must vanish at r = ∞ .
While the first and the second conditions are given by the delta distribution in Eq. (12) and provide relations for the coefficients A l and B l in the regions r 0 < r 1 2 , < R , the third and fourth conditions provide the extra conditions by stating which u ( s 1 s 2 ) l coefficients must vanish.
From now on we will use the notations F r (˜ ) j ≡ F ( j ) and F R ( ˜ ) ≡ F ( R ) . The continuity conditions at r 1 = r 0 imply
$$A _ { l } ^ { ( - - ) } K _ { l + n } ^ { ( 0 ) } + B _ { l } ^ { ( - - ) } I _ { l + n } ^ { ( 0 ) } = 0, \ l \geq 1,$$
$$A _ { l } ^ { ( - - ) } K _ { l + n - 1 } ^ { ( 0 ) } - B _ { l } ^ { ( - - ) } I _ { l + n - 1 } ^ { ( 0 ) } = 0, \ l \leq 0,$$
$$A _ { l } ^ { ( + + ) } K _ { l + n + 1 } ^ { ( 0 ) } - B _ { l } ^ { ( + + ) } I _ { l + n + 1 } ^ { ( 0 ) } = 0, \ l \geq 0,$$
$$A _ { l } ^ { ( + + ) } K _ { l + n } ^ { ( 0 ) } + B _ { l } ^ { ( + + ) } I _ { l + n } ^ { ( 0 ) } = 0, \ l \leq - 1.$$
Additionally, the continuity conditions at r 1 = R demand:
$$A _ { l } ^ { \prime ( - - ) } K _ { l + n } ^ { ( R ) } + B _ { l } ^ { \prime ( - - ) } I _ { l + n } ^ { ( R ) } = 0, \ l \leq 0,$$
$$A _ { l } ^ { \prime ( - - ) } K _ { l + n - 1 } ^ { ( R ) } - B _ { l } ^ { \prime ( - - ) } I _ { l + n - 1 } ^ { ( R ) } = 0, \ l - 1 \geq 0,$$
$$A _ { l } ^ { \prime ( + + ) } K _ { l + n + 1 } ^ { ( R ) } - B _ { l } ^ { \prime ( + + ) } I _ { l + n + 1 } ^ { ( R ) } = 0, \ l + 1 \leq 0,$$
$$A _ { l } ^ { \prime ( + + ) } K _ { l + n } ^ { ( R ) } + B _ { l } ^ { \prime ( + + ) } I _ { l + n } ^ { ( R ) } = 0, \ l \geq 0.$$
On the other hand, the two conditions at r 1 = r 2 are
$$A _ { l } ^ { ( \pm \pm ) } K _ { l + n } ^ { ( 2 ) } + B _ { l } ^ { ( \pm \pm ) } I _ { l + n } ^ { ( 2 ) } - A _ { l } ^ { \prime ( \pm \pm ) } K _ { l + n } ^ { ( 2 ) } - B _ { l } ^ { \prime ( \pm \pm ) } I _ { l + n } ^ { ( 2 ) } = 0,$$
$$\cdots _ { l } \cdots _ { l + n } \cdot \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot \cdot _ { l + n } \cdot \cdot \cdot _ { l + n } \cdot \cdot \cdot _ { l + n } \cdot \cdot \cdot _ { l + n } \cdot \cdot \cdot _ { l + n } \cdot \cdot \cdot _ { l + n } \cdot \cdot \cdot _ { l + n } \cdot \cdot \cdot _ { l + n } \cdot \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \cdot _ { l + n } \dots$$
The particular Green functions are thus found by solving for the coefficients given in Eqs. (19) -(28). Using the definitions
$$t _ { l } ^ { ( x ) } \equiv \frac { K _ { l } ( x ) } { I _ { l } ( x ) } \,,$$
and θ 12 ≡ θ 1 -θ 2 , and writing n as n = k + , where ν k and ν are the integer and fractional parts of n respectively, we find
$$\cdot \cdot \cdot \cdot \cdot _ { r } \cdot \cdot \cdot \cdot \cdot _ { v }, \cdot \cdot \cdot \cdot \cdot \cdot \cdot \\ g _ { + + } & = \frac { m } { 2 \pi } e ^ { - i k \theta _ { 1 2 } } \left [ \sum _ { l } e ^ { i l \theta _ { 1 2 } } I _ { l + \nu } ^ { ( < ) } K _ { l + \nu } ^ { ( > ) } + \Delta _ { + + } ^ { ( \nu, \tilde { r } _ { 0 }, \tilde { R } ) } \right ], \\ a & = \frac { m } { } e ^ { - i k \theta _ { 1 2 } } \left [ \sum _ { l } e ^ { i l \theta _ { 1 2 } } I ^ { ( < ) } K ^ { ( > ) } + \Lambda ^ { ( \nu, \tilde { r } _ { 0 }, \tilde { R } ) } \right ]$$
where the notation ( < ) and ( > ) means that we choose between the smaller and larger between r 1 and r 2 respectively and
$$g _ { - - } & = \frac { m } { 2 \pi } e ^ { -$$
$$\text{where the notation (} < ) \, \text{and} \, ( > )$$
The functions g ±∓ are found using Eqs. (13) and (14).
/negationslash
From this point we should emphasize the difference between the cases n ∈ Z ( ν = 0) and n / ∈ Z ( ν = 0). Mathematically, when l ∈ Z , I l ( x ) = I -l ( x ) but I -( + l ν ) ( x ) = I l + ν ( x ) + d l ν K l + ν ( x ), with d l ν ≡ 2 π ( -1) sin( l πν ).
/negationslash
Up to a phase factor, the term m 2 π ∑ l e ilθ 12 I ( < ) l + ν K ( > ) l + ν in last equations reproduces the contribution for the unperturbed plasma g 0 ss ′ ( r 1 , r 2 ) = m 2 π K 0 ( ˜ | r 1 -˜ r 2 | ) in the case ν = 0. For ν = 0 we separate this term by writing
$$\begin{array} {$$
The remaining contributions, defined as ∆ ( ν,r ˜ 0 ,R ˜ ) s 1 s 2 , depend on ν , r 0 and R and describe effects purely generated by the impurity.
## 4 Density Profile
The density of the positive and negative particles in the plasma can easily be evaluated using the Green's functions found in the previous section. Using Eq. (6), we can verify that
$$\rho _ { \pm } ( { \mathbf r } ) = m g _ { \$$
We can also define the charge density (divided by e ), ρ t ( r ), as the difference between the density of positive and negative particles; it is given by
$$\rho _ { t } ( { \mathbf r } ) = m [ \, g _$$
/negationslash
Since the case ν = 0 is a particular situation of the more general case ν = 0, we will assume the most general situation and use the limit ν → 0 in order to find the solutions for the particular case in which the impurity has an integer charge. Remember that in the unperturbed case ρ 0 ± ( r ) ≡ ρ 0 = a → 0 m 2 2 π ( ln 2 ma -γ ) + O (1) [1], where a is the cutoff and γ is the Euler-Mascheroni constant.
## 4.1 General Case
The density profile can be found using Eqs. (32) -(36). For computational reasons it is better to eliminate the negative modes in Eqs. (32) and (33) in favor of only positive ones. Using I -( + l ν ) ( x ) = I l + ν ( x ) + d K l ν l + ν ( x ) we notice that
$$[ t ^ { ( 0 ) } _ { - l + \nu } ] ^ { - 1 } = [ t ^ { ( 0 ) } _ { l - \nu } ] ^ { - 1 } - d ^ { l } _ { \nu }, \quad \ t ^ { ( R ) } _ { - l + \nu + 1 } = \frac { t ^ { ( R ) } _ { l - \nu - 1 } } { 1 + d ^ { l } _ { \nu } t ^ { ( R ) } _ { l - \nu - 1 } }.$$
We use a definition for ν in which -1 < ν < 1. For example, if n = 7 2 we have that / k = 3 and ν = 1 2, but for / n = -7 / 2 we have k = -3 and ν = -1 / 2. We thus use the
notation ν ± to indicate the cases when n ≥ 0 and n < 0 respectively. Then we find
$$\text{notation} \, \nu _ { \pm } \, \text{ to indicate the cases when } n \geq 0 \, \text{ and } n < 0 \, \text{ respectively. } I \, \text{ then we hnd} \\ \frac { 2 \pi } { m ^ { 2 } } \rho _ { + } ^ { ( \nu _ { + }, \tilde { r } _ { 0 }, \tilde { R } ) } ( \mathbf r ) & = \frac { 2 \pi } { m ^ { 2 } } \tilde { \rho } _ { \nu } ^ { 0 } - \sum _ { l = 0 } ^ { k - 1 } \frac { [ \mathbf r _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } [ K _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + [ \mathbf r _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } \mathbf r _ { l + \nu } ^ { ( R ) } - \sum _ { l = 0 } ^ { k } \mathbf r _ { l + \nu } ^ { ( r ) } + \frac { 2 [ \mathbf r _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } I _ { l + \nu } ^ { ( r ) } K _ { l + \nu } ^ { ( r ) } - [ \, I _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + [ \mathbf r _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } \mathbf r _ { l + \nu } ^ { ( R ) } + 1 } \\ & \quad - \sum _ { l = 1 } ^ { \infty } \frac { [ \mathbf r _ { l - \nu } ^ { ( 0 ) } ] ^ { - 1 } [ K _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + [ \mathbf r _ { l - \nu } ^ { ( 0 ) } ] ^ { - 1 } \mathbf r _ { l + \nu } ^ { ( R ) } - \sum _ { l = 1 } ^ { \infty } \mathbf r _ { l - \nu } ^ { ( R ) } \frac { 2 [ \mathbf r _ { l - \nu } ^ { ( 0 ) } ] ^ { - 1 } I _ { l + \nu } ^ { ( r ) } K _ { l + \nu } ^ { ( r ) } - [ \, I _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + [ \mathbf r _ { l - \nu } ^ { ( 0 ) } ] ^ { - 1 } \mathbf r _ { l + \nu } ^ { ( R ) } - 1 } \\ & \quad + \sum _ { l = k } ^ { \infty } \frac { [ \mathbf r _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } [ K _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + [ \mathbf r _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } \mathbf r _ { l + \nu } ^ { ( R ) } - \sum _ { l = k } ^ { \infty } \mathbf r _ { l + \nu } ^ { ( R ) } \frac { 2 [ \mathbf r _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } I _ { l + \nu } ^ { ( r ) } K _ { l + \nu } ^ { ( r ) } + [ \, I _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + [ \mathbf r _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } \mathbf r _ { l + \nu } ^ { ( R ) } }, \quad ( 3 8 ) \\ 2 \pi _ { \mathbf r _ { - } ( \nu _ { + }, \tilde { r } _ { 0 }, \tilde { R } ) } \, \text{--} \, \sum _ { - } ^ { 2 \pi _ { \nu } } \, \text{--} \, \frac { k } { \nu _ { + } } [ \mathbf r _ { l + \nu - 1 } ^ { ( 0 ) } ] ^ { - 1 } [ K _ { l + \nu } ^ { ( r ) } ] ^ { 2 } \quad \, \, \text{--} \, \frac { k } { \nu _ { + } } ( R ) ^ { 2 } [ \mathbf r _ { l + \nu - 1 } ^ { ( 0 ) } ] ^ { - 1 } I _ { l + \nu } ^ { ( r ) } K _ { l + \nu } ^ { ( r ) } + [ \, I _ { l + \nu } ^ { ( r ) } ] ^ { 2 } }$$
$$\cdot, ( 3 9 )$$
$$& \underbrace { \left \lfloor l = 1 } _ { l = 1 } \ 1 + [ t _ { l } ^ { ( 0 ) } _ { l - \nu } ] ^ { - 1 } t _ { l - \nu - 1 } ^ { ( k ) } } \ \right \rfloor _ { l = 1 } ^ { l - \nu - 1 } \ 1 + [ t _ { l - \nu } ^ { ( 0 ) } ] ^ { - 1 } t _ { l - \nu } ^ { ( k ) } ] \\ & + \sum _ { l = k } ^ { \infty } \frac { [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } [ K _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + [ t _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu } ^ { ( k ) } } - \sum _ { l = k } ^ { \infty } t _ { l + \nu } ^ { ( R ) } \frac { 2 [ t _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } [ l _ { l + \nu } ^ { ( r ) } K _ { l + \nu } ^ { ( r ) } + [ l _ { l + \nu } ^ { ( l ) } ] ^ { 2 } }, \quad ( 3 8 ) \\ & \frac { 2 \pi } { m ^ { 2 } } \rho _ { - } ^ { ( \nu _ { + }, \tilde { r } _ { 0 }, \tilde { R } ) } ( r ) = \frac { 2 \pi } { m ^ { 2 } } \rho _ { \nu } ^ { 0 } + \sum _ { l = 0 } ^ { k } \frac { [ t _ { l + \nu - 1 } ^ { ( 0 ) } ] ^ { - 1 } [ K _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + [ t _ { l + \nu - 1 } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu } ^ { ( k ) } } - \sum _ { l = 0 } ^ { k } t _ { l + \nu } ^ { ( R ) } \frac { 2 [ t _ { l + \nu - 1 } ^ { ( 0 ) } ] ^ { - 1 } [ l _ { l + \nu } ^ { ( r ) } K _ { l + \nu } ^ { ( r ) } + [ l _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } \\ & + \sum _ { l = 1 } ^ { \infty } \frac { [ t _ { l } ^ { ( 0 ) } _ { l + 1 } ] ^ { - 1 } [ K _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + [ t _ { l - \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } t _ { l - \nu } ^ { ( k ) } } - \sum _ { l = 1 } ^ { \infty } t _ { l - \nu } ^ { ( R ) } \frac { 2 [ t _ { l + \nu - 1 } ^ { ( 0 ) } ] ^ { - 1 } [ l _ { l + \nu } ^ { ( r ) } K _ { l + \nu } ^ { ( r ) } + [ l _ { l - \nu } ^ { ( r ) } ] ^ { 2 } } \\ & - \sum _ { l = k + 1 } ^ { \infty } \frac { [ t _ { l } ^ { ( 0 ) } _ { l + 1 } ] ^ { - 1 } [ K _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + [ t _ { l + \nu - 1 } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu - 1 } ^ { ( k ) } } - \sum _ { l = k + 1 } ^ { \infty } t _ { l + \nu - 1 } ^ { ( R ) } \frac { 2 [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } [ l _ { l + \nu } ^ { ( r ) } K _ { l + \nu } ^ { ( r ) } - [ l _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + [ t _ { l + \nu - 1 } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu - 1 } ^ { ( R ) } }, ( 3 9 ) \\ & \\ & \rho _ { - } ^ { \infty } \ \infty$$
with
$$\tilde { \rho } _ { \nu } ^ { 0 } & = \frac { m ^ { 2 } } { 2 \pi } \sum _ { l = 0 } ^ { \infty } I _ { l + \nu } ^ { ( r ) } K _ { l + \nu } ^ { ( r ) } + \frac { m ^ { 2 } } { 2 \pi } \sum _ { l = 1 } ^ { \infty } I _ { l - \nu } ^ { ( r ) } K _ { l - \nu } ^ { ( r ) }. \\ & \times 0 \, \text{$row$ on $other$ to continue $one$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$} \, \text{$row$ to $two$.}$$
When ν < 0 we can obtain the solutions using the symmetry relations ρ + ( ν + , k ) = ρ -( -ν -, -k ) and ρ + ( -ν -, -k ) = ρ -( ν + , k ). Notice that ˜ ρ 0 0 = ρ 0 when ν = 0.
Figures 1, 2, 3 and 4 show the density of positive and negative particles for different values of charge impurity n > 0. The density of positive particles vanishes near r 0 due to the repulsive interaction with the impurity carrying a positive charge. On the contrary the density of negative particles increases in an effort to screen the impurity. Far from the impurity both densities converge to the bulk density ρ 0 . For n < 1, the behavior of the negative density in the presence of a fractional impurity presents an interesting effect which takes place near r 0 approaching it from above (see figure 4). The amount of negative ions slightly decreases after taking a maximum value in the neighborhood of the impurity. This behavior is more notable for smaller values of ν . This effect, however, seems to be a property satisfied only by charges n < 1 and not by any non-integer impurity; as can be seen in figure 2, such behavior is absent for n = 2 1. The electrostatic . interaction between the impurity and a negative charge of the plasma can be caracterized by a coupling constant given by Γ n = nβe 2 (see table 1 for more details). If the impurity radius is very small r 0 → 0, then when n ≥ 1 (Γ n ≥ 2) there will be a collapse of the negative particles against the charged impurity, therefore a very large value of ρ -( r ) as r → 0. On the other hand, if n < 1 the system is stable against that collapse and so the
Figure 1: Density of the positive charge for four different integer impurities located at the origin. The size of the impurity is mr 0 = 0 05 and the size of the disk is . mR = 30. The density becomes zero as r → r 0 . We used m -2 ρ 0 = 0 87 . .
Figure 2: Density of the negative charge for four different impurities located at the origin. The size of the impurity is mr 0 = 0 05 and the size of the disk is . mR = 30. We used m -2 ρ 0 = 0 87 . .
Figure 3: Density of the positive charge for four different fractional impurities located at the origin. The size of the impurity is mr 0 = 0 05 and the size of the disk is . mR = 30. We used m -2 ρ 0 = 0 87 . .
value of the negative density at the origin decreases. (Such behavior is described in table 2.) A fingerprint of these two regimes can be seen in figures 2, 4 and 5 which show a change of behavior in the density when the impurity charge changes from n < 1 to n > 1. A similar observation was made in the analysis when r 0 = 0 and Γ < 2 [23]. Although such results are restricted for Γ < 2 and r 0 = 0, the negative density for small r in the case Γ → 2 -would take the form [23]
$$\rho _ { - } ( r ) \underset { r \to 0 } { = } A _ { 1 } r ^ { - 2 n } + O ( r ^ { 2 ( 1 - n ) } ), \quad \text{ for } n < 1 \,,$$
where A 1 is a constant. The results shown in table 1 match last predictions in the case n < 1 and r 0 → 0, ie. ρ -( r ) ∝ r -2 n . When n > 1, this behavior changes, in part due to the collapse of the negative ions into the charged impurity.
## 4.2 Asymptotic limits
It is interesting to study the limiting cases in which r 0 → 0 and R →∞ . When r 0 → 0 we have that [ t (0) l + ν ] -1 ∼ 2 Γ( + l ν )Γ( + l ν +1) ( r 0 2 ) 2( + l ν ) for l > 0 and [ t (0) 0 ] -1 ∼ 2(1 -δ 0 ν ) Γ( ν )Γ( ν +1) ( r 0 2 ) 2 ν + m 2 2 π ρ δ 0 ν 0 ( r 0 ) , where ρ 0 ( r 0 ) = ρ 0 ( a = r 0 ), for l = 0. Although such terms vanish as r 0 → 0, a term of the form [ t (0) l + ν ] -1 [ K ( r ) l + ν ] 2 does not necessarily cancel for small values of r because K ( r ) l + ν diverges. These contributions can be neglected, however, when r /greatermuch r 0 ; in this limit
Figure 4: Density of the negative charge for four different fractional impurities located at the origin. The size of the impurity is mr 0 = 0 05 and the size of the disk is . mR = 30. We used m -2 ρ 0 = 0 87 . .
we have
$$\text{we have} \frac { 2 \pi } { m ^ { 2 } } \rho _ { + } ^ { ( \nu +, 0, \tilde { R } ) } ( \mathbf r ) & = \frac { 2 \pi } { m ^ { 2 } } \tilde { p } _ { \nu } ^ { 0 } + \sum _ { l = 1 } ^ { \infty } t _ { l - \nu - 1 } t _ { l - \nu - 1 } ^ { ( R ) } [ \, I _ { l - \nu } ] ^ { 2 } + \sum _ { l = 1 } ^ { k - 1 } t _ { l + \nu + 1 } [ \, I _ { l + \nu } ] ^ { 2 } - \sum _ { l = k } ^ { \infty } t _ { l + \nu } ^ { ( R ) } [ \, I _ { l + \nu } ^ { ( r ) } ] ^ { 2 } } \\ & - ( 1 - \delta _ { 0 k } ) \frac { m ^ { 2 } \frac { \delta _ { \mu } } { 2 \pi } \rho _ { ( \mathbf r ) } [ K _ { 0 } ^ { ( r ) } ] ^ { 2 } } { 1 + \frac { m ^ { 2 } \frac { \delta _ { \mu } } { 2 \pi } \rho _ { ( \mathbf r ) } t _ { 1 } ^ { ( R ) } } - ( 1 - \delta _ { 0 k } ) t _ { 1 + \nu } \frac { m ^ { 2 } \frac { m ^ { 2 } } { 2 \pi } \frac { \delta _ { \mu } } { 2 \pi } \rho _ { ( \mathbf r ) } T _ { 0 } ^ { ( r ) } }, \, ( 4 2 ) \\ & - ( 1 - \delta _ { 0 k } ) \frac { m ^ { 2 } \frac { \delta _ { \mu } } { 2 \pi } \delta _ { \mu } } { 1 + \frac { m ^ { 2 } \frac { \delta _ { \mu } } { 2 \pi } \rho _ { ( \mathbf r ) } t _ { 1 } ^ { ( R ) } } - \sum _ { l = 1 } ^ { \infty } t _ { l - \nu } ^ { ( R ) } [ \, I _ { l - \nu } ] ^ { 2 } - \sum _ { l = 2 } ^ { k } t _ { l + \nu } [ \, I _ { l + \nu } ] ^ { 2 } + \sum _ { l = k + 1 } ^ { \infty } t _ { l + \nu - 1 } [ \, I _ { l + \nu } ] ^ { 2 } \\ & - t _ { \nu } ^ { ( R ) } \frac { 2 d _ { \nu } ^ { 0 } } { 1 + \nu } ^ { ( r ) } K _ { \nu } ^ { ( r ) } + [ I _ { \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + d _ { \nu } ^ { 0 } t _ { \nu } ^ { ( R ) } } - ( 1 - \delta _ { 0 k } ) t _ { 1 + \nu } ^ { ( R ) } \frac { 2 m ^ { 2 } \frac { \delta _ { \mu } } { 2 \pi } \rho _ { ( \mathbf r ) } T _ { 0 } ^ { ( r ) } } { 1 + \frac { m ^ { 2 } \frac { \delta _ { \mu } } { 2 \pi } \rho _ { ( \mathbf r ) } t _ { 1 } ^ { ( R ) } } \\ & + \frac { d ^ { 0 } [ K _ { \nu } ^ { ( r ) } ] ^ { 2 } } { 1 + d _ { \nu } ^ { 0 } t _ { \nu } ^ { ( R ) } } + ( 1 - \delta _ { 0 k } ) \frac { m ^ { 2 } \frac { m ^ { 2 } } { 2 \pi } \rho _ { ( \mathbf r ) } [ K _ { 1 } ^ { ( r ) } ] ^ { 2 } } { 1 + \frac { m ^ { 2 } \frac { \delta _ { \mu } } { 2 \pi } \rho _ { ( \mathbf r ) } t _ { 1 } ^ { ( R ) } }. \\ \text{Another interesting limit is the case $R$} \to \infty. \, \text{In this limit} \, t _ { l + \nu + 1 } ^ { ( R ) } \to 0 \, \text{ but terms of the form} \\ \tilde { ( R ) } _ { 1 } \, \mathbf r ( r ) _ { 1 } ^ { 2 } \, \mathbf a \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \mathbf w \, \matrix }$$
Another interesting limit is the case R →∞ . In this limit t ( R ) l + ν → 0 but terms of the form t ( R ) l [ I ( r ) l ] 2 do not necessarily vanish as r → ∞ . However, if we consider the case r /lessmuch R
Table 1: Slopes of the plots ln( m -2 ρ -( r )) vs. ln( r/r 0 ). The slope describes the linear behavior present in the curves shown, for instance, in figure 4; we considered different values for the charge of the impurity n . While the slope takes the value Γ n = -nβe 2 = -2 n for n < 1 and r 0 → 0, the slope does not satisfy this relation as n > 1 or r 0 is large enough. The larger n the lesser the slope approximates to -2 n . For n = 1 the slope also satisfies this condition. Our results match the predictions stated in [23] found when Γ < 2 and r 0 → 0. We used m -2 ρ 0 = 0 87 and . mR = 30.
| mr 0 | n | n | n | n | n | n | n | n | n | n | n | n |
|--------|----------|----------|----------|----------|----------|----------|----------|----------|----------|----------|----------|----------|
| | 7 | 5 . 2 | 4 | 3 . 5 | 2 . 7 | 2 | 1 . 4 | 1 | 0 . 8 | 0 . 6 | 0 . 4 | 0 . 2 |
| 10 - 1 | - 8 . 92 | - 7 . 25 | - 5 . 88 | - 5 . 33 | - 4 . 37 | - 3 . 43 | - 2 . 47 | - 1 . 65 | - 1 . 23 | - 0 . 81 | - 0 . 40 | - 0 . 12 |
| 10 - 3 | - 9 . 11 | - 7 . 42 | - 6 . 13 | - 5 . 52 | - 4 . 65 | - 3 . 81 | - 2 . 80 | - 2 . 00 | - 1 . 60 | - 1 . 20 | - 0 . 80 | - 0 . 40 |
| 10 - 6 | - 9 . 15 | - 7 . 44 | - 6 . 20 | - 5 . 53 | - 4 . 66 | - 3 . 90 | - 2 . 80 | - 2 . 00 | - 1 . 60 | - 1 . 20 | - 0 . 80 | - 0 . 40 |
we find
$$\frac { 2 \pi } { m ^ { 2 } } \rho _ { + } ^ { ( \nu _ { + }, \tilde { r } _ { 0 }, \infty ) } ( \mathbf r ) & = \frac { 2 \pi } { m ^ { 2 } } \hat { \rho } _ { \nu } ^ { 0 } - \sum _ { l = 1 } ^ { \infty } [ t _ { l - \nu } ^ { ( 0 ) } ] ^ { - 1 } [ K _ { l - \nu } ^ { ( r ) } ] ^ { 2 } - \sum _ { l = 0 } ^ { k - 1 } [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } [ K _ { l + \nu } ^ { ( r ) } ] ^ { 2 } \\ & \quad + \sum _ { l = k } ^ { \infty } [ t _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } [ K _ { l + \nu } ^ { ( r ) } ] ^ { 2 }, \\ \frac { 2 \pi } { \lambda } \varrho _ { + } ^ { ( \nu _ { + }, \tilde { r } _ { 0 }, \infty ) } ( \mathbf r ) & = \frac { 2 \pi } { \lambda ^ { 0 } } \varrho _ { + } \sum _ { l = 1 } ^ { \infty } [ \mathbf r ^ { ( 0 ) } ) \quad \, 1 ^ { - 1 } [ K ^ { ( r ) } ] ^ { 2 } \, + \sum _ { l = 1 } ^ { k } [ \mathbf r ^ { ( r ) } ] ^ { 2 }$$
$$\overline { l = k } ^ { \overline { l } = k } \\ \frac { 2 \pi } { m ^ { 2 } } \rho _ { - } ^ { ( \nu _ { + }, \tilde { r } _ { 0 }, \infty ) } ( \mathbf r ) & = \frac { 2 \pi } { m ^ { 2 } } \hat { \rho } _ { \nu } ^ { 0 } + \sum _ { l = 1 } ^ { \infty } [ t _ { l - \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } [ K _ { l - \nu } ^ { ( r ) } ] ^ { 2 } + \sum _ { l = 0 } ^ { k } [ t _ { l + \nu - 1 } ^ { ( 0 ) } ] ^ { - 1 } [ K _ { l + \nu } ^ { ( r ) } ] ^ { 2 } \\ & \quad - \sum _ { l = k + 1 } ^ { \infty } [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } [ K _ { l + \nu } ^ { ( r ) } ] ^ { 2 }. \\ \text{Finally, we can combine both limits } r _ { 0 } \to 0 \text{ and } R \to \infty. \text{ In the interval } r _ { 0 } \ll r \ll R \text{ we}$$
Finally, we can combine both limits r 0 → 0 and R →∞ . In the interval r 0 /lessmuch r /lessmuch R we get
$$\rho _ { + } ^ { ( \nu _ { + }, 0, \infty ) } ( \mathbf r ) & = \hat { \rho } _ { \nu } ^ { 0 } - \left ( \frac { m ^ { 2 } } { 2 \pi } \right ) ^ { 2 } \frac { ( 1 - \delta _ { 0 k } ) \delta _ { 0 \nu } } { \rho ^ { 0 } ( r _ { 0 } ) } [ K _ { 0 } ^ { ( r ) } ] ^ { 2 }, \\ \omega _ { + } \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \quad \, m ^ { 2 } \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda \, \Lambda.$$
$$\rho _ { - } ^ { ( \nu _ { + }, 0, \infty ) } ( { \mathbf r } ) = \hat { \rho } _ { \nu } ^ { 0 } + \left ( \frac { m ^ { 2 } } { 2 \pi } \right ) ^ { 2 } \frac { ( 1 - \delta _ { 0 k } ) \delta _ { 0 \nu } } { \rho ^ { 0 } ( r _ { 0 } ) } [ K _ { 1 } ^ { ( r ) } ] ^ { 2 } + \frac { m ^ { 2 } } { 2 \pi } d _ { \nu } ^ { 0 } [ K _ { \nu } ^ { ( r ) } ] ^ { 2 }.$$
Table 2: Value of ln( m -2 ρ -( r )) at r = r 0 for different values of r 0 and n . As expected, the smaller r 0 , the larger the negative density at r = r 0 because the collapse of the counterions is stronger. For n > 1 the values of the negative density at r 0 take a very large value as r 0 → 0. For n < 1 such values are also large but notably decrease because of the decrease of negative ions already described. We used m -2 ρ 0 = 0 87 and . mR = 30.
| mr 0 | n | n | n | n | n | n | n | n | n |
|---------|--------|--------|--------|--------|--------|--------|---------|-----------------|-----------------|
| | 7 | 5 . 3 | 3 . 5 | 2 | 1 | 0 . 7 | 0 . 5 | 0 . 3 | 0 . 1 |
| 10 - 1 | 6 . 51 | 5 . 91 | 5 . 01 | 3 . 66 | 1 . 94 | 1 . 15 | 0 . 586 | 3 . 76 × 10 - 2 | - 0 . 464 |
| 10 - 2 | 11 . 1 | 10 . 5 | 9 . 55 | 8 . 16 | 5 . 82 | 4 . 11 | 2 . 75 | 1 . 37 | 7 . 11 × 10 - 3 |
| 10 - 3 | 15 . 7 | 15 . 1 | 14 . 2 | 12 . 7 | 10 . 0 | 7 . 29 | 5 . 07 | 2 . 82 | 0 . 607 |
| 10 - 4 | 20 . 3 | 19 . 7 | 18 . 8 | 17 . 3 | 14 . 4 | 10 . 5 | 7 . 37 | 4 . 23 | 1 . 22 |
| 10 - 5 | 24 . 9 | 24 . 3 | 23 . 4 | 21 . 9 | 18 . 7 | 13 . 7 | 9 . 67 | 5 . 62 | 1 . 80 |
| 10 - 6 | 29 . 5 | 28 . 9 | 28 . 0 | 26 . 6 | 23 . 2 | 16 . 9 | 12 . 0 | 7 . 01 | 2 . 34 |
| 10 - 7 | 34 . 1 | 33 . 5 | 32 . 6 | 31 . 1 | 27 . 6 | 20 . 2 | 14 . 3 | 8 . 39 | 2 . 87 |
| 10 - 8 | 38 . 7 | 38 . 1 | 37 . 2 | 35 . 7 | 32 . 1 | 23 . 4 | 16 . 6 | 9 . 77 | 3 . 37 |
| 10 - 9 | 43 . 3 | 42 . 8 | 41 . 8 | 40 . 3 | 36 . 6 | 26 . 6 | 18 . 9 | 11 . 2 | 3 . 87 |
| 10 - 10 | 48 . 0 | 47 . 4 | 46 . 4 | 44 . 9 | 41 . 1 | 29 . 8 | 21 . 2 | 12 . 5 | 4 . 35 |
From Eqs. (46) and (47) we can obtain the particular cases
$$\rho _ { + } ^ { ( \nu _ { + }, 0, \infty ) } ( \mathbf r ) = \begin{cases} \begin{array} { c c } \rho ^ { 0 }, & n = 0 \\ \frac { \rho _ { + + } ^ { ( 2 ) 0 } ( r, 0 ) } { \tilde { \rho } _ { \nu } ^ { 0 }, & n \geq 1, n \in \mathbb { Z } } \\ \tilde { \rho } _ { \nu } ^ { 0 }, & n \notin \mathbb { Z } \\ \end{array} \\ \tilde { \rho } _ { \nu } ^ { 0 }, & n \notin \mathbb { Z } \\ \end{cases} \quad. \quad \end{cases}$$
where ρ (2)0 + ± ( r, 0) are the two-point correlation functions for the unperturbed plasma [1]. Notice that for n ≥ 1 and n ∈ Z both the positive and negative densities do not depend on n (their respective densities are the same for any integer charge). This takes place because an amount of k -1 counter-ions screen the impurity as their collapse cannot be avoided in the limit r 0 → 0 leaving a charge n = 1 unscreened.
$$\rho _ { + } ^ { ( \nu _ { + }, \i, \infty ) } ( \mathbf r ) & = \begin{cases} \frac { \rho _ { + + } ^ { ( r, 0 ) } ( r, 0 ) } { \rho _ { \nu } ^ { 0 }, } & n \geq 1, n \in \mathbb { Z } & ( 4 8 ) \\ \hat { \rho } _ { \nu } ^ { 0 }, & n \notin \mathbb { Z } & \\ \rho _ { - } ^ { ( \nu _ { + }, 0, \infty ) } ( \mathbf r ) & = \begin{cases} \rho _ { 0 } ^ { 0 }, & n = 0 \\ \frac { \rho _ { + + } ^ { ( 2 ) 0 } ( r, 0 ) } { \rho _ { 0 } ^ { 0 } ( r _ { 0 } ) }, & n \geq 1, n \in \mathbb { Z } & ( 4 9 ) \\ \hat { \rho } _ { \nu } ^ { 0 } + \frac { m ^ { 2 } } { 2 \pi } d _ { \nu } ^ { 0 } [ K _ { \nu } ^ { ( r ) } ] ^ { 2 }, & n \notin \mathbb { Z } & \\ \end{cases} \\ o _ { + + } ^ { ( 2 ) 0 } ( r, 0 ) \, \text{are the two-point correlation functions for the unperturbed plasma} [ 1 ]. \text{that for $n>1$ and $n\in \mathbb{ Z}$ both the positive and negative densities do not depend}$$
ln
Figure 5: Density of the negative charge for four different fractional impurities located at the origin. The size of the impurity is mr 0 = 0 4 and the size of the disk is . mR = 30. Notice that the negative density close to the impurity decreases due to the reduction of counter-ions when r 0 increases. A large impurity avoids the strong collapse of counter-ions due to its hard core effects; this effect does not occur as r 0 → 0. We used m -2 ρ 0 = 0 87 . .
## 5 Integrated Charge
We now focus on finding the integrated charge of the system contained in an annulus region from r 0 up to a distance r > r 0 . Let e ± ( r, r 0 ) = ∫ r r 0 ∫ 2 π 0 ρ ± ( r ) d 2 r . We now define the following functions
$$a _ { j } ( r, r _ { 0 } ) & = \frac { m ^ { 2 } } { 2 \pi } \int _ { r _ { 0 } } ^ { r } \int _ { 0 } ^ { 2 \pi } d ^ { 2 } r ^ { \prime } [ K _ { j } ^ { ( r ^ { \prime } ) } ] ^ { 2 } = \frac { x ^ { 2 } } { 2 } \left [ [ K _ { j } ^ { ( x ) } ] ^ { 2 } - K _ { j - 1 } ^ { ( x ) } K _ { j + 1 } ^ { ( x ) } \right ] \right ] _ { \tilde { r } _ { 0 } } ^ { \tilde { \ } }, \\ b _ { j } ( r, r _ { 0 } ) & = \frac { m ^ { 2 } } { \varpi } \, \int _ { \ } ^ { r } \, \int _ { \ } ^ { 2 \pi } d ^ { 2 } r ^ { \prime } [ I _ { i } ^ { ( r ^ { \prime } ) } ] ^ { 2 } = \frac { x ^ { 2 } } { \varpi } \left [ [ I _ { i } ^ { ( x ) } ] ^ { 2 } - I _ { i - 1 } ^ { ( x ) } I _ { i + 1 } ^ { ( x ) } \right ] \right ] ^ { \tilde { r } },$$
$$\nu _ { j } \vee, \, \varphi _ { j } & = 2 \pi \, J _ { r _ { 0 } } \, J _ { 0 } ^ { \quad \lambda \, \varphi \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \rightarrow \, - } \, 2 \, \downarrow \, ^ { \downarrow \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \, j \} \\ c _ { j } ( r, r _ { 0 } ) & = \frac { m ^ { 2 } } { 2 \pi } \int _ { r _ { 0 } } ^ { r } \int _ { 0 } ^ { 2 \pi } d ^ { 2 } r ^ { \prime } I _ { j } ^ { ( r ^ { \prime } ) } K _ { j } ^ { ( r ^ { \prime } ) }, \\ \varphi _ { j } & = \int _ { j } ^ { r } \int _ { \varphi \, \varphi \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downnull \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downarrow \, \downless$$
$$a _ { j } ( r, r _ { 0 } ) & = \frac { } { 2 \pi } \int _ { r _ { 0 } } \int _ { 0 } ^ { \ a c r \left [ \Lambda _ { j } ^ { \prime } \right ] ^ { \lambda } = \frac { } { 2 } \left [ \left [ \Lambda _ { j } ^ { \prime } \right ] ^ { \lambda } = \Lambda _ { j - 1 } ^ { \prime } \Lambda _ { j + 1 } \right ] \right | _ { \tilde { r } _ { 0 } }, \\ b _ { j } ( r, r _ { 0 } ) & = \frac { m ^ { 2 } } { 2 \pi } \int _ { r _ { 0 } } ^ { r } \int _ { 0 } ^ { 2 \pi } d ^ { 2 } r ^ { \prime } [ \, I _ { j } ^ { ( r ^ { \prime } ) } ] ^ { 2 } = \frac { x ^ { 2 } } { 2 } \left [ [ I _ { j } ^ { ( x ) } ] ^ { 2 } - I _ { j - 1 } ^ { ( x ) } I _ { j + 1 } ^ { ( x ) } \right ] \right | _ { \tilde { r } _ { 0 } } ^ { \tilde { r } }, \\ c _ { j } ( r, r _ { 0 } ) & = \frac { m ^ { 2 } } { 2 \pi } \int _ { r _ { n } } ^ { r } \int _ { 0 } ^ { 2 \pi } d ^ { 2 } r ^ { \prime } I _ { j } ^ { ( r ^ { \prime } ) } K _ { j } ^ { ( r ^ { \prime } ) },$$
$$e _ { 0 } ( r, r _ { 0 } ) = \int _ { r _ { 0 } } ^ { r } \int _ { 0 } ^ { 2 \pi } d ^ { 2 } r ^ { \prime } \rho ^ { 0 } = \frac { \tilde { r } ^ { 2 } - \tilde { r } _ { 0 } ^ { 2 } } { 2 } \left ( \ln \frac { 2 } { m a } - \gamma \right ). \\ \text{...} \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad.$$
The total positive and negative charges in the annulus region can easily be found by integrating the densities found in the previous section and using Eqs. (50)-(53). The
integrated charge, from r 0 up to a distance r , is given by e t ( r, r 0 ) = ∫ r r 0 ∫ 2 π 0 d 2 r ′ ( ρ + ( r ′ ) -ρ -( r ′ )). Its general expression is given by
$$( 5 4 )$$
$$\text{integrated charge, from } r _ { 0 } \, \rightarrow \, \text{to a distance } r, \, \text{is given by } e _ { t } ( r, r _ { 0 } ) = \int _ { r _ { 0 } } ^ { r } \int _ { 0 } ^ { 2 \pi } d ^ { 2 } r ^ { \prime } ( \rho _ { + } ( r ^ { \prime } ) - \rho _ { - } ( r ^ { \prime } ) ). \, \text{Its general expression is given by} \\ e _ { t } ( \nu _ { +, r _ { 0 }, \bar { R } } ) ( r ) & = - \sum _ { l = 0 } ^ { k - 1 } \frac { [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } A _ { l + \nu } ( r, r _ { 0 } ) } { t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu + 1 } } - \sum _ { l = 0 } ^ { k - 1 } t _ { l + \nu + 1 } ^ { ( R _ { l } ) } \frac { 2 [ t _ { l + + \nu } ^ { ( 0 ) } ] ^ { - 1 } C _ { l + \nu } ( r, r _ { 0 } ) - B _ { l + \nu } ( r, r _ { 0 } ) } { 1 + [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu + 1 } ^ { ( R _ { l } ) } - t _ { l + \nu + 1 } ^ { ( R _ { l } ) } \\ & \quad - \sum _ { l = 1 } ^ { \infty } \frac { [ t _ { l } ^ { ( 0 ) } \nu _ { + 1 } ] ^ { - 1 } A _ { l - \nu } ( r, r _ { 0 } ) } { 1 + [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu } ^ { ( R _ { l } ) } - \sum _ { l = k } ^ { \infty } t _ { l + \nu } ^ { ( R _ { l } ) } \frac { 2 [ t _ { l } ^ { ( 0 ) } \nu _ { + 1 } ] ^ { - 1 } C _ { l - \nu } ( r, r _ { 0 } ) + B _ { l - \nu } ( r, r _ { 0 } ) } { 1 + [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu } ^ { ( R _ { l } ) } - t _ { l + \nu } ^ { ( R _ { l } ) } } \\ & \quad + \sum _ { l = k } ^ { \infty } \frac { [ t _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } A _ { l + \nu } ( r, r _ { 0 } ) } { 1 + [ t _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu } ^ { ( R _ { l } ) } - \sum _ { l = k } ^ { \infty } t _ { l + \nu } ^ { ( R _ { l } ) } \frac { 2 [ t _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } C _ { l + \nu } ( r, r _ { 0 } ) + B _ { l + \nu } ( r, r _ { 0 } ) } { 1 + [ t _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu } ^ { ( R _ { l } ) } } \\ & \quad - \frac { [ t _ { l + \nu - 1 } ^ { ( 0 ) } ] ^ { - 1 } a _ { \nu } ( r, r _ { 0 } ) } { 1 + [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu } ^ { ( R _ { l } ) } } \\ & \quad - \frac { [ t _ { l + \nu - 1 } ^ { ( 0 ) } ] ^ { - 1 } a _ { 1 - \nu } ( r, r _ { 0 } ) } { 1 + [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu } ^ { ( R _ { l } ) } - \sum _ { l = k } ^ { \infty } t _ { l + \nu } ^ { ( R _ { l } ) } - \frac { 2 [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } c _ { 1 - \nu } ( r, r _ { 0 } ) - b _ { 1 - \nu } ( r, r _ { 0 } ) } { 1 + [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu } ^ { ( R _ { l } ) } } \\ \text{where we used the definitions} \\ a _ { l + \nu } ( r, r _ { 0 } ) + a _ { l + \nu + 1 } ( r, r _ { 0 } ) = - x K _ { l + \nu } ^ { ( x _ { l + \nu } ) } K _ { l + \nu + 1 } ^ { ( x _ { l + \nu }, \nu _ { + 1 } ) } ^ { ( x _ { l + \nu } ) } ^ { ( 0 ) } \, \equiv A _ { l + \nu } ( r, r _ { 0 } ),$$
where we used the definitions
$$\intertext { use w as u m u m u m u m } a _ { l + \nu } ( r, r _ { 0 } ) + a _ { l + \nu + 1 } ( r, r _ { 0 } ) & = - x K _ { l + \nu } ^ { ( x ) } K _ { l + \nu + 1 } ^ { ( x ) } \Big | _ { r _ { 0 } } ^ { r } \equiv A _ { l + \nu } ( r, r _ { 0 } ), \\ b _ { l + \nu } ( r, r _ { 0 } ) + b _ { l + \nu + 1 } ( r, r _ { 0 } ) & = x I _ { i, \dots } ^ { ( x ) } I _ { i, \dots + 1 } ^ { ( x ) } \Big | ^ { r } \ \equiv B _ { l + \nu } ( r, r _ { 0 } ),$$
$$b _ { l + \nu } ( r, r _ { 0 } ) + b _ { l + \nu + 1 } ( r, r _ { 0 } ) & = x I _ { l + \nu } ^ { ^ { \prime } } I _ { l + \nu + 1 } ^ { ^ { \prime } } \Big | _ { r _ { 0 } } \equiv B _ { l + \nu } ( r, r _ { 0 } ), & ( 0 ) \\ c _ { l + \nu } ( r, r _ { 0 } ) - c _ { l + \nu + 1 } ( r, r _ { 0 } ) & \equiv C _ { l + \nu } ( r, r _ { 0 } ). & ( 5 7 )$$
$$a _ { l + \nu } ( r, r _ { 0 } ) + a _ { l + \nu + 1 } ( r, r _ { 0 } ) & = - x K _ { l + \nu } ^ { ( x ) } K _ { l + \nu + 1 } ^ { ( x ) } \Big | _ { r _ { 0 } } ^ { r } \equiv A _ { l + \nu } ( r, r _ { 0 } ), \\ b _ { l + \nu } ( r, r _ { 0 } ) + b _ { l + \nu + 1 } ( r, r _ { 0 } ) & = x I _ { l + \nu } ^ { ( x ) } I _ { l + \nu + 1 } ^ { ( x ) } \Big | _ { r _ { 0 } } ^ { r } \equiv B _ { l + \nu } ( r, r _ { 0 } ), \\ c _ { l + \nu } ( r, r _ { 0 } ) - c _ { l + \nu + 1 } ( r, r _ { 0 } ) & \equiv C _ { l + \nu } ( r, r _ { 0 } ).$$
## 5.1 Integrated Charge at the boundary
An interesting fact is evaluating the total integrated charge at the boundary of the system, i.e., at r = R . We should expect this charge to vanish because, excluding the impurity, the system is neutral. We can easily check that in the limits ˜ R /greatermuch 1 and ˜ r 0 /lessmuch 1 we have
$$A _ { l + \nu } ( R, r _ { 0 } ) = \tilde { r } _ { 0 } K _ { l + \nu + 1 } ( \tilde { r } _ { 0 } ) K _ { l + \nu } ( \tilde { r } _ { 0 } ) + O ( e ^ { - \tilde { R } } ),$$
$$B _ { l + \nu } ( R, r _ { 0 } ) = \tilde { R } I _ { l + \nu + 1 } ( \tilde { R } ) I _ { l + \nu } ( \tilde { R } ) + O ( \tilde { r } _ { 0 } ^ { 2 ( l + \nu + 1 ) } ).$$
Since t ( R ) l + +1 ν ∼ e -2 ˜ R , the contributions coming from the terms C R,r l ( 0 ) are exponentially small and so can be neglected. We thus have
$$& \text{small and so can be neglected. We thus have} \\ & e _ { t } ^ { ( \nu _ { + }, \bar { r } _ { 0 }, \bar { R } ) } ( R ) = - \sum _ { l = 1 } ^ { \infty } \frac { \tilde { r } _ { 0 } I _ { - \nu + 1 } ^ { ( 0 ) } K _ { l - \nu } ^ { ( 0 ) } } { 1 + [ t _ { l - \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } t _ { l - \nu } ^ { ( R ) } } - \sum _ { l = 0 } ^ { k - 1 } \frac { \tilde { r } _ { 0 } I _ { + \nu } ^ { ( 0 ) } K _ { l + \nu + 1 } ^ { ( 0 ) } } { 1 + [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu + 1 } ^ { ( R ) } } + \sum _ { l = k } ^ { \infty } \frac { \tilde { r } _ { 0 } I _ { + \nu + 1 } ^ { ( 0 ) } K _ { l + \nu } ^ { ( 0 ) } } { 1 + [ t _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu } ^ { ( R ) } } \\ & \quad + \sum _ { l = 1 } ^ { \infty } \frac { \tilde { R } K _ { l - \nu } ^ { ( R ) } I _ { - \nu } ^ { ( R ) } I _ { - \nu + 1 } ^ { ( R ) } } { 1 + [ t _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } t _ { l - \nu } ^ { ( R ) } } + \sum _ { l = 0 } ^ { k - 1 } \frac { \tilde { R } K _ { l + \nu + 1 } ^ { ( R ) } I _ { + \nu } ^ { ( R ) } } { 1 + [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu + 1 } ^ { ( R ) } } - \sum _ { l = k } ^ { \infty } \frac { \tilde { R } K _ { l + \nu } ^ { ( R ) } I _ { + \nu } ^ { ( R ) } } { 1 + [ t _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } t _ { l + \nu } ^ { ( R ) } } \\ & \quad - \frac { \tilde { r } _ { 0 } I _ { 1 - \nu } ^ { ( 0 ) } K _ { \nu } ^ { ( 0 ) } + d _ { 0 } ^ { 0 } a _ { \nu } ( R, r _ { 0 } ) } { 1 + t _ { \nu } ^ { ( 0 ) } ( d _ { \nu } ^ { 0 } + [ t _ { 1 - \nu } ^ { ( 0 ) } ] ^ { - 1 } ) } + \frac { \tilde { R } K _ { \nu } ^ { ( R ) } I _ { 1 - \nu } ^ { ( R ) } } { 1 + t _ { \nu } ^ { ( R ) } ( d _ { \nu } ^ { 0 } + [ t _ { 1 - \nu } ^ { ( 0 ) } ] ^ { - 1 } ) } + O ( e ^ { - \tilde { R }, r _ { 0 } ^ { 2 } ( l + 1 ) } ). ( 6 0 )$$
Figure 6: Integrated charge in the plasma for three different impurities located at the origin. The size of the impurity is mr 0 = 0 05 and the size of the disk is . mR = 10.
In the same approximation all the denominators behave like ∼ 1 1+ O e ( -˜ R ) = 1+ O e ( -˜ R ). So
$$\ast \cdots \nu _ { \bullet } \quad, \\ e _ { t } ^ { ( \nu _ { + }, \bar { r } _ { 0 }, \bar { R } ) } ( R ) & = - \tilde { r } _ { 0 } \sum _ { l = 0 } ^ { k - 1 } \left [ I _ { l - \nu + 1 } ^ { ( 0 ) } K _ { l - \nu } ^ { ( 0 ) } + I _ { l + \nu } ^ { ( 0 ) } K _ { l + \nu + 1 } ^ { ( 0 ) } \right ] + \tilde { r } _ { 0 } \sum _ { l = k } ^ { \infty } \left [ I _ { l + \nu + 1 } ^ { ( 0 ) } K _ { l + \nu } ^ { ( 0 ) } - I _ { l - \nu + 1 } ^ { ( 0 ) } K _ { l - \nu } ^ { ( 0 ) } \right ] \\ & \quad + \tilde { R } \sum _ { l = 0 } ^ { k - 1 } \left [ K _ { l - \nu } ^ { ( R ) } I _ { l - \nu + 1 } ^ { ( R ) } + K _ { l + \nu + 1 } ^ { ( R ) } I _ { l + \nu } ^ { ( R ) } \right ] - \tilde { R } \sum _ { l = k } ^ { \infty } \left [ K _ { l + \nu } ^ { ( R ) } I _ { l + \nu + 1 } ^ { ( R ) } - K _ { l - \nu } ^ { ( R ) } I _ { l - \nu + 1 } ^ { ( R ) } \right ] \\ & \quad - d _ { \nu } ^ { 0 } a _ { \nu } ( R, r _ { 0 } ) + O ( e ^ { - \tilde { R } }, \tilde { r } _ { 0 } ^ { 2 ( l + 1 ) } ).$$
/negationslash
When ν = 0, it is easy to see, using the Wronskian relations for the Bessel functions [24], that e (0 ˜ ,r 0 ,R ˜ ) t + ( R ) = 0, so the total charge of the system is ne (the charge of the impurity). This means that the charge distribution generated by the impurity close to the origin is compensated by a charge distribution of opposite sign in the boundary. For ν = 0 the charge at r = R does not simplify and so numerical solutions must be found; in this case we also find e ( ν,r ˜ 0 ,R ˜ ) t ( R ) = 0.
/negationslash
Figures 6 and 7 show the integrated charge for different values of the impurity charge n . By definition it starts with a zero value at r = r 0 = 0. Then it decreases to reach a valley with a value approaching -n as expected since the plasma screens the impurity. Then, close the outer boundary ( r → R ), the integrated charge increases to reach the zero value, as discussed above.
/Minus
1.0
Figure 7: Integrated charge in the plasma for four different impurities located at the origin. The size of the impurity is mr 0 = 0 05 and the size of the disk is . mR = 10.
## 5.2 Asymptotic Limits
We will evaluate the same asymptotic cases than those analyzed in section 4. For r 0 → 0 and r /greatermuch r 0 the net density becomes
$$& \text{and} \ r \gg r _ { 0 } \text{ the net density becomes} \\ & e _ { t } ^ { ( \nu _ { + }, 0, \bar { R } ) } ( r ) = & - k + \delta _ { 0 \nu } \left ( 1 - \frac { 1 } { 1 + \frac { m ^ { 2 } } { 2 \pi } \frac { \delta _ { 0 \nu } } { \rho ^ { 0 } ( r _ { 0 } ) } t _ { 1 + \nu } ^ { ( R ) } } \right ) - ( 1 - \delta _ { 0 k } ) ^ { t _ { 1 + \nu } } _ { t _ { 1 + \nu } } \frac { 2 m ^ { 2 } } { \rho ^ { \rho } ( r _ { 0 } ) } c _ { 0 } ( r _ { 0 } ) - B _ { \nu } ( r _ { 0 } ) } { 1 + \frac { m ^ { 2 } } { 2 \pi } \frac { \delta _ { 0 \nu } } { \rho ^ { 0 } ( r _ { 0 } ) } t _ { 1 + \nu } ^ { t _ { 1 + \nu } } } \\ & \quad + r \sum _ { l = 1 } ^ { k - 1 } t _ { l + \nu + 1 } ^ { ( r ) } I _ { l + \nu + 1 } ^ { ( r ) } I _ { l + \nu } ^ { ( r ) } + r \sum _ { l = 1 } ^ { \infty } t _ { l - \nu } ^ { ( r ) } I _ { l - \nu } ^ { ( r ) } I _ { l - \nu + 1 } ^ { ( r ) } I _ { l - \nu } ^ { ( r ) } - r \sum _ { l = k } ^ { \infty } t _ { l + \nu } ^ { ( r ) } I _ { l + \nu + 1 } ^ { ( r ) } I _ { l + \nu } ^ { ( r ) } \\ & \quad - \frac { d _ { \nu } ^ { 0 } a _ { \nu } ( r, 0 ) } { 1 + d _ { \nu } ^ { 0 } t _ { \nu } ^ { ( R ) } } + t _ { \nu } ^ { ( R ) } \frac { 2 d _ { \nu } ^ { 0 } c _ { \nu } ( r, 0 ) + b _ { \nu } ( r, 0 ) + b _ { 1 - \nu } ( r, 0 ) } { 1 + d _ { \nu } ^ { 0 } t _ { \nu } ^ { ( R ) } } \\ & \text{When} \ R \to \infty \text{ and} \ r \ll R \text{ we have}$$
When R →∞ and r /lessmuch R we have
$$e _ { t } ^ { ( \nu _ { + }, \bar { r } _ { 0 }, \infty ) } ( r ) & = - \sum _ { l = 0 } ^ { k - 1 } [ t _ { l + \nu } ^ { ( 0 ) } ] ^ { - 1 } A _ { l + \nu } ( r, r _ { 0 } ) - \sum _ { l = 1 } ^ { \infty } [ t _ { l - \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } A _ { l - \nu } ( r, r _ { 0 } ) \\ & \quad + \sum _ { l = k } ^ { \infty } [ t _ { l + \nu + 1 } ^ { ( 0 ) } ] ^ { - 1 } A _ { l + \nu } ( r, r _ { 0 } ) - [ t _ { \nu - 1 } ^ { ( 0 ) } ] ^ { - 1 } a _ { \nu } ( r, r _ { 0 } ) - [ t _ { 1 - \nu } ^ { ( 0 ) } ] ^ { - 1 } a _ { 1 - \nu } ( r, r _ { 0 } ) \quad ( 6 3 ) \\ \text{In the combined limits $r_{0} \to 0$ and $R\to \infty$, we get, in the interval $r_{0} \ll r\ll R$}$$
In the combined limits r 0 → 0 and R →∞ , we get, in the interval r 0 /lessmuch r /lessmuch R
$$e _ { t } ^ { ( \nu _ { + }, 0, \infty ) } ( r ) = - \, k - d _ { \nu } ^ { 0 } a _ { \nu } ( r, 0 )$$
/negationslash
/negationslash
Figure 8: Integrated charge in the plasma for four different fractional impurities located at the origin. The impurity is a treated as point particle ( r 0 = 0) according to Eq. (62). The size of the disk is mR = 10.
/negationslash
/negationslash
Figs. 8 and 9 show the integrated charge in the case when r 0 = 0. Fig. 8 is, in essence, similar to Fig. 7 when r 0 = 0. However, Fig. 9 shows a new interesting effect, which is not present in the case when r 0 = 0 (Fig. 6). When n is an integer, the accumulated charge e t ( r ) starts at r = r 0 = 0 with a value equal to -n , instead of zero, as it was the case when r 0 = 0. This means that exactly n ions of charge -e of the plasma (counter-ions) have collapsed at the position r 0 = 0 where the impurity is, and they have completely neutralized it. On the other hand, when n = k + ν is not an integer ( ν = 0), k negative ions will partially neutralize the impurity by collapsing into it at position r 0 = 0, but a fractional charge eν remains to be screened, which cannot be completely neutralized by collapsing into it, because the charges of ions the plasma are integers. Nevertheless, this remaining charge is screened by a difuse layer of plasma ions, of tipical length given by the screening length m -1 . This situation is illustrated with the cases n = 1 7 and . n = 2 4 . in Fig. 9.
/negationslash
Fig. 10 shows the integrated charge for r 0 = 0 and R → ∞ . Since the boundary of the large disk R is now at infinity, the difuse layer of remaining charge that usually accumulates close that boundary, observed in the previous figures, is now receded to infinity with the boundary and it cannot be observed anymore in Fig. 10. Fig. 11 shows the combined effects of taking r 0 = 0 and R →∞ .
## 6 Grand Potential and Pressure
In this section we will calculate the grand potential and the partition function for the system we have discussed. Using Eq. (3) we see that the eigenvectors that generate the
Figure 9: Integrated charge in the plasma for four different impurities located at the origin. The impurity is a treated as point particle ( r 0 = 0) according to Eq. (62). The size of the disk is mR = 10.
Figure 10: Integrated charge in the plasma for five different impurities located at the origin. The impurity has a size mr 0 = 0 05 and we consider the limit . R →∞ according to Eq. (63). The density of the gas is constant with value -n for large enough distances.
/Minus
2.5
Figure 11: Integrated charge in the plasma for five different impurities located at the origin. We consider the limits r 0 → 0 and R →∞ according to Eq. (64).
partition function satisfy the equations
$$m _ { + } ( \mathbf r ) g ( \mathbf r ) & = 2 \lambda \partial _ { z } f ( \mathbf r ), \\ m _ { - } ( \mathbf r ) f ( \mathbf r ) & = 2 \lambda \partial _ { \bar { z } } g ( \mathbf r ).$$
Remember that for this problem m ± ( r ) = m ( r ) r L ± 2 n , where
$$& \quad \begin{array} { c } \cdots \cdot \\ \text{is problem } m _ { \pm } ( \mathbf r ) = m ( \mathbf r ) \left ( \frac { r } { L } \right ) ^ { \pm 2 n }, \text{ where } \\ & & \\ m ( \mathbf r ) & = & \begin{cases} 0, & r < r _ { 0 } \\ m, & r > r _ { 0 } \end{cases}. \\ < R \text{ it is more convenient to define the functions } \bar { a } ( \mathbf r ) = e ^ { - V ( \mathbf r ) } a ( \mathbf r ) \end{array}$$
In the region r 0 < r < R it is more convenient to define the functions ¯( g r ) = e -V ( r ) g ( r ) and ¯ ( f r ) = e V ( r ) f ( r ). Decomposing ¯( ) as a Fourier series of the form ¯ ( g r g l r ) = ¯ ( g l r e ) ilθ , we find each mode satisfies the differential equation
$$\left \{ \rho ^ { 2 } \frac { d ^ { 2 } } { d \rho ^ { 2 } } + \rho \frac { d } { d \rho } - \left [ ( l + n ) ^ { 2 } + \rho ^ { 2 } \right ] \right \} \bar { g } _ { l } ( \rho ) = 0, \\ \colon \bar { \underline { r } } \, \text{The canonical collections are given} \, \text{w}$$
with ρ ≡ ˜ r λ . The general solutions are given by
$$- \lambda \cdot \text{**} \, \text{$8$} \, \text{sources} \, \text{sources} \, \text{$5$} \, \text{s} \\ g _ { l } ( \mathbf r ) & = \left ( \frac { r } { L } \right ) ^ { - n } e ^ { i l \theta } \left ( B _ { l } I _ { l + n } ( \rho ) + A _ { l } K _ { l + n } ( \rho ) \right ), \\ f _ { l } ( \mathbf r ) & = \left ( \frac { r } { L } \right ) ^ { n } e ^ { i ( l + 1 ) \theta } \left ( B _ { l } I _ { l + n + 1 } ( \rho ) - A _ { l } K _ { l + n + 1 } ( \rho ) \right ). \\ \text{first region} \, ( r < r _ { 0 } ) \text{ and the third region} \, ( r > R ), \, \text{we have the general solutions}$$
In the first region ( r < r 0 ) and the third region ( r > R ), we have the general solutions
$$g _ { l } ^ { ( 1, 3 ) } ( \mathbf r ) & = u _ { l } ^ { ( 1, 3 ) } e ^ { i l \theta } \left ( \frac { m L } { \lambda } \right ) ^ { n } \left ( \frac { \tilde { r } } { \lambda } \right ) ^ { l }, \\ f _ { i } ^ { ( 1, 3 ) } ( \mathbf r ) & = v _ { i } ^ { ( 1, 3 ) } e ^ { i ( l + 1 ) \theta } \left ( \frac { \lambda } { \alpha } \right ) ^ { n } \left ( \frac { \tilde { r } } { \alpha } \right ) ^ { - l - 1 }.$$
$$f _ { l } ^ { ( 1, 3 ) } ( \mathbf r ) = v _ { l } ^ { ( 1, 3 ) } e ^ { i ( l + 1 ) \theta } \overset { \cdot \cdot \cdot } { \cdot } \overset { \cdot \cdot } { \cdot } \overset { \cdot \cdot } { \cdot } \overset { \cdot \cdot } { \left ( \frac { \lambda } { m L } \right ) } ^ { n } \left ( \frac { \tilde { r } } { \lambda } \right ) ^ { - l - 1 }. \\ \infty$$
(The coefficients 1, 3 refer to the first and third regions.) The boundary conditions imply again that both g (1) l ( r ) and f (1) l ( r ) must be finite at r = 0 while g (3) l ( r ) and f (3) l ( r ) must vanish at r →∞ . This generates the conditions
$$\int _ { \substack { r \searrow \dots \colon \\ r \searrow \dots \colon } } \left ( \begin{array} { c c } - K _ { l + n + 1 } ( \tilde { r } _ { 0 } / \lambda ) & I _ { l + n + 1 } ( \tilde { r } _ { 0 } / \lambda ) \\ K _ { l + n } ( \tilde { R } / \lambda ) & I _ { l + n } ( \tilde { R } / \lambda ) \end{array} \right ) \left ( \begin{array} { c } A _ { l } \\ B _ { l } \end{array} \right ) = \left ( \begin{array} { c } 0 \\ 0 \end{array} \right )$$
for l ≥ 0, and
$$\left ( \begin{array} { c c } K _ { l + n } ( \tilde { r } _ { 0 } / \lambda ) & I _ { l + n } ( \tilde { r } _ { 0 } / \lambda ) \\ - K _ { l + n + 1 } ( \tilde { R } / \lambda ) & I _ { l + n + 1 } ( \tilde { R } / \lambda ) \end{array} \right ) \left ( \begin{array} { c } A _ { l } \\ B _ { l } \end{array} \right ) = \left ( \begin{array} { c } 0 \\ 0 \end{array} \right ) \\ \vdots \quad 1 \quad \tau \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \varphi \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \alpha \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \xi \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \eta \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \lambda \skip \} } \end{array}$$
for l ≤ -1. The solutions are non-vanishing if the determinant of previous matrices is zero. Thus, we have the conditions (after shifting indices and defining z = λ -1 )
$$K _ { l + n + 1 } ( \tilde { r } _ { 0 } z ) I _ { l + n } ( \tilde { R } z ) + K _ { l + n } ( \tilde { R } z ) I _ { l + n + 1 } ( \tilde { r } _ { 0 } z ) = 0, \ \ l \geq 0,$$
$$K _ { l + n - 1 } ( \tilde { r } _ { 0 } z ) I _ { l + n } ( \tilde { R } z ) + K _ { l + n } ( \tilde { R } z ) I _ { l + n - 1 } ( \tilde { r } _ { 0 } z ) = 0, \ l \leq 0.$$
We now again write n as n = k + , with ν k ∈ Z and | ν | < 1. Our conditions now become
$$K _ { l + \nu + 1 } ( \tilde { r } _ { 0 } z ) I _ { l + \nu } ( \tilde { R } z ) + K _ { l + \nu } ( \tilde { R } z ) I _ { l + \nu + 1 } ( \tilde { r } _ { 0 } z ) = 0, \ l \geq k,$$
$$K _ { l - \nu + 1 } ( \tilde { r } _ { 0 } z ) I _ { - l + \nu } ( \tilde { R } z ) + K _ { l - \nu } ( \tilde { R } z ) I _ { - l + \nu - 1 } ( \tilde { r } _ { 0 } z ) = 0, \ l \geq - k.$$
After eliminating the negative l -modes in favor of the positive ones we obtain
$$K _ { l + | \nu | + 1 } ( \tilde { r } _ { 0 } z ) I _ { l + | \nu | } ( \tilde { R } z ) + K _ { l + | \nu | } ( \tilde { R } z ) I _ { l + | \nu | + 1 } ( \tilde { r } _ { 0 } z ) = 0, \ \ l \geq | k |$$
$$K _ { l + | \nu | - 1 } ( \tilde { r } _ { 0 } z ) I _ { l + | \nu | } ( \tilde { R } z ) + K _ { l + | \nu | } ( \tilde { R } z ) I _ { l + | \nu | - 1 } ( \tilde { r } _ { 0 } z ) = 0, \ | k | \geq l \geq 0$$
$$K _ { l - | \nu | + 1 } ( \tilde { r } _ { 0 } z ) I _ { l - | \nu | } ( \tilde { R } z ) + K _ { l - | \nu | } ( \tilde { R } z ) I _ { l - | \nu | + 1 } ( \tilde { r } _ { 0 } z ) = 0, \ l \geq 1$$
From now on we will assume ν > 0 and k > 0. Let us define the functions
$$h _ { l } ^ { ( 1 ) } ( z ) & = \frac { \tilde { r } _ { 0 } ^ { l + \nu + 1 } z } { \tilde { R } ^ { l + \nu } } \Big [ K _ { l + \nu + 1 } ( \tilde { r } _ { 0 } z ) I _ { l + \nu } ( \tilde { R } z ) + K _ { l + \nu } ( \tilde { R } z ) I _ { l + \nu + 1 } ( \tilde { r } _ { 0 } z ) \Big ], \ \ l \geq k, \quad \ ( 8 0 ) \\ \psi _ { l } ^ { ( 2 ) } / \sim \, - \, \begin{array} { c } \tilde { R } ^ { l + \nu } z \, \Big [ \tau _ { 0 }, \quad \left ( \tilde { r } _ { 0 } \sim \tau _ { 0 }, \quad \left ( \tilde { D } _ { 0 } \sim \tau _ { 0 }, \quad \left ( \tilde { r } _ { 0 } \sim \wedge \end{array} \Big ] \quad \psi _ { l } \times \imath \subset \imath \circ \quad \ ( \tau _ { 0 } \, \cdot \, \end{array}$$
$$h _ { l } ^ { ( 3 ) } ( z ) & = \frac { \overset { \cdot \upsilon } { \tilde { r } } _ { 0 } ^ { l - \nu + 1 } z } { \tilde { R } ^ { l - \nu } } \left [ K _ { l - \nu + 1 } ( \tilde { r } _ { 0 } z ) I _ { l - \nu } ( \tilde { R } z ) + K _ { l - \nu } ( \tilde { R } z ) I _ { l - \nu + 1 } ( \tilde { r } _ { 0 } z ) \right ], \ \ l \geq 1. \quad \ \ 8 2 ) \\ \text{Using the well known properties of the Bessel functions } [ 2 4 ], \, \text{we can check the conditions}$$
$$h _ { l } ^ { ( 2 ) } ( z ) & = \frac { \tilde { R } ^ { l + \nu } z } { \tilde { r } _ { 0 } ^ { l + \nu - 1 } } \left [ K _ { l + \nu - 1 } ( \tilde { r } _ { 0 } z ) I _ { l + \nu } ( \tilde { R } z ) + K _ { l + \nu } ( \tilde { R } z ) I _ { l + \nu - 1 } ( \tilde { r } _ { 0 } z ) \right ], \ \ k \geq l \geq 0, \quad & \\, ( 3 ) / \vee \quad & \tilde { r } _ { 0 } ^ { l - \nu + 1 } z \uparrow \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \tilde { \tau } _ { \tilde { n } } \vee \lim _ { \nu } \tilde { \tau } _ { \tilde { n } } \vee \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ { \nu } \lim _ }$$
Using the well known properties of the Bessel functions [24], we can check the conditions h ( ) i l (0) = 1, h ′ ( ) i l (0) = 0, and h ( ) i l ( z ) = h ( ) i l ( -z ). As similarly as performed in [1, 10, 11, 12], those properties allow us to decompose our grand potential as a Weierstrass product over its zeroes. The grand potential for our system is
$$\beta \Omega = - \sum _ { i = 1 } ^ { 3 } \sum _ { l } \ln h _ { l } ^ { ( i ) } ( - 1 ), \\ \dots$$
where l lies in the intervals [ k, ∞ ), [0 , k ], and [1 , ∞ ) for i = 1 2, and 3, respectively. ,
Therefore, we get
$$\beta \Omega = \beta \Omega _ { 0 } + \beta \Omega ( n, \tilde { R } ) + \beta \Omega ( k, \nu, \tilde { r } _ { 0 } ) + \beta \Omega ^ { \nu } ( \tilde { R } ) + O ( e ^ { - 2 \tilde { R } } ),$$
where [1]
$$\beta \Omega _ { 0 } = - \pi R ^ { 2 } \, \frac { m ^ { 2 } } { 2 \pi } \Big { [ } \ln \frac { 2 } { m a } - \gamma + \frac { 1 } { 2 } \Big { ] } + 2 \pi R \, m \Big { [ } \frac { 1 } { 4 } - \frac { 1 } { 2 \pi } \Big { ] } + \frac { 1 } { 6 } \ln ( m R ) + O ( 1 ) \quad \ \ ( 8 5 ) \\ \text{is the ground notential of the unmertured nlasma. and } ( \text{remember that $k+\nu=n$} )$$
is the grand potential of the unperturbed plasma, and (remember that k + ν = n )
$$\beta \Omega ( n, \tilde { R } ) = n \ln ( m R ) - \ln [ \, \Gamma ( n + 1 ) \, ] - n \ln 2 + O ( 1 / \tilde { R } ),$$
$$\beta \Omega ( n, \tilde { R } ) & = n \ln ( m R ) - \ln [ \Gamma ( n + 1 ) ] - n \ln 2 + O ( 1 / \tilde { R } ), \\ \beta \Omega ( k, \nu, \tilde { r } _ { 0 } ) & = - \sum _ { l = k } ^ { \infty } \ln \left [ \left ( \tilde { r } _ { 0 } ^ { \, } \right ) ^ { l + \nu + 1 } \frac { 2 \, K ^ { ( 0 ) } _ { l + \nu + 1 } } { \Gamma ( l + \nu + 1 ) } \right ] - \sum _ { l = 1 } ^ { \infty } \ln \left [ \left ( \tilde { r } _ { 0 } ^ { \, } \right ) ^ { l - \nu + 1 } \frac { 2 \, K ^ { ( 0 ) } _ { l - \nu + 1 } } { \Gamma ( l - \nu + 1 ) } \right ] \\ & \quad - \sum _ { l = 0 } ^ { k } \ln \left [ \frac { \tilde { R } ^ { 2 ( l + \nu ) } } { \tilde { r } _ { 0 } ^ { \, } 0 \nu - 1 } \frac { K ^ { ( 0 ) } _ { l + \nu - 1 } } { 2 ^ { l + \nu } ( l + \nu + 1 ) } \right ], \\ \beta \Omega ^ { \nu } ( \tilde { R } ) & = - \sum _ { l = 0 } ^ { \infty } \ln \left [ \frac { I ^ { ( R ) } _ { l + \nu } } { I ^ { l } _ { l } } \frac { \Gamma ( l + \nu + 1 ) } { \Gamma ( l + 1 ) } \left ( \frac { 2 } { R } \right ) ^ { \nu } \right ] - \sum _ { l = 1 } ^ { \infty } \ln \left [ \frac { I ^ { l } _ { l - \nu } } { I ^ { l } _ { l } } \frac { \Gamma ( l - \nu + 1 ) } { \Gamma ( l + 1 ) } \left ( \frac { 2 } { R } \right ) ^ { - \nu } \right ] \\ & \quad - \ln \left [ \frac { I ^ { ( R ) } _ { \nu } } { I ^ { ( R ) } _ { 0 } } \right ]. \\ \text{Last expressions cannot be evaluated exactly. For } \tilde { r } _ { 0 } \ll 1 \, \text{and } \tilde { R } \gg 1 \, \text{we can use the Euler} \\. \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \tilde {. } \ t.$$
Last expressions cannot be evaluated exactly. For ˜ r 0 /lessmuch 1 and ˜ R /greatermuch 1 we can use the Euler Mc-Laurin formula to transform discrete sums into integrals and expand K (0) l in powers of ˜ r 0 using the fact that ˜ r 0 /lessmuch 1. After some algebra, adding all the different contributions so we can write the grand potential as β Ω = β Ω + 0 β Ω ( 1 k, ν ), and neglecting O (1) terms we find
$$\beta \Omega _ { 1 } ( 0, 0 ) & = \frac { \tilde { r _ { 0 } ^ { 2 } } } { 2 } \ln \left ( \frac { R } { a } \right ) + O ( \tilde { r } _ { 0 } ^ { 2 } \ln \tilde { r } _ { 0 } ), \\ \beta \Omega _ { 1 } ( n, 0 ) & = - \ln \left [ \ln \left ( \frac { 2 } { \tilde { r } _ { 0 } } \right ) - \gamma \right ] + n ( n - 1 ) \ln \tilde { r } _ { 0 } - n ^ { 2 } \ln ( m R ) + \frac { \tilde { r _ { 0 } ^ { 2 } } } { 2 } \ln \left ( \frac { R } { a } \right ) + O ( \tilde { r } _ { 0 } ^ { 2 } \ln \tilde { r } _ { 0 } ),$$
and
$$\beta \Omega _ { 1 } ( 0, \nu ) & = - \nu ^ { 2 } \ln ( m R ) + \frac { \tilde { r } _ { 0 } ^ { 2 } } { 2 } \ln \left ( \frac { R } { a } \right ) + O ( \tilde { r } _ { 0 } ^ { 2 ( 1 - \nu ) } ), \\ \beta \Omega _ { 1 } ( k, \nu ) & = \left [ n ( n - 1 ) + \nu ( 1 - \nu ) \right ] \ln \tilde { r } _ { 0 } - n ^ { 2 } \ln ( m R ) + \frac { \tilde { r } _ { 0 } ^ { 2 } } { 2 } \ln \left ( \frac { R } { a } \right ) + O ( \tilde { r } _ { 0 } ^ { 2 ( 1 - \nu ) }, \tilde { r } _ { 0 } ^ { 2 \nu } ). ( 9 2 )$$
The results found for the grand potential are consistent with what we expect. First of all, the bulk pressure and the surface tension are not modified by the presence of the impurity, because we are working in the thermodynamic limit and one single impurity cannot alter extensive quantities, or quantities proportional to the boundary length of the system. Second, we expect to find terms proportional to ln ˜ r 0 that diverge when r 0 → 0 and increase with n because of the mentioned particle collapse at Γ = 2. However, notice that such divergence does not appear when n = 0 because such condition only applies for charged particles; the only contribution in the case n = 0 is a hard core effect that depends on the cutoff a and vanishes as r 0 → 0, Eq. (89), which is expected. Notice that such hard core effect is present for all values of n .
The term proportional to ln( mR ) is a universal finite-size correction term that is related to the topology of the considered system. Generally, for a two-dimensional conformal field theory, confined in a domain of typical length R , in the limit R →∞ , the free energy of the system exhibits a finite size correction given by β ∆ F = -1 6 cχ ln( R ), where c is the conformal anomaly number (central charge) and χ the Euler characteristic of the domain containing the system [25]. As explained in [26, 27], Coulomb systems should exhibit a similar correction, with c = 1, due to the long range of the electric potential correlations, however with a change of sign, due to the fact that, in the partition function, one integrates over the fluctuations of the density and not directly over the fluctuations of the electric potential. Now, in the present system, we have shown that the finite size correction to the grand potential is changed from 1 6 ln( mR ) to 1 -6 n 2 6 ln( mR ). The presence of the charged impurity modifies the central charge from c = 1 to c = 1 -6 n 2 . This is in agreement with what is expected from conformal field theory, as a simple deformation of the minimal free boson conformal field theory ( c = 1) is obtained by spreading out a charge α / 0 2 = n at infinity to obtain a conformal field theory with c = 1 -24 α 2 0 = 1 -6 n 2 [28, 29]. For the Coulomb system studied here, the external charge is not spread at infinity, but located at the origin. Nevertheless it has the same effect of shifting the central charge from c = 1 to c = 1 -6 n 2 .
One difference between the contributions of integer and non-integer charges to the grand potential is in the terms associated with ln ˜ . r 0 While the term ∆Ω = n n ( -1) ln ˜ r 0 is common in both cases, there is an additional contribution that differs in the integer and the non-integer cases. For integer charges, this contribution is β ∆Ω int = -ln [ ln ( 2 ˜ r 0 ) -γ ] , which clearly differs from the contribution for non-integer charges, given by β ∆Ω non -int = ν (1 -ν ) ln ˜ . r 0
## 7 Conclusions
We have studied a two-dimensional two-component plasma at Γ = 2 with an electric impurity confined in a large disk of radius R . Particularly, we found analytical expressions for the density and correlation functions and the grand potential, which provide
information for the pressure and the superficial tension. When the impurity is located at the origin of the confined plasma there is rotational invariance and the equations can be solved analytically using the method described in [1]. The counter-ions and co-ions accumulate close to the origin and the boundary respectively.
/negationslash
/negationslash
The case in which the electric charge of the impurity is an integer multiple of the charges in the plasma ± e and the case where the impurity charge is not an integer multiple of ± e were analyzed independently, to highlight some interesting differences due to the discrete nature of the electric charges of ions of the plasma. Although we could have expected quite different behaviors in the cases where the charge takes an 'integer' and a 'non-integer' value, we found that the effects on the charge redistribution are similar when the radius of the impurity r 0 = 0. But for a point-like impurity, r 0 = 0, important differences can be observed between the 'integer' and 'non-integer' cases. When the impurity charge has an integer value and r 0 → 0 there is an integer number of counterions that can cancel such effect when they collapse at the site where the impurity is located. On the other hand, finite-size effects for finite r 0 avoid the cancellation of the charge of the impurity and so the charge is differently redistributed close to the impurity. When r 0 > 0 the charges redistribute in such way that there are accumulated charges -n and n around the impurity and the boundary respectively. In the case r 0 → 0, and a non integer value of the charge n = k + ν ( ν = 0), k counter-ions collapse into the impurity, and the charge accumulated around the impurity is only -νe .
This problem can be extended with the introduction of an additional impurity. Unfortunately, the rotational symmetry in this situation is broken and it might not be possible to find analytical results. When two impurities are included, we expect to obtain the effect of each individual particle (ignoring the effects of the other) plus an additional term that is related to the interaction between the two particles.
The results for the grand potential are as expected. Since the size of the impurity is negligible compared to the the size of the disk, we do not expect the pressure and the superficial tension to be modified. Nonetheless, the central charge of the system is modified because of the presence of the impurity. In this case, we found that the topological term takes the form ( 1 6 -n 2 ) ln( mR ), with n the charge of the impurity, indicating a change in the central charge from c = 1 to c = 1 -6 n 2 .
Partial financial support from Fondo de Investigaciones, Facultad de Ciencias, Universidad de los Andes (project 2014-2 'Impurezas cargadas en plasmas y electrolitos'), and ECOS-Nord/COLCIENCIAS-MEN-ICETEX is acknowledged.
## References
- [1] F. Cornu and B. Jancovici, The electrical double layer: A solvable model, J. Chem. Phys. 90 :2444 (1989).
| [2] | L. ˇ Samaj and I. Travˇnec, e Thermodynamic properties of the two-dimensional two- component plasma, J. Stat. Phys. 101 :713 (2000). |
|-------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [3] | L. ˇ Samaj and B. Jancovici, Surface tension of a metal - electrolyte boundary: exactly solvable model, J. Stat. Phys. 103 :717 (2001). |
| [4] | L. ˇ Samaj, Surface tension of an ideal dielectric - electrolyte boundary: exactly solv- able model, J. Stat. Phys. 103 :737 (2001). |
| [5] | J. M. Kosterlitz, D. J. Thouless, Ordering, metastability and phase transitions in two-dimensional systems, J. Phys. C 6 :1181 (1973). |
| [6] | P. Kalinay, L. Samaj, Thermodynamic properties of the two-dimensional coulomb gas in the low-density limit, J. Stat. Phys. 106 :857-874 (2002). |
| [7] | G. T´llez, e Equation of state in the fugacity format for the two-dimensional Coulomb gas, J. Stat. Phys. 126 :281 (2007) |
| [8] | M. L. Rosinberg, J. L. Lebowitz, L. Blum, Solvable model for localized adsorption in a Coulomb system, J. Stat. Phys. 44 :153 (1986). |
| [9] | F. Cornu, Two-dimensional models for an electrode with adsorption sites, J. Stat. Phys. 54 :681 (1989). |
| [10] | G. T´llez, e L. Merch´n, a Solvable model for electrolytic soap films: the two-dimensional two-component plasma, J. Stat. Phys. 108 :495 (2002). |
| [11] | L. Merch´n, a G. T´llez, e Confined Coulomb systems with adsorbing boundaries: the two-dimensional two-component plasma, J. Stat. Phys. 114 :735 (2004). |
| [12] | A. Ferrero, G. T´llez, e Two-dimensional two-component plasma with adsorbing im- purities, J. Stat. Phys. 129 :759 (2007). |
| [13] | P. Debye and E. H¨ckel, u The theory of electrolytes. I. Lowering of freezing point and related phenomena, Phys. Z. 24 :185 (1923). |
| [14] | R. M. Fuoss, A. Katchalsky, S. Lifson, The potential of an infinite rod-like molecule and the distribution of the counter ions, Proc. Natl Acad. Sci. USA 37 :579 (1951). |
| [15] | G. S. Manning, Limiting laws and counterion condensation in polyelectrolyte solu- tions I. Colligative properties, J. Chem. Phys. 51 :924 (1969). |
| [16] | I. Rouzina, V. A. Bloomfield, Macroion attraction due to electrostatic correlation between screening counterions. 1. Mobile surface-adsorbed ions and diffuse ion cloud, J. Phys. Chem. 100 :9977 (1996). |
| [17] | A. Naji and R.R. Netz, Attraction of like-charged macroions in the strong-coupling limit, Eur. Phys. J. E 13 :43 (2004). |
|--------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [18] | A. Naji and S. Jungblut, A.G. Moreira and R.R. Netz, Electrostatic interactions in strongly-coupled soft matter, Physica A 352 :131 (2005) |
| [19] | L. ˇ Samaj, E. Trizac, Counterions at highly charged interfaces: From one plate to like-charge attraction, Phys. Rev. Lett. 106 :078301 (2011). |
| [20] | L. ˇ Samaj, E. Trizac, Wigner-crystal formulation of strong-coupling theory for coun- terions near planar charged interfaces, Phys. Rev. E 84 :041401 (2011). |
| [21] | J. P. Mallarino, G. T´llez, e E. Trizac, Counter-ion density profile around charged cylinders: the strong-coupling needle limit, J. Phys. Chem. B 117 :12702 (2013). |
| [22] | L. ˇ Samaj, Anomalous effects of 'guest' charges immersed in electrolyte: Exact 2D results, J. Stat. Phys. 120 :125 (2005) |
| [23] | G. T´llez, e Short-distance expansion of correlation functions for the charge-symmetric two-dimensional two-component plasma: exact results, J. Stat. Mech. P10001 (2005). |
| [24] | I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and Products , Academic Press, Fifth Edition (1994). |
| [25] | J. L. Cardy, I. Peschel, Finite-size dependence of the free energy in two-dimensional critical systems, Nucl. Phys. B 300 :377-392 (1988). |
| [26] | B. Jancovici, G. Manificat, C. Pisani, Coulomb systems seen as critical systems: Finite-size effects in two dimensions, J. Stat. Phys. 76 :307-329 (1994). |
| [27] | B. Jancovici, G. T´llez, e Coulomb systems seen as critical systems: Ideal conductor boundaries, J. Stat. Phys. 82 , 609-632 (1996). |
| [28] | Vl. S. Dotsenko, Serie de cours sur la th´orie e conforme , Universit´ e de Paris VI - Universit´ e de Paris VII (2004). |
| [29] | P. Ginsparg, Applied conformal field theory , Les Houches, Session XLIX, Fields, strings and critical phenomena, ed. by E. Br´zin e and J. Zinn-Justin, Elsevier (1989). |
|
10.1088/1742-5468/2014/11/P11021
|
[
"Alejandro Ferrero",
"Gabriel Téllez"
] |
2014-08-05T19:55:16+00:00
|
2014-08-05T19:55:16+00:00
|
[
"cond-mat.stat-mech"
] |
Screening of an electrically charged particle in a two-dimensional two-component plasma at $Γ=2$
|
We consider the thermodynamic effects of an electrically charged impurity
immersed in a two-dimensional two-component plasma, composed by particles with
charges $\pm e$, at temperature $T$, at coupling $\Gamma=e^2/(k_B T)=2$,
confined in a large disk of radius $R$. Particularly, we focus on the analysis
of the charge density, the correlation functions, and the grand potential. Our
analytical results show how the charges are redistributed in the circular
geometry considered here. When we consider a positively charged impurity, the
negative ions accumulate close to the impurity leaving an excess of positive
charge that accumulates at the boundary of the disk. Due to the symmetry under
charge exchange, the opposite effect takes place when we place a negative
impurity. Both the cases in which the impurity charge is an integer multiple of
the particle charges in the plasma, $\pm e$, and a fraction of them are
considered; both situations require a slightly different mathematical
treatments, showing the effect of the quantization of plasma charges. The bulk
and tension effects in the plasma described by the grand potential are not
modified by the introduction of the charged particle. Besides the effects due
to the collapse coming from the attraction between oppositely charged ions, an
additional topological term appears in the grand potential, proportional to
$-n^2\ln(mR)$, with $n$ the dimensionless charge of the particle. This term
modifies the central charge of the system, from $c=1$ to $c=1-6n^2$, when
considered in the context of conformal field theories.
|
1408.1081v2
|
## Super-Sample Signal
Yin Li, 1, 2 Wayne Hu, 2 and Masahiro Takada 3
1 Department of Physics, University of Chicago, Chicago, Illinois 60637, U.S.A. 2 Kavli Institute for Cosmological Physics, Department of Astronomy & Astrophysics, Enrico Fermi Institute, University of Chicago, Chicago, Illinois 60637, U.S.A. 3 Kavli Institute for the Physics and Mathematics of the Universe (WPI), Todai Institutes for Advanced Study, The University of Tokyo, Chiba 277-8583, Japan
When extracting cosmological information from power spectrum measurements, we must consider the impact of super-sample density fluctuations whose wavelengths are larger than the survey scale. These modes contribute to the mean density fluctuation δ b in the survey and change the power spectrum in the same way as a change in the cosmological background. They can be simply included in cosmological parameter estimation and forecasts by treating δ b as an additional cosmological parameter enabling efficient exploration of its impact. To test this approach, we consider here an idealized measurement of the matter power spectrum itself in the ΛCDM cosmology though our techniques can readily be extended to more observationally relevant statistics or other parameter spaces. Using sub-volumes of large-volume N -body simulations for power spectra measured with respect to either the global or local mean density, we verify that the minimum variance estimator of δ b is both unbiased and has the predicted variance. Parameter degeneracies arise since the response of the matter power spectrum to δ b and cosmological parameters share similar properties in changing the growth of structure and dilating the scale of features especially in the local case. For matter power spectrum measurements, these degeneracies can lead in certain cases to substantial error degradation and motivates future studies of specific cosmological observables such as galaxy clustering and weak lensing statistics with these techniques.
## I. INTRODUCTION
The statistical properties of large scale structure provide a wealth of cosmological information on fundamental physics, including cosmic acceleration, neutrino masses and inflation. The simplest statistic is the two-point correlation function or power spectrum of the matter density field which underlies observables such as weak lensing and galaxy clustering. To extract cosmological information from the power spectrum of upcoming wide-area galaxy surveys (e.g. [1]), we need to properly model its own statistical properties, one of which is the impact of modes whose wavelengths are larger than the survey scale, the so-called super-sample modes.
stochasticity, or as additional signal from which the mean density in a given volume may be recovered [13]. In the former view, the super-sample effect introduces a covariance to power spectrum estimates since each realization of δ b coherently changes the power spectrum according to the calibrated response. Ref. [14] used the covariance matrix of power spectra in subsampled simulations to verify the separate universe response itself.
While these super-sample modes are not directly observable, they impact the evolution of sub-sample modes in an observable way due to nonlinear mode coupling [215]. Ref. [13] developed a simple, unified approach that describes the impact of the super-sample modes as the response of the power spectrum to a change in the mean density in the finite-volume region. Ref. [14] then utilized the so-called separate universe approach to calibrate this response in N -body simulations. Here, the mean density fluctuation δ b is absorbed into a change in cosmological parameters (e.g. [16-20]). The separate universe approach is also useful for gaining a physical understanding of the response. In a coherently overdense region structure grows more quickly and regions expand less quickly. The response is therefore a change in the amplitude or growth of structure and in the scale or dilation of features in the power spectrum.
The impact of the mean density fluctuation can also be viewed in two ways: as additional noise due to its
Here we develop the alternative view that in each realization the super-sample effect biases the measured power spectrum or equivalently introduces an extra parameter, δ b , upon which it depends. The mean density fluctuation leaves a signal in the power spectrum which can be used to recover its value providing constraints on super-sample modes that are not directly observable in the survey (see also [12, 21]). Moreover, this view facilitates studies of the impact of super-sample modes on cosmological parameter estimation by treating the two on an equal footing.
While here we only test this view in the idealization that the matter power spectrum is directly observed, these ideas can be readily extended to more observationally relevant power spectrum observables once their response to δ b is calibrated.
The structure of this paper is as follows. In § II we review the super-sample effects on the power spectrum and test its interpretation as signal by explicitly constructing an estimator of the mean density fluctuation in simulations. In § III we study the similarities between the power spectrum response to the mean density fluctuation and to cosmological parameters which propagate into parameter degeneracies in the Fisher information matrix. We discuss these results in § IV. Appendix A provides details
FIG. 1. Power spectrum response to the mean density fluctuation in a survey region, calibrated with the separate universe simulations through a change in background parameters (see Ref. [14]). Solid curve represents the response of the power spectrum referenced to the true or global mean density, while the dashed curve to the mean survey or local density. The two are related by an additive constant 2, which changes the shape of the response.
of the simulations and the numerical techniques.
## II. SUPER-SAMPLE SIGNAL
In § II A, we first briefly review the impact of supersample modes on the power spectrum covariance which was developed in Refs. [13, 14]. Next we show the effect can alternately be treated as a signal that allows us to construct an unbiased minimum variance estimator of the mean density fluctuation in § II B. We test this estimator against cosmological simulations in § II C and show that its statistical properties are well-characterized by the separate universe response of the power spectrum to the mean density and the covariance of the power spectrum in the absence of the mean density mode.
## A. Super-sample modes
In this section we characterize the effect of supersample modes on power spectrum measurements in a finite survey volume in a manner that brings out their role as both signal and noise. Let us denote the average density fluctuation in the survey region from these modes as δ b and assume that the measurement of the power spectrum of sub-survey modes ˆ ( P k δ ; b ) is sample variance limited. Let us bin these measurements into a data vector
$$\hat { \mathcal { P } } _ { i } ( \delta _ { \text{b} } ) \equiv \frac { k _ { i } ^ { 3 } \hat { P } ( k _ { i } ; \delta _ { \text{b} } ) } { 2 \pi ^ { 2 } }, \quad \ \ \ ( 1 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
and consider the impact of δ b on the mean and covariance of these data. We use the dimensionless power spectrum since it is independent of the units with which the wavenumber is measured and thus allows a cleaner separation of effects on the amplitude and scale of structure (see Ref. [14]).
Even if we define an estimator which yields the true power spectrum P k ( i ) when averaged over realizations of all modes, including the super-sample ones
$$\langle \hat { \mathcal { P } } _ { i } \rangle \equiv \mathcal { P } _ { i } = \frac { k _ { i } ^ { 3 } P ( k _ { i } ) } { 2 \pi ^ { 2 } },$$
the average over realizations of different sub-survey modes at fixed δ b , which we denote by 〈 . . . 〉 b , is biased
$$\langle \hat { \mathcal { P } } _ { i } \rangle _ { b } \equiv \mathcal { P } _ { i } ( \delta _ { b } ) = \mathcal { P } _ { i } ( 1 + T _ { i } ^ { \delta _ { b } } \delta _ { b } ). \quad \quad ( 3 )$$
Here we have linearized the response of the power spectrum assuming | δ b | glyph[lessmuch] 1
$$T _ { i } ^ { \delta _ { b } } \equiv \frac { \partial \ln \mathcal { P } _ { i } } { \partial \delta _ { b } }.$$
As demonstrated in Ref. [14] through its effects on the power spectrum covariance below, we can use the separate universe approach to calibrate the power spectrum response, T δ b i , for a given fiducial cosmological model. In this approach the mean density fluctuation of the survey δ b is absorbed into the local background density by redefining cosmological parameters.
Fig. 1 summarizes the results for the ΛCDM model (see Tab. II) at z = 0 using N -body simulations to calibrate the response deep into the nonlinear regime. Note that the technique itself is more general than this particular implementation. Given a state-of-the-art simulation that includes baryonic effects, e.g. star formation and feedback [22, 23], the impact of super-sample modes can still be calibrated by runs where the background parameters are changed to absorb δ b .
These results apply to power spectrum measurements where the global mean density is known through cosmological parameters, e.g. in the case of weak lensing power spectra. If the power spectrum is estimated with respect to the local mean density within the survey region, which is the case for galaxy surveys, the relevant observable is
$$\hat { \mathcal { P } } ^ { W } = \frac { \hat { \mathcal { P } } } { ( 1 + \delta _ { \text{b} } ) ^ { 2 } },$$
where W denotes the survey volume or window here and below. The power spectrum response is therefore modified to be
$$T _ { i } ^ { \delta _ { b } } \Big | _ { \text{local} } \equiv \frac { \partial \ln \mathcal { P } _ { i } ^ { W } } { \partial \delta _ { b } } \approx \frac { \partial \ln \mathcal { P } _ { i } } { \partial \delta _ { b } } - 2. \quad \quad ( 6 )$$
Because the formalism for super-sample effects is otherwise identical, we use the term 'local' data or response when replacing the 'global' versions of Eqs. (4) and (5),
in relevant formula below. Both responses are shown in Fig. 1.
Since δ b is a random zero mean variable, its variance turns into a covariance of the power spectrum estimators. In this sense the super-sample effect contributes as additional noise. We can model this noise by treating the bias as a purely systematic additive shift in the power spectrum per realization
$$\hat { \mathcal { P } } _ { i } ( \delta _ { b } ) = \hat { \mathcal { P } } _ { i } ( 0 ) + \mathcal { P } _ { i } T _ { i } ^ { \delta _ { b } } \delta _ { b }. \quad \ \ ( 7 ) \quad \text{fro} \\. \quad \. \ \. \ \. \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
The covariance matrix of the power spectrum data is then given by
$$\langle \hat { \mathcal { P } } _ { i } \hat { \mathcal { P } } _ { j } \rangle - \mathcal { P } _ { i } \mathcal { P } _ { j } = C _ { i j } + \mathcal { P } _ { i } \mathcal { P } _ { j } T _ { i } ^ { \delta _ { b } } T _ { j } ^ { \delta _ { b } } \sigma _ { b } ^ { 2 }, \quad ( 8 ) \quad \text{eters} \quad \text{Mon}$$
where σ 2 b is the variance of the mean density field in the survey window, defined as
$$\sigma _ { b } ^ { 2 } \equiv \langle \delta _ { b } ^ { 2 } \rangle = \frac { 1 } { V _ { W } ^ { 2 } } \int \frac { d ^ { 3 } { \mathbf q } } { ( 2 \pi ) ^ { 3 } } | \tilde { W } ( { \mathbf q } ) | ^ { 2 } P ( q ). \quad ( 9 ) \quad \text{ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
Here V W is the survey volume and ˜ ( W q ) is the survey window function which acts as a low pass filter. For a sufficiently large survey volume, the super-sample modes are in the linear regime and therefore σ b can be accurately calculated with the linear power spectrum either by explicit computation of Eq. (9) or Gaussian realizations of the linear density field. The matrix C ij in Eq. (8) is defined as
$$C _ { i j } = \langle \hat { \mathcal { P } } _ { i } ( 0 ) \hat { \mathcal { P } } _ { j } ( 0 ) \rangle - \mathcal { P } _ { i } \mathcal { P } _ { j }. \quad \quad ( 1 0 ) \quad \text{with}$$
Note that this is the covariance of the power spectrum estimators in the absence of the super-sample effect. It can be readily calibrated from a suite of smallvolume N -body simulations for a given cosmological model [24, 25] (see Appendix A and Ref. [14] for implementation specifics). The sum of the two contributions of Eq. (8) reproduces the super-sample covariance derived in Ref. [14] from trispectrum considerations.
An alternate model would be to consider the bias in the power spectrum estimators to be multiplicative with respect to ˆ (0), P i
$$\hat { \mathcal { P } } _ { i } = \hat { \mathcal { P } } _ { i } ( 0 ) ( 1 + T _ { i } ^ { \delta _ { b } } \delta _ { b } ). \quad \quad ( 1 1 )$$
This model yields an additional contribution to the second term of Eq. (8)
$$\mathcal { P } _ { i } \mathcal { P } _ { j } T _ { i } ^ { \delta _ { b } } T _ { j } ^ { \delta _ { b } } \sigma _ { b } ^ { 2 } \to ( \mathcal { P } _ { i } \mathcal { P } _ { j } + C _ { i j } ) T _ { i } ^ { \delta _ { b } } T _ { j } ^ { \delta _ { b } } \sigma _ { b } ^ { 2 } \quad ( 1 2 )$$
and hence only a small change for well measured bins where the covariance is much smaller than the product of the means. We hereafter adopt the additive model.
## B. Signal vs. noise
In a given realization of the survey volume, the supersample effect systematically changes the power spectrum of sub-survey modes just like a cosmological parameter p c . Indeed, in the additive model of Eq. (7), the analogy is precise: δ b is simply a parameter that changes the mean power spectrum
$$\hat { \mathcal { P } } _ { i } ( \delta _ { b } ) - \mathcal { P } _ { i } ( \delta _ { b }, \mathbf p _ { c } ) = \hat { \mathcal { P } } _ { i } ( 0 ) - \mathcal { P } _ { i } ( \mathbf p _ { c } ). \quad ( 1 3 )$$
Thus the data in the presence of δ b has the same statistical properties as in its absence. They are both drawn from a distribution with covariance C ij and only the mean is shifted. Parameter estimation then proceeds in the usual way by treating δ b as a parameter p = { δ b , p c } . For example, the posterior probability of model parameters including δ b can be estimated using Markov Chain Monte Carlo techniques based on the likelihood function constructed from the covariance matrix C ij with the model parameterized by p . In this view, the supersample effect is a signal which allows the mean density fluctuation to be recovered rather than a source of additional noise.
To test this interpretation with simulations, we can construct an explicit estimator of δ b based on linear response. Here we shall assume that cosmological parameters are fixed and hence suppress their appearance in the expressions. A general unbiased linear estimator of δ b takes the form
$$\hat { \delta } _ { b } = \sum _ { i } w _ { i } ( \hat { \mathcal { P } } _ { i } - \mathcal { P } _ { i } ),$$
where the weight
$$\sum _ { i } w _ { i } \mathcal { P } _ { i } T _ { i } ^ { \delta _ { \text{b} } } = 1,$$
is constrained by the condition 〈 ˆ δ b b 〉 = δ b . The remaining freedom in choosing the weights is fixed by minimizing the variance of the estimator
$$\sigma _ { \delta _ { b } } ^ { 2 } \equiv \langle \hat { \delta } _ { b } ^ { 2 } \rangle _ { b } - \delta _ { b } ^ { 2 } = \sum _ { i j } w _ { i } w _ { j } C _ { i j }, \quad \ \ ( 1 6 )$$
subject to the Lagrange multiplier constraint yielding
$$w _ { i } = \frac { \sum _ { j } T _ { j } ^ { \delta _ { b } } \mathcal { P } _ { j } \left [ \mathbf C ^ { - 1 } \right ] _ { i j } } { \sum _ { j k } T _ { j } ^ { \delta _ { b } } \mathcal { P } _ { j } \left [ \mathbf C ^ { - 1 } \right ] _ { j k } T _ { k } ^ { \delta _ { b } } \mathcal { P } _ { k } } \quad \quad ( 1 7 )$$
and
$$\sigma _ { \delta _ { \text{b} } } ^ { 2 } = \left ( \sum _ { j k } T _ { j } ^ { \delta _ { \text{b} } } \mathcal { P } _ { j } \left [ \text{C} ^ { - 1 } \right ] _ { j k } T _ { k } ^ { \delta _ { \text{b} } } \mathcal { P } _ { k } \right ) ^ { - 1 }. \quad ( 1 8 )$$
If instead of the additive bias model, the multiplicative bias model of Eq. (11) is correct, this estimator remains unbiased but its variance changes. Thus comparing the predicted variance from Eq. (16) with the variance obtained from simulations tests the accuracy of our additive model of super-sample effects as well as that of the response T δ b i and covariance C ij calibration.
FIG. 2. Estimated ( ˆ δ b ) vs. true ( δ b ) mean density fluctuations for the set of 3584 simulation subvolumes. The means ( × , +, slightly shifted for clarity) of the estimators using both globally and locally referenced power spectra show no trace of a bias for bins with sufficient statistics that the standard deviation (errorbars) can be estimated. This standard deviation is significantly smaller than ± σ b (gray band) for the global case and comparable for the local case.
## C. Density estimation
In this section, we test the ˆ δ b estimator with largevolume simulations where the true mean density fluctuation δ b is known in each subvolume of the simulation. We summarize here the relevant simulation details presented in Appendix A. Specifically 7 large-volume simulations, each with a 4 Gpc /h box length are divided into a total of N s = 7 × 8 3 = 3584 subvolumes of size 500 h -1 Mpc each. For reference the variance of the mean density fluctuation of the subvolumes is
$$\hat { \sigma } _ { b } ^ { 2 } = & \frac { 1 } { N _ { s } } \sum _ { a = 1 } ^ { N _ { s } } \delta _ { b, a } ^ { 2 } = ( 0. 0 1 2 6 3 ) ^ { 2 } \quad \quad ( 1 9 )$$
which matches well the computation from Eq. (9), σ b = 0 01258. . We measure the power spectrum ˆ P i in each subvolume separately. Note that this power spectrum is of the density fluctuation from the large-box mean and hence characterizes the 'global' data in the language of § II A. For the 'local' data, we use the true average density in the subvolume to rescale the global data according to Eq. (5).
To calibrate the mean power spectrum P i and the covariance matrix C ij , we run another suite of N s small box simulations each of the same size as the subvolume. All power spectra are binned in 80 k -bins per decade and we
-
FIG. 3. Distribution or scatter of the density estimator ˆ δ b around the true value (histogram) versus the prediction from Eq. (16) with a Gaussian distribution (curve). In both the global and local cases, the predictions are within 7% of the simulation results in the standard deviation justifying their use in parameter forecasts below.
utilize measurements out to k glyph[lessorsimilar] 2 h Mpc -1 up to which we verified the response calibration is accurate to several percent or better, based on higher-resolution simulations [14].
From ˆ , P i P i , and C ij we form the ˆ δ b estimator of Eq. (14) for both the global and local cases and compare it with the true value δ b in each of N s subvolumes. Fig. 2 shows the results binned in δ b with width σ / b 4. In both the global and local estimations, there is no trace of a bias in the estimator to a small fraction of its standard deviation. Moreover, in Fig. 3 we test the predicted distribution of the estimator combining all bins with the prediction from Eq. (16) for the variance under the assumption of a Gaussian distribution. The agreement in both cases is very good. For a more quantitative assessment we measure the variance of the estimator with our N s samples in the usual way
$$\hat { \sigma } _ { \delta _ { b } } ^ { 2 } = & \frac { 1 } { N _ { s } } \sum _ { a = 1 } ^ { N _ { s } } ( \hat { \delta } _ { b, a } - \delta _ { b, a } ) ^ { 2 }.$$
The result is ˆ σ δ b = 0 29 . σ b for the global case and 1 4 . σ b for the local case. In both cases the result is only ∼ 7% larger in standard deviation than the prediction of Eq. (16). Even in the local case where the standard deviation is comparable to σ b there is non-negligible extra information provided by the estimator.
These results validate the use of our additive model to predict the impact of the super-sample effect in other cases of interest. In Fig. 4, we use this model to explore the dependence of the standard deviation of the estimator on the maximum k used. Note that in the local case, σ δ b is only a weak function of k max . This is because,
the intrinsic covariance between bins C ij induces very similar changes to the power spectrum as T δ b local for k glyph[greaterorsimilar] 0 1 . h Mpc -1 . The difference in shape between the local and global responses causes an improvement in the standard deviation of the latter for k glyph[greaterorsimilar] 1 h Mpc -1 .
The above results apply to estimates of δ b when all other parameters that change the power spectrum are known a priori. With joint estimator of parameters from the survey, the δ b mode will degrade results on cosmological parameters and vice versa if their impact on the power spectrum is sufficiently similar to cause degeneracies.
Finally in the additive model, the scaling of these results with the volume of the survey is also simple. Since C ij characterizes the covariance of subsurvey modes in the absence of the super-sample effect, it scales with volume as
$$C _ { i j } ( V ) = \frac { V _ { 0 } } { V } C _ { i j } ( V _ { 0 } ) \quad \quad ( 2 1 ) \quad \text{FI} ( \begin{matrix} \\ 2 \\ \text{anc} \end{matrix} )$$
where V 0 = (500 h -1 Mpc) , the volume of the simulation 3 test. Hence
$$\sigma _ { p _ { \mu } } ( V ) = \sqrt { \frac { V _ { 0 } } { V } } \sigma _ { p _ { \mu } } ( V _ { 0 } ) \quad \ \ ( 2 2 ) \quad \ \ ^ { \text{me} } _ { 1 \, h \, l }$$
for any parameter estimated from power spectra data including δ b . The ratio σ δ b /σ b does in general depend on volume. However in ΛCDM the quantity σ V b 1 / 2 varies weakly with V , typically by ∼ √ 2 across a V/V 0 = 100. Thus we expect that the relative impact of the supersample effect will be only weakly dependent on volume for the cubic geometry we consider. In Ref. [13] it was also shown that for a cylindrical geometry the scaling holds although σ b ( V 0 ) itself is smaller for the same volume.
## III. PARAMETER FORECASTS
In this section we use the Fisher information matrix formalism to study the impact of the super-sample effect on cosmological parameter estimation in the idealized context of direct matter power spectrum measurements. After defining the Fisher matrix in § III A, we study similarities in the power spectrum response between the super-sample mode and cosmological parameters in § III B. These similarities lead to degeneracies which degrade errors when parameters are jointly estimated in § III C and are themselves limited by prior information on the variance of the super-sample mode III D.
## A. Fisher matrix
As discussed in § II B, the additive model δ b can be thought of as an additional parameter of the model power spectrum so that the full parameter vector is
FIG. 4. Standard deviation of the δ b estimator, σ δ b , in the local and global cases as a function of the maximum k -bin compared with the rms density fluctuation σ b . In the local case the standard deviation approaches the rms in the nonlinear regime and remains nearly constant to 2 h Mpc -1 . In the global case, it drops below the rms in the nonlinear regime and continues to improve beyond 1 h Mpc -1 .
p ≡ { δ b , p c } , where p c are cosmological parameters. In the Fisher approximation, the information from the power spectrum mean is added to any prior information F prior µν on the parameters,
$$F _ { \mu \nu } = \sum _ { i j } \mathcal { P } _ { i } T _ { i } ^ { \mu } [ \mathbf C ^ { - 1 } ] _ { i j } \mathcal { P } _ { j } T _ { j } ^ { \nu } + F _ { \mu \nu } ^ { \text{prior} }, \quad ( 2 3 )$$
where
$$T _ { i } ^ { \mu } \equiv \frac { \partial \ln \mathcal { P } _ { i } } { \partial p _ { \mu } }.$$
The inverse Fisher matrix is an approximation to the covariance matrix of the parameters
$$[ \mathbf F ^ { - 1 } ] _ { \mu \nu } \approx \langle \hat { p } _ { \mu } \hat { p } _ { \nu } \rangle _ { \mathbf b } - p _ { \mu } p _ { \nu }.$$
To make contact with § II C, note that in the limit that the parameter space includes only δ b and there is no prior information on it, Eq. (25) yields the variance of the estimator given in Eq. (18). This is because the Fisher approximation involves the same linearization of the response to the parameters that permits the construction of a linear minimum variance unbiased estimator. With additional parameters and no external prior, degeneracies where the errors strongly covary appear if the responses take a similar form. Thus to understand the impact of super-sample modes on parameter estimation we must compare the various responses T µ i .
FIG. 5. Power spectrum response to cosmological parameters { ln A , n , h s s } as calibrated from simulations. Similarities between these responses and Fig. 1 cause degeneracies which degrade parameter errors once uncertainties in the super-sample mode δ b are marginalized.
## B. Parameter responses
Given that the separate universe technique described in § II A and detailed in Ref. [14] involves modeling the super-sample effect by changes in background cosmological parameters, we expect the response to δ b and cosmological parameters will contain similarities that can create parameter degeneracies. Namely, these parameters change the amplitude of power in the spectrum and the scale at which features like baryon acoustic oscillations (BAO) appear. We will call these effects growth or 'g' and dilation or 'd' respectively.
Let us begin by examining this decomposition for δ b and a power spectrum measurement with respect to the local mean density of the survey, defining
$$T ^ { \delta _ { b } } \Big | _ { \text{local} } \equiv 2 \frac { \partial \ln D } { \partial \delta _ { b } } T ^ { \delta _ { b }, g } - \frac { 1 } { 3 } T ^ { \delta _ { b }, d }. \quad \ ( 2 6 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
Here and below we omit the k -bin index i where no confusion should arise. The first term is due to the enhancement of the growth of structure in a coherently overdense region. Absorbing this fluctuation into a redefinition of the background implies a change in the linear growth function of density fluctuations D with respect to the local mean density [19]
$$\frac { \partial \ln D } { \partial \delta _ { b } } \approx \frac { 1 3 } { 2 1 }. \quad \quad \ ( 2 7 ) \quad \ \text{th} \\ \quad \text{or}$$
Hence, with this normalization, lim k → 0 T δ b , g = 1.
The second term of Eq. (26) is due to the fact that an overdense region expands less quickly than the global universe. This changes the comoving scale of physical features in the power spectrum according to a dilation template
$$T ^ { \text{d} } \equiv \frac { \partial \ln \mathcal { P } } { \partial \ln k }.$$
The factor of 1 / 3 arises since an equal time comparison is at equal physical mean density and so the scale factor is adjusted by (1 + δ b ) 1 / 3 . By removing this rescaling with a choice of simulation box size introduced in Ref. [14] (see also Appendix A), we can determine T δ b , g independently of the full response T δ b as shown in Fig. 6. The difference between these responses gives the dilation template T δ b , d which is compared with the k -derivative of the mean power spectra T d in Fig. 7. The agreement is good and as discussed in the Appendix, the response difference provides a more accurate way of calibrating dilation than differencing noisy power spectrum data.
Finally, when referenced to the true global mean density, the response becomes (see Eq. 6)
$$T ^ { \delta _ { \text{b} } } = T ^ { \delta _ { \text{b} } } \Big | _ { \text{local} } + 2. \text{ \quad \quad \ \ } ( 2 9 )$$
Note that the additional factor flattens the response of the power spectrum as shown in Fig. 1. We shall see that this addition is important for understanding parameter degeneracies (cf. [26]).
Now let us compare this response with those of cosmological parameters. The full set of flat ΛCDM parameters are A , n s s which jointly define the primordial curvature power spectrum
$$\mathcal { P } _ { \mathcal { R } } = A _ { s } \left ( \frac { k } { 0. 0 5 \, \text{Mpc} ^ { - 1 } } \right ) ^ { n _ { s } - 1 } ; \quad \quad ( 3 0 )$$
the baryon density Ω b h 2 ; the cold dark matter density Ω c h 2 and the dimensionless Hubble constant h . The parameters Ω b h 2 and Ω c h 2 are well determined by the cosmic microwave background (CMB) data and additionally are not degenerate with pure growth and dilation due to the changes in the BAO and matter radiation equality features that they induce. We therefore study responses to changes in the parameter set p c = { ln A , n , h s s } and keep the remaining parameters fixed to their fiducial values. We calibrate each response function as discussed in Appendix A2b and show the results in Fig. 5.
The response to h shares the same features as δ b . Since h is varied at fixed Ω b h 2 and Ω c h 2 in a flat universe its impact on the power spectrum in the linear regime at a fixed scale in Mpc comes solely from changing the growth function D due to the change in Ω Λ . However because observations at z = 0 determine a scale in h Mpc -1 , observable features in the power spectrum shift. This is the same effect that allows BAO to measure h at z = 0 or the expansion rate and angular diameter distance at higher redshift. The result is that the response can be decomposed as
$$T ^ { h } = 2 \frac { \partial \ln D } { \partial h } T ^ { h, g } + \frac { 1 } { h } T ^ { h, d }. \quad \quad ( 3 1 )$$
FIG. 6. Growth component of the power spectrum response to δ b and h compared with the response to the initial amplitude of power ln A s . The two growth responses are nearly identical whereas the ln A s one differs in the nonlinear regime. The difference can be attributed to a change in halo scale radii between models with the same linear power spectrum.
FIG. 7. Dilation component of the power spectrum response to δ b and h compared with that derived from the slope of the power spectrum T d . All three responses are nearly identical with the T d exhibiting more noise and resolution dependence from numerical differencing.
For reference, in the chosen cosmology ∂ ln D/∂h ≈ -0 668. . As shown in Fig. 6, the growth pieces of the δ b and h responses are nearly indistinguishable. Likewise, defining the dilation piece as the difference of these responses agrees with the dilation defined from δ b and the k -derivative of the mean power spectrum as shown in Fig. 7. In an extended parameter space, we expect the response to parameters such as the dark energy equation of state or curvature can be modeled accurately with T h, g and their impact on the linear growth function D .
Next ln A s directly controls the amount of power in the spectrum at the initial epoch. At z = 0 a change in ln A s and a change in growth can produce the same linear power spectrum and hence in the linear regime the two responses are indistinguishable. In Fig. 6 we compare these response and show that they begin to differ in the nonlinear regime. If the nonlinear power spectrum were a functional of the linear power spectrum then the responses would be identical. For example, in a halo model description, if the mass function were universal and the scale radii of halos were fixed then the linear power spectrum determines the nonlinear power spectrum directly with no further reference to the past history of structure formation. We have verified that the differences in response are qualitatively modeled by a change in concentration of halos with respect to a fixed scale [27]. The concentration of halos retains information about the mean density of the Universe at their formation epoch, and so one would not expect a change in the initial conditions and late-time growth that leaves the linear power invariant to yield the same concentration. In principle, then ln A s is not degenerate with other parameters associated with growth. In practice, uncertainty in the concentration-mass relation due to baryonic physics can restore this degeneracy.
Small changes in the tilt can also restore this degeneracy across a limited range in k . Fig. 5 also shows the response to tilt, which changes the power spectrum in opposite directions around the pivot point k = 0 05 Mpc . -1 = 0 071 . h Mpc -1 . Given the greater statistical power of measurements in the nonlinear regime, a small amount of tilt can compensate the differences due to changes in the concentration.
In summary, the response functions show that for power spectrum measurements with respect to the local mean, there are 3 parameters { δ b , h, ln A s } whose power spectrum response is characterized mainly by linear combinations of two templates, T g and T d once minor changes in tilt or halo concentration are factored in. We therefore expect a strong degeneracy in the local case. For power spectra measured with respect to the global mean, the addition of 2 to the response breaks this degeneracy yielding three templates for three parameters. We shall now see that these features are reflected in the forecasted parameter errors.
## C. Parameter constraints without δ b prior
We begin with parameter constraints for the case where there is no external prior on δ b so that any information about it must be recovered from the power spectrum measured by the survey. Fig. 8 shows an overview of the impact of marginalizing δ b on cosmological parameter estimation. The left panel shows the impact on the
FIG. 8. Errors on cosmological parameters as a function of the maximum k -bin for δ b fixed (dot-dashed), marginalized in the local case (dashed) and in the global case (solid). Left panel: single cosmological parameter estimation with the remaining two held fixed. Right panel: joint cosmological parameter estimation.
3 cosmological parameters considered one at a time with the other 2 fixed whereas the right panel shows the result with the other two marginalized. In the former case, the degradation is the most severe when the two templates are most similar. For example in the global case, the flattening of the response in the nonlinear regime relative to local in Fig. 1 makes it more similar to ln A s around k ∼ 1 -2 h Mpc -1 and causes a larger degradation in errors for such choices of k max . Parameter degeneracies also increase the importance of having a sufficiently large k max to distinguish the responses compared with a naive quantification of information through F µµ [28, 29] (see also [6, 7]).
Similar statements apply for the case where the two other cosmological parameters are also marginalized. Here what is important for determining degeneracies is whether a linear combination of the four responses can compensate each other. From Fig. 8 (right) we see that the flattening of the global vs. local response now has the opposite effect on ln A s . By k max = 2 h Mpc -1 most degeneracies are broken for the global case whereas they remain strong and impact all three cosmological parameters in the local case.
To see how these results arise geometrically, we show various 2D 68% confidence regions for the 4 parameters for the local case in Fig. 9 (left) with k max = 2 h Mpc -1 . In particular, the contours involving δ b show a strong degeneracy with all three of the other parameters. This implies that without a prior on δ b , constraints on cosmological parameters will be severely degraded by marginalizing δ b . This degradation is quantified in Tab. I (see
Appendix A3 for details and notation). With the Fisher matrix, we can use the eigenvector with the largest variance to identify the degenerate direction
$$\mathbb { e } \quad \pi _ { \degen } = 3. 5 \, \delta _ { b } + 2. 0 \, \delta h - 4. 9 \, \delta \ln A _ { s } + 1. 5 \, \delta n _ { s }, \quad ( 3 2 )$$
where we have normalized the vector such that the estimator of this mode has unit variance. This direction is displayed as the dashed lines in Fig. 9 (left).
Fig. 10 shows the response of the power spectrum to this combination. This linear combination nulls the response in the nonlinear region where changes in the tilt and halo concentration can compensate each other. The change in tilt implies a change in the linear regime which variations in h partially compensate at the expense of leaving a residual response in the BAO scale.
Next in Fig. 9 (right) we show the 2D error contours for the global case. As discussed in the previous section, the change in the δ b response from the local case eliminates the near degeneracy provided by the growth and dilation responses. Consequently the degradation in parameter errors from marginalizing δ b is much smaller as shown in Tab. I. The largest degradation occurs for ln A s and involves mainly the direction
$$\underset { 1 } { r } s \quad \pi _ { \deg e n } = 5. 5 \, \delta _ { b } - 0. 2 2 \, \delta h - 2 0 \, \delta \ln A _ { s } + 6. 4 \, \delta n _ { s }. \quad ( 3 3 )$$
Fig. 10 shows that this direction does not allow an approximate nulling of the response in the nonlinear region. Instead it yields a flat response in the nonlinear region which no longer requires significant variations in h to compensate in the linear regime.
FIG. 9. Error contours (68%) and distributions for joint δ b and cosmological parameter estimation with k max = 2 h Mpc -1 with (inner, black) and without (outer, blue) an external prior from σ b . Left panel: local case, with dashed lines representing the degeneracy direction from Eq. (32). Right panel: global case.
TABLE I. Error degradation from marginalizing δ b , with or without prior, for the local and global cases, the fiducial V 0 and 100 V 0 volumes and k max = 2 h Mpc -1 . Note that 'max' is the combination of cosmological parameters that is maximally degraded by marginalizing δ b and that degradation is numerically equal to that on the errors of δ b by marginalizing cosmological parameters.
| | local | local | global | global |
|-------------|----------|-----------|----------|-----------|
| parameter | no prior | σ b prior | no prior | σ b prior |
| V = V 0 | | | | |
| ln A s | 4.14 | 1.17 | 1.51 | 1.24 |
| n s | 2.37 | 1.05 | 1.09 | 1.04 |
| h | 2.54 | 1.06 | 1.10 | 1.04 |
| max, δ b | 5.02 | 1.24 | 4.34 | 2.91 |
| V = 100 V 0 | | | | |
| ln A s | 4.14 | 1.09 | 1.51 | 1.15 |
| n s | 2.37 | 1.03 | 1.09 | 1.03 |
| h | 2.54 | 1.03 | 1.10 | 1.03 |
| max, δ b | 5.02 | 1.13 | 4.34 | 2.37 |
Although these degeneracy studies, which identify the worst constrained directions, reveal the overall impact on the individual cosmological parameters, it is also interesting to consider the impact of marginalizing δ b on the best constrained directions. For example, if the CMB or other cosmological probe constrains a different combination of these parameters and breaks the degeneracy, then the impact of the best constrained directions may be revealed. Moreover, the impact of marginalizing δ b is generally larger on the best rather than worst constrained di- rections. In Appendix A 3, we formalize these statements by identifying the combination of cosmological parameters whose errors are most degraded by marginalizing δ b ,
$$\pi _ { \max } = 1 2 7 \, \delta h + 1 3 8 \, \delta \ln A _ { s } + 1 8 3 \, \delta n _ { s }, \text{ global,}$$
$$\pi _ { \max } = 6 7. 9 \, \delta h + 1 0 6 \, \delta \ln A _ { s } + 1 1 6 \, \delta n _ { s }, \text{ } \text{ } \text{local, } \text{ } ( 3 4 )$$
which in practice are in directions very similar to the best constrained direction without marginalization
$$\pi _ { \text{best} } = 1 2 5 \, \delta h + 1 3 7 \, \delta \ln A _ { s } + 1 9 9 \, \delta n _ { s }, \quad ( 3 5 )$$
especially in the global case. Here again we normalize each mode so that its estimator has unity variance, conditional on δ b held fixed. The amount of the degradation is numerically equal to that on δ b upon marginalizing over cosmological parameters. In Tab. I we compare the various degradation in errors. Note that even in the global case, the maximal degradation is large and comparable to the local case.
Finally given that the covariance matrix scales with survey volume according to Eq. (21), we can account for a change in survey volume by simply rescaling all parameter errors according to Eq. (22) so that the relative impact of δ b remains the same.
In summary, even though the local case involves a smaller response to δ b , it can have a much bigger impact on cosmological parameter estimation compared with the global case due to the ability to construct near perfect degeneracies in the nonlinear regime. On the other hand, errors in the combination of cosmological parameters along the direction that is best constrained without δ b are substantially degraded in both cases.
FIG. 10. Power spectrum responses in most degenerate directions in the local and global cases, Eq. (32) and (33), respectively. In the local case, the direction effectively nulls the response in the nonlinear regime creating a near perfect degeneracy. In the global case, no direction nulls the response and the degeneracy is much weaker. Responses are normalized to represent 1 σ deviations in degenerate parameter combination.
## D. Parameter constraints with δ b prior
If the linear power spectrum were perfectly predicted by external information on cosmological parameters such as the CMB measurements, then we would possess prior knowledge that δ b is distributed as a Gaussian with variance σ 2 b . In this limiting case
$$F _ { \mu \nu } ^ { \text{prior} } = \frac { 1 } { \sigma _ { b } ^ { 2 } } \delta _ { \delta _ { b } \mu } ^ { K } \delta _ { \delta _ { b } \nu } ^ { K }, \quad \quad \ ( 3 6 ) \quad \ \text{cai} \\ \ m c$$
where δ K µν is the Kronecker delta. Here we use Eq. (9) to evaluate σ b . In practice, cosmological parameter uncertainties should be propagated into this prior but this simple prior is useful to study as a best case scenario. Note that the view of super-sample effects as signal makes it simple to incorporate uncertainties in σ b as opposed to the excess covariance approach where it would enter through the cosmological parameter dependence of the covariance matrix.
The impact of this prior is qualitatively different in the local and global cases. In the local case, with all 4 parameters jointly estimated, intrinsic parameter degeneracies are so strong that the net degradation in parameter estimation from marginalizing δ b is prior limited as shown in Fig. 9 (left) and Tab. I. For the global case, the degradation is only marginally changed by the prior. In both cases, with the prior, the remaining impact of marginalizing δ b is mainly on ln A s and represents approximately a factor of 1 2 degradation in its errors. .
The reason that these degradations are small in both cases is the remaining degeneracies among the cosmological parameters themselves come to dominate the parameter errors. This fact however does not mean that marginalizing δ b has little effect on combination of cosmological parameters that are well constrained in its absence. The direction of maximal degradation is the same with or without the σ b prior and so is close to the best constrained direction in Eq. (35). In the global case, this degradation is nearly a factor of 3 and is substantially larger than the local case. This degradation can be important for errors on individual cosmological parameters if other measurements break the intrinsic cosmological parameter degeneracies.
Finally, with a δ b prior, parameter errors do not simply scale with volume as V -1 / 2 unless σ b ∝ V -1 / 2 . To test the impact of volume scaling we consider a case where the volume is increased by V/V 0 = 100. In this case σ V b 1 / 2 is reduced by a factor of 1 4. . For cases that are not limited by the prior, the volume scaling of C ij ensures that the degradation factors are the same. In general for ΛCDM, increasing the volume tends to slightly strengthen the relative impact of the prior.
## IV. DISCUSSION
Super-sample density fluctuations systematically change the observable power spectrum of sub-sample modes. In this paper, we have developed and tested the interpretation of this effect as a signal due to the dependence of the observed power spectrum on the mean density fluctuation in the survey volume δ b . This dependence can be calibrated efficiently by using the separate universe technique that absorbs the fluctuation into a change in the background parameters.
This interpretation has the advantage that the effect can be incorporated into parameter estimation without modification of traditional procedures or pipelines. The form of the likelihood of the power spectrum data as a function of the model takes the same form except that the model gains a parameter δ b in addition to cosmological parameters. Its impact on cosmological parameter estimation comes through parameter degeneracies. Contrast this with the alternate but equivalent view that when ensemble averaged over many realizations of the survey volume and δ b , the effect induces a covariance in the power spectrum that modifies the form of the likelihood function.
The super-sample signal allows δ b itself to be estimated from power spectrum data. The amount of information on δ b depends on whether the power spectrum is measured with respect to the global or local mean density, which is relevant, e.g., for weak lensing or galaxy clustering, respectively. For a wide range of survey volumes, the global case contains substantial extra information in the nonlinear regime on top of the prior expectation that it be limited by the rms σ b predicted by linear theory. For the local case, the extra information is comparable to the prior for k max = 2 h Mpc -1 .
If cosmological parameters are jointly estimated, this extra information can be lost to degeneracies. Likewise,
marginalization of δ b can degrade errors on cosmological parameters. The degradation takes on different values depending on the the parameter space and priors considered, whether the local or global power spectra are used, and the maximum wavenumber utilized. In general, the strongest degeneracies arise from compensated changes in the growth of structure and dilation of features induced by the parameters, which provides a physical basis for which to extend our results beyond the ΛCDM parameter space. Without prior information on σ b , the degradation of errors for the local case can be more than a factor of 4 for k max = 2 h Mpc -1 and likewise in the global case but only for the combination of cosmological parameters that is best measured in the absence of δ b and maximally degraded by its marginalization. Even with a prior that reflects perfect knowledge of the linear power spectrum, the maximal degradation in the global case can reach a factor of 2-3. Fortunately all of these cases and more can be simply and rapidly considered given a single calibration of the power spectrum responses.
While we have only considered the effects on the matter power spectrum through N -body simulations in ΛCDM, our separate universe and growth-dilation techniques can be extended to other parameter spaces or observables and can incorporate baryonic effects and galaxy formation. While other uncertainties in modeling the data, e.g. redshift space distortions and galaxy bias for galaxy surveys, may dominate the error budget for current measurements, the super-sample mode provides an intrinsic limitation to extracting cosmological information that is degenerate with its effects.
Acknowledgments.We thank M. Becker, M. Busha, B. Erickson, G. Evrard, N. Gnedin, A. Kravtsov, D. Rudd, R. Wechsler, and the University of Chicago Research Computing Center for running, storing, and allowing us to use the large-volume simulations in this study. We also thank M. Becker, D. Rudd., N. Gnedin, A. Kravtsov and F. Schmidt for useful discussions. YL and WH were supported by U.S. Dept. of Energy contract DE-FG02-13ER41958 and the Kavli Institute for Cosmological Physics at the University of Chicago through grants NSF PHY-0114422 and NSF PHY-0551142. MT was supported by World Premier International Research Center Initiative (WPI Initiative), MEXT, Japan, by the FIRST program Subaru Measurements of Images and Redshifts (SuMIRe), CSTP, Japan, and by Grantin-Aid for Scientific Research from the JSPS Promotion of Science (Nos. 23340061 and 26610058). This work was also supported in part by the National Science Foundation under Grant No. PHYS-1066293 and the hospitality of the Aspen Center for Physics, where this work was completed. Plots in this work are made with matplotlib [30].
TABLE II. Parameters of baseline flat ΛCDM model used throughout.
| Ω m | Ω b | h | n s | σ 8 |
|-------|-------|-----|-------|-------|
| 0.286 | 0.047 | 0.7 | 0.96 | 0.82 |
## Appendix A: Numerical Implementation
In this Appendix we provide details on various numerical calculations in the main paper. In § A1, we describe the cosmological simulations used in the super-sample signal studies. We calibrate the power spectrum response to various parameters in § A2. In § A3, we establish the formalism for calculating the degradation of parameter errors upon marginalizing over δ b .
## 1. Simulations
Here we summarize the salient features of the simulations and power spectrum analysis from Ref. [14] and used in § II to test super-sample effects with subvolumes of large-volume simulations. For the large-volume simulations, we take a suite of 7 realizations of the fiducial cosmology given in Tab. II, originally made for the Dark Energy Survey. Each of these has a 4 h -1 Gpc box length evolved from initial conditions at a i = 0 02 that are pro-. vided by CAMB [31, 32] and 2LPTIC [33], using L-Gadget2 [34] with 2048 3 particles and 3072 3 (Tree-)PM grid. We then assign the particles to a (8 × 1920) 3 grid with a cloudin-cell (CIC) scheme, before subdividing each large box into 8 3 = 512 subvolumes of size
$$V _ { 0 } = ( 5 0 0 h ^ { - 1 } \text{Mpc} ) ^ { 3 } \text{ \quad \ \ } ( A 1 )$$
for a total of N s = 3584 subboxes.
In each subvolume, we extract the mean density fluctuation δ b and the power spectrum by FFT. For the power spectrum, we then deconvolve the CIC window and bin the result to 80 logarithmically spaced k -bins per decade to form ˆ P sub i . Each bin is positioned at the average k i weighted by the number of modes. This binning scheme is used throughout the paper.
To calibrate the mean power spectrum P i and its covariance matrix C ij , in the absence of δ b , we use the same number N s of simulations of the same size as the subboxes but with 256 3 particles and 512 3 (Tree-)PM grid. We measure the power spectrum ˆ P sm of each small-box simulation in the same way with a 1920 3 grid. All the numerical settings match the 1 / 8 scaling of the large box dimensions except for the (Tree-)PM grid.
The mean power spectrum of the subboxes differs from that of the small boxes in two ways. On large scales the difference is dominated by convolution bias from the subbox window, and on small scales there is a 1% difference due mainly to using different resolutions for the (Tree)PM grid in simulations. We debias the estimator as in
Ref [14] by rescaling
$$\hat { \mathcal { P } } _ { i } = \frac { \mathcal { P } _ { i } ^ { \text{sm} } } { \mathcal { P } _ { i } ^ { \text{sub} } } \hat { \mathcal { P } } _ { i } ^ { \text{sub} }, \quad \quad \quad ( A 2 ) \quad \begin{pmatrix} \sum \\ \delta \Omega \\ \text{as} \\ \text{is} \end{pmatrix}$$
where we have defined
$$\mathcal { P } _ { i } ^ { X } \equiv \frac { 1 } { N _ { s } } \sum _ { a = 1 } ^ { N _ { s } } \hat { \mathcal { P } } _ { i } ^ { X, a } \quad \quad \ ( A 3 ) \quad \text{in} \quad \ \text{in} _ { \ g l c }$$
as the average over the N s samples. Thus
$$\langle \hat { \mathcal { P } } _ { i } \rangle = \mathcal { P } _ { i } ^ { \text{sm} } \equiv \mathcal { P } _ { i }. \quad \quad \ \ ( \text{A4} ) \quad \text{sp} \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
For the power spectra referenced to the subbox mean
$$\hat { \mathcal { P } } _ { i } ^ { W } = \frac { \hat { \mathcal { P } } _ { i } } { ( 1 + \delta _ { b } ) ^ { 2 } }, \quad \quad ( A 5 ) \quad \begin{array} { c } \text{$\text{$se$} \\ \text{$Gi$} \\ \text{is} \end{array} }$$
where δ b is the average density fluctuation in the same box.
Next we estimate the covariance matrix in the absence of δ b as
$$C _ { i j } = \frac { N _ { s } } { N _ { s } - 1 } \left [ \frac { \sum _ { a = 1 } ^ { N _ { s } } \hat { \mathcal { P } } _ { i } ^ { s m, a } \hat { \mathcal { P } } _ { j } ^ { s m, a } } { N _ { s } } - \mathcal { P } _ { i } ^ { s m } \mathcal { P } _ { j } ^ { s m } \right ]. \ \ ( A 6 ) \quad \begin{pmatrix} \text{$\text{$\text{$to$the}$} \end{pmatrix} \\ \text{$\text{$\text{$Separ}$} }. \end{pmatrix}$$
Finally unless otherwise specified, the simulations used to construct the response functions in the next section follow the prescription for the small box simulations in order to preserve the same mass and force resolution.
## 2. Response calibrations
In this section we provide some detail on the calibration of the power spectrum response to δ b and the cosmological parameters. We review the δ b response calibrated in Ref. [14] and illustrate the cosmological parameter calibration with h as it demonstrate all the important concept and techniques.
## a. δ b response
Following the separate universe technique developed in Ref. [14] (see also [18-20]), a nonzero mean density fluctuation δ b at z = 0 can be absorbed into the background by a redefinition of the cosmological parameters compared to those of a global ΛCDM universe
$$\frac { \delta \Omega _ { m } } { \Omega _ { m } } \approx \frac { \delta \Omega _ { \Lambda } } { \Omega _ { \Lambda } } \approx - 2 \frac { \delta h } { h } \approx \frac { 5 \Omega _ { m } } { 3 } \frac { \delta _ { b } } { D _ { 0 } } \quad \ ( A 7 ) \quad \ \text{The} \quad \text{spo} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$} \quad \text{$\Phi$}$$
where the D 0 = D z ( = 0) and the linear growth function is normalized as
$$\lim _ { a \to 0 } D = a. \quad \quad \ \ ( A 8 ) \quad \ \ s p$$
glyph[negationslash]
Note that even if the global universe is flat, the separate universe would have a nonzero spatial curvature δ Ω m + δ Ω Λ = 0. Finally, as discussed in § III B, the scale factor associated with a given value of a in the global cosmology is shifted by
$$\frac { \delta a } { a } \approx - \frac { 1 } { 3 } \frac { D } { D _ { 0 } } \delta _ { b },$$
in the separate universe. For example, at z = 0 in the global universe, z L = δ b / 3 in the local universe.
With the separate universe cosmological parameters set, we conduct N -body simulations to calibrate the response of the power spectrum by finite difference of models with δ b = ± 0 01 evaluated at . z L . Here and below we always difference simulations with the same initial seeds to suppress the stochasticity from sample variance. Given that the mean density of the separate universe is reinterpreted as the local density of the finite survey, power spectra extracted from these simulations are always referenced to the local mean, P W in Eq. (5).
To calibrate the growth response T δ b , g , we fix the simulation box in comoving Mpc and difference the results from ± δ b . This procedure only includes the impact of δ b on the growth of structure and omits the fact that due to the difference in redshift, the physical scale associated with a given comoving scale differs by the dilation factor. Separately, we also calculate the total response T δ b | local by instead fixing the physical scale in Mpc of the simulations at the final redshift with the same initial seeds in box coordinates. Finally we average over 64 pairs of realizations to reduce the remaining stochasticity to a level that is negligible for our purposes, with standard errors of the mean of a few percent or better, and at sub-percent level in the nonlinear regime.
To test the precision of our results, we have employed simulations with twice the mass and PM resolutions, to verify that at z = 0 for k glyph[lessorsimilar] 2 h /Mpc the responses have converged to several percent or better. We refer the readers to Ref. [14] for more details of the calibration pipeline.
With T δ b , g and T δ b | local calibrated from simulations we can construct the dilation response from Eq. (26)
$$T ^ { \delta _ { b }, d } = - 3 \left ( T ^ { \delta _ { b } } \Big | _ { \text{local} } - 2 \frac { \partial \ln D } { \partial \delta _ { b } } T ^ { \delta _ { b }, g } \right ). \quad ( A 1 0 )$$
We compare this constructed dilation response in Fig. 7 with the response calibrated directly as the slope of a cubic spline fitted to the mean power spectrum of N s small box simulations
$$T ^ { \text{d} } = \frac { \partial \ln \mathcal { P } } { \partial \ln k }. \quad \quad \text{(A11)}$$
The differences in fact reflect that the constructed response reduces the stochasticity from sample variance and the sensitivity to systematic changes in the slope from finite resolution. We have verified using higher resolution simulations that the constructed response is both more precise and more accurate than the slope-based response.
FIG. 11. Power spectrum response to h from finite differences of simulations with δh = ± 0 02 and scales fixed in . h -1 Mpc. Thin gray lines show differences of 64 pairs of realizations, with solid black lines representing their means with standard errors. A single calibration from one pair of simulations is highlighted in blue to illustrate that stochasticity around the mean are highly correlated across nonlinear k . In the nonlinear regime where most information is located, correction from the second derivative in dotted line is much smaller compared to the response itself, demonstrating that linear response serves as a good approximation for the power spectrum variation.
## b. h response
As an example of cosmological parameter response calibration, we choose h here as it demonstrate all the important concept and techniques, including the growthdilation split, and the test of the linear approximation. We start with the baseline cosmology in Tab. II and utilize a suite of 64 simulations from the small box simulations of § A1. As in the δ b calibration we then simulate pairs of models with h ′ = h ± δh with the same seeds to form a triplet of simulations at fixed ln A s , n s , (Ω b h 2 ) ′ = Ω b h 2 and (Ω c h 2 ) ′ = Ω c h 2 in a flat universe. We choose δh = 0 02 in accordance with the 68% con-. fidence limits constrained by Planck and WMAP polarization data.
For the total response T h , we set the comoving size of simulation boxes in h -1 Mpc to be the same, i.e.
$$L ^ { \prime } = 5 0 0 \, \text{Mpc} / h ^ { \prime }, \text{ \quad \ \ } ( \text{A12} )$$
whereas for extracting the growth component, we set the box scale in Mpc to be the same
$$L ^ { \prime } = 5 0 0 \, \text{Mpc/} h = 5 0 0 \frac { h ^ { \prime } } { h } \, \text{Mpc/} h ^ { \prime }. \quad \ \ ( A 1 3 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
We then difference the binned power spectrum ˆ P in box coordinates in each case to form the required deriva- tive
$$\hat { T } ^ { h } \approx \frac { \ln \hat { \mathcal { P } } ( + \delta h ) - \ln \hat { \mathcal { P } } ( - \delta h ) } { 2 \delta h }. \quad \quad ( A 1 4 )$$
For sufficiently small δh , this finite difference converges to the derivative as required for a Fisher matrix calculation. Each of the 64 pairs provides a separate estimate of the response. In Fig. 11 we show the individual estimates and the mean of the suite T h . Note that run-to-run deviations from the mean strongly covary across k in the nonlinear regime.
Finally, to test the linearity of the response, we calibrate the change in the response from the second derivative
$$\delta T ^ { h } = \frac { \ln \mathcal { P } ( + \delta h ) - 2 \ln \mathcal { P } ( 0 ) + \ln \mathcal { P } ( - \delta h ) } { \delta h }. \quad ( A 1 5 )$$
using the averages over the 64 triplets. In Fig. 11, we demonstrate that this second derivative error term | δT h | glyph[lessmuch] T h . Specifically, in the fully nonlinear regime where the statistical power lies, the error correction is glyph[lessorsimilar] 5%. While we could further reduce this error by choosing a smaller δh , this test demonstrates that even for current uncertainties on this parameter, the Fisher approximation should suffice.
To test the resolution dependence, we have employed 16 pairs of higher resolution simulations with 512 3 particles and 1024 3 (Tree-)PM grid to verify that at k glyph[lessorsimilar] 2 h Mpc -1 our response results have converged to percent level or better.
With both T h and T h, g calibrated in this manner, we can construct the dilation response using Eq. (31),
$$T ^ { h, d } = h \left ( T ^ { h } - 2 \frac { \partial \ln D } { \partial h } T ^ { h, g } \right ). \quad \text{ (A16)}$$
Similar to the δ b case discussed above, this construction yields a more accurate and precise dilation response than the slope-based response.
The calibration of T ln A s and T n s are simpler as no scale dilation is involved, so that we can use the same box size for simulations. We take δ ln A s = 0 03 and . δn s = 0 01 . again according to the CMB prior, and the rest of the procedures are the same as for h . For T ln A s and T n s , the linear response assumption is an excellent approximation, with second order corrections at the percent level or smaller.
## 3. Error Degradation
In the main text, we quote the degradation in the errors in a given cosmological parameter and the maximal degradation for any linear combination of parameters caused by marginalizing δ b . Here we give details for those calculations.
The covariance matrix of cosmological parameters with δ b marginalized is simply the 3 × 3 subblock of the 4 × 4
inverse Fisher matrix that contains them. We can formalize the extraction of this matrix by defining a 4 × 3 projection matrix
$$\mathbf P = \left ( \begin{array} { c } 0 \\ \mathbf I _ { 3 } \end{array} \right ), \quad \quad \ \ ( A 1 7 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\quad \text{$\text{ca}$} }$$
where 0 = (0 0 0) and , , I 3 is the 3 × 3 identity matrix. Thus
$$C _ { \text{mar} } \equiv \text{P} ^ { \text{T} } \text{F} ^ { - 1 } \text{P}. \quad \quad ( A 1 8 ) \quad \text{or}$$
Conversely, if δ b is fixed then the covariance matrix is instead the inverse of the projected Fisher matrix
$$C _ { \text{fix} } \equiv [ \text{P} ^ { \text{T} } \text{F} \text{P} ] ^ { - 1 }. \quad \quad ( A 1 9 ) \quad \text{Fo}$$
In the same linear approximation employed in the Fisher analysis, the deviation in any derived parameter from its fiducial value can be thought of as a linear combination of changes in fundamental parameters,
$$\pi = \sum _ { \mu } \frac { \partial \pi } { \partial p _ { \mu } } \, \delta p _ { \mu } \equiv \mathbf v _ { \pi } \cdot \delta \mathbf p, \text{ \quad \ \ } ( A 2 0 )$$
where Greek indices in this section run over the 3 cosmological parameters and parameter vectors v π lie in this space. Thus the degradation in errors d π upon marginalizing δ b is given by
$$d _ { \pi } ^ { 2 } \equiv \frac { \sigma _ { \text{mar} } ^ { 2 } ( \pi ) } { \sigma _ { \text{fix} } ^ { 2 } ( \pi ) } = \frac { \mathbf v _ { \pi } ^ { \text{T} } \text{C} _ { \text{mar} } \mathbf v _ { \pi } } { \mathbf v _ { \pi } ^ { \text{T} } \text{C} _ { \text{fix} } \mathbf v _ { \pi } }. \quad \text{(A21)} \quad \text{ is n}$$
We can most easily study the degradation in errors in a new basis, called the Karhunen-Lo`ve basis. e The 3 basis vectors v k are linearly independent solutions of the generalized eigenvector equation
$$C _ { \text{mar} } \mathbf v _ { k } = \lambda _ { k } C _ { \text{fix} } \mathbf v _ { k }. \quad \quad ( A 2 2 ) \quad \text{an} \quad.$$
Note that λ k = d 2 k , the degradation in the variance of a parameter defined by v k and we can normalize these statistically independent vectors as
$$v _ { k } ^ { T } C _ { \text{mar} } \mathbf v _ { k ^ { \prime } } = \lambda _ { k } \delta _ { k k ^ { \prime } } ^ { K }, \ \mathbf v _ { k } ^ { T } C _ { \text{fix} } \mathbf v _ { k ^ { \prime } } = \delta _ { k k ^ { \prime } } ^ { K }. \quad ( A 2 3 )$$
These vectors are not typically orthonormal in the usual sense v k · v k ′ = δ K kk ′ .
glyph[negationslash]
[1] M. Takada, R. S. Ellis, M. Chiba, J. E. Greene, H. Aihara, N. Arimoto, K. Bundy, J. Cohen, O. Dor´ e, G. Graves, J. E. Gunn, T. Heckman, C. M. Hirata, P. Ho, J.-P. Kneib, O. L. F`vre, L. Lin, S. More, H. Mue rayama, T. Nagao, M. Ouchi, M. Seiffert, J. D. Silverman, L. Sodr´ e, D. N. Spergel, M. A. Strauss, H. Sugai, Y. Suto, H. Takami, and R. Wyse, PASJ 66 , R1 (2014), arXiv:1206.0737 [astro-ph.CO].
The Karhunen-Lo` eve basis is typically used when the two matrices in Eq. (A22) are the signal and noise covariance respectively and in that context the eigenvectors are called signal-to-noise eigenvectors. In this case, they represent degradation eigenvectors.
The eigenvalues λ k are particularly easy to find in our case given the relationship between the two matrices implies
$$( \mathbf P ^ { T } \mathbf F \mathbf P ) ( \mathbf P ^ { T } \mathbf F ^ { - 1 } \mathbf P ) \mathbf v _ { k } = \lambda _ { k } \mathbf v _ { k }, \quad ( A 2 4 )$$
or in terms of components
$$\sum _ { \nu } \left ( \delta _ { \mu \nu } ^ { K } - F _ { \delta _ { b } \mu } [ \mathbf F ^ { - 1 } ] _ { \delta _ { b } \nu } \right ) [ \mathbf v _ { k } ] _ { \nu } = \lambda _ { k } [ \mathbf v _ { k } ] _ { \mu }. \quad ( \mathbf A 2 5 ) \\ \text{For source in the future divoctime that are orthogonal to}$$
For vectors in the two directions that are orthogonal to [ F -1 ] δ b µ , the second term on the left hand side vanishes and λ k = 1. In these directions, marginalization of δ b has no impact since these elements of F -1 represent a covariance of the associated direction with δ b . The remaining eigenvector is in the direction of
$$[ \mathbf v _ { \max } ] _ { \mu } \in F _ { \delta _ { \mathbf b } \mu }, \quad \quad ( A 2 6 )$$
which is not in general parallel to [ F -1 ] δ b µ or orthogonal to the 2D space spanned by the other vectors. Thus the direction of maximal degradation and the combination of cosmological parameters associated with it
$$\pi _ { \max } = \mathbf v _ { \max } \cdot \delta \mathbf p \text{ \quad \ \ } ( A 2 7 )$$
is not in general the same as those of maximal covariance or degeneracy. The maximal degradation itself is given by the eigenvalue
$$\lambda _ { \max } = d _ { \pi _ { \max } } ^ { 2 } = F _ { \delta _ { b } \delta _ { b } } [ \mathbf F ^ { - 1 } ] _ { \delta _ { b } \delta _ { b } } \quad \ ( A 2 8 )$$
and is exactly the degradation in σ δ b on marginalizing over cosmological parameters. A general linear combination of cosmological parameters suffers a degradation whose value cannot exceed the maximum since it is composed partially of the 2 Karhunen-Lo`ve directions that e are not degraded [35].
- [2] A. J. S. Hamilton, C. D. Rimes, and R. Scoccimarro, Mon. Not. Roy. Astron. Soc. 371 , 1188 (2006), arXiv:astro-ph/0511416.
- [3] E. Sefusatti, M. Crocce, S. Pueblas, and R. Scoccimarro, Phys. Rev. D 74 , 023522 (2006), arXiv:astro-ph/0604505.
- [4] W. Hu and A. V. Kravtsov, Astrophys. J. 584 , 702 (2003), arXiv:astro-ph/0203169.
- [5] M. Takada and S. Bridle, New Journal of Physics 9 , 446
- (2007), arXiv:arXiv:0705.0163.
- [6] M. Takada and B. Jain, Mon. Not. Roy. Astron. Soc. 395 , 2065 (2009), arXiv:0810.4170.
- [7] M. Sato, T. Hamana, R. Takahashi, M. Takada, N. Yoshida, T. Matsubara, and N. Sugiyama, Astrophys. J. 701 , 945 (2009), arXiv:0906.2237 [astro-ph.CO].
- [8] R. Takahashi, N. Yoshida, M. Takada, T. Matsubara, N. Sugiyama, I. Kayo, A. J. Nishizawa, T. Nishimichi, S. Saito, and A. Taruya, Astrophys. J. 700 , 479 (2009), arXiv:0902.0371 [astro-ph.CO].
- [9] R. de Putter, C. Wagner, O. Mena, L. Verde, and W. Percival, JCAP 1204 , 019 (2012), arXiv:1111.6596 [astro-ph.CO].
- [10] M. D. Schneider, S. Cole, C. S. Frenk, and I. Szapudi, Astrophys.J. 737 , 11 (2011), arXiv:1103.2767 [astroph.CO].
- [11] I. Kayo, M. Takada, and B. Jain, Mon. Not. Roy. Astron. Soc. 429 , 344 (2013), arXiv:1207.6322 [astroph.CO].
- [12] M. Takada and D. N. Spergel, ArXiv e-prints (2013), arXiv:1307.4399 [astro-ph.CO].
- [13] M. Takada and W. Hu, Phys. Rev. D 87 , 123504 (2013), arXiv:1302.6994 [astro-ph.CO].
- [14] Y. Li, W. Hu, and M. Takada, Phys. Rev. D 89 , 083519 (2014), arXiv:1401.0385.
- [15] E. Schaan, M. Takada, and D. N. Spergel, ArXiv e-prints (2014), arXiv:1406.3330.
- [16] G. Tormen and E. Bertschinger, Astrophys. J. 472 , 14 (1996), astro-ph/9512131.
- [17] S. Cole, Mon. Not. Roy. Astron. Soc. 286 , 38 (1997), astro-ph/9604046.
- [18] E. Sirko, Astrophys. J. 634 , 728 (2005), arXiv:astroph/0503106.
- [19] T. Baldauf, U. Seljak, L. Senatore, and M. Zaldarriaga, JCAP 1110 , 031 (2011), arXiv:1106.5507 [astro-ph.CO].
- [20] N. Y. Gnedin, A. V. Kravtsov, and D. H. Rudd, Astrophys.J.Suppl. 194 , 46 (2011), arXiv:1104.1428 [astroph.CO].
- [21] C.-T. Chiang, C. Wagner, F. Schmidt, and E. Komatsu, ArXiv e-prints (2014), arXiv:1403.3411 [astro-ph.CO].
- [22] D. H. Rudd, A. R. Zentner, and A. V. Kravtsov, Astrophys.J. 672 , 19 (2008), arXiv:astro-ph/0703741 [ASTRO-PH].
- [23] M. P. van Daalen, J. Schaye, C. M. Booth, and C. Dalla Vecchia, Mon. Not. Roy. Astron. Soc. 415 , 3649 (2011), arXiv:1104.1174 [astro-ph.CO].
- [24] R. Scoccimarro, M. Zaldarriaga, and L. Hui, Astrophys. J. 527 , 1 (1999), arXiv:astro-ph/9901099.
- [25] A. Meiksin and M. White, Mon. Not. Roy. Astron. Soc. 308 , 1179 (1999), astro-ph/9812129.
- [26] I. Mohammed and U. Seljak, (2014), arXiv:1407.0060 [astro-ph.CO].
- [27] Fixed as opposed to a parameter dependent scale such as the virial scale. For example r 180 where the spherical overdensity is 180 times the mean matter density.
- [28] C. D. Rimes and A. J. S. Hamilton, Mon. Not. Roy. Astron. Soc. 360 , L82 (2005), arXiv:astro-ph/0502081.
- [29] C. D. Rimes and A. J. S. Hamilton, Mon. Not. Roy. Astron. Soc. 371 , 1205 (2006), arXiv:astro-ph/0511418.
- [30] J. D. Hunter, Computing In Science & Engineering 9 , 90 (2007).
- [31] A. Lewis, A. Challinor, and A. Lasenby, Astrophys. J. 538 , 473 (2000), astro-ph/9911177.
- [32] C. Howlett, A. Lewis, A. Hall, and A. Challinor, JCAP 1204 , 027 (2012), arXiv:1201.3654 [astro-ph.CO].
- [33] http://cosmo.nyu.edu/roman/2LPT/.
- [34] V. Springel, S. D. White, A. Jenkins, C. S. Frenk, N. Yoshida, et al. , Nature 435 , 629 (2005), arXiv:astroph/0504097 [astro-ph].
- [35] K. M. Smith, W. Hu, and M. Kaplinghat, Phys.Rev. D74 , 123002 (2006), arXiv:astro-ph/0607315 [astro-ph].
|
10.1103/PhysRevD.90.103530
|
[
"Yin Li",
"Wayne Hu",
"Masahiro Takada"
] |
2014-08-05T19:55:25+00:00
|
2014-12-02T15:24:16+00:00
|
[
"astro-ph.CO"
] |
Super-Sample Signal
|
When extracting cosmological information from power spectrum measurements, we
must consider the impact of super-sample density fluctuations whose wavelengths
are larger than the survey scale. These modes contribute to the mean density
fluctuation $\delta_b$ in the survey and change the power spectrum in the same
way as a change in the cosmological background. They can be simply included in
cosmological parameter estimation and forecasts by treating $\delta_b$ as an
additional cosmological parameter enabling efficient exploration of its impact.
To test this approach, we consider here an idealized measurement of the matter
power spectrum itself in the $\Lambda$CDM cosmology though our techniques can
readily be extended to more observationally relevant statistics or other
parameter spaces. Using sub-volumes of large-volume $N$-body simulations for
power spectra measured with respect to either the global or local mean density,
we verify that the minimum variance estimator of $\delta_b$ is both unbiased
and has the predicted variance. Parameter degeneracies arise since the response
of the matter power spectrum to $\delta_b$ and cosmological parameters share
similar properties in changing the growth of structure and dilating the scale
of features especially in the local case. For matter power spectrum
measurements, these degeneracies can lead in certain cases to substantial error
degradation and motivates future studies of specific cosmological observables
such as galaxy clustering and weak lensing statistics with these techniques.
|
1408.1082v2
|
## Interpretations of anomalous LHC events with electrons and jets
Bogdan A. Dobrescu /star and Adam Martin /diamondmath
- /star Theoretical Physics Department, Fermilab, Batavia, IL 60510, USA /diamondmath Department of Physics, University of Notre Dame, Notre Dame, IN 46556, USA
August 5, 2014; revised September 7, 2014
## Abstract
The CMS Collaboration has recently reported some excess events in final states with electrons and jets, in searches for leptoquarks and W ′ bosons. Although these excesses may be due to some yet-to-be-understood background mismodeling, it is useful to seek realistic interpretations involving new particles that could generate such events. We show that resonant pair production of vector-like leptons that decay to an electron and two jets leads to kinematic distributions consistent with the CMS data.
## I. INTRODUCTION
The success of the Standard Model (SM) in describing the large number of final states analyzed so far at the LHC is remarkable. At the same time, there are many hypothetical particles whose signatures have not been searched for yet, especially when two or more of these particles are produced in the same process. An example is provided by the pair production of some new heavy particle whose decays are controlled by an approximate symmetry. The dominant decays of such a particle may involve peculiar flavor combinations of many SM particles, while the constraints from low-energy flavor processes are avoided if the breaking of the symmetry that controls the decay parameters is sufficiently soft.
Nevertheless, the ensuing signals could show up in some of the existent searches, even though the event selection is not chosen for those signals.
Recently, the CMS Collaboration has reported excess events involving electrons and jets, while similar events involving muons and jets are well described by the SM. In the search for a W ′ boson decaying to the eejj final state [1], the leading two electrons and two jets form an invariant mass peak near 2.1 TeV; even though the statistical significance is only 2 8 . σ , this signal is sufficiently clean to warrant further scrutiny. Based on some kinematic distributions, the CMS Collaboration concluded that the signal is not consistent with the W ′ hypothesis.
In the CMS search for first-generation leptoquarks [2] the leading two electrons and two jets pair up so that the ej invariant mass exhibits a broad excess with a maximum at about 500 GeV. The statistical significance of this excess is 2 4 . σ when a cut on the invariant mass of the ej system with smaller m ej is imposed ( m min ej > 360 GeV). When that cut is removed, the deviation from the SM increases substantially, despite larger systematic uncertainties. The leptoquark hypothesis would imply a very narrow ej invariant mass peak, which is again not consistent with the observed excess.
The leptoquark search [2] has also led to an excess of events with one electron, two jets and missing transverse energy ( / E T ), roughly consistent with the hypothesis that one leptoquark decays into ej and the other one into νj . This eνjj signal has a statistical significance of 2 6 . σ , and entails a broad ej peak at around 700 GeV.
Given that the event selections are sufficiently complicated (especially for the leptoquark searches), it may be that there are correlations between the events that contribute to these three excesses, in which case the combined statistical significance my not be as large as it seems. It is also possible that the background modeling is not sufficiently accurate, but this would be surprising for the very high p T events involved in the W ′ and leptoquark searches.
Here we investigate the possibility that the eejj and eνjj excesses are due to some new particles. The resonant production of leptoquarks, investigated in Ref. [3], leads to an eejj invariant mass peak but does not appear to be consistent with the broad ej peaks. By contrast, our starting point is resonant signatures involving four (or more) jets plus an e + e -pair. Such final states lead to combinatoric ambiguities due to multiple choices for the
leading two jets selected by the CMS leptoquark search. As a result the ej invariant mass distributions become very wide. Furthermore, we show that although the ee +2 j invariant mass peak has a flatter distribution within these ee +4 j final states compared to a resonant eejj signal, the remaining edge-like feature provides a good fit to the excess in the CMS W ′ search.
In Section II we describe the details of the CMS searches with electrons and jets, and select the type of particles that may be responsible for the observed excesses. In Section III we propose a few models involving a Z ′ boson that is produced in the s channel and decays into a pair of heavy vector-like leptons. The decays of the vector-like leptons lead to signatures that may explain the CMS events, as shown in Section IV. We discuss additional tests of these models, and then in Section V provide a brief outlook.
## II. EVENT SELECTIONS AND SIGNAL FEATURES
We start by reviewing the analysis cuts imposed in each of the three CMS searches with electrons and jets, and then by drawing some conclusions about the type of new particles that could cause the disagreement between data and SM expectation.
In the first-generation leptoquark searches [2], both the single- and di-electron searches require at least two jets, p T > 45 GeV , | η | < 2 4, with the leading jet satisfying . p T > 125 GeV. The analysis uses the antik T jet algorithm with size R = 0 5. . A minimum jet-electron separation of 0 3 is also imposed. . Electron candidates must have p T > 45 GeV, and | η | < 2 5; . exactly two candidates are required for the di-electron channel, while in the single electron channel there is an additional requirement of large missing energy that is azimuthally well separated from the electron and leading jet: / E T > 55 GeV, ∆ φ / ( E ,e T ) > 0 8 . , ∆ ( φ / E ,j T ) > 0 5. . In both channels, any events containing a muon p T > 10 GeV are vetoed.
Further cuts, chosen to optimize the analyses for a given leptoquark mass, are subsequently applied in each channel. For the di-electron channel, the cuts used for optimization are the scalar sum of the p T of the two leptons and the leading two jets S T , the di-electron mass m ee and the minimum electron-jet invariant mass m min ej . The m min ej is calculated by first pairing each electron in the event with one of the hardest two jets; there are two ways
to pair the objects and the choice that is most consistent with production of a pair of equalmass ej resonances is made. Once the pairing has been decided, the minimum mass of the two ej systems becomes m min ej . The optimization cuts for a 650 GeV leptoquark signal are S T > 850 GeV , m ee > 155 GeV and m min ej > 360 GeV. For these cut values, 36 events are observed compared with 20 49 . ± 2 4 . ± 2 45(syst.) . events expected from SM backgrounds ( Z +jets , tt ¯ , and multijet QCD). The excess has been examined for jet-flavor content and does not appear to contain b -jets. When the m min ej cut is removed, Figure 11 of Ref. [2] shows that there are additional 49 events observed for a background in the range of 23-33 events.
The fact that the m min ej distribution is so broad suggests that there is no on-shell particle decaying to ej . In Section III we construct a model where a very heavy scalar leptoquark, integrated out, induces 3-body decays of a parent fermion, leading to a broad m min ej distribution. To have the peak of that distribution at ∼ 500 GeV it is necessary for the new fermion (which decays to ejj ) to be somewhat heavier than that. We also construct two models where there is no particle coupled to an electron and a jet, so that any pairing of one electron and one jet in the final state is non-resonant; a broad m min ej peak is instead induced by cuts and kinematic limits.
In the eνjj analysis of the leptoquark search [2], the optimization cuts are on the missing energy / E T , S T (defined now as the scalar sum of the missing energy with the p T of the electron and of the two leading jets), the mass of the electron-jet system m ej , and the transverse mass of the / E T -electron system m T,eν . As with m min ej above, there are two possible choices for m ej . Here, the choice that minimizes | m T,ej -m T,νj | is used. For the same leptoquark signal assumption as in the di-electron channel, the optimization cut values are S T > 1040 GeV, / E T > 145 GeV, m ej > 555 GeV and m T,eν > 270 GeV, under which 18 events are observed with 7 54 . ± 1 20 . ± 1 07(syst.) expected. .
It is intriguing that the peak of the m ej distribution in the eνjj search is at a value comparable to that of the m min ej distribution in the 'leptoquark' eejj search. This suggests that one of the electrons observed in the eejj search originates from a vertex involving the lepton doublet ( ν e L , e L ). A further supporting fact is that the cross sections for the two excesses are comparable, of order 1 fb.
The excess di-electron events observed in the W ′ → eejj search [1] have been subjected
to a different, though not orthogonal set of analysis cuts. There, the electron cuts are staggered: p T > 60 GeV , | η | < 2 5 ( . p T > 40 GeV , | η | < 2 5) . for the leading (subleading) electron. At least two jets are required, as before, though at a slightly lower p T threshold of 40 GeV. For all events that satisfy m ee > 200 GeV, the two electrons and two leading jets are combined, and the total invariant mass is calculated. While m eejj data and the SM prediction agree at lower values, there is an excess of events quoted as 2 8 . σ deviation in the 1 8 TeV . < m eejj < 2 2 TeV bin (roughly 14 events observed with 4 expected). . The similar ATLAS search for W ′ → /lscriptN → /lscript/lscriptj j used only 2.1 fb -1 of 7 TeV data [4], and is not sensitive to masses of 2 TeV.
The concentration of events in one high-mass bin strongly suggests a resonant production of the entire final state. While a new spin-1 particle (a W ′ boson as in [1], a coloron as in [3], or a Z ′ boson as in our models presented in Section III) with sizable couplings to the quark-antiquark initial states can easily generate the required 2 TeV peak in m eejj , the fit to additional kinematic distributions is difficult to achieve. The W ′ → eN → eejj hypothesis, although well motivated [5], is not consistent with some (unspecified) kinematic distributions, as mentioned by the CMS Collaboration [1]. It is then useful to try all possible decay patterns leading to eejj final states [6]. In our case, given that we try to find a common origin with the signal seen in the leptoquark search, we attempt to find models where the signal responsible for the m eejj peak gives an m min ej distribution as in [2]. To that end we have analyzed a model where Z ′ → ¯ qχ , where χ is a vector-like quark decaying to e + e -q ; the m min ej distribution turns out to be too flat in that case, with a peak at a m min ej value that is too large, almost independently of the χ mass.
A more promising hypothesis, consistent with the features of the 'leptoquark signal', is that the s -channel resonance decays to e + e -and more than two jets. That can be consistent with the ' W ′ signal' because the 1 8 TeV . < m eejj < 2 2 TeV bin where the events . are concentrated is wide enough. In Section IV we will show that ee +4 j final states, arising from the Z ′ models presented next, indeed satisfy this requirement. Given that some of the energy released in the Z ′ decay is taken away by the third and fourth jets, the Z ′ mass has to be slightly larger than 2.2 TeV.
## III. RESONANT PRODUCTION OF VECTOR-LIKE LEPTONS
Let us consider the SM plus a U (1) B gauge group, spontaneously broken by the VEV of a scalar φ B . The new heavy gauge boson, Z ′ , has a mass M Z ′ , which should be in the 2 2 . -2 5 TeV range in order to produce a resonant signal consistent with the CMS . eejj excess. We impose that all SM quarks have same U (1) B charge (chosen to be 1/3), and all SM leptons are U (1) B neutral, so that the constraints from FCNCs and di-lepton resonance searches are avoided. Besides the SM fermions we include a vector-like lepton that carries U (1) B charge z vl and is an SU (2) W singlet. (More generally, the U (1) B charges of its left-and right-handed components could differ, but this possibility would not modify our conclusions; in either case, the anomaly cancellation conditions for U (1) B require additional vector-like fermions [8-10].)
We consider the cases where the vector-like fermion is electrically neutral and is labelled by N , or carries electric charge -1 and is labelled by E . We assume that its mass ( m N or m E ) is less than half the Z ′ mass M Z ′ , and that its Yukawa couplings to the SM leptons and the Higgs doublet are negligible. The normalization for g z used in what follows is that where the Z ′ couplings to SM quarks are given by ( g / z 6) Z qγ ′ µ ¯ µ q .
The branching fraction for Z ′ → NN ¯ (or for Z ′ → E E + -, with m E replacing m N ) is
$$B ( Z ^ { \prime } \rightarrow N \bar { N } ) = \frac { z _ { v l } ^ { 2 } \beta _ { N } } { 2 ( 1 + \alpha _ { s } / \pi ) + z _ { v l } ^ { 2 } \beta _ { N } } \$$
where β N = (1 -4 m /M 2 N 2 Z ′ ) 1 2 / , α s is the strong coupling constant, and we have ignored electroweak corrections. For concreteness, we will use z vl = 1.
For M Z ′ = 2 3 TeV and . m N = 700 GeV we find B Z ( ′ → NN ¯ ) ≈ 26%, and the leadingorder cross sections (computed with MadGraph 5 [7]) for NN production at the LHC with √ s = (8 13 14) TeV are , , g 2 z × (7 3 . , 47 61) fb, respectively. , For M Z ′ = 2 4 TeV and . m E = 800 GeV we find B Z ( ′ → E E + -) ≈ 25%, and the leading-order cross sections for E E + -LHC production with √ s = (8 13 14) TeV are , , g 2 z × (3 6 . , 9 3 . , 12) fb, respectively.
QCD effects at next-to-leading order (NLO) typically increase the Z ′ production by about 30%. Given that we use only leading order production cross section, the NLO effects may be taken into account through a reduction in g z by about 15%.
An upper limit on g z follows from di-jet resonance searches; for M Z ′ = 2 3 TeV, the limit . derived as in [8] is given by
$$g _ { z } \left ( 1 - B ( Z ^ { \prime } \to N \bar { N } ) \right ) ^ { 1 / 2 } < 1. 6 \$$
based on the ATLAS search with 20.3 fb -1 [11]. With B Z ( ′ → NN ¯ ) = 26%, the upper limit on the Z ′ gauge coupling is g z < 1 8. . For M Z ′ = 2 4 TeV the limit is weaker. .
There are several possibilities for the decays of N or E . We concentrate on three models, described in what follows.
## A. Z N ′ Model
In the first one, referred to as the Z N ′ Model, N has 3-body decays into e -ud ¯ or νdd ¯ via the dimension-7 operator
$$\frac { y _ { N q } y _ { d l } } { M _ { L Q } ^ { 3 } } \phi _ { B } ( \bar { \mathcal { Q } } _ { L } ^ { 1 } N _ { R } ) i \sigma _ { 2 } ( \bar { \mathcal { L } } _ { L } ^ { 1 } d _ { R } )$$
where Q 1 L = ( u , d L L ) and L 1 L = ( ν e L , e L ) are the quark and lepton doublets of the first generation, M LQ /greatermuch M N is related to the mass of a particle whose exchange generates the dimension-7 operator, and y Nq , y dl are its couplings to fermions. We assigned U (1) B charge -1 to the scalar φ B responsible for spontaneous U (1) B breaking.
A renormalizable UV completion is provided by two scalar leptoquarks, ˜ q 0 and ˜ , which q 1 carry U (1) B charge 0 and +1 respectively, and have a trilinear term ˜ ˜ q q φ 1 † 0 B . The Yukawa couplings y Nq ˜ q 0 ¯ L 1 L d R and y dl ˜ q † 1 ¯ Q 1 L N R then induce the operator (3.3) at tree level, once the heavy ˜ q 0 and ˜ q 1 are integrated out.
The flavor structure of the operator can be approximately enforced by a discrete symmetry that distinguishes the first generation of quark and leptons. Such a symmetry is broken by the mixing of the quark doublets required to generate the CKM matrix, but this effect is not problematic from the point of view of FCNCs because y Nq and y dl can be very small. The only requirement for producing a large signal with electrons and jets is that the 3-body decays of N into an electron and a quark-antiquark pair of the first or second generation has a large enough branching fraction.
FIG. 1. Resonant production of vectorlike neutral leptons, followed by 3-body decays within the Z N ′ Model (left diagram), or by decays through a scalar doublet ( S + , S 0 ) within the Z NS ′ Model (right diagram). In both models the final states are e + e -+4 , j e +4 + j / E T , or 4 j + / E T .
The process relevant for the ee -plus-jets final state is pp → Z ′ → NN ¯ followed by N → e -jj and ¯ N → e + jj , so that there is a e -e + + 4 j resonance at M Z ′ , as shown in the left diagram of Fig. 1. The eν -plus-jets final state arises from the same resonant NN ¯ production followed by N → νjj and ¯ N → e + jj , or N → e -jj and ¯ N → ¯ νjj . Assuming that the branching fractions for decays of N involving µ τ , , b or t are negligible, the branching fractions for N → ejj and N → νjj are equal to 50%. Thus, the total cross section for the pp → Z ′ → e + e -+4 j and pp → Z ′ → eν +4 j processes, for a narrow Z ′ , are
$$\sigma ( e e + 4 j ) = \frac { 1 } { 2 } \sigma ( e \nu + 4 j ) = \frac { 1 } { 4 } \sigma ( p p \to Z ^ { \prime } X ) B ( Z ^ { \prime } \to N \bar { N } ) \ \.$$
For M Z ′ = 2 3 TeV and . m N = 700 GeV we find σ ee ( +4 ) = j g 2 z × 1 8 fb at the 8 TeV LHC. .
## B. Z NS ′ Model
In the second model, referred to as Z NS ′ , instead of the dimension-7 operator (3.3) there is an SU (2) W -doublet scalar, S = ( S + , S 0 ), which is neutral under U (1) B and has a Yukawa-like dimension-5 interaction
$$y _ { S } \frac { \phi _ { B } } { M _ { S ^ { \prime } } } \bar { \mathcal { L } } ^ { 1 } _ { L } N _ { R } i \sigma _ { 2 } S ^ { * } + \text{H.c.}$$
where y S is a dimensionless parameter, and M S ′ is the mass of a very heavy particle that has been integrated out, for example an S ′ scalar of U (1) B charge +1. The above dimension-5
operator, with φ B replaced by its VEV, induces the N → S ν 0 and N → S + e -decays. For simplicity, we take the masses of S 0 and S + to be equal, i.e. , they get contributions only from SU (2) W invariant terms. Furthermore, we include a large positive squared mass for S and forbid any term in the scalar potential that is linear in S , so that S has no tree-level VEV.
We assume that S 0 and S + decay predominantly into a quark-antiquark pair of the second or first generation, via dimension-5 operators of the type
$$\frac { \varphi } { m _ { \psi } } \bar { \mathcal { Q } } _ { L } ^ { 2 } c _ { R } i \sigma _ { 2 } S ^ { * } + \text{H.c.}$$
where m ψ is the mass, in the multi-TeV range, of a vectorlike quark ψ that has been integrated out, and ϕ is a gauge-singlet scalar that gets a VEV much smaller than m ψ . After replacing ϕ by its VEV, this leads to small effective Yukawa couplings of the type S + ¯ c R L s and S c 0 ¯ R L c . Nevertheless, the decays of the S scalars are prompt as long as the effective Yukawa couplings are larger than O (10 -7 ).
At 1-loop, the effective Yukawa couplings generate an SH † term (where H is the SM Higgs doublet) in the effective Lagrangian, leading to a tiny, inconsequential VEV for S 0 . Couplings to third generation quarks may be suppressed by a Z 2 symmetry under which S and c R are odd; the S 0 → ct ¯ and S + → cb ¯ decays are still allowed, but have CKM suppressed amplitudes. Therefore, S 0 and S ± decays into a pair of nonb jets have branching fractions close to 100%.
The process pp → Z ′ → NN ¯ followed by NN ¯ → S + e -S -e + → e + e -+ 4 j or NN ¯ → S ± ∓ e S 0( ) † ν → eν +4 j then generates the relevant signals (see right diagram of Fig. 1). The branching fraction for each of the N → S ν 0 and N → S + e -decays can each be very close to 50%; competing decays such as N → e ± W ∓ , N → νZ , N → νh are suppressed by a very small mixing angle squared, while N → µS and N → τS can be almost forbidden by assigning first-generation lepton number +1 to N . Thus, Eq. (3.4) is valid for both the Z NS ′ and Z N ′ Models.
## C. Z ES ′ Model
The third model, referred to as Z ES ′ , has the same particle content as the Z NS ′ Model, except that N is replaced by a vectorlike lepton E (of electric charged -1). The scalar doublet has an SU (2) W invariant mass M S , but electroweak symmetry breaking splits the masses of S 0 and S + through the ( S H † )( H S † ) term in the Lagrangian. For simplicity we take the CP-even and -odd components of S 0 to have the same mass.
We impose that E carries the lepton number of the first generation. The Yukawa-like coupling involving S , E and the first generation lepton doublet takes the form
$$y _ { S } \frac { \phi _ { B } } { M _ { S ^ { \prime } } } \bar { \mathcal { L } } _ { L } ^ { 1 } E _ { R } i \sigma _ { 2 } S + \text{H.c.}$$
similarly to the operator (3.5). For M S + > M S 0 , the charged S scalar decays through S + → S W 0 + , or through a virtual W boson into a 3-body final state for M S + -M S 0 < M W . To evade constraints on the ∆ ρ isospin-violating parameter it is actually preferable to have a small mass splitting within the S doublet, so that we consider only the 3-body S + decays. The S + decay is prompt provided M S + -M S 0 > O (1) GeV.
We assume that the neutral scalar decays into two gluons through the loop-induced dimension-6 operator
$$\frac { c _ { S } \alpha _ { s } } { m _ { \chi } ^ { 2 } } S H ^ { \dagger } G ^ { \mu \nu } G _ { \mu \nu } + \text{H.c.}$$
where m χ is the mass of a heavy particle running in the loop, and c S is a dimensionless coefficient that depends on the loop integral and on the product of χ couplings to the S and H doublets. An SH † term is also generated at one loop, leading as in the Z NS ′ Model to a tiny S 0 VEV.
The E E + -production is followed by the cascade decays shown in Fig. 2. The ee + 4 j final state proceeds through an intermediate S 0 -S 0 † pair and involves only gluon jets. The eν +jets final state involves four gluon jets and a (possibly off-shell) W boson. The hadronic decays of the W boson then lead to a eν +6 j final state.
The branching fractions of E + → e + S 0 and E ± → νS ± are slightly different due to the
S + -S 0 mass difference, while other decay modes of E ± can be neglected. We obtain
$$B ( E \to e S ^ { 0 } ) & = \frac { ( m _ { E } ^ { 2 } - M _ { S ^ { 0 } } ^ { 2 } ) ^ { 2 } } { ( m _ { E } ^ { 2 } - M _ { S ^ { 0 } } ^ { 2 } ) ^ { 2 } + ( m _ { E } ^ { 2 } - M _ { S ^ { + } } ^ { 2 } ) ^ { 2 } } \\ & \approx \frac { 1 } { 2 } + \frac { 2 M _ { S } } { m _ { E } ^ { 2 } } \, ( M _ { S ^ { + } } - M _ { S ^ { 0 } } ) + O \left ( ( M _ { S ^ { + } } - M _ { S ^ { 0 } } ) ^ { 2 } / m _ { E } ^ { 2 } \right )$$
and B E ( ± → νS ± ) = 1 -B E ( → eS 0 ).
The branching fractions of S 0 → gg and S + → S W 0 +( ) ∗ are nearly 100%. However, the event selection discards final states where the W boson (whether on- or off-shell) decays involves an electron or muon, so the relevant branching fraction is B S ( + → hadrons) ≈ 75%, where we included both W → jets and W → τν with hadronic τ decays. Thus, in the Z ES ′ model,
$$\sigma ( e e + 4 j ) = \sigma ( p p \to Z ^ { \prime } X ) B ( Z ^ { \prime } \to E ^ { + } E ^ { - } ) B ( E \to e S ^ { 0 } ) ^ { 2 } \$$
$$\frac { \sigma ( e \nu + \text{hadrons} ) } { \sigma ( e e + 4 j ) } = 2 \frac { B ( E ^ { + } \to \nu S ^ { + } ) } { B ( E \to e S ^ { 0 } ) } B ( S ^ { + } \to \text{hadrons} ) \ \.$$
For M Z ′ = 2 4 TeV, . m E = 800 GeV, M S 0 = 400 GeV and M S + = 440 GeV we obtain B E ( → eS 0 ) = 54%, σ ee ( +4 ) = j g 2 z × 1 0 fb and . σ eν ( + hadrons) = g 2 z × 1 3 fb at the 8 . TeV LHC.
-
FIG. 2. Resonant production of vectorlike leptons, followed by decays through a scalar ( S + , S 0 ) within the Z ES ′ Model. The final states are e + e -+4 j (left diagram) or e +4 + j W ( ∗ ) + / E T (right diagram).
TABLE I. Gauge charges and masses for the new particles included in at least one of the three models described in Section III, and involved on-shell in the processes shown in Figs. 1 and 2.
| new particles | U (1) B | SU (2) W | U (1) Y | Z ′ N model | Z ′ NS model | Z ′ ES model |
|--------------------------|-----------|------------|-----------|---------------|----------------|----------------|
| vector-like fermion N | +1 | 1 | 0 | 700 GeV | 800 GeV | - |
| vector-like fermion E | +1 | 1 | - 1 | - | - | 800 GeV |
| scalar S = ( S + , S 0 ) | 0 | 2 | +1/2 | - | 400 GeV | (440, 400) GeV |
| gauge boson Z ′ | 0 | 1 | 0 | 2.3 TeV | 2.4 TeV | 2.4 TeV |
## IV. KINEMATIC DISTRIBUTIONS
To study the LHC signatures of the vector-like lepton models described in the previous section, we turn to Monte Carlo simulations. We implement the Z N, Z NS ′ ′ , and Z ES ′ models into MadGraph 5 [7] using the FeynRules [12] package. These models have several new parameters in the form of couplings and masses. We choose the masses of the new particles, as in Table I, to roughly fit the eejj and eνjj excesses. With these masses set, the overall pp → NN,E E ¯ + -rate can be adjusted by dialing the Z ′ gauge coupling g z . The scales and rates of the observed excesses determine the parameters of the Z N ′ model. In the Z NS ′ , and Z ES ′ , there is additional freedom, in the form of the S mass, and the S + -S 0 mass splitting, to change the morphology of the signal.
We generate parton-level signal events with MadGraph 5, which are then passed to PYTHIA6.4 [13] for showering, hadronization, and decay. The events are subsequently directed through the DELPHES [14] package to incorporate detector geometry and response effects. Post-detector level events are then analyzed using the cuts described in Sec. II.
The eejj invariant mass from the models of resonant vector-like lepton production with a cross section that roughly fits the CMS data is shown in Fig. 3. The rates used there are given by Eq. (3.4) with g z = 1 0 for . Z N ′ and g z = 1 05 for . Z NS ′ , and by Eq. (3.10) with g z = 1 2 for . Z ES ′ . Note that the distribution has been truncated below 1 5 TeV. .
eejj
FIG. 3. The m eejj distribution, formed from the two electrons and two leading jets, for the Z N ′ (red, solid line), Z NS ′ (blue, dotted line) and Z ES ′ (black, dashed line) model. The data points with error bars are taken from [1], after SM background subtraction. The new particle masses are listed in Table I. The values of g z used here are 1.0 for Z N ′ , 1.05 for Z NS ′ , and 1.2 for Z ES ′ .
Comparing this figure with Fig. 2 from Ref. [1], all three models of vector-like leptons describe the data better than a W ′ . The key to the improved fit is the presence of more than two jets in the vector-like lepton signal; the Z ′ in our signal is actually an ee +4 j resonance. As only the leading two jets are included in the eejj distribution, the signal shows an edge, rather than the full Breit-Wigner line shape. As a result, the signal predicts fewer events at high m eejj , in agreement with the data. For the Z N ′ model, the improved m eejj fit is automatic once we have chosen the rate. For the Z ES ′ and Z NS ′ models, the agreement depends on the mass of S . For M S /lessmuch m ,m E N , the S scalars emerge boosted from the N/E decay, bringing their decay products (two quarks or gluons) closer together. In events where the S decay products each merge into a single jet, the entire Z ′ energy will be captured in the electrons and two leading jets; in this case, m eejj has a more characteristic resonance shape, leading to too many events at high m eejj . To match the eejj excess in the 1 8 . -2 2 TeV bin, .
as well as the agreement between data and the SM in the 2 2 . -4 0 TeV bin, scalar masses . around M S ∼ 400 GeV work well. Recall, as discussed in Sec. III, that S 0 lies in an SU (2) W doublet with the charged state slightly heavier than the neutral one, such that S + decays to S 0 via an off-shell W + .
With the signal rate and S mass set to match the m eejj excess of the W ′ analysis, we now examine how the vector-like lepton signals appear under the 'leptoquark' event selection. Applying the leptoquark analysis cuts optimized for a 650 GeV leptoquark, the signal m min ej and m ej distributions are shown in Fig. 4 for the same parameters used in Fig. 3. First, we emphasize that the range of m min ej and m ej shown in Fig. 4 are larger than what CMS uses to determine the significance. As the vector-like lepton models lead to more events in the 200 GeV < m min ej < 800 GeV region than the SM, the fit is improved. The improvement of the fit is less clear in the single-electron plus missing energy channel, where the data is lower than the SM at low m ej and higher than the SM at high m ej , transitioning at m ej /similarequal 500 GeV. In all cases, the vector-like lepton m min ej , m ej distributions are wide and do not reveal a sharp peak at m N . This is again due to the presence of extra jets: the N/E decay to an electron (or neutrino) plus a pair of jets and are hence not reconstructed accurately if only one jet is included.
In all of the channels discussed, the statistics are low, so it is possible that the true nature of the excess - if due to new physics - requires more data to reveal its shape. As more data will take time to collect, it is worth considering what other distributions to examine in the existing 8 TeV data, both in events with the leptoquark/ W ′ selections and elsewhere, to better gauge the veracity of the excess and its properties. For the vector-like lepton signals discussed here, a first place to look is in the jet multiplicity distribution; for all three models shown here, this distribution peaks at four or more jets. A second distribution is m ejj , the invariant mass of either electron when combined with two of the jets in the event, which should reveal a bump at m ,m N E . To more reliably extract an intermediate N/E state, it is useful to require four jets, matching up each electron with two jets such that the two ejj clusters have the smallest mass difference. The cleanliness of this approach, though, is reduced by NLO QCD effects, which (especially through initial state radiation) could lead to jets that are more energetic than some of the four jets originating from the Z ′ decay chain.
ej
FIG. 4. The m min ej (left panel) and m ej (right panel) distributions for eejj and eνjj events, respectively, passing the 'leptoquark' selection without the m min ej or m ej cut. The signal colors and parameters are the same as in Fig. 3. The black data points are taken from [2], and the shaded histogram is the quoted systematic uncertainty band on the SM background.
Depending on which vector-like model is behind the excess - should it persist - there are other signals to be hunted for, both in the existing data and at the future 13 TeV run of the LHC. In all of the models we've discussed, the vector-like leptons are resonantly produced so that a smoking-gun signal for this type of interpretation is a resonant bump in the ee +4 j distribution. Additionally, all models contain a companion signal νν ¯ + jets which should be produced with a rate comparable to that for ee +4 , of the order of 1 fb, and should appear j in jets plus / E T searches. The Z NS ′ and Z ES ′ models contain an S 0 particle that decays to two jets; given that the S 0 gg coupling is very small, the rates for single and doubleS production via gluon fusion are too small to be useful for di-jet or paired-dijet searches. A more tractable option is to search for pairs of di-jet resonances within the eν + jets or e + e -+jets events.
## V. CONCLUSIONS
We have presented alternative interpretations of the excess di-electron-plus-jets and electron-plus-jets events recently reported by the CMS Collaboration, in terms of resonantly produced vector-like leptons. While it is possible that the excesses are the result of statistical fluctuations that will shrink with the addition of more data, alternative interpretations are useful; they show what type of new physics would be necessary to describe the data, and provide a list of cross-checks. Furthermore, our simple renormalizable models and their possible extensions motivate additional searches.
The models we explored vary in detail, but all have a set of basic features in common; resonant pair production, through a Z ′ boson, of a color-singlet fermion (vector-like lepton) that decay into three or more SM particles. The decays of the new fermion into a lepton and two or more jets (including N → ejj , E + → νS + → ν +4 , etc.) j are the key to why these models fit the observed excess better than the CMS signal hypotheses. Regardless of whether the CMS excesses persist, multi-particle resonance searches, including mixtures of leptons and jets like the ejj state studied here, should be added to the toolbox of future LHC analyses.
Acknowledgments . We would like to thank John Paul Chou, Patrick Fox, Steve Mrenna and Felix Yu for helpful comments and conversations. We are grateful to Pavel Fileviez P´rez e for pointing out that the interactions displayed in an earlier version violated U (1) B . The work of AM was partially supported by the National Science Foundation under Grant No. PHY14-17118.
[1] V. Khachatryan et al. [CMS Collaboration], 'Search for heavy neutrinos and W bosons with right-handed couplings in proton-proton collisions at √ s = 8 TeV,' arXiv:1407.3683 [hep-ex].
[2] CMS Collaboration, 'Search for Pair-production of First Generation Scalar Leptoquarks in pp Collisions at √ s = 8 TeV,' note CMS-PAS-EXO-12-041, July 2014.
- [3] Y. Bai and J. Berger, 'Coloron-assisted Leptoquarks at the LHC,' arXiv:1407.4466 [hep-ph].
- [4] G. Aad et al. [ATLAS Collaboration], 'Search for heavy neutrinos and right-handed W bosons in events with two leptons and jets in pp collisions at √ s = 7 TeV', Eur. Phys. J. C 72 , 2056 (2012) [arXiv:1203.5420 [hep-ex]].
- [5] M. Heikinheimo, M. Raidal and C. Spethmann, 'Testing Right-Handed Currents at the LHC,' arXiv:1407.6908 [hep-ph].
- F. F. Deppisch, T. E. Gonzalo, S. Patra, N. Sahu and U. Sarkar, 'A Signal of Right-Handed Charged Gauge Bosons at the LHC?,' arXiv:1407.5384 [hep-ph].
- [6] M. Abdullah, et al , 'Systematically using Topological Models,' Phys. Rev. D 89 , 095002 (2014) [arXiv:1401.1462 [hep-ph]].
- B. Knuteson and S. Mrenna, 'BARD: Interpreting new frontier energy collider hep-ph/0602101.
- Searching for New Resonances at the Energy Frontier physics,'
- [7] J. Alwall, et al , 'The automated computation of tree-level and next-to-leading order differential cross sections, and their matching to parton shower simulations,' JHEP 1407 , 079 (2014) [arXiv:1405.0301 [hep-ph]].
- [8] B. A. Dobrescu and F. Yu, 'Coupling-mass mapping of dijet peak searches,' Phys. Rev. D 88 , no. 3, 035021 (2013) [arXiv:1306.2629 [hep-ph]].
- [9] B. A. Dobrescu and C. Frugiuele, 'Hidden GeV-scale interactions of quarks,' Phys. Rev. Lett. 113 , 061801 (2014) [arXiv:1404.3947 [hep-ph]].
- [10] M. S. Carena, A. Daleo, B. A. Dobrescu and T. M. P. Tait, ' Z ′ gauge bosons at the Tevatron,' Phys. Rev. D 70 , 093009 (2004) [hep-ph/0408098]. M. Duerr, P. Fileviez Perez and M. B. Wise, 'Gauge Theory for Baryon and Lepton Numbers with Leptoquarks,' Phys. Rev. Lett. 110 , 231801 (2013) [arXiv:1304.0576 [hep-ph]]. Ohmer and H. H. Patel, 'Minimal Theory for Lepto-Baryons,' Phys.
- P. Fileviez Perez, S. Lett. B 735 , 283 (2014) [arXiv:1403.8029 [hep-ph]].
- [11] G. Aad et al. [ATLAS Collaboration], 'Search for new phenomena in the dijet mass distribution using pp collision data at √ s = 8 TeV,' arXiv:1407.1376 [hep-ex].
- [12] A. Alloul, N. D. Christensen, C. Degrande, C. Duhr and B. Fuks, 'FeynRules 2.0 - A complete toolbox for tree-level phenomenology,' Comput. Phys. Commun. 185 , 2250 (2014)
[arXiv:1310.1921 [hep-ph]].
- [13] T. Sjostrand, S. Mrenna and P. Z. Skands, 'PYTHIA 6.4 Physics and Manual,' JHEP 0605 , 026 (2006) [hep-ph/0603175].
- [14] S. Ovyn, X. Rouby and V. Lemaitre, 'DELPHES, a framework for fast simulation of a generic collider experiment,' arXiv:0903.2225 [hep-ph].
|
10.1103/PhysRevD.91.035019
|
[
"Bogdan A. Dobrescu",
"Adam Martin"
] |
2014-08-05T19:55:29+00:00
|
2014-09-07T20:48:43+00:00
|
[
"hep-ph",
"hep-ex"
] |
Interpretations of anomalous LHC events with electrons and jets
|
The CMS Collaboration has recently reported some excess events in final
states with electrons and jets, in searches for leptoquarks and $W'$ bosons.
Although these excesses may be due to some yet-to-be-understood background
mismodeling, it is useful to seek realistic interpretations involving new
particles that could generate such events. We show that resonant pair
production of vector-like leptons that decay to an electron and two jets leads
to kinematic distributions consistent with the CMS data.
|
1408.1084v1
|
## Nonlinear structure formation in Nonlocal Gravity
Alexandre Barreira, 1, 2, ∗ Baojiu Li, 1 Wojciech A. Hellwing, 1,3 Carlton M. Baugh, 1 and Silvia Pascoli 2
1 Institute for Computational Cosmology, Department of Physics, Durham University, Durham DH1 3LE, U.K. 2 Institute for Particle Physics Phenomenology, Department of Physics, Durham University, Durham DH1 3LE, U.K. 3 Interdisciplinary Centre for Mathematical and Computational Modeling (ICM),
University of Warsaw, ul. Pawi´ nskiego 5a, Warsaw, Poland
We study the nonlinear growth of structure in nonlocal gravity models with the aid of N-body simulation and the spherical collapse and halo models. We focus on a model in which the inverse-squared of the d'Alembertian operator acts on the Ricci scalar in the action. For fixed cosmological parameters, this model differs from ΛCDM by having a lower late-time expansion rate and an enhanced and time-dependent gravitational strength ( ∼ 6% larger today). Compared to ΛCDM today, in the nonlocal model, massive haloes are slightly more abundant (by ∼ 10% at M ∼ 10 14 M /h glyph[circledot] ) and concentrated ( ≈ 8% enhancement over a range of mass scales), but their linear bias remains almost unchanged. We find that the Sheth-Tormen formalism describes the mass function and halo bias very well, with little need for recalibration of free parameters. The fitting of the halo concentrations is however essential to ensure the good performance of the halo model on small scales. For k glyph[greaterorsimilar] 1 h/ Mpc , the amplitude of the nonlinear matter and velocity divergence power spectra exhibits a modest enhancement of ∼ 12% to 15% , compared to ΛCDM today. This suggests that this model might only be distinguishable from ΛCDM by future observational missions. We point out that the absence of a screening mechanism may lead to tensions with Solar System tests due to local time variations of the gravitational strength, although this is subject to assumptions about the local time evolution of background averaged quantities.
## I. INTRODUCTION
It has been almost a century since Einstein proposed his theory of General Relativity (GR) which is still considered one of the main pillars of modern physics. The outstanding success of GR comes mostly from its ability to pass a number of stringent tests of gravity performed in the Solar System [1]. When applied on cosmological scales, however, GR seems to lose some of its appeal as it requires the presence of some unknown form of dark energy in order to explain the observed accelerated expansion of the Universe. The simplest candidate for dark energy is a cosmological constant, Λ , but the value of Λ required to explain the observations lacks theoretical support. This has provided motivation for the proposal of alternative gravity models which attempt to reproduce cosmic acceleration without postulating the existence of dark energy. Furthermore, the fact that the laws of gravity have never been tested directly on scales larger than the Solar System justifies the exploration of such modifications to GR on cosmological scales. By understanding better the various types of observational signatures that different modified gravity models can leave on cosmological observables, one can improve the chance of identifying any departures from GR, or alternatively, extend the model's observational success into a whole new regime. Currently, the study of modified gravity models is one of the most active areas of research in both theoretical and observational cosmology [2-5].
terms typically involve the inverse of the d'Alembertian operator, glyph[square] -1 , acting on curvature tensors. To ensure causality, such terms must be defined with the aid of retarded Green functions (or propagators). However, it is well known that such retarded operators cannot be derived from standard action variational principles (see e.g. Sec. 2 of Ref. [6] for a discussion). One way around this is to specify the model in terms of its equations of motion and not in terms of its action. One may still consider a nonlocal action to derive a set of causal equations of motion, so long as in the end one replaces, by hand, all of the resulting operators by their retarded versions. Both of these approaches, however, imply that nonlocal models of gravity must be taken as purely phenomenological and should not be interpreted as fundamental theories. In general, one assumes that there is an unknown fundamental (local) quantum field theory of gravity, and the nonlocal model represents only an effective way of capturing the physics of the fundamental theory in some appropriate limit.
Here, we focus on a class of model that has attracted much attention recently, which is known as nonlocal gravity [6]. In these models, the modifications to gravity arise via the addition of nonlocal terms (i.e. which depend on more than one point in spacetime) to the Einstein field equations. These
∗ Electronic address: [email protected]
It was in the above spirit that Ref. [7] proposed a popular nonlocal model of gravity capable of explaining cosmic acceleration. In this model, which has been extensively studied (see e.g. Refs. [6, 8-17] and references therein), one adds the term Rf ( glyph[square] -1 R ) to the Einstein-Hilbert action, where R is the Ricci scalar and f is a free function. As described in Ref. [6], the function f can be constructed in such a way that it takes on different values on the cosmological background and inside gravitationally bound systems. In particular, at the background level, f can be tuned to reproduce ΛCDM -like expansion histories, but inside regions like the Solar System, one can assume that f vanishes, thus recovering GR completely. This model, however, seems to run into tension with data sensitive to the growth rate of structure on large scales [16, 17].
More recently, nonlocal terms have also been used to construct theories of massive gravity. An example of this is obtained by adding directly to the Einstein field equations a term
like m 2 ( g µν glyph[square] -1 R ) T [18-21], where m is a mass scale and T means the extraction of the transverse part (see also Refs. [2225] for models in which glyph[square] -1 acts on the Einstein and Ricci tensors). This model has no ΛCDM limit for the background evolution, but it can still match the current background expansion and growth rate of structure data with a similar goodnessof-fit [21]. Furthermore, Ref. [20] has investigated spherically symmetric static solutions in this model, concluding that it does not suffer from instabilities that usually plague theories of massive gravity. A similar model was proposed by Ref. [26], which is characterized by a term ∝ m R 2 glyph[square] -2 R in the action (see Eq. (1)). Reference [27] showed that this model can reproduce current type Ia Supernovae (SNIa) data, although it also has no ΛCDM limit for the background expansion. The time evolution of linear matter density fluctuations in this model also differs from that in ΛCDM , but the work of Ref. [27] suggests that the differences between these two models are small enough to be only potentially distinguishable by future observational missions.
Here, we extend the previous work done for the model of Refs. [26, 27] by examining its predictions in the nonlinear regime of structure formation. We achieve this by running a set of N-body simulations, which we use to analyse the model predictions for the nonlinear matter and velocity divergence power spectra, and also halo properties such as their abundance, bias and concentration. To the best of our knowledge this is the first time N-body simulations have been used to study the nonlinear regime of structure formation in nonlocal gravity cosmologies. N-body simulations are however computationally expensive to run. To overcome this, it is common to try to devise semi-analytical formulae that, motivated by simple physical assumptions, aim to reproduce the results from the simulations with the calibration of free parameters kept to a minimum. A popular example is given by the ShethTormen halo mass function [28-30] and its use in the halo model approach for the nonlinear matter power spectrum [31]. The performance of the halo model is well understood within ΛCDM , but less so in alternative gravity scenarios. In particular, the free parameters in these formulae might need substantial recalibration in models that differ significantly from ΛCDM (see e.g. Ref. [32]). One of our goals is to assess the performance of these analytical formulae in nonlocal gravity models.
This paper is organized as follows. In Sec. II we present the model and layout the equations relevant for the background evolution. We also derive the equations of motion for spherically symmetric configurations under the quasi-static and weak-field approximations. In Sec. III we present the formulae relevant for the calculation of the nonlinear matter power spectrum in the halo model formalism. In particular, we describe the Sheth-Tormen expressions for the halo mass function and bias, and define the Navarro-Frenk-White (NFW) halo concentration parameter. We also present the equations relevant for the linear evolution and spherical collapse of matter overdensities. Our results are presented in Sec. IV, where we discuss the results from the N-body simulations and compare them with the predictions from the analytical formulae. Wealso comment on the role that Solar System tests of gravity could play in setting the observational viability of this model. We summarize our findings in Sec. V.
In this paper, we work with the metric signature (+ , - - -, , ) and use units in which the speed of light c = 1 . Latin indices run over 1 2 3 , , and Greek indices run over 0 1 2 3 , , , . We use κ = 8 πG = 1 /M 2 Pl interchangeably, where M Pl is the reduced Planck mass and G is Newton's constant.
## II. THE R glyph[square] -2 R NONLOCAL GRAVITY MODEL
## A. Action and field equations
We consider the nonlocal gravity model of Refs. [26, 27], whose action is given by
$$A = \frac { 1 } { 2 \kappa } \int \mathrm d x ^ { 4 } \sqrt { - g } \left [ R - \frac { m ^ { 2 } } { 6 } R \Box ^ { - 2 } R - \mathcal { L } _ { m } \right ], \quad ( 1 )$$
where g is the determinant of the metric g µν , L m is the Lagrangian density of the matter fluid, R is the Ricci scalar and glyph[square] = ∇ ∇ µ µ is the d'Alembertian operator. To facilitate the derivation of the field equations, and to solve them afterwards, it is convenient to introduce two auxiliary scalar fields defined as
$$U = - \Box ^ { - 1 } R,$$
$$S = - \Box ^ { - 1 } U = \Box ^ { - 2 } R.$$
The solutions to Eqs. (2) and (3) can be obtained by evaluating the integrals
$$U & \equiv - \Box ^ { - 1 } R & ( 4 ) \\ & = U _ { \hom } ( x ) - \int \mathrm d ^ { 4 } y \sqrt { - g ( y ) } G ( x, y ) R ( y ), \\ S & \equiv - \Box ^ { - 1 } U & ( 5 ) \\ & = S _ { \hom } ( x ) - \int \mathrm d ^ { 4 } y \sqrt { - g ( y ) } G ( x, y ) U ( y ),$$
where U hom and S hom are any solutions of the homogeneous equations glyph[square] U = 0 and glyph[square] S = 0 , respectively, and G x, y ( ) is any Green function of glyph[square] . The choice of the homogeneous solutions and of the Green function specify the meaning of the operator glyph[square] -1 . To ensure causality, one should use the retarded version of the Green function, i.e., the solutions of U (or S ) should only be affected by the values of R (or U ) that lie in its past light-cone. The homogeneous solutions can be set to any value, which is typically zero, without any loss of generality. In principle, the model predictions can be obtained by solving Eqs. (4) and (5). However, it is convenient to use the fields U and S to cast the nonlocal action of Eq. (1) in the form of a local scalar-tensor theory [10, 33, 34] as
$$\mathring { \nu } _ { \cdot } \quad A = \frac { 1 } { 2 \kappa } \int \mathrm d x ^ { 4 } \sqrt { - g } \left [ R - \frac { m ^ { 2 } } { 6 } R S - \xi _ { 1 } \left ( \Box U + R \right ) \\ \nu _ { \cdot } \quad - \xi _ { 2 } \left ( \Box S + U \right ) - \mathcal { L } _ { m } \right ], \quad \left ( 6 \right )$$
where ξ 1 and ξ 2 are Lagrange multipliers. The field equations can then be written as
$$G _ { \mu \nu } - \frac { m ^ { 2 } } { 6 } K _ { \mu \nu } = \kappa T _ { \mu \nu }, \quad \quad \quad ( 7 ) \\ \neg r \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \ \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \cdot \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \quad \tau \ < \tilde { \alpha } }$$
$$\Box U = - R,$$
$$\Box S = - U,$$
with
$$K _ { \mu \nu } \equiv & \ 2 S G _ { \mu \nu } - 2 \nabla _ { \mu } \nabla _ { \nu } S - 2 \nabla _ { ( \mu } S \nabla _ { \nu ) } U \\ & + \left ( 2 \Box S + \nabla _ { \alpha } S \nabla ^ { \alpha } U - \frac { U ^ { 2 } } { 2 } \right ) g _ { \mu \nu }, \quad ( 1 0 )$$
and where T µν = (2 / √ -g ) δ ( L m √ -g ) /δg µν is the energymomentum tensor of the matter fluid. The use of the scalar fields U and S therefore allows one to obtain the solutions by solving a set of coupled differential equations, instead of the more intricate integral equations associated with the inversion of a differential operator. These two formulations are, however, not equivalent as explained with detail in many recent papers (see e.g. Refs.[9, 18, 19, 23, 34-36]): Eqs. (7), (8) and (9) admit solutions that are not solutions of the original nonlocal problem. For instance, if U ∗ is a solution of Eq. (8), then U ∗ + U hom is also a solution for any U hom , since glyph[square] U hom = 0 (the same applies for the field S and Eq. (9)). If one wishes the differential equations (7), (8) and (9) to describe the nonlocal model, then one must solve them with the one and only choice of initial conditions that is compatible with the choice of homogeneous solutions in Eqs. (4) and (5). All other initial conditions lead to spurious solutions and should not be considered.
## B. Background equations
Throughout, we always work with the perturbed Friedmann-Roberston-Walker (FRW) line element in the Newtonian gauge,
$$\mathrm d s ^ { 2 } = \left ( 1 + 2 \Psi \right ) \mathrm d t ^ { 2 } - a ( t ) ^ { 2 } \left ( 1 - 2 \Phi \right ) \gamma _ { i j } \mathrm d x ^ { i } \mathrm d x ^ { j }, \, \left ( 1 1 \right ) \quad m \text{ ref}$$
where a = 1 (1 + / z ) is the cosmic scale factor ( z is the redshift) and the gravitational potentials Φ Ψ , are assumed to be functions of time and space. γ ij is the spatial sector of the metric, which is taken here to be flat.
At the level of the cosmological background ( Φ = Ψ = 0 ), the two Friedmann equations can be written as
$$3 H ^ { 2 } = \kappa \bar { \rho } _ { m } + \kappa \bar { \rho } _ { d e } \quad \quad ( 1 2 ) \quad \text{sy}$$
$$- 2 \dot { H } - 3 H ^ { 2 } = \kappa \bar { p } _ { m } + \kappa \bar { p } _ { d e }, \text{ \quad \ \ } ( 1 3 ) \text{ \ \ } \bar { \text{Eq.} }$$
where we have encapsulated the effects of the nonlocal term into an effective background "dark energy" density, ¯ ρ de , and pressure ¯ p de , which are given, respectively, by
$$\kappa \bar { \rho } _ { d e } = \frac { m ^ { 2 } } { 6 } \left [ 6 \bar { S } H ^ { 2 } + 6 H \dot { \bar { S } } - \dot { \bar { U } } \dot { \bar { S } } - \frac { \bar { U } ^ { 2 } } { 2 } \right ], \quad ( 1 4 )$$
$$\kappa \bar { p } _ { d e } = - \frac { \mu ^ { 2 } } { 6 } \left [ 2 \bar { S } \left ( 2 \dot { H } + 3 H ^ { 2 } \right ) + \ddot { \bar { S } } \\ + 4 H \dot { \bar { S } } + \dot { \bar { U } } \dot { \bar { S } } - \frac { \bar { U } ^ { 2 } } { 2 } \right ].$$
$$( 1 5 )$$
Additionally, Eqs. (8) and (9) yield
$$\ddot { \bar { U } } + 3 H \dot { \bar { U } } = 6 \left ( \dot { H } + 2 H ^ { 2 } \right ), \text{ \quad \ \ } ( 1 6 )$$
$$\ddot { \bar { S } } + 3 H \dot { \bar { S } } = - \bar { U }.$$
(18)
In the above equations, a dot denotes a partial derivative w.r.t. physical time, t , an overbar indicates that we are considering only the background average and H = ˙ a/a is the Hubble expansion rate.
The background evolution in the R glyph[square] -2 R model has to be obtained numerically. The differential equations are evolved starting from deep into the radiation dominated era ( z = 10 6 ) with initial conditions for the auxiliary fields ¯ U = ˙ ¯ U = ¯ = S ˙ ¯ = 0 S . Note that, in the radiation era, the Ricci scalar vanishes ( ¯ = 6 ˙ R H +12 H 2 = 0 ). Hence, from Eqs. (4) and (5) one sees that these initial conditions are indeed compatible with the choice U hom = S hom = 0 . The value of the parameter m is determined by a trial-and-error scheme to yield the value of ¯ ρ de 0 that makes the Universe spatially flat, i.e., ¯ ρ r 0 + ¯ ρ m 0 + ¯ ρ de 0 = ¯ ρ c 0 ≡ 3 H /κ 2 0 , where the subscripts r , m refer to radiation and matter, respectively, the subscript 0 denotes present-day values, and H 0 = 100 h km s Mpc / / is the present-day Hubble rate.
## C. Spherically symmetric nonlinear equations
By assuming that the potentials Φ and Ψ are spherically symmetric, one can write the (0 , 0) and ( r, r ) components of Eq. (7), and Eqs. (8) and (9), respectively, as
$$\frac { 2 } { r ^ { 2 } } \left ( r ^ { 2 } \Phi _ { r } \right ), _ { r } - \frac { m ^ { 2 } } { 6 } \left [ 6 S H ^ { 2 } + \frac { 4 S } { r ^ { 2 } } \left ( r ^ { 2 } \Phi _ { r } \right ), _ { r } - \frac { 2 } { r ^ { 2 } } \left ( r ^ { 2 } S _ { r } \right ), _ { r } + 2 S _ { r } \Phi _ { r } - S _ { r } U _ { r } - \frac { U ^ { 2 } } { 2 } \right ] = \kappa \bar { \rho } _ { m } \delta a ^ { 2 },$$
$$& \frac { 2 } { r } \left ( \Phi _ { r } - \Psi _ { r } \right ) - \frac { m ^ { 2 } } { 6 } \left [ 4 S \dot { H } a ^ { 2 } + 6 S H ^ { 2 } a ^ { 2 } + \frac { 4 S } { r } \left ( \Phi _ { r } - \Psi _ { r } \right ) + 4 S _ { r } \Phi _ { r } - 2 S _ { r } \, \Psi _ { r } - \frac { S _ { r } } { r } + 2 S _ { r } \, U _ { r } - \frac { U ^ { 2 } } { 2 } \right ] = 0, \ \ ( 2 0 ) \\ & \frac { 1 } { r ^ { 2 } } \left ( r ^ { 2 } U _ { r } \right ), r ^ { \, + } U _ { r } \left ( \Psi _ { r } - \Phi _ { r } \right ) = 2 \frac { 1 } { r ^ { 2 } } \left ( r ^ { 2 } \Psi _ { r } \right ), r ^ { \, - } 4 \frac { 1 } { r ^ { 2 } } \left ( r ^ { 2 } \Phi _ { r } \right ), r \,, \\ & \frac { 1 } { r ^ { 2 } } \left ( r ^ { 2 } S _ { r } \right ), r ^ { \, + } S _ { r } \left ( \Psi _ { r } - \Phi _ { r } \right ) = U,$$
$$r$$
where , r denotes a partial derivative w.r.t. the comoving radial coordinate r . When writing Eqs. (19)-(22), we have already employed the following simplifying assumptions:
- 1. We have assumed the so-called quasi-static limit, under which one neglects the time derivatives of perturbed quantities, e.g., ˙ S = ˙ ¯ + S ˙ δS ≈ ˙ ¯ S , where δS is the perturbed part of the auxiliary field;
- 2. We have also employed the so-called weak-field limit, which accounts for neglecting terms that involve Φ and Ψ , and their first spatial derivatives, over those that involve their second spatial derivatives. For example, (1 -2Φ)Φ , rr ≈ Φ , rr and Φ , r Φ , r glyph[lessmuch] Φ , rr .
The above equations still contain terms with Ψ , r and Φ , r , because these terms contain the fields U and S , and up to now, we have not discussed the validity of applying these approximations to the auxiliary fields. However:
- 1. Equation (21) tells us that the U field is of the same order as the scalar potentials, U ∼ Φ Ψ , . Consequently, the above approximations also hold for U ;
- 2. Equation (22) tells us that S, rr ∼ Φ Ψ , , which means we can also neglect all terms containing S S, , r and S, rr .
Under these considerations, the above equations simplify drastically. In particular, the only equation that remains relevant for the study of the spherical collapse of matter overdensities is Eq. (19), which can be written as:
$$\frac { 1 } { r ^ { 2 } } \left ( r ^ { 2 } \Phi, _ { r } \right ), _ { r } = 4 \pi G _ { \text{eff} } \bar { \rho } _ { m } \delta a ^ { 2 }, \quad \ \ ( 2 3 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
where
$$G _ { \text{eff} } = G \left [ 1 - \frac { m ^ { 2 } \bar { S } } { 3 } \right ] ^ { - 1 }. \quad \quad ( 2 4 ) \quad \text{plic}$$
Equation (23) is the same as in standard gravity, but with Newton's constant replaced by the time-dependent gravitational strength, G eff . This time dependence follows directly from the term 2 SG µν in the field equations, Eq. (7), which in turn follows from the variation of the term ∝ SR in the action Eq. (6). The fact that G eff depends only on time tells us that gravity is modified with equal strength everywhere, regardless of whether or not one is close to massive bodies or in high-density regions. This may bring into question the ability of this model to pass the stringent Solar System tests of gravity [1, 37, 38]. We come back to this discussion in Sec. IV C. We note also that from Eq. (20), it follows that Φ = Ψ in the quasi-static and weak-field limits.
FIG. 1. CMB temperature power spectrum of the ΛCDM (black) and R glyph[square] -2 R (blue) models for the cosmological parameters of Table I. The data points with errorbars show the power spectrum measured by the Planck satellite [39, 40].
## D. Model parameters
The results presented in this paper are for the cosmological parameter values listed in Table I. These are the best-fitting ΛCDM parameters to a dataset that comprises the CMB data from the Planck satellite (both temperature and lensing) [3941], and the BAO data from the 6df [42], SDDS DR7 [43] and BOSSDR9[44] galaxy redshift surveys. The parameters were found by following the steps outlined in Ref. [45], although in the latter, neutrino masses are also varied in the constraints. In this paper, however, we treat neutrinos as massless for simplicity.
The CMB temperature power spectra of the ΛCDM and R glyph[square] -2 R models for the parameters listed in Table I are shown in Fig. 1. The R glyph[square] -2 R model predictions were obtained with a suitably modified version of the CAMB code [46]. The derivation of the perturbed equations that enter the calculations in CAMB follows the steps of Ref. [47], to which we refer the interested reader for details. The results in Fig. 1 shows that the R glyph[square] -2 R model is able to fit the CMB data with a goodness-of-fit that is similar to that of ΛCDM . In fact, the R glyph[square] -2 R model is in slightly better agreement with the data at low- , which is mostly determined by the Integrated l Sachs-Wolfe (ISW) effect. However, the larger errorbars on these scales due to cosmic variance do not allow stringent constraints to be derived.
TABLEI. Cosmological parameter values adopted in this paper. Ω r 0 , Ω b 0 , Ω c 0 , Ω de 0 , h n , s , and τ are, respectively, the present day fractional energy density of radiation ( r ), baryons ( b ), cold dark matter ( c ) and dark energy ( de ), the dimensionless present day Hubble expansion rate, the primordial scalar spectral index and the optical depth to reionization. The scalar amplitude at recombination A s refers to a pivot scale k = 0 05Mpc . -1 . These are the ΛCDM parameters that best-fit the CMB temperature and lensing data from the Planck satellite [39-41], and the BAO data from the 6df [42], SDDS DR7 [43] and BOSS DR9 [44] galaxy redshift surveys. The parameters were determined by following the strategy outlined in Ref. [45], although in the latter neutrino masses are also varied in the constraints. For the purpose of this paper we can assume neutrinos to be effectively massless. The R glyph[square] -2 R model parameter m is derived by the condition to make the Universe spatially flat, i.e. , 1 = Ω r 0 +Ω b 0 +Ω c 0 +Ω de 0 ( m ) .
| Parameter | Planck (temperature+lensing) + BAO |
|------------------------------------------------------|--------------------------------------|
| Ω r 0 h 2 Ω b 0 h 2 Ω c 0 h 2 h n s τ log 10 [ 10 10 | 4 . 28 × 10 - 5 |
| | 0 . 02219 |
| | 0 . 1177 |
| | 0 . 6875 |
| | 0 . 968 |
| | 0 . 0965 |
| A s ] | 3 . 097 |
| Ω de 0 | 0 . 704 |
| m | 0 . 288 |
In Sec. IV we shall compare the results of the R glyph[square] -2 R model with those of standard ΛCDM . In this paper, we are mostly interested in the phenomenology driven by the modifications to gravity in the R glyph[square] -2 R model. This is why we shall use the same cosmological parameters for both models. Aformal exploration of the constraints on the parameter space in the R glyph[square] -2 R model is beyond the scope of the present paper (see Ref. [48]).
## III. HALO MODEL OF THE NONLINEAR MATTER POWER SPECTRUM
In this section, we describe the halo model of the nonlinear matter power spectrum, as well as all of its ingredients. In particular, we define the halo mass function, linear halo bias and halo density profiles. We also present the equations that govern the linear growth and spherical collapse of structures.
## A. Halo model
In the halo model approach, one assumes that all matter in the Universe lies within gravitationally bound structures (see Ref. [31] for a review). As a result, the two-point correlation function of the matter field can be decomposed into the contributions from the correlations between elements that lie in the same halo (the 1-halo term) and in different haloes (the 2-halo term). The power spectrum can also be decomposed in a similar way, and one can write
$$P _ { k } = P _ { k } ^ { 1 \text{h} } + P _ { k } ^ { 2 \text{h} }, \text{ \quad \ \ } ( 2 5 )$$
where
$$P _ { k } ^ { 1 \text{h} } & = \int \text{d} M \frac { M } { \bar { \rho } _ { m 0 } ^ { 2 } } \frac { \text{d} n ( M ) } { \text{d} \ln M } | u ( k, M ) | ^ { 2 }, \\ P _ { k } ^ { 2 \text{h} } & = I ( k ) ^ { 2 } P _ { k, \text{lin} },$$
are, respectively, the 1- and 2-halo terms, with
$$I ( k ) = \int \mathrm d M \frac { 1 } { \bar { \rho } _ { m 0 } } \frac { \mathrm d n ( M ) } { \mathrm d \ln M } b _ { \mathrm i \ln } ( M ) | u ( k, M ) |. \quad ( 2 7 )$$
In Eqs.(25)-(27), ¯ ρ m 0 is the present-day background (total) matter density; P k, lin is the matter power spectrum obtained using linear theory; k is the comoving wavenumber; d n M / ( ) dln M is the mass function , which describes the comoving number density of haloes per differential logarithmic interval of mass; u k, M ( ) is the Fourier transform of the density profile of the haloes truncated at their size and normalized such that u k ( → 0 , M ) → 1 ; b lin ( M ) is the linear halo bias parameter. We model all these quantities in the remainder of this section, in which we follow the notation of Refs. [32, 49].
## B. Halo mass function
We define the halo mass function as
$$\frac { d n ( M ) } { d \ln M } \text{d} \ln M = \frac { \bar { \rho } _ { m 0 } } { M } f ( S ) \text{d} S, \text{ \quad \ \ } ( 2 8 )$$
where S is the variance of the linear density field filtered on a comoving length scale R ,
$$S ( R ) \equiv \sigma ^ { 2 } ( R ) = \frac { 1 } { 2 \pi ^ { 2 } } \int k ^ { 2 } P _ { k, \text{lin} } \tilde { W } ^ { 2 } \left ( k, R \right ) \text{d} k.$$
Here, ˜ W k,R ( ) = 3(sin( kR ) -kR cos( kR )) / ( kR ) 3 is the Fourier transform of the filter, which we take as a top-hat in real space. The total mass enclosed by the filter is given by
$$M = 4 \pi \bar { \rho } _ { m 0 } R ^ { 3 } / 3.$$
In Eq. (28), f ( S )d S describes the fraction of the total mass that resides in haloes whose variances lie within [ S, S +d ] S (or equivalently, whose masses lie within [ M -d M,M ] ) 1 . Here, we use the Sheth-Tormen expression [28-30],
1 Note that the quantities S , R and M can be related to one another via Eqs. (29) and (30). In this paper, we use these three quantities interchangeably when referring to the scale of the haloes.
$$f ( S ) = A \sqrt { \frac { q } { 2 \pi } } \frac { \delta _ { c } } { S ^ { 3 / 2 } } \left [ 1 + \left ( \frac { q \delta _ { c } ^ { 2 } } { S } \right ) ^ { - p } \right ] \exp \left [ - q \frac { \delta _ { c } ^ { 2 } } { 2 S } \right ], \quad \text{We}$$
where A is a normalization constant fixed by the condition ∫ f ( S )d S = 1 . δ c ≡ δ c ( z ) is the critical initial overdensity for a spherical top-hat to collapse at redshift z , extrapolated to z = 0 with the Λ CDM linear growth factor. This extrapolation is done purely to ensure that the resulting values of δ c can be more readily compared to values in ΛCDM . Note that for consistency, P k, lin in Eq. (29) is also the initial power spectrum of the specific model, evolved to z = 0 with the Λ CDM linear growth factor.
The Press-Schechter mass function [50] is obtained by taking ( q, p ) = (1 0) , in Eq. (31). This choice is motivated by the spherical collapse model. However, Refs. [28-30] showed that the choice of parameters ( q, p ) = (0 75 0 3) . , . (motivated by the ellipsoidal, instead of spherical collapse) leads to a better fit to the mass function measured from N-body simulations of Λ CDMmodels (see also Ref. [51]). For alternative models, such as those with modified gravity, it is not necessarily true that the standard ST parameters ( ( q, p ) = (0 75 0 3) . , . ) also provide a good description of the simulation results. For example, in Ref. [32], we demonstrated that the ST mass function can only provide a good fit to the simulation results of Galileon gravity models [47, 52-54] after a recalibration of the ( q, p ) parameters. In Sec. IV D, we shall investigate the need for a similar recalibration in the R glyph[square] -2 R model.
## C. Linear halo bias
The linear halo bias parameter b M ( ) [55] relates the clustering amplitude of haloes of mass M to that of the total matter field on large length scales ( k glyph[lessmuch] 1 h/ Mpc ),
$$\delta _ { \text{halo} } ( M ) = b ( M ) \delta _ { \text{matter} }, \quad \ \ ( 3 2 ) \quad \text{Th} \epsilon$$
where δ halo and δ matter are the density contrast of the distribution of haloes of mass M and of the total matter field, respectively. On smaller length scales, where the matter overdensities become larger δ matter glyph[greaterorsimilar] 1 , Eq. (32) requires higher order corrections (see e.g. [56]).
Following the same derivation steps as in Ref. [49], it is straightforward to show that the ST linear halo bias parameter can be written as
$$b ( M ) = 1 + g ( z ) \left ( \frac { q \delta _ { c } ^ { 2 } / S - 1 } { \delta _ { c } } + \frac { 2 p / \delta _ { c } } { 1 + ( q \delta _ { c } ^ { 2 } / S ) ^ { p } } \right ), ( 3 3 ) \quad \text{$\text{$as we$}$}$$
with g z ( ) = D ΛCDM ( z = 0) /D Model ( z ) , where D z ( ) is the linear growth factor of a specific model defined as δ matter ( z ) = D z δ ( ) matter ( z i ) /D z ( i ) .
## D. Halo density profiles
We adopt the NFW formula [57] to describe the radial density profile of dark matter haloes
$$\rho _ { \text{NFW} } ( r ) = \frac { \rho _ { s } } { r / r _ { s } \left [ 1 + r / r _ { s } \right ] ^ { 2 } },$$
where ρ s and r s are often called the characteristic density and the scale radius of the halo. The mass of the NFW halo, M ∆ , is obtained by integrating Eq. (34) up to some radius R ∆ (the meaning of the subscript ∆ will become clear later)
$$M _ { \Delta } & = \int _ { 0 } ^ { R _ { \Delta } } d r 4 \pi r ^ { 2 } \rho _ { \text{NFW} } ( r ) & ( 3 5 ) \\ & = 4 \pi \rho _ { s } \frac { R _ { \Delta } ^ { 3 } } { c _ { \Delta } ^ { 3 } } \left [ \ln \left ( 1 + c _ { \Delta } \right ) - \frac { c _ { \Delta } } { 1 + c _ { \Delta } } \right ],$$
where c ∆ = R /r ∆ s is the concentration parameter.
In our simulations, we define the halo mass as
$$M _ { \Delta } = \frac { 4 \pi } { 3 } \Delta \bar { \rho } _ { c 0 } R _ { \Delta } ^ { 3 },$$
i.e., M ∆ is the mass that lies inside a comoving radius R ∆ , within which the mean density is ∆ times the critical density of the Universe at the present day, ¯ ρ c 0 . In this paper, we take ∆ = 200 . By combining the two mass definitions of Eqs. (35) and (36), one gets ρ s as a function of c ∆ :
$$\rho _ { s } = \frac { 1 } { 3 } \Delta \bar { \rho } _ { c 0 } c _ { \Delta } ^ { 3 } \left [ \ln { ( 1 + c _ { \Delta } ) } - \frac { c _ { \Delta } } { 1 + c _ { \Delta } } \right ] ^ { - 1 }. \quad ( 3 7 )$$
The NFW profile then becomes fully specified by the values of r s , which are determined by direct fitting to the density profiles of the haloes from the simulations. Equivalently, and as is common practice in the literature, one can specify the concentration-mass relation c ∆ ( M ∆ ) , instead of r s ( M ∆ ) . In the context of ΛCDM cosmologies, the (mean) concentrationmass relation is well fitted by a power law function over a certain mass range [58-61]. The same is true, for instance, for Galileon gravity models [32], although with very different fitting parameters. A proper assessment of the performance of the halo model prescription therefore requires us to fit the concentration-mass relation of the R glyph[square] -2 R simulations as well. This is done in Sec. IV F. Given the relation c ∆ ( M ∆ ) , then the NFW density profile becomes completely specified by the halo mass M ∆ .
Finally, since it is the Fourier transform of the profiles, u k, M ( ) , and not the profiles themselves, that enter Eqs. (26) and (27), we simply mention that it is possible to show that
$$u _ { \text{NFW} } ( k, M ) = \int _ { 0 } ^ { R _ { \Delta } } d r 4 \pi r ^ { 2 } \frac { \sin k r } { k r } \frac { \rho _ { \text{NFW} } ( r ) } { M _ { \Delta } } \quad \text{ for.} \\ = 4 \pi \rho _ { s } r _ { s } ^ { 3 } \left \{ \frac { \sin \left ( k r _ { s } \right ) } { M } \left [ \text{Si} \left ( \left [ 1 + c _ { \Delta } \right ] k r _ { s } \right ) - \text{Si} \left ( k r _ { s } \right ) \right ] \\ & + \frac { \cos \left ( k r _ { s } \right ) } { M } \left [ \text{Ci} \left ( \left [ 1 + c _ { \Delta } \right ] k r _ { s } \right ) - \text{Ci} \left ( k r _ { s } \right ) \right ] \\ & - \frac { \sin \left ( c _ { \Delta } k r _ { s } \right ) } { M \left ( 1 + c _ { \Delta } \right ) k r _ { s } } \right \}, \\ \text{where} \, \text{Si} ( x ) = \int _ { \real ^ { x } } ^ { t } d t s \sin ( t ) / t \, \text{and} \, \text{Ci} ( x ) = - \int _ { \real ^ { \infty } } ^ { \infty } d t \cos ( t ) / t.$$
where Si( x ) = ∫ x 0 d sin( ) t t /t and Ci( x ) = -∫ ∞ x d cos( ) t t /t . Note that u k ( → 0 , M ) → 1 , as required by its normalization.
## E. Linear growth factor and spherical collapse dynamics
In order to use the ST formulae for the mass function and linear halo bias one still needs to specify and solve the equations that determine the threshold density δ c and the evolution of the linear overdensities. For scales well within the horizon, the linear (small) density contrast δ lin is governed by
$$\ddot { \delta } _ { \text{lin} } + 2 H \dot { \delta } _ { \text{lin} } - 4 \pi G _ { \text{eff} } ( a ) \bar { \rho } _ { m } \delta _ { \text{lin} } = 0, \quad ( 3 9 ) \quad \text{in} ^ { 1 } _ { \text{felt} 1 }$$
or equivalently, by changing the time variable to N = ln a , by
$$D ^ { \prime \prime } + \left ( \frac { E ^ { \prime } } { E } + 2 \right ) D ^ { \prime } - \frac { 3 } { 2 } \frac { G _ { \text{eff} } ( a ) } { G } \frac { \Omega _ { m 0 } e ^ { - 3 N } } { E ^ { 2 } } = 0, \ ( 4 0 ) \quad \text{plac} ^ { \cdot } _ { \text{mode} }$$
where we have used that δ lin ( a ) = D a δ ( ) lin ( a i ) /D a ( i ) and a prime denotes a derivative w.r.t. N . The initial conditions are set up at z i = 300 using the known matter dominated solution D a ( i ) = D a ′ ( i ) = a i . The R glyph[square] -2 R model changes the way structure grows on large scales compared to ΛCDM via its modifications to E a ( ) ≡ H a /H ( ) 0 and G eff /G .
We have defined δ c ( z ) as the linearly extrapolated value (using the Λ CDMlinear growth factor) of the initial overdensity of a spherical region for it to collapse at a given redshift, z . To determine δ c , we consider the evolution equation of the physical radius ζ = a t r ( ) of the spherical halo at time t , which satisfies the Euler equation
$$\frac { \ddot { \zeta } } { \zeta } - \left ( \dot { H } + H ^ { 2 } \right ) = - \frac { \Phi, \zeta } { \zeta } = - \frac { G _ { \text{eff} } ( a ) } { G } \frac { H _ { 0 } ^ { 2 } \Omega _ { m 0 } \delta a ^ { - 3 } } { 2 } ( 4 1 ) \quad \text{the P} \{ \\ \text{user-s} }$$
where the last equality follows from integrating Eq. (23) over ∫ r 2 d r . Changing the time variable to N and defining y t ( ) = ζ ( ) t / ( aR ) , Eq. (41) becomes
$$y ^ { \prime \prime } & + \left ( \frac { E ^ { \prime } } { E } + 2 \right ) y ^ { \prime } \\ & + \frac { G _ { \text{eff} } ( a ) } { G } \frac { \Omega _ { m 0 } e ^ { - 3 N } } { 2 E ^ { 2 } } \left ( y ^ { - 3 } - 1 \right ) y = 0, \quad ( 4 2 ) \quad \frac { ^ { 2 } \text{Ex} _ { 1 } } { y ^ { - 1 } }.$$
TABLE II. Summary of the three models we simulate in this paper. All models share the cosmological parameters of Table I. The QCDM model has the same expansion history as the R glyph[square] -2 R model, but with GR as the theory of gravity (cf. Sec. IV A).
| Model | H ( a ) | G eff /G |
|----------------------------------------|--------------------------------------------------------------------------|--------------|
| "Full" R glyph[square] - 2 R QCDM ΛCDM | H ( a ) R glyph[square] - 2 R H ( a ) R glyph[square] - 2 R H ( a ) ΛCDM | Eq. (24) 1 1 |
where we have used δ = y -3 -1 , which follows from mass conservation 2 . The initial conditions are set up as y a ( i ) = 1 -δ lin i , / 3 and y ′ ( a i ) = δ lin i , / 3 (here, δ lin i , is the linear density contrast at the initial time). The value of δ c is then determined by finding the value of the initial density δ lin i , that leads to collapse ( y = 0 , δ → ∞ ) at redshift z , evolving this afterwards until today using the Λ CDMlinear growth factor.
As we have noted above, in the R glyph[square] -2 R model, the modifications to gravity are time dependent only. In other words, the value of G eff is the same on large and on small scales. This is different, for instance, from the case of Galileon gravity. In the latter, the nonlinearities of the Vainshtein screening mechanism suppress the effective gravitational strength felt by a test particle that lies within a certain radius (known as Vainshtein radius) from a matter source. In Eq. (42), this would be simply taken into account by replacing G eff ( a ) with G eff ( a, δ = y -3 -1) [49]. The picture becomes more complicated in the case of modified gravity models which employ chameleon-type screening mechanisms [62, 63]. In these models, G eff also depends on the size (or mass) of the halo and on the gravitational potentials in the environment where the halo forms. This requires a generalization of the spherical collapse formalism for these models, which has been developed by Ref. [64].
## IV. RESULTS
## A. N-body simulations summary
Our simulations were performed with a modified version of the publicly available RAMSES N-body code [65]. RAMSES is an Adaptive Mesh Refinement (AMR) code, which solves the Poisson equation on a grid that refines itself when the effective number of particles within a given grid cell exceeds a user-specified threshold, N th . Our modifications to the code consist of (i) changing the routines that compute the background expansion rate to interpolate the R glyph[square] -2 R model expansion rate from a pre-computed table generated elsewhere; (ii) re-scaling the total force felt by the particles in the simu-
2 Explicitly, one has ¯ ρ m a R 3 3 = (1 + δ ) ¯ ρ m ζ 3 ⇒ δ = ( aR/ζ ) 3 -1 = y -3 -1 .
lation by G eff ( a /G ) , whose values are also interpolated from a table generated beforehand.
In the following sections we show the N-body simulation results obtained for three models. We simulate the "full" R glyph[square] -2 R model of action Eq. (1), whose expansion history and G eff /G are given by Eqs. (12) and (24), respectively. We also simulate a standard ΛCDM model and a model with the same expansion history as the R glyph[square] -2 R model, but with G eff /G = 1 . We call the latter model QCDM , and comparing its results to ΛCDM allows us to pinpoint the impact of the modified H a ( ) alone on the growth of structure. The specific impact of the modified G eff can then be measured by comparing the results from the "full" R glyph[square] -2 R model simulations with those from QCDM . Table II summarizes the models we consider in this paper.
We simulate all models on a cubic box of size L = 200 Mpc /h with N p = 512 3 dark matter particles. We take N th = 8 as the grid refinement criterion. The initial conditions are set up at z = 49 , using the ΛCDM linear matter power spectrum with the parameters of Table I. For each model, we simulate five realizations of the initial conditions (generated using different random seeds), which we use to construct errorbars for the simulation results by determining the variance across the realizations.
Finally, we simply note that the modifications to RAMSES needed to simulate the R glyph[square] -2 R model are trivial compared to those that are necessary to simulate models such as f ( R ) [6671], Symmetron [72, 73] or Galileons [74-79]. In these, because of the screening mechanisms (which introduce density and scale dependencies of the total force), additional solvers are needed for the (nonlinear) equations of the extra scalar degrees of freedom. In the ECOSMOG code [66] (also based on RAMSES ), these equations are solved via Gauss-Seidel relaxations on the AMR grid, which makes the simulations significantly more time consuming. We note also that recently, Ref. [80] has proposed a new and faster scheme to simulate screened modified gravity in the mildly non-linear regime. This scheme uses the linear theory result, but combines it with a screening factor computed analytically assuming spherical symmetry, which helps speed up the calculations without sacrificing the accuracy on mildly nonlinear scales too much.
## B. Linear growth and δ c curves
Before discussing the results from the simulations, it is instructive to look at the model predictions for the linear growth rate of structure and for the time dependence of the critical density δ c ( z ) .
From top to bottom, Fig. 2 shows the time evolution of the fractional difference of the expansion rate relative to ΛCDM , H/H ΛCDM -1 , the effective gravitational strength G eff /G and the fractional difference of the squared linear density contrast relative to ΛCDM ( , δ/δ ΛCDM ) 2 -1 . The expansion rate in the R glyph[square] -2 R model is lower than in ΛCDM for a glyph[greaterorsimilar] 0 1 . . This reduces the amount of Hubble friction and therefore boosts the linear growth rate. The gravitational strength in the R glyph[square] -2 R model starts growing after a glyph[greaterorsimilar] 0 2 . , being approx-
FIG. 2. The upper panel shows the evolution of the expansion rate, plotted as the fractional difference w.r.t. the ΛCDM (black) result, H a /H ( ) ΛCDM -1 as a function of the expansion factor, a . H a ( ) is the same for the R glyph[square] -2 R (blue) and QCDM (red) models. The middle panel shows the evolution of the effective gravitational strength, G eff /G . This is unity in the ΛCDM and QCDM models at all times. The lower panel shows the evolution of the squared linear density contrast, δ 2 , plotted as the fractional difference w.r.t. the ΛCDM prediction.
TABLE III. Values of the critical initial overdensity for the collapse of a spherical top-hat to occur at a = 0 6 . , a = 0 8 . , a = 1 0 . , extrapolated to a = 1 0 . with the Λ CDM linear growth factor. This extrapolation is done to allow the resulting values of δ c to be more easily compared to values in ΛCDM .
| Model | a = 0 . 6 δ c | a = 0 . 8 δ c | a = 1 . 0 δ c |
|-----------------------|-----------------|-----------------|-----------------|
| R glyph[square] - 2 R | 2 . 316 | 1 . 859 | 1 . 622 |
| QCDM | 2 . 320 | 1 . 866 | 1 . 632 |
| ΛCDM | 2 . 362 | 1 . 913 | 1 . 678 |
imately 6% larger than in GR at the present day. This also boosts the linear growth of structure, but has a smaller impact compared to the effect of the lower expansion rate. This is seen by noting that the differences between QCDM and ΛCDM in the bottom panel are larger than the differences between QCDM and the R glyph[square] -2 R model.
Figure 3 shows the time dependence of δ c . In the top panel, all models exhibit the standard result that δ c decreases with time, i.e., the initial overdensity of the spherical top-hat should be smaller, if the collapse is to occur at later times. Compared to ΛCDM , at late times ( a glyph[greaterorsimilar] 0 3 . ), the QCDM and R glyph[square] -2 R
FIG. 3. The upper panel shows the time evolution of the critical initial density for a spherical top-hat halo to collapse at scale factor a (linearly extrapolated to the present-day using the ΛCDM linear growth factor), for the ΛCDM (black), QCDM (red) and R glyph[square] -2 R (blue) models. The lower panel shows the fractional difference w.r.t. ΛCDM .
models predict lower values for δ c . This is as expected since structure formation is boosted at late times in these models, and as a result, this needs to be compensated by smaller values of the initial overdensities for the collapse to occur at the same epoch as in ΛCDM . Just like in the case of the linear growth rate, the differences w.r.t. the ΛCDM results are mainly affected by the lower expansion rate, and not by the larger values of G eff /G . At earlier times ( a glyph[lessorsimilar] 0 3 . ), all models have essentially the same expansion rate and gravitational strength, and as a result, the values of δ c are roughly the same. Table III shows the values of δ c at a = 0 60 . , a = 0 80 . and a = 1 00 . , for the three models.
## C. Interpretation of the constraints from Solar System tests of gravity
The absence of a screening mechanism in the R glyph[square] -2 R model may raise concerns about the ability of the model to satisfy Solar System constraints [1]. For instance, for the parameters of Table I, the R glyph[square] -2 R model predicts that the rate of change of the gravitational strength today, ˙ G eff /G , is
$$\frac { \dot { G } _ { \text{eff} } } { G } = H _ { 0 } \frac { \text{d} } { \text{d} N } \left ( \frac { G _ { \text{eff} } } { G } \right ) \approx 9 2 \times 1 0 ^ { - 1 3 } \, y r s ^ { - 1 }, \ \ ( 4 3 ) \ \ \ _ { \stackrel { 3 \text{ per} } { \text{ for} } }$$
which is at odds with the observational contraint ˙ G eff /G = (4 ± 9) × 10 -13 yr -1 , obtained from Lunar Laser Ranging experiments [81]. Hence, it seems that this type of local constraints can play a crucial role in determining the observational viability of the R glyph[square] -2 R model, potentially ruling it out (see e.g. Refs. [37, 38] for a similar conclusion, but in the context of other models).
It is interesting to contrast this result with that of the nonlocal model of Ref. [7], which we call here the f ( X ) model (for brevity), where f ( X ) is a free function that appears in the action and X = glyph[square] -1 R . The equations of motion of this model can be schematically written as
$$G _ { \mu \nu } \left [ 1 + \chi ( X ) \right ] + \Delta G _ { \mu \nu } = T _ { \mu \nu }, \text{ \quad \ \ } ( 4 4 )$$
where ∆ G µν encapsulates all the extra terms that are not proportional to G µν and the factor χ X ( ) is given by
$$\chi = f ( X ) + \Box ^ { - 1 } \left [ R \frac { \mathrm d f } { \mathrm d X } ( X ) \right ]. \quad \quad ( 4 5 )$$
For the purpose of our discussion, it is sufficient to look only at the effect of χ in Eq. (44). This rescales the gravitational strength as
$$\frac { G _ { \text{eff} } } { G } = \left \{ 1 + \chi \right \} ^ { - 1 },$$
which is similar to the effect of S in the R glyph[square] -2 R model. There is, however, one very important difference associated with the fact that in the case of the f ( X ) model, one has the freedom to choose the functional form of the terms that rescale G eff . To be explicit, we write the argument of f as
$$X = \Box ^ { - 1 } R = \Box ^ { - 1 } \bar { R } + \Box ^ { - 1 } \delta R, \text{ \quad \ \ } ( 4 7 )$$
where ¯ R and δR are, respectively, the background and spatially perturbed part of R . As explained in Ref. [6], the relative size of ¯ R and δR is different in different regimes. At the background level, glyph[square] -1 δR = 0 and so the operator glyph[square] -1 acts only on ¯ R . On the other hand, within gravitationally bound objects we have glyph[square] -1 δR > glyph[square] -1 ¯ R . Now recall that the covariant glyph[square] operator acts with different signs on purely timeand space-dependent quantities 3 . As a result, the sign of X on the background differs from that within bound systems, such as galaxies or our Solar System. This can be exploited to tune the function f in such a way that it vanishes when the sign of X is that which corresponds to bound systems. In this way, χ = 0 and one recovers GR completely 4 . When X takes the sign that corresponds to the background, then the function f is tuned to reproduce a desired expansion history, typically ΛCDM . In the case of the R glyph[square] -2 R model, S is fixed
3 For instance, in flat four-dimensional Minkowski space we have glyph[square] = + ∂ 2 ∂t 2 -∂ 2 ∂x 2 -∂ 2 ∂y 2 -∂ 2 ∂z 2 .
4 In Eq. (44), ∆ G µν also vanishes if χ = 0 .
TABLE IV. Best-fitting Sheth-Tormen ( q, p ) parameters to the simulation results at a = 0 6 . , a = 0 8 . and a = 1 0 . . The uncertainty in the values of q and p is ∆ = 3 5 q . × 10 -3 and ∆ = 1 5 p . × 10 -3 , respectively. These parameters are those that minimize the quantity ∑ i | n sims ( > M i ) /n ST ( > M ,q,p i ) - | 1 , in which n sims is the cumulative mass function measured from the simulations and n ST is the analytical result given by the Sheth-Tormen mass function of Eqs. (28) and (31). Here, the index i runs over the number of bins used in the simulation results.
| Model | a = 0 . 6 ( q, p ) | a = 0 . 8 ( q, p ) | a = 1 . 0 ( q, p ) |
|-----------------------|----------------------|----------------------|----------------------|
| Standard | (0 . 750 , 0 . 300) | (0 . 750 , 0 . 300) | (0 . 750 , 0 . 300) |
| ΛCDM | (0 . 713 , 0 . 323) | (0 . 756 , 0 . 326) | (0 . 756 , 0 . 341) |
| QCDM | (0 . 727 , 0 . 321) | (0 . 756 , 0 . 331) | (0 . 763 , 0 . 344) |
| R glyph[square] - 2 R | (0 . 720 , 0 . 321) | (0 . 741 , 0 . 326) | (0 . 756 , 0 . 336) |
to be S = glyph[square] -2 ( ¯ + R δR ) and one does not have the freedom to set it to zero inside bound objects. Consequently, the timedependent part of S is always present in Eq. (24), which could potentially lead to a time-dependent gravitational strength that is at odds with the current constraints.
For completeness, one should be aware of a caveat. In the above reasoning, we have always assumed that the line element of Eq. (11) is a good description of the geometry of the Solar System. The question here is whether or not the factor a t ( ) 2 should be included in the spatial sector of the metric when describing the Solar System. This is crucial as the presence of a t ( ) in Eq. (11) determines if S varies with time or not. If a t ( ) is considered, then S varies with time and G eff is time-varying as well. In this way, the model fails the Solar System tests. On the other hand, if one does not consider a t ( ) in the metric, then G eff is forcibly constant, and there are no apparent observational tensions. Such a static analysis was indeed performed by Refs. [20, 26], where it was shown that the model can cope well with the local constraints.
This boils down to determining the impact of the global expansion of the Universe on local scales. It is not clear to us that if a field is varying on a time-evolving background, then it should not do so in a small perturbation around that background. However, we acknowledge this is an open question to address, and such study is beyond the scope of the present paper. In what follows, we limit ourselves to assuming that Eq. (24) holds on all scales, but focus only on the cosmological (rather than local) interpretation of the results.
## D. Halo mass function
Our results for the cumulative mass function of the ΛCDM (black), QCDM (red) and R glyph[square] -2 R (blue) models are shown in Fig. 4 at a = 0 60 . , a = 0 80 . and a = 1 00 . . The symbols show the simulation results obtained with the halo catalogues we built using the Rockstar halo finder [82]. The results in the figure correspond to catalogues with subhaloes filtered out. The lines show the ST analytical prediction (Eqs. (28) and (31)) computed for the fitted (solid lines) and standard (dashed lines) ST ( q, p ) parameters of Table IV. The ( q, p ) parameters were fitted for all of the epochs shown, using the corresponding values of δ c in Table III. From the figure, one notes that although performing the fitting helps to improve the accuracy of the analytical formulae, overall the use of the standard values for ( q, p ) provides a fair estimate of the halo abundances in the R glyph[square] -2 R model, and of its relative difference w.r.t. ΛCDM . This is not the case, for instance, in Galileon gravity models, for which Ref. [32] has found that it is necessary to recalibrate substantially the values of ( q, p ) if the ST mass function is to provide a reasonable estimate of the effects of the modifications to gravity. In the case of the R glyph[square] -2 R model, the fact that the standard values of ( q, p ) = (0 75 0 30) . , . work reasonably well means that the modifications in the R glyph[square] -2 R model, relative to ΛCDM , are mild enough for its effects on the mass function to be well captured by the differences in δ c ( z ) .
At a = 1 00 . , the mass function of the QCDM model shows an enhancement at the high-mass end ( M glyph[greaterorsimilar] 5 × 10 12 M /h glyph[circledot] ), and a suppression at the low-mass end ( M glyph[lessorsimilar] 5 × 10 12 M /h glyph[circledot] ), relative to ΛCDM . This is what one expects in hierarchical models of structure formation if the growth rate of structure is boosted, as smaller mass objects are assembled more efficiently to form larger structures, leaving fewer of them. The effects of the enhanced G eff /G maintain this qualitative picture, but change it quantitatively. More explicitly, the mass scale below which the mass function drops below that of ΛCDM is smaller than the mass range probed by our simulations; and the enhancement of the number density of massive haloes is more pronounced. In particular, compared to ΛCDM , haloes with masses M ∼ 10 14 M /h glyph[circledot] are ∼ 5% and ∼ 15% more abundant in the QCDM and R glyph[square] -2 R models, respectively. Figure 4 also shows that the relative differences w.r.t. ΛCDM do not change appreciably with time after a ∼ 0 80 . . At earlier times ( a ∼ 0 60 . ), the halo abundances in the QCDM and R glyph[square] -2 R models approach one another, and their relative difference to ΛCDM decreases slightly, compared to the result at later times.
## E. Halo bias
The linear halo bias predictions for the ΛCDM (black), QCDM (red) and R glyph[square] -2 R (blue) models are shown in Fig. 5 at a = 0 6 . , a = 0 8 . and a = 1 0 . . The symbols show the simulation results, which were obtained by measuring the ratio
$$b ( k, M ) = \frac { P _ { \text{hm} } ( k, M ) } { P _ { k } },$$
where P k is the total matter power spectrum and P hm ( k, M ) is the halo-matter cross spectrum for haloes of mass M . These were measured with the aid of a Delaunay Tessellation field estimator code [83, 84]. We measure the cross spectrum, instead of the halo-halo power spectrum, to reduce the amount of shot noise. The linear halo bias parameter is then given by the large scale limit of b k, M ( ) , i.e., b M ( ) = b k ( glyph[lessmuch] 1 , M ) .
FIG. 4. The cumulative mass function of dark matter haloes (upper panels) for the ΛCDM (black), QCDM (red) and R glyph[square] -2 R (blue) models, at three epochs a = 0 6 . , a = 0 8 . a = 1 0 . , as labelled. The lower panels show the difference w.r.t. the ΛCDM model results. The symbols show the simulation results, and the errorbars indicate twice the variance across the five realizations of the initial conditions. We have used the phase-space friends-of-friends Rockstar code [82] to build the halo catalogues (without subhalos) used to compute the halo abundances. We only show the results for haloes with mass M 200 > 100 × M p ∼ 5 × 10 11 M /h glyph[circledot] , where M p = ρ m 0 L /N 3 p is the particle mass in the simulations. The lines correspond to the ST mass function of Eqs. (28) and (31) obtained using the fitted (solid lines) and the standard (dashed lines) ( q, p ) parameters listed in Table IV.
We only consider the halo mass bins for which b k, M ( ) has clearly saturated to its asymptotic value on large scales.
The simulation results show that, within the errorbars, the linear halo bias parameter for the three models is indistinguishable at all epochs shown. This shows that the modifications to gravity in the R glyph[square] -2 R model are not strong enough to modify substantially the way that dark matter haloes trace the underlying density field. The ST formula, Eq. (33), reproduces the simulations results very well. Note also that there is little difference between the curves computed using the fitted (solid lines) and the standard (dashed lines) ( q, p ) parameters of Table IV. We conclude the same as in the case of the mass function that, in the context of the R glyph[square] -2 R model, there is no clear need to recalibrate the ( q, p ) parameters in order to reproduce the bias results from the simulations.
TABLE V. Concentration-mass relation best-fitting ( α, β ) parameters in the parametrization log 10 ( c 200 ) = α + β log 10 ( M 200 / [ 10 12 M /h glyph[circledot] ]) to the simulation results at a = 0 6 . , a = 0 8 . and a = 1 0 . . The uncertainty in the values of α and β is ∆ α = ∆ β = 0 001 . . These are the parameters that minimize the quantity ∑ i ( c sims 200 ( M i ) -c param 200 ( M ,α,β i )) 2 / (2∆ c sims 200 ( M i )) 2 , where c sims 200 ( M i ) is the mean halo concentration measured from the simulations, ∆ c sims 200 ( M i ) is the variance of the mean across the five realizations and c param 200 ( M ,α,β i ) is the concentration given by the parametrization. Here, the index i runs over the number of mass bins.
## F. Halo concentration
Figure 6 shows the halo concentration-mass relation for the ΛCDM (black), QCDM (red) and R glyph[square] -2 R (blue) models, at a = 0 60 . , a = 0 80 . and a = 1 00 . . The symbols correspond to the mean values of c 200 identified in the same halo catalogues used in Fig. 4. For all models, and at all epochs and mass
| Model | a = 0 . 6 ( α, β ) | a = 0 . 8 ( α, β ) | a = 1 . 0 ( α, β ) |
|-----------------------|-----------------------|-----------------------|-----------------------|
| ΛCDM | (0 . 729 , - 0 . 066) | (0 . 813 , - 0 . 084) | (0 . 863 , - 0 . 093) |
| QCDM | (0 . 726 , - 0 . 068) | (0 . 814 , - 0 . 087) | (0 . 866 , - 0 . 100) |
| R glyph[square] - 2 R | (0 . 737 , - 0 . 067) | (0 . 834 , - 0 . 086) | (0 . 898 , - 0 . 095) |
scales shown, one sees that the halo concentrations are well fitted by the power law function (solid lines),
$$\vdots \quad \log _ { 1 0 } ( c _ { 2 0 0 } ) = \alpha + \beta \log _ { 1 0 } \left ( M _ { 2 0 0 } / \left [ 1 0 ^ { 1 2 } M _ { \odot } / h \right ] \right ), \, ( 4 9 )$$
FIG. 5. Linear halo bias in the ΛCDM (black), QCDM (red) and R glyph[square] -2 R (blue) models, for three epochs a = 0 6 . , a = 0 8 . and a = 1 0 . , as labelled. The symbols show the asymptotic value of the halo bias on large scales measured from the simulations as b M ( ) = P hm ( k → 0 , M ) /P k ( ) , considering only haloes (and not subhaloes) with mass M 200 > 100 × M p ∼ 5 × 10 11 M /h glyph[circledot] , where M p = ρ m 0 L /N 3 p is the particle mass. Only the mass bins for which the values of P hm ( k, M ) /P k ( ) have reached a constant value on large scales are shown. The errorbars show twice the variance across the five realizations of the initial conditions. The solid and dashed lines show the prediction from the ST formula, Eq. (33), computed, respectively, with the best-fitting and standard ( q, p ) parameters listed in Table IV.
FIG. 6. Halo concentration-mass relation in the ΛCDM (black), QCDM (red) and R glyph[square] -2 R (blue) models, for three epochs a = 0 6 . , a = 0 8 . and a = 1 0 . , as labelled. The symbols show the mean halo concentration in each mass bin, considering only haloes (and not subhaloes) with mass M 200 > 1000 × M p ∼ 5 × 10 12 M /h glyph[circledot] , where M p = ρ m 0 L /N 3 p is the particle mass. In the a = 0 6 . panel, we omit the results from the two highest mass bins due to their few number of objects. The errorbars show twice the variance of the mass-binned mean concentration across the five realizations of the initial conditions. The solid lines show the best-fitting power law relations of Table V.
.
with the best-fitting ( α, β ) parameters given in Table V. In the R glyph[square] -2 R and QCDM models, one recovers the standard ΛCDM result that halo concentration grows with time at fixed mass, and that, at a given epoch, the concentration decreases with halo mass.
At early times ( a glyph[lessorsimilar] 0 6 . ), all models predict essentially the same concentration-mass relation. At later times, however, the halo concentrations in the R glyph[square] -2 R model become increasingly larger compared to ΛCDM . In particular, at a = 1 00 . , the halos are ≈ 8% more concentrated in the R glyph[square] -2 R model, compared to ΛCDM , for the entire mass range probed by the simulations. This can be attributed to a combination of two effects. Firstly, the enhanced structure formation in the
R glyph[square] -2 R model may cause the haloes to form at earlier times. This leads to higher concentrations since the haloes form at an epoch when the matter density in the Universe was higher. Secondly, the increasingly larger value of G eff is also expected to play a role via its effect in the deepening of the gravitational potentials. In other words, even after the halo has formed, the fact that gravity keeps getting stronger with time may also help to enhance the concentration of the haloes (see also Refs. [85-88]). In the case of the QCDM model, one finds that the halo concentrations are hardly distinguishable (within errorbars) from those in the ΛCDM model, at all times and for all mass scales. This suggests that the differences between the expansion history of the QCDM and
ΛCDM models (cf. Fig. 2) are not large enough to have an impact on the formation time of the haloes. Once the haloes have formed in these two models, one can think as if the clustering inside these haloes decouples from the expansion. As a result, and since the gravitational strength is the same (cf. Table II), one sees no significant differences in the concentration of the haloes from the QCDM and ΛCDM simulations.
## G. Nonlinear matter power spectrum
Figure 7 shows our results for the nonlinear matter power spectrum. The power spectrum from the simulations was measured using the POWMES code [89]. The solid (dashed) lines show the halo model prediction obtained using Eq. (25) with the fitted (standard) ( q, p ) parameters of Table IV. The dotted lines show the predictions obtained using linear theory. Next, we discuss these results separately for large, intermediate and small length scales.
Large scales. On scales k glyph[lessorsimilar] 0 1 . h/ Mpc , the halo model is dominated by the 2-halo term, which is practically indistinguishable from the linear matter power spectrum. This is because, in the limit in which k → 0 , one has that
$$I ( k ) \sim \int \mathrm d M \frac { 1 } { \bar { \rho } _ { m 0 } } \frac { \mathrm d n ( M ) } { \mathrm d \mathrm l n M } b _ { \mathrm l i n } ( M ) = 1, \quad ( 5 0 ) \underset { \substack { \bmod { \mathrm l } \\ \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \mathrm \cdot } } { \bmod { \mathrm l } \sum } }$$
where we have used the fact that | u k ( → 0 , M ) | → 1 and the last equality holds by the definition of the ST mass function and halo bias [31]. In fact, in the standard halo model approach, replacing P 2 h k by P k, lin in Eq. (25) leads to practically no difference. As a result, the agreement between the halo model and the simulation results on large (linear) scales is always guaranteed.
Intermediate scales. On scales 0 1 . h/ Mpc glyph[lessorsimilar] k glyph[lessorsimilar] 1 h/ Mpc , the halo model underpredicts slightly the power spectrum measured from the simulations, for all models and at all epochs shown. This is due to a fundamental limitation of the halo model on these scales, which follows from some simplifying assumptions about the modelling of halo bias on these intermediate scales (see e.g. Sec. IV.F of Ref. [32] for a simple explanation). In fact, the so-called halofit model arises as an alternative to the halo model that is more accurate on these intermediate scales [90-92]. Nevertheless, in terms of the relative difference to ΛCDM , the halo model limitations cancel to some extent, which leads to a better agreement with the simulation results. Focusing on a = 1 00 . , the halo model predictions for the QCDM model reproduce very well the results from the simulations. The predictions for the R glyph[square] -2 R model, although not as accurate as in QCDM , still provide a fair estimate of the enhancement of the clustering power relative to ΛCDM . For example, at k ≈ 1 h/ Mpc and a = 1 00 . , the simulations show an increase of ≈ 11% in the power relative to ΛCDM , whereas the halo model predicts an enhancement of ≈ 15% , which is similar. Finally, it is worth mentioning that the performance of the halo model when ones uses the standard ( q, p ) = (0 75 0 30) . , . values (dashed lines) is com- parable to the case where one uses the values that best fit the mass function results (solid lines).
Small scales. On scales of k glyph[greaterorsimilar] 1 h/ Mpc , the halo model predictions are dominated by the 1-halo term, whose agreement with the simulations becomes better than on intermediate scales, especially at a = 1 00 . . There are still some visible discrepancies at a = 0 60 . , which are similar to those found in Ref. [32] for Galileon gravity models. These discrepancies are however likely to be related with some of the assumptions made in the halo model approach, namely that all matter in the Universe lies within bound structures, which is not true in the simulations. However, similarly to what happens on intermediate scales, the halo model performs much better when one looks at the relative difference w.r.t. ΛCDM . The predictions obtained by using the standard ( q, p ) parameter values (dashed lines), although not as accurate as the results obtained by using the fitted ( q, p ) values (solid lines), are still able to provide a good estimate of the effects of the modifications to gravity in the R glyph[square] -2 R model on the small-scale clustering power.
In the QCDM model, the relative difference w.r.t. ΛCDM becomes smaller with increasing k . In particular, for k glyph[greaterorsimilar] 10 h/ Mpc at a = 1 0 . , the clustering amplitude of these two models becomes practically indistinguishable. This result can be understood with the aid of the halo model expression for the 1-halo term, P 1 h k , (cf. Eq. (26)), which depends on the halo mass function and concentration-mass relation. Firstly, one notes that for smaller length scales, the integral in P 1 h k becomes increasingly dominated by the lower mass end of the mass function. Consequently, the fact that the mass function of the QCDM model approaches that of ΛCDM at low masses (becoming even smaller for M glyph[lessorsimilar] 5 × 10 12 M /h glyph[circledot] at a = 1 00 . ), helps to explain why the values of ∆ P /P k k, ΛCDM decrease for k glyph[greaterorsimilar] 1 h/ Mpc . Secondly, according to Fig. 6, the halo concentrations are practically the same in the ΛCDM and QCDM models. In other words, this means that inside small haloes (those relevant for small scales), matter is almost equally clustered in these two models, which helps to explain why ∆ P /P k k, ΛCDM is compatible with zero for k glyph[greaterorsimilar] 10 h/ Mpc (Ref. [93] finds similar results for k -mouflage gravity models).
The same reasoning also holds for the R glyph[square] -2 R model, which is why one can also note a peak in ∆ P /P k k, ΛCDM at k ∼ 1 h/ Mpc . However, in the case of the R glyph[square] -2 R model, the mass function is larger at the low-mass end and the halo concentrations are also higher, compared to QCDM and ΛCDM . These two facts explain why ∆ P /P k k, ΛCDM does not decrease in the R glyph[square] -2 R model, being roughly constant at a = 1 00 . for k glyph[greaterorsimilar] 1 h/ Mpc . In particular, we have explicitly checked that if one computes the halo model predictions of the R glyph[square] -2 R model, but using the concentration-mass relation of ΛCDM , then one fails to reproduce the values of ∆ P /P k k, ΛCDM on small scales. This shows that a good performance of the halo model on small scales is subject to a proper modelling of halo concentration, which can only be accurately determined in N-body simulations.
FIG. 7. The nonlinear matter power spectrum (upper panels) in the ΛCDM (black), QCDM (red) and R glyph[square] -2 R (blue) models, at three epochs a = 0 6 . , a = 0 8 . and a = 1 0 . , as labelled. The lower panels show the different w.r.t. ΛCDM . The symbols show the simulation results, where the errorbars show twice the variance across the five realizations of the initial conditions. The solid lines show the halo model prediction obtained using Eq. (25), with the best-fitting ( q, p ) parameters listed in Table IV. The dashed lines show the power spectrum when using the standard ST ( q, p ) = (0 75 0 30) . , . parameter values. The dotted lines show the result from linear perturbation theory. These lines are indistinguishable in the upper panels.
## H. Nonlinear velocity divergence power spectrum
Figure 8 shows the nonlinear velocity divergence power spectrum, P θθ , 5 for the three models of Table II and for a = 0 60 . , a = 0 80 . and a = 1 00 . . The computation was done by first building a Delaunay tessellation using the particle distribution of the simulations [83, 84], and then interpolating the density and velocity information to a fixed grid to measure the power spectra. The upper panels show that on scales k glyph[lessorsimilar] 0 1 . h/ Mpc , the results from the simulations of all models approach the linear theory prediction, which is given by flows that exist on larger scales. This leads to an overall suppression of the divergence of the velocity field compared to the linear theory result for scales k glyph[greaterorsimilar] 0 1 . h/ Mpc , as shown in the upper panels.
$$P _ { \theta \theta } ^ { \text{linear} } = a ^ { 2 } \left ( \frac { H } { H _ { 0 } } \right ) ^ { 2 } f ^ { 2 } P _ { k } ^ { \text{linear} }, \quad \ \ ( 5 1 ) \quad \text{Hen} \\ \ g e n \,.$$
where P linear k is the linear matter power spectrum and f = dln δ lin / dln a . On smaller scales, the formation of nonlinear structures tends to slow down the coherent (curl-free) bulk
5 Here, θ is the Fourier mode of the divergence of the peculiar physical velocity field v , defined as θ glyph[vector] ( x ) = ∇ v glyph[vector] ( x /H ) 0 .
In the lower panels, the simulation results also agree with the linear theory prediction for k glyph[lessorsimilar] 0 1 . h/ Mpc . On these scales, the time evolution of the power spectrum of all models is scale independent and the relative difference encapsulates the modifications to the time evolution of P linear k , H and f , in Eq. (51). On smaller scales, the values of ∆ P θθ k /P θθ k, ΛCDM decay w.r.t. the linear theory result until approximately k = 1 h/ Mpc . This suppression follows from the fact that the formation of nonlinear structures is enhanced in the QCDM and R glyph[square] -2 R models, relative to ΛCDM (cf. Figs. 4 and 7). Hence, on these scales, the suppression in the velocity divergence caused by nonlinear structures is stronger in the QCDM and R glyph[square] -2 R model, compared to ΛCDM . Finally, on scales k glyph[greaterorsimilar] 2 -3 h/ Mpc , the relative difference to ΛCDM grows back to values comparable to the linear theory prediction. On these scales, one does not expect haloes to contribute considerably to P θθ k for two main reasons. First, as haloes virialize, the motion of its particles tends to become more random, which helps to reduce the divergence of the velocity field there. Secondly, and perhaps more importantly, P θθ k is computed from a volume-weighted field, and as a result, since
FIG. 8. The nonlinear peculiar velocity divergence power spectrum (upper panels) in the ΛCDM (black), QCDM (red) and R glyph[square] -2 R (blue) models, for three epochs a = 0 6 . , a = 0 8 . and a = 1 0 . , as labelled. The lower panels show the difference w.r.t. ΛCDM . The symbols show the simulation results, where the errorbars show the variance across the five realizations of the initial conditions. The dashed lines only link the symbols to help the visualization. The dotted lines in the bottom panels show the prediction of linear perturbation theory.
haloes occupy only a small fraction of the total volume, they are not expected to contribute significantly to the total velocity divergence power spectrum. On the other hand, considerable contributions may arise from higher-volume regions such as voids, walls or filaments, where coherent matter flows exist. For instance, matter can flow along the direction of dark matter filaments, or inside a large wall or void that is expanding (see e.g. [94-96]). These small scale flows are larger in the QCDM and R glyph[square] -2 R models at a fixed time, as shown by the growth of the values of ∆ P θθ k /P θθ k, ΛCDM on small scales.
On scales k glyph[greaterorsimilar] 2 -3 h/ Mpc , one may find it odd that the QCDM model predicts roughly the same matter power spectrum as ΛCDM (cf. Fig. 7), but has a different velocity divergence power spectrum. This has to do with the weight with which different structures contribute to P k and P θθ k . For instance, P k is computed from a mass-weighted density field, and hence, it is dominated by the highest density peaks, which are due to dark matter haloes. In other words, it is very insensitive to the behavior of the clustering of matter in voids, walls or filaments due to their lower density. On the contrary, P θθ k , which is computed from a volume-weighted field, is forcibly less sensitive to dark matter haloes due to their low volume fraction. The values of P θθ k are then mostly determined by the velocity field inside voids, walls and filaments. These structures are typically larger than haloes and therefore they are more sensitive to the background expansion of the Universe. Consequently, they are more likely to be affected by modifi- cations to H a ( ) , compared to haloes which detach from the overall expansion sooner. This can then explain the differences in the sizes of the modifications to P k and P θθ k on small scales in the QCDM model, relative to ΛCDM . To test this we have computed P θθ k by artificially setting θ glyph[vector] ( x ) = 0 in regions where the density contrast exceeds δ = 50 . This should roughly exclude the contribution from haloes to the values of P θθ k . We have found no visible difference w.r.t. the results of Fig. 8, which shows that the small scale behavior of the velocity divergence is not affected by what happens inside dark matter haloes. We have performed the same calculation, but by setting θ glyph[vector] ( x ) = 0 whenever δ < 0 , to exclude the contribution from voids. We have found that at a = 1 , the relative difference of QCDM to ΛCDM at k ∼ 10 h/ Mpc drops from ∼ 9% (as in Fig. 8) to ∼ 7% . This seems to suggest that the dominant effect in the small scale behavior of P θθ k comes from walls and/or filaments. The velocity divergence in these structures is typically large (see e.g. Fig. 2 of Ref. [97]) and they also occupy a sizeable fraction of the total volume as well. A more detailed investigation of these results is beyond the scope of the present paper.
Focusing at a = 1 , at k ∼ 10 h/ Mpc and relative to ΛCDM , the velocity power spectrum in the R glyph[square] -2 R model is enhanced by ∼ 12% , and the matter power spectrum by ∼ 15% . On large (linear) scales the same figures are ∼ 12% and ∼ 7% , respectively. The size of the modifications to the matter and velocity divergence power spectrum are rather
similar, but the latter might be easier to measure as they are typically less sensitive to assumptions about baryonic processes such as galaxy bias. As an example, redshift space distortions (RSD) [98-100] are sensitive to the boost of the velocity field on large scales, and therefore can be used to test modified gravity models (see e.g. Refs. [77, 101]). The work of Refs. [102, 103] illustrated how the velocity distribution of infalling galaxies around massive clusters can be used to detect modifications to gravity (see also [104]). More recently, Ref. [105] demonstrated that modified gravity models can leave particularly strong signatures in the velocity dispersion of pairs of galaxies on a broad range of distance scales. The level of precision of the data from future observational missions should prove sufficient to disentangle the differences depicted in Fig. 8. Such a forecast study would involve running simulations with better resolution and larger box sizes, and as such, we leave it for future work.
## V. SUMMARY
We have studied the nonlinear regime of structure formation in nonlocal gravity cosmologies using N-body simulations, and also in the context of the semi-analytical ellipsoidal collapse and halo models. To the best of our knowledge, this is the first time the nonlinear growth of structure in nonlocal cosmologies has been studied. In particular, we investigated the impact that the modifications to gravity in nonlocal models have on the halo mass function, linear halo bias parameters, halo concentrations and on the statistics of the density and velocity fields of the dark matter.
The action or equations of motion of nonlocal gravity models are typically characterized by the inverse of the d'Alembertian operator acting on curvature tensors. Here, we focused on the model of Refs. [26, 27], in which the standard Einstein-Hilbert action contains an extra term proportional to R glyph[square] -2 R (cf. Eq. (1)). The constant of proportionality is fixed by the dark energy density today, and hence this model contains the same number of free parameters as ΛCDM , although it has no ΛCDM limit for the background dynamics or gravitational interaction.
Our goal was not to perform a detailed exploration of the cosmological parameter space in the R glyph[square] -2 R model. Instead, for the R glyph[square] -2 R model we used the same cosmological parameters as ΛCDM (cf. Table I). In this way one isolates the impact of the modifications to gravity from the impact of having different cosmological parameter values. Nevertheless, although a formal exploration of the parameter space in the R glyph[square] -2 R model is left for future work (see Ref. [48]), the comparison presented in Fig. 1 suggests that the model fits the CMB temperature data as well as ΛCDM . Our main results can be summarized as follows:
- · The expansion rate in the R glyph[square] -2 R model is smaller than in ΛCDM at late times, and the gravitational strength is enhanced by a time-dependent factor (cf. Fig. 2). Both effects help to boost the linear growth of structure (cf. Fig. 2) and also speed up the collapse of spherical matter overdensities
- (cf. Fig. 3). In particular, at the present day, the amplitude of the linear matter (velocity divergence) power spectrum is enhanced by ≈ 7% ( ≈ 12% ) in the R glyph[square] -2 R model, compared to ΛCDM . These results are in agreement with Ref. [27]. The critical density for collapse today, δ c ( a = 1) , is ≈ 3% smaller in the R glyph[square] -2 R model, relative to ΛCDM (cf Fig. 3). For these results, the modified expansion history plays the dominant role in driving the differences w.r.t. ΛCDM , compared to the effect of the enhanced G eff .
- · At late times ( a > 0 6 . ), the number density of haloes with masses M glyph[greaterorsimilar] 10 12 M /h glyph[circledot] is higher in the R glyph[square] -2 R model, compared to ΛCDM . The difference becomes more pronounced at the high-mass end of the mass function. In particular, at a = 1 , haloes with mass M ∼ 10 14 M /h glyph[circledot] are ≈ 10% more abundant in the R glyph[square] -2 R model than in ΛCDM . At M = 10 12 M /h glyph[circledot] this difference is only ≈ 2% . The effects of the modified H a ( ) and G eff on the enhancement of the high-mass end of the mass function are comparable.
The ST mass function describes well the absolute values of the halo number densities as well as the relative differences w.r.t. ΛCDM , for all of the epochs studied (cf. Fig. 4). We find that the use of the standard ( q, p ) = (0 75 0 30) . , . ST parameter values provides a fair estimate of the modifications to the mass function in the R glyph[square] -2 R model. However, recalibrating these parameters to the simulation results helps to improve the accuracy of the fit (cf. Table IV).
- · The linear halo bias parameter in the R glyph[square] -2 R model is barely distinguishable from that in ΛCDM for all masses and epochs studied (cf. Fig. 5). In other words, the modifications to gravity in the R glyph[square] -2 R model play a negligible role in the way dark matter haloes trace the underlying density field. The ST halo bias formula provides therefore a good description of the simulation results. There is also almost no difference between the semi-analytical predictions for the bias computed using the best-fitting and standard values for the ( q, p ) ST parameters.
- · The halo concentration-mass relation is well-fitted by a power law function (cf. Fig. 6), but with fitting parameters that differ from those of ΛCDM (see Table V). For a glyph[lessorsimilar] 0 6 . , the concentration of the haloes in the R glyph[square] -2 R model is roughly the same as in ΛCDM , but it increases with time. In particular, at a = 1 0 . ( a = 0 8 . ) and for all masses, haloes are ≈ 8% ( ≈ 4% ) more concentrated in the R glyph[square] -2 R model, compared to ΛCDM . This is likely to be mainly due to the enhanced G eff on small scales, which helps to make the gravitational potential continuously deeper inside the haloes. On the other hand, the effects of the modifications to the expansion history in the R glyph[square] -2 R play a negligible role in changing the concentration of the haloes. This can be explained by the fact that the modifications to H a ( ) are small enough not to have a significant impact on the formation time of the haloes. Consequently, once the haloes form, they detach from the expansion of the Universe and no longer "feel" the dynamics of the background.
- · The modifications to gravity in the R glyph[square] -2 R model lead only to a modest enhancement of the clustering power. For instance, at a = 1 0 . ( a = 0 8 . ) the amplitude of the nonlinear matter power spectrum is never larger than ≈ 15% ( ≈ 10% ) on all scales (cf. Fig. 7). These differences might be hard to disentangle using data from galaxy clustering given the known uncertainties in modelling galaxy bias. On small scales, k glyph[greaterorsimilar] 1 h/ Mpc , the differences w.r.t. ΛCDM are completely determined by the enhanced G eff , and not by the modifications to H a ( ) .
At a = 1 0 . , the halo model describes the simulation results on large ( k glyph[lessorsimilar] 0 1 . h/ Mpc ) and small ( k glyph[greaterorsimilar] 1 h/ Mpc ) scales very well. On intermediate scales and also at earlier times, the performance becomes worse due to known limitations of the halo model [31, 32]. These follow from a number of approximations in the derivation of the halo model formalism, which sacrifice some accuracy in favour of analytical convenience. In terms of the relative difference w.r.t. ΛCDM , these limitations cancel out and the halo model describes the simulation results reasonably well for all epochs and scales. Moreover, the performance of the halo model formulae in describing the simulation results does not depend critically on the fitting of the ( q, p ) parameters to the mass function. However, we have checked that the good performance of the halo model on small scales is subject to a correct modelling of the halo-concentration mass relation, which can only be properly determined via N-body simulation.
- · Similarly to the case of the matter power spectrum, the modifications in the R glyph[square] -2 R model lead only to modest changes in the amplitude of the nonlinear velocity divergence power spectrum. In particular, at a = 1 0 . ( a = 0 8 . ) the enhancement relative to ΛCDM is kept below ≈ 12% ( ≈ 10% ) on all scales. However, measurements of RSD and/or galaxy infall dynamics are less subject to galaxy bias uncertainties, and therefore, might stand a better chance of distinguishing between these two models.
- · The R glyph[square] -2 R model possesses no screening mechanism to suppress the modifications to gravity on small scales. As a result, Solar System tests of gravity can be used to constrain the model. For example, the R glyph[square] -2 R model predicts that ˙ G eff /G ≈ 92 × 10 -13 yrs -1 , which is incompatible with the current bound from Lunar Laser Ranging experiments, ˙ G eff /G = (4 ± 9) × 10 -13 yr -1 [81]. The local time variation of G eff follows from the background evolution of the auxiliary scalar field S , and it seems nontrivial to devise a mechanism that can suppress it around massive objects or in high-density regions [37, 38]. In this paper, we focused only on a particular choice of cosmological parameters. As a result, it might be possible that certain parameter combinations can be made compatible with Solar System tests, whilst still being able to
yield viable cosmological solutions. Nevertheless, it seems clear that these tests should be taken into account in future constraint studies, as they might have the potential to rule out these models observationally.
In conclusion, the R glyph[square] -2 R model, although it has no ΛCDM limit for the dynamics of the background and gravitational interaction, exhibits changes of only a few percent in observables sensitive to the nonlinear growth of structure. Some of these effects are degenerate with baryonic mechanisms such as AGN feedback or galaxy bias, or even with massive neutrinos [45, 106-108]. This makes it challenging to distinguish this model from ΛCDM , but the precision of upcoming observational missions such as Euclid [109, 110], DESI [111] or LSST [112] should make this possible. From an observational point of view, however, future nonlinear studies are only warranted provided the model is able to fit successfully the CMB data from Planck (see Ref. [48]), in a way that is also compatible with the constraints from Lunar Laser Ranging experiments. The latter requirement might be hard to satisfy due to the absence of a screening mechanism.
## ACKNOWLEDGMENTS
We thank Philippe Brax, Marius Cautun, Yves Dirian, Stefano Foffa, Ruth Gregory, Nima Khosravi, Michele Maggiore, Michele Mancarella, Alex Kehagias, Martin Kunz and Patrick Valageas for useful comments and discussions. We also thank Lydia Heck for invaluable numerical support. This work was supported by the Science and Technology Facilities Council [grant number ST/L00075X/1]. This work used the DiRAC Data Centric system at Durham University, operated by the Institute for Computational Cosmology on behalf of the STFC DiRAC HPC Facility (www.dirac.ac.uk). This equipment was funded by BIS National E-infrastructure capital grant ST/K00042X/1, STFC capital grant ST/H008519/1, and STFC DiRAC Operations grant ST/K003267/1 and Durham University. DiRAC is part of the National EInfrastructure. AB is supported by FCT-Portugal through grant SFRH/BD/75791/2011. BL is supported by the Royal Astronomical Society and Durham University. WAH is supported by the ERC Advanced Investigator grant of C. S. Frenk, COSMIWAY, and by the Polish National Science Center through grant DEC-2011/01/D/ST9/01960. The research leading to these results has received funding from the European Research Council under the European Union's Seventh Framework Programme (FP/2007-2013) / ERC Grant NuMass Agreement n. [617143]. This work has been partially supported by the European Union FP7 ITN INVISIBLES (Marie Curie Actions, PITN- GA-2011- 289442) and STFC.
- et al. (2013), arXiv:1309.5385 [astro-ph.CO]
- [4] B. Jain, A. Joyce, R. Thompson, A. Upadhye, J. Battat, et al. (2013), arXiv:1309.5389 [astro-ph.CO]
- [5] A. Joyce, B. Jain, J. Khoury, and M. Trodden(2014), arXiv:1407.0059 [astro-ph.CO]
- [6] R. Woodard, Found.Phys. 44 , 213 (2014), arXiv:1401.0254 [astro-ph.CO]
- [7] S. Deser and R. Woodard, Phys.Rev.Lett. 99 , 111301 (2007), arXiv:0706.2151 [astro-ph]
- [8] C. Deffayet and R. Woodard, JCAP 0908 , 023 (2009), arXiv:0904.0961 [gr-qc]
- [9] S. Deser and R. Woodard, JCAP 1311 , 036 (2013), arXiv:1307.6639 [astro-ph.CO]
- [10] S. Nojiri and S. D. Odintsov, Phys.Lett. B659 , 821 (2008), arXiv:0708.0924 [hep-th]
- [11] S. Jhingan, S. Nojiri, S. Odintsov, M. Sami, I. Thongkool, et al. , Phys.Lett. B663 , 424 (2008), arXiv:0803.2613 [hep-th]
- [12] T. Koivisto, Phys.Rev. D77 , 123513 (2008), arXiv:0803.3399 [gr-qc]
- [13] T. S. Koivisto, Phys.Rev. D78 , 123505 (2008), arXiv:0807.3778 [gr-qc]
- [14] E. Elizalde, E. O. Pozdeeva, and S. Y. Vernov, Phys. Rev. D 85 , 044002 (Feb. 2012), arXiv:1110.5806 [astro-ph.CO]
- [15] E. Elizalde, E. O. Pozdeeva, and S. Y. Vernov, Classical and Quantum Gravity 30 , 035002 (Feb. 2013), arXiv:1209.5957 [astro-ph.CO]
- [16] S. Park and S. Dodelson, Phys. Rev. D 87 , 024003 (Jan. 2013), arXiv:1209.0836 [astro-ph.CO]
- [17] S. Dodelson and S. Park(2013), arXiv:1310.4329 [astroph.CO]
- [18] M. Maggiore, Phys.Rev. D89 , 043008 (2014), arXiv:1307.3898 [hep-th]
- [19] S. Foffa, M. Maggiore, and E. Mitsou, Int.J.Mod.Phys. A, to appear(2014), arXiv:1311.3435 [hep-th]
- [20] A. Kehagias and M. Maggiore, JHEP, to appear(2014), arXiv:1401.8289 [hep-th]
- [21] S. Nesseris and S. Tsujikawa(2014), arXiv:1402.4613 [astroph.CO]
- [22] M. Jaccard, M. Maggiore, and E. Mitsou, Phys.Rev. D88 , 044033 (2013), arXiv:1305.3034 [hep-th]
- [23] S. Foffa, M. Maggiore, and E. Mitsou, Phys.Lett. B733 , 76 (2014), arXiv:1311.3421 [hep-th]
- [24] L. Modesto and S. Tsujikawa, Phys.Lett. B727 , 48 (2013), arXiv:1307.6968 [hep-th]
- [25] P. G. Ferreira and A. L. Maroto, Phys.Rev. D88 , 123502 (2013), arXiv:1310.1238 [astro-ph.CO]
- [26] M. Maggiore and M. Mancarella, Phys.Rev. D90 , 023005 (2014), arXiv:1402.0448 [hep-th]
- [27] Y. Dirian, S. Foffa, N. Khosravi, M. Kunz, and M. Maggiore, JCAP 1406 , 033 (2014), arXiv:1403.6068 [astro-ph.CO]
- [28] R. K. Sheth and G. Tormen, Mon.Not.Roy.Astron.Soc. 308 , 119 (1999), arXiv:astro-ph/9901122 [astro-ph]
- [29] R. K. Sheth, H. Mo, and G. Tormen, Mon.Not.Roy.Astron.Soc. 323 , 1 (2001), arXiv:astroph/9907024 [astro-ph]
- [30] R. K. Sheth and G. Tormen, Mon.Not.Roy.Astron.Soc. 329 , 61 (2002), arXiv:astro-ph/0105113 [astro-ph]
- [31] A. Cooray and R. K. Sheth, Phys.Rept. 372 , 1 (2002), arXiv:astro-ph/0206508 [astro-ph]
- [32] A. Barreira, B. Li, W. A. Hellwing, L. Lombriser, C. M. Baugh, and S. Pascoli, JCAP 4 , 029 (Apr. 2014), arXiv:1401.1497
- [33] S. Capozziello, E. Elizalde, S. Nojiri, and S. D. Odintsov, Phys.Lett. B671 , 193 (2009), arXiv:0809.1535 [hep-th]
- [34] N. Koshelev, Grav.Cosmol. 15 , 220 (2009), arXiv:0809.4927 [gr-qc]
- [35] T. S. Koivisto, AIP Conf.Proc. 1206 , 79 (2010), arXiv:0910.4097 [gr-qc]
- [36] A. O. Barvinsky, Phys.Rev. D85 , 104018 (2012), arXiv:1112.4340 [hep-th]
- [37] E. Babichev, C. Deffayet, and G. Esposito-Farese, Phys.Rev.Lett. 107 , 251102 (2011), arXiv:1107.1569 [gr-qc]
- [38] R. Kimura, T. Kobayashi, and K. Yamamoto, Phys. Rev. D 85 , 024023 (Jan. 2012), arXiv:1111.6749 [astro-ph.CO]
- [39] P. Ade et al. (Planck Collaboration)(2013), arXiv:1303.5075 [astro-ph.CO]
- [40] P. Ade et al. (Planck Collaboration)(2013), arXiv:1303.5076 [astro-ph.CO]
- [41] P. Ade et al. (Planck Collaboration)(2013), arXiv:1303.5077 [astro-ph.CO]
- [42] F. Beutler, C. Blake, M. Colless, D. H. Jones, L. StaveleySmith, et al. , Mon.Not.Roy.Astron.Soc. 416 , 3017 (2011), arXiv:1106.3366 [astro-ph.CO]
- [43] N. Padmanabhan, X. Xu, D. J. Eisenstein, R. Scalzo, A. J. Cuesta, et al. , Mon.Not.Roy.Astron.Soc. 427 , 2132 (2012), arXiv:1202.0090 [astro-ph.CO]
- [44] L. Anderson, E. Aubourg, S. Bailey, D. Bizyaev, M. Blanton, et al. , Mon.Not.Roy.Astron.Soc. 427 , 3435 (2013), arXiv:1203.6594 [astro-ph.CO]
- [45] A. Barreira, B. Li, C. Baugh, and S. Pascoli(2014), arXiv:1406.0485 [astro-ph.CO]
- [46] A. Lewis, http://camb.info/
- [47] A. Barreira, B. Li, C. M. Baugh, and S. Pascoli, Phys.Rev. D86 , 124016 (2012), arXiv:1208.0600 [astro-ph.CO]
- [48] Y. Dirian, S. Foffa, M. Kunz, M. Maggiore, and V. Pettorino, in preparation
- [49] A. Barreira, B. Li, C. M. Baugh, and S. Pascoli, JCAP 1311 , 056 (2013), arXiv:1308.3699 [astro-ph.CO]
- [50] W. H. Press and P. Schechter, Astrophys. J. 187 , 425 (Feb. 1974)
- [51] M. Maggiore and A. Riotto, Astrophys. J. 717 , 515 (Jul. 2010), arXiv:0903.1250 [astro-ph.CO]
- [52] A. Nicolis, R. Rattazzi, and E. Trincherini, Phys. Rev. D 79 , 064036 (2009)
- [53] A. De Felice and S. Tsujikawa, Phys.Rev.Lett. 105 , 111301 (2010), arXiv:1007.2700 [astro-ph.CO]
- [54] C. Deffayet, S. Deser, and G. Esposito-Farese, Phys.Rev. D80 , 064015 (2009), arXiv:0906.1967 [gr-qc]
- [55] H. Mo and S. D. White, Mon.Not.Roy.Astron.Soc. 282 , 347 (1996), arXiv:astro-ph/9512127 [astro-ph]
- [56] J. N. Fry and E. Gaztanaga, Astrophys.J. 413 , 447 (1993), arXiv:astro-ph/9302009 [astro-ph]
- [57] J. F. Navarro, C. S. Frenk, and S. D. White, Astrophys.J. 490 , 493 (1997), arXiv:astro-ph/9611107 [astro-ph]
- [58] J. S. Bullock, T. S. Kolatt, Y. Sigad, R. S. Somerville, A. V. Kravtsov, et al. , Mon.Not.Roy.Astron.Soc. 321 , 559 (2001), arXiv:astro-ph/9908159 [astro-ph]
- [59] A. F. Neto, L. Gao, P. Bett, S. Cole, J. F. Navarro, et al. , Mon.Not.Roy.Astron.Soc. 381 , 1450 (2007), arXiv:0706.2919 [astro-ph]
- [60] A. V. Maccio', A. A. Dutton, and F. C. d. Bosch(2008), arXiv:0805.1926 [astro-ph]
- [61] F. Prada, A. A. Klypin, A. J. Cuesta, J. E. Betancort-Rijo, and J. Primack(2011), arXiv:1104.5130 [astro-ph.CO]
- [62] J. Khoury and A. Weltman, Phys.Rev.Lett. 93 , 171104 (2004), arXiv:astro-ph/0309300 [astro-ph]
- [63] J. Khoury and A. Weltman, Phys.Rev. D69 , 044026 (2004),
arXiv:astro-ph/0309411 [astro-ph]
- [64] B. Li and G. Efstathiou, Mon.Not.Roy.Astron.Soc. 421 , 1431 (2012), arXiv:1110.6440 [astro-ph.CO]
- [65] R. Teyssier, Astron.Astrophys. 385 , 337 (2002), arXiv:astroph/0111367 [astro-ph]
- [66] B. Li, G.-B. Zhao, R. Teyssier, and K. Koyama, JCAP 1201 , 051 (2012), arXiv:1110.1379 [astro-ph.CO]
- [67] E. Puchwein, M. Baldi, and V. Springel(2013), arXiv:1305.2418 [astro-ph.CO]
- [68] C. Llinares, D. F. Mota, and H. A. Winther(2013), arXiv:1307.6748 [astro-ph.CO]
- [69] H. Oyaizu, Phys.Rev. D78 , 123523 (2008), arXiv:0807.2449 [astro-ph]
- [70] G.-B. Zhao, B. Li, and K. Koyama, prd 83 , 044007 (Feb. 2011), arXiv:1011.1257 [astro-ph.CO]
- [71] W. A. Hellwing, B. Li, C. S. Frenk, and S. Cole, Mon.Not.Roy.Astron.Soc. 435 , 2806 (Nov. 2013), arXiv:1305.7486 [astro-ph.CO]
- [72] A.-C. Davis, B. Li, D. F. Mota, and H. A. Winther, Astrophys. J. 748 , 61 (Mar. 2012), arXiv:1108.3081 [astro-ph.CO]
- [73] C. Llinares and D. F. Mota(2013), arXiv:1312.6016 [astroph.CO]
- [74] K. C. Chan and R. Scoccimarro, PRD 80 , 104005 (Nov. 2009), arXiv:0906.4548 [astro-ph.CO]
- [75] F. Schmidt, Phys.Rev. D80 , 043001 (2009), arXiv:0905.0858 [astro-ph.CO]
- [76] B. Li, G.-B. Zhao, and K. Koyama, JCAP 1305 , 023 (2013), arXiv:1303.0008 [astro-ph.CO]
- [77] M. Wyman, E. Jennings, and M. Lima, Phys.Rev. D88 , 084029 (2013), arXiv:1303.6630 [astro-ph.CO]
- [78] A. Barreira, B. Li, W. A. Hellwing, C. M. Baugh, and S. Pascoli, JCAP 2013 , 027 (2013), arXiv:1306.3219 [astro-ph.CO]
- [79] B. Li, A. Barreira, C. M. Baugh, W. A. Hellwing, K. Koyama, et al. JCAP11 , 012 (2013), arXiv:1308.3491 [astro-ph.CO]
- [80] H. A. Winther and P. G. Ferreira(2014), arXiv:1403.6492 [astro-ph.CO]
- [81] J. G. Williams, S. G. Turyshev, and D. H. Boggs, Phys.Rev.Lett. 93 , 261101 (2004), arXiv:gr-qc/0411113 [grqc]
- [82] P. S. Behroozi, R. H. Wechsler, and H.-Y. Wu, Astrophys.J. 762 , 109 (2013), arXiv:1110.4372 [astro-ph.CO]
- [83] M. C. Cautun and R. van de Weygaert, ArXiv e-prints(May 2011), arXiv:1105.0370 [astro-ph.IM]
- [84] W. E. Schaap and R. van de Weygaert, A&A 363 , L29 (Nov. 2000), arXiv:astro-ph/0011007
- [85] W. A. Hellwing and R. Juszkiewicz, Phys. Rev. D 80 , 083522 (Oct. 2009), arXiv:0809.1976
- [86] W. A. Hellwing, S. R. Knollmann, and A. Knebe, MNRAS 408 , L104 (Oct. 2010), arXiv:1004.2929 [astro-ph.CO]
- [87] W. A. Hellwing, Annalen der Physik 522 , 351 (Mar. 2010), arXiv:0911.0573 [astro-ph.CO]
- [88] W. A. Hellwing, M. Cautun, A. Knebe, R. Juszkiewicz, and S. Knollmann, J. Cosmology Astropart. Phys. 10 , 012 (Oct. 2013), arXiv:1111.7257 [astro-ph.CO]
- [89] S. Colombi, A. H. Jaffe, D. Novikov, and C. Pichon(2008),
- arXiv:0811.0313 [astro-ph]
- [90] R. Smith et al. (Virgo Consortium), Mon.Not.Roy.Astron.Soc. 341 , 1311 (2003), arXiv:astro-ph/0207664 [astro-ph]
- [91] R. Takahashi, M. Sato, T. Nishimichi, A. Taruya, and M. Oguri, Astrophys.J. 761 , 152 (2012), arXiv:1208.2701 [astro-ph.CO]
- [92] G.-B. Zhao, Astrophys.J.Suppl. 211 , 23 (2014), arXiv:1312.1291 [astro-ph.CO]
- [93] P. Brax and P. Valageas(2014), arXiv:1403.5424 [astroph.CO]
- [94] M. Cautun, R. van de Weygaert, B. J. T. Jones, and C. S. Frenk, Mon.Not.Roy.Astron.Soc. 441 , 2923 (Jul. 2014), arXiv:1401.7866
- [95] M. Cautun, R. van de Weygaert, and B. J. T. Jones, Mon.Not.Roy.Astron.Soc. 429 , 1286 (Feb. 2013), arXiv:1209.2043 [astro-ph.CO]
- [96] N. I. Libeskind, Y. Hoffman, and S. Gottlöber, Mon.Not.Roy.Astron.Soc. 441 , 1974 (Jul. 2014), arXiv:1310.5706
- [97] B. Li, W. A. Hellwing, K. Koyama, G.-B. Zhao, E. Jennings, and C. M. Baugh, Mon.Not.Roy.Astron.Soc. 428 , 743 (Jan. 2013), arXiv:1206.4317 [astro-ph.CO]
- [98] R. Scoccimarro, Phys.Rev. D70 , 083007 (2004), arXiv:astroph/0407214 [astro-ph]
- [99] E. Jennings, C. M. Baugh, and S. Pascoli, Mon.Not.Roy.Astron.Soc. 410 , 2081 (Jan. 2011), arXiv:1003.4282 [astro-ph.CO]
- [100] E. Jennings, Mon.Not.Roy.Astron.Soc. 427 , L25 (Nov. 2012), arXiv:1207.1439 [astro-ph.CO]
- [101] E. Jennings, C. M. Baugh, B. Li, G.-B. Zhao, and K. Koyama, mnras 425 , 2128 (Sep. 2012), arXiv:1205.2698 [astro-ph.CO]
- [102] T. Y. Lam, T. Nishimichi, F. Schmidt, and M. Takada, Physical Review Letters 109 , 051301 (Aug. 2012), arXiv:1202.4501 [astro-ph.CO]
- [103] T. Y. Lam, F. Schmidt, T. Nishimichi, and M. Takada, Phys.Rev. D88 , 023012 (2013), arXiv:1305.5548 [astroph.CO]
- [104] Y. Zu, D. Weinberg, E. Jennings, B. Li, and M. Wyman(2013), arXiv:1310.6768 [astro-ph.CO]
- [105] W. A. Hellwing, A. Barreira, C. S. Frenk, B. Li, and S. Cole, Phys.Rev.Lett. 112 , 221102 (2014), arXiv:1401.0706 [astroph.CO]
- [106] M. Baldi, F. Villaescusa-Navarro, M. Viel, E. Puchwein, V. Springel, et al. (2013), arXiv:1311.2588 [astro-ph.CO]
- [107] J. Shim, J. Lee, and M. Baldi(2014), arXiv:1404.3639 [astroph.CO]
- [108] A. Barreira, B. Li, C. Baugh, and S. Pascoli(2014), arXiv:1404.1365 [astro-ph.CO]
- [109] R. Laureijs et al. (EUCLID Collaboration)(2011), arXiv:1110.3193 [astro-ph.CO]
- [110] L. Amendola et al. (Euclid Theory Working Group)(2012), arXiv:1206.1225 [astro-ph.CO]
- [111] M. Levi et al. (DESI collaboration)(2013), arXiv:1308.0847 [astro-ph.CO]
- [112] LSST Dark Energy Science Collaboration, ArXiv eprints(Nov. 2012), arXiv:1211.0310 [astro-ph.CO]
|
10.1088/1475-7516/2014/09/031
|
[
"Alexandre Barreira",
"Baojiu Li",
"Wojciech A. Hellwing",
"Carlton M. Baugh",
"Silvia Pascoli"
] |
2014-08-05T19:59:54+00:00
|
2014-08-05T19:59:54+00:00
|
[
"astro-ph.CO"
] |
Nonlinear structure formation in Nonlocal Gravity
|
We study the nonlinear growth of structure in nonlocal gravity models with
the aid of N-body simulation and the spherical collapse and halo models. We
focus on a model in which the inverse-squared of the d'Alembertian operator
acts on the Ricci scalar in the action. For fixed cosmological parameters, this
model differs from $\Lambda{\rm CDM}$ by having a lower late-time expansion
rate and an enhanced and time-dependent gravitational strength ($\sim 6\%$
larger today). Compared to $\Lambda{\rm CDM}$ today, in the nonlocal model,
massive haloes are slightly more abundant (by $\sim 10\%$ at $M \sim 10^{14}
M_{\odot}/h$) and concentrated ($\approx 8\%$ enhancement over a range of mass
scales), but their linear bias remains almost unchanged. We find that the
Sheth-Tormen formalism describes the mass function and halo bias very well,
with little need for recalibration of free parameters. The fitting of the halo
concentrations is however essential to ensure the good performance of the halo
model on small scales. For $k \gtrsim 1 h/{\rm Mpc}$, the amplitude of the
nonlinear matter and velocity divergence power spectra exhibits a modest
enhancement of $\sim 12\%$ to $15\%$, compared to $\Lambda{\rm CDM}$ today.
This suggests that this model might only be distinguishable from $\Lambda{\rm
CDM}$ by future observational missions. We point out that the absence of a
screening mechanism may lead to tensions with Solar System tests due to local
time variations of the gravitational strength, although this is subject to
assumptions about the local time evolution of background averaged quantities.
|
1408.1085v1
|
## An integrative approach for modeling and simulation of Heterocyst pattern formation in Cyanobacteria strands
Alejandro Torres-S´nchez a 1 2 , , Jes´s G´mez-Garde˜es u o n 1 3 , , Fernando Falo 1 3 , , ∗
- 1 Departamento de F´ ısica de la Materia Condensada, Universidad de Zaragoza, E-50009 Zaragoza, Spain.
- 2 Laboratori de C`lcul Num`ric, Universitat de Polit`cnica de Catalunya, E-08034 a e e Barcelona, Spain.
- 3 Instituto de Biocomputaci´n y F´ ısica de Sistemas Complejos (BIFI), Universidad de o Zaragoza, E-50018 Zaragoza, Spain.
- ∗ E-mail: [email protected]
## Abstract
A comprehensive approach to cellular differentiation in cyanobacteria is developed. To this aim, the process of heterocyst cell formation is studied under a systems biology point of view. By relying on statistical physics techniques, we translate the essential ingredients and mechanisms of the genetic circuit into a set of differential equations that describes the continuous time evolution of combined nitrogen, PatS, HetR and NtcA concentrations. The detailed analysis of these equations gives insight into the single cell dynamics. On the other hand, the inclusion of diffusion and noisy conditions allows simulating the formation of heterocysts patterns in cyanobacteria strains. The time evolution of relevant component concentrations are calculated allowing for a comparison with experiments. Finally, we discuss the validity and the possible improvements of the model.
## Author Summary
Understanding the underlying mechanisms favoring the cooperative behavior exhibited by different cell types constitutes the first step to explain the emergence of specialized cells in the early life stages of multicellular systems. Multicellularity appeared on Earth some billion years ago, and cyanobacteria are one of the first organisms that developed this feature. In fact, being differentiation processes the cornerstone of multicellularity, cyanobacteria strands are paradigmatic examples of prokaryotic cellular differentiation and cooperative pattern formation. When a strand of cyanobacteria cells is deprived of combined nitrogen, some of the vegetative cells start differentiating into heterocysts, which are terminally differentiated nitrogen-fixing cells. However, not all vegetative cells differentiate as ( ) heterocysts loose i their photosynthetic capacity so they need vegetative cells around to be provided of a source of fixed carbon and ( ii ) cell division, i.e. reproduction, is only accomplished by vegetative cells. From such paradigmatic example it is clear that differentiation processes are the result of the interplay of complex regulatory networks acting inside the cell and external stimuli, from both the neighboring cells and the environment. Thus, in this article we present an integrative approach that combines the study of the internal regulatory processes, diffusion and noisy environments in order to capture the key mechanisms leading to the differentiation of vegetative cyanobacteria into heterocysts and the subsequent pattern formation.
## Introduction
One of the most striking and complex problems tackled by biology is the formation of multicellular organisms. Multicellular organisms rely on cell differentiation, a mechanism by which cells become more specialized. More precisely, differentiation processes lead to precise alterations in gene expression that result in changes in cell morphology and function. These processes are highly dynamical, directed by complex regulatory networks involving cell-to-cell interactions, and often triggered by external stimuli. As a result of the differentiation, the organism separates functions in different cell types establishing a rich cooperative pattern that increases its complexity and adaptability. Due to the large number of scales involved, ranging from protein binding to diffusion of specific elements throughout the organism, a correct mathematical modeling of differentiation processes and biological pattern formation demands an integrated approach combining tools from statistical mechanics and the theory of dynamical systems (see for instance [1,2]).
A landmark process of (prokaryotic) cellular differentiation and cooperative pattern formation is the heterocyst differentiation in cyanobacteria strands [3, 4]. Cyanobacteria are thought to be one of the first organisms in developing multicellularity some (2 -3) billion years ago [5]. These bacteria perform oxygenic photosynthesis releasing oxygen to the environment. On the other hand, nitrogenase, the enzyme that performs nitrogen fixation, is deactivated by oxygen so that nitrogen fixation cannot occur in its presence [6]. Cyanobacteria solve the incompatibility of incorporating both oxygenic photosynthesis and nitrogen fixation by separating these processes ( ) temporally, such as in the unicellular i Cyanothece sp. strain ATCC 51142 that presents photosynthetic activity during the day and fixes nitrogen during the night [7], or ( ii ) spatially, by the generation of non-photosynthetic nitrogen-fixing cells.
When provided of combined nitrogen (such as nitrate, nitrite, ammonium or urea), most cyanobacteria (with Anabaena PCC strain 7120 the most representative example) form long filaments of photosynthetic vegetative cells. However, in the absence of combined nitrogen (cN), a part of the vegetative cells differentiate into heterocysts, which are terminally differentiated nitrogen-fixing cells. By differentiating, heterocysts lose their photosysthetic capacity, so they need for an external source of fixed carbon [8, 9]. To this aim, each forming heterocyst sends a signal, by means of the diffusion of some particular chemical along the chain, to its neighboring cells to avoid their differentiation. A cooperative pattern is thus established: heterocysts provide cN to the cellular chain while non-differentiated cells remain supplying fixed carbon to the system. As a result, heterocysts appear interspersed within around 10 vegetative cells, depending on the species, forming a semiregular pattern that remains approximately constant in the chain regardless cell division [4]. This pattern forms one of the simplest and most primitive examples of multicellular organism as a product of the interdependence between heterocysts and vegetative cells. Interestingly, an isolated cyanobacterium does not differentiate but it first divides so that one of the descendants differentiates. This is necessary since ( ) a sole heterocyst would lack of a source of fixed i carbon and ( ii ) it would not reproduce since it is a terminally differenciated cell [10].
Some quantitative modeling has been done to explain the dynamical and equilibrium properties of heterocyst pattern formation. In references [11,12] Rutenberg and coworkers analyzed a model to explain heterocyst patterns by means of the study of combined nitrogen diffusion along a cyanobacterial strain. On the other hand, Gerdtzen et al. [13] modeled cyanobacterial filaments based on a time-discrete dynamical system exhibiting the main interactions between the most important proteins that take part in heterocyst formation.
In this work, we develop a simple mathematical model, by integrating the recent experimental results on the gene regulatory network of cyanobacteria into the theoretical machinery of system biology. To this aim, we take into account the fundamental genetic paths that underlie the differentiation of heterocysts considering the interactions that come into play in this process. Furthermore, our model connects the diffusion of combined nitrogen with the dynamical properties of the underlying genetic circuit of each cyanobacterium, linking pattern formation and maintenance. Our model shows that noise enables the
Figure 1. Main components and interactions involved in the reaction to nitrogen deprivation in cyanobacteria. Rectangular boxes represent genes ( ntcA , hetR and patS ) while rounded boxes and circles represent transcription factors (NtcA, HetR and PatS) and smaller molecules (2-OG and cN) respectively. Normal-tipped and flat-tipped arrows stand for up-regulating and down-regulating processes respectively. Dashed lines stand for indirect or imperfectly understood interactions. The accumulation of 2-OG enhances the DNA-binding activity of NtcA, which in turn up-regulates the transcription of ntcA and hetR . HetR activates ntcA and hetR (composing the central NtcA-HetR autoregulatory loop), the inhibitor patS and other genes that lead to nitrogen fixation and the morphological changes involved in heterocyst differentiation. 2-OG and cN levels are linked through the GS/GOGAT cycle (see Fig. 2).
development of the characteristic heterocysts patterns for a wide range of model parameters, revealing, in a quantitative way, that cyanobacteria strains may have evolved towards an efficient response mechanism to the noisy conditions that characterize the natural environment.
The work is structured as follows. First we present the main actors of the basic regulatory network and the different dynamical interactions that take place during the differentiation process. Then we develop a mathematical model for the unicellular reaction to nitrogen deprivation. Although a unicellular model does not suffice to understand heterocyst formation, we analyze the main features that arise from the dynamical behavior of the system to gain insight about cell dynamics under different external conditions. Finally, we round off the paper by introducing the spatial model consisting of a filament of cyanobacteria, each one characterized by the dynamical circuit developed previously, that interact by means of protein diffusion.
## Results
## Description of the main Genes and their basic Genetic Circuit
Heterocyst development begins with sensing combined-nitrogen (cN) limitation and ends with nitrogen fixation in mature heterocysts. This process is usually completed after a time of about 20 hours at 30 C [9]. ◦ In Fig. 1 we show a basic scheme of the genetic circuit including the most relevant elements and their respective interactions. Here we explain the main features of this genetic circuit.
The process is initiated with the accumulation of the enzyme 2-oxoglutarate (2-OG) as a consequence of nitrogen deprivation [9, 14]. 2-OG interacts with ammonium through the GS/GOGAT cycle [15-17] (see Fig. 2). Under cN starvation, the GS/GOGAT cycle breaks down, leading to the afore-mentioned
Figure 2. 2-OG and cN indirectly interact through the GS/GOGAT cycle. Glutamine is transformed into glutamate by means of 2-OG through the 2-OG amidotransferase (GOGAT) while cN converts glutamate into glutamine through the glutamine synthetase (GS). The importance of the cycle in heterocyst differentiation is twofold. From one side, it constitutes the early one-cell sensor to nitrogen starvation: the absence of cN breaks the cycle down and 2-OG starts to accumulate, whose action leads to the cascade of processes that provoke the differentiation (see Fig. 1). On the other hand, later during the differentiation, it process the cN created by the heterocysts decreasing the levels of 2-OG. The latter is crucial for the formation of the heterocyst pattern (see Fig. 3).
accumulation of 2-OG inside the cell [9]. In its turn, 2-OG stimulates the DNA-binding activity of NtcA, an important transcription factor for heterocyst development [16,18,19]. Furthermore, the transcription of the genes targeted by NtcA does not start in the absence of 2-OG [20, 21]. NtcA presents an autoregulatory activity [20, 22, 23] and indirectly activates the key gene that controls cell differentiation and pattern formation: hetR [24-26]. For its binding activity NtcA needs for homodimerization [27, 28]. Therefore, the accumulation of 2-OG is the factor that triggers differentiation. In agreement with this idea, artificial increased levels of 2-OG result in heterocyst development even in the presence of ammonium [14, 16, 29].
As already mentioned, the next step of the genetic circuit is the activation of hetR . Importantly, null mutants of hetR do not produce heterocysts whereas an overexpression of hetR leads to an increased heterocyst frequency [25, 30, 31]. The transcription of hetR is induced by NtcA through the action of an intermediate, nrrA [26]. The DNA-binding activity of HetR requires its homodimerization [32, 33]. Multiple transcription factors related to heterocyst formation are up-regulated by HetR, including hetR itself [33], ntcA [34] and patS [33].
The up-regulatory loop composed of NtcA and HetR is central for heterocyst differentiation [34-37]. However, the action of NtcA and HetR alone cannot account for pattern formation. Another transcription factor, PatS, inhibits the DNA-binding activity of HetR [8,33,38,39]. This inhibitory behavior is essential for the communication with adjacent cells and thus for achieving the observed patterns of vegetative cells and heterocysts in cyanobacteria strains (see Fig. 3). Furthermore, patS is strongly expressed in differentiating cells and mature heterocysts due to its upregulation by HetR [8]. A strain without patS develops multiple contiguous heterocysts (about a 30% of all cells as compared to the usual 10% in the wild-type strain). On the other hand, an overexpression of patS suppresses heterocyst differentiation [9]. In fact, the addition to the growth medium of a synthetic peptide composed of the last five residues (RGSGR) of PatS (PatS5) inhibits heterocyst development, suggesting that PatS5 may be a diffusive mature form of PatS that stops the differentiation of the rest of vegetative cells of the strain [38].
Figure 3. Schematic representation of the diffusion processes that sustain the heterocyst pattern. Heterocysts produce cN and PatS. cN diffuses along the strain where, due to the action of the GS/GOGAT cycle (see Fig. 2), decreases the levels of 2-OG breaking the autoregulatory core NtcA-HetR. On the other hand, early during the differentiation, PatS (or other derivative of it, see the text) diffuses along the strain inhibiting HetR. Both processes combined prevent the differentiation of the rest of vegetative cells and explain the formation of the pattern.
The last stages of heterocyst development cause the physiological changes of the cell aimed at creating an anaerobic environment that sustains nitrogen fixation. To this end, two new membrane layers are biosynthesized to decrease the entry of oxygen into the cell [40]. The morphogenesis of these two layers is controlled by two family of genes, hep and hgl , that are indirectly up-regulated by HetR [33]. After these morphological changes the genes in charge of nitrogen fixation, nif genes, are expressed. These genes encode, among others, the nitrogenase enzyme that performs nitrogen fixation.
The fixed nitrogen of the new heterocysts, when transferred to other cells of the strain, acts as an inhibitor of their differentiation together with the transferred PatS [41]. Thus, the diffusion of both PatS and cN from heterocysts along the strain plays a key role in its pattern maintenance (see Fig. 3). Finally, after the differentiation process is finished, a cooperative system between heterocysts and vegetative cells is established so to ensure the survival of the strain. In particular, heterocysts produce fixed nitrogen from N 2 of the atmosphere and they interchange this nitrogen with the oxygen derivatives produced by the vegetative cells.
## Regulatory equations
In this section we translate the genetic circuit described previously into a set of differential equations, for which we follow the derivation in [42-44]. Details are left to supplementary information (SI). To simplify notation, constants related to NtcA, HetR, PatS and cN are denoted with the letters a , r , s , and n respectively.
We start looking at the transcription of ntcA , which is regulated by HetR and NtcA itself. We assume that the probability that NtcA binds the promoter in the absence of 2-OG can be neglected. Taking into account that both HetR and NtcA dimerize to bind DNA we find:
$$v _ { a } = L _ { a } + \frac { v _ { a } ^ { a } \kappa _ { a } ^ { a } [ 2 { - } \text{OG} ] [ N t c A ] ^ { 2 } + v _ { a } ^ { r } \kappa _ { a } ^ { r } [ \text{Het} ] ^ { 2 } + v _ { a } ^ { a r } \kappa _ { a } ^ { a } \kappa _ { a } ^ { r } [ 2 { - } \text{OG} ] [ N t c A ] ^ { 2 } [ \text{Het} R ] ^ { 2 } } { ( 1 + \kappa _ { a } ^ { a } [ 2 { - } \text{OG} ] [ N t c A ] ^ { 2 } ) ( 1 + \kappa _ { a } ^ { r } [ \text{Het} R ] ^ { 2 } ) }, \quad ( 1 )$$
where v a measures the production rate of NtcA in units of concentration per time, v a a , v a r and v a ar are the rates when only NtcA, only HetR or both are bound to DNA respectively, and κ a ∗ are the inverse of the effective dissociation constants of the compounds that bind DNA. On the other hand, L a , the so-called leak term, measures the basal production of ntcA in the absence of regulation. Subscripts and superscripts identify the binding site and the transcription factor for which the constants are given respectively.
Similarly we can obtain the transcription velocity for HetR. We assume that hetR is regulated by NtcA by means of a usual Hill function although the real process presents an intermediate, nrrA . This
can be done if the mediator element(s) are not regulated through another factor of the genetic circuit under consideration and if they relax rapidly to their limiting values. On the other hand, PatS affects the autoregulatory loop of HetR. It has been suggested that PatS binds the binding site of HetR in the promoter of hetR preventing that HetR binds it [33]. These facts, along with the influence of the levels 2-OG, provide with an expression for the transcription velocity:
$$v _ { r } = L _ { r } + \frac { v _ { r } ^ { a } \kappa _ { r } ^ { a } [ 2 { - } { O } { G } ] [ N t c { A } ] ^ { 2 } ( 1 + \kappa _ { r } ^ { s } [ P a t { S } ] ) + v _ { r } ^ { r } \kappa _ { r } ^ { r } [ H e t { R } ] ^ { 2 } + v _ { r } ^ { a r } \kappa _ { r } ^ { a } \kappa _ { r } ^ { r } [ 2 { - } { O } { G } ] [ N t c { A } ] ^ { 2 } [ H e t { R } ] ^ { 2 } } { ( 1 + \kappa _ { r } ^ { a } [ 2 { - } { O } { G } ] [ N t c { A } ] ^ { 2 } ) \, ( 1 + \kappa _ { r } ^ { r } [ H e t { R } ] ^ { 2 } + \kappa _ { r } ^ { s } [ P a t { S } ] ) } \quad ( 2 )$$
HetR regulates most processes of the genetic circuit. It governs, among others, the transcription of ntcA , patS , hep , hgl and nif genes that lead to most of structural changes of the cell and to nitrogen fixation.
The inhibitor PatS is regulated by HetR, and we assume no other influence. This gives the simple transcription velocity:
$$v _ { s } = L _ { s } + \frac { v _ { s } ^ { r } \kappa _ { s } ^ { r } [ \text{HetR} ] ^ { 2 } } { 1 + \kappa _ { s } ^ { r } [ \text{HetR} ] ^ { 2 } }$$
Finally, we have to relate nitrogenase concentration [Ni] to that of combined Nitrogen [cN], both regulated by HetR and the levels of 2-OG [2-OG]. Let us begin by examining nitrogenase concentration, which is directly controlled by nif genes. Although this is not a direct process, we can assume, as we did for the NtcA-regulation by hetR , that nif genes are functionally governed by [HetR] following a typical Hill function. The nitrogenase production rate is given by:
$$\frac { d [ N i ] } { d t } = L _ { N i } + \frac { v _ { N i } ^ { r } \kappa _ { N i } ^ { r } [ \text{HetR} ] ^ { 2 } } { 1 + \kappa _ { N i } ^ { r } [ \text{HetR} ] ^ { 2 } } - \delta _ { N i } [ N i ].$$
where δ Ni represents the degradation rate of nitrogenase. We can effectively account for the lag introduced by intermediate processes not taken into account explicitly in the model by increasing the value of δ Ni so that [Ni] relaxes more slowly. Assuming that nitrogenase produces fixed nitrogen at a constant rate, we arrive at the equation that governs cN levels in cyanobacteria:
$$\frac { d [ c N ] } { d t } = L _ { n } ^ { \prime } + v _ { n } ^ { \prime } [ N \mathbf i ] - \delta _ { n } ^ { \prime } [ c N ],$$
where L ′ n represents the flux of cN from the exterior of the cell. Assuming that the levels of cN relax rapidly we solve Equation (5) for the steady state. Substituting in Equation (4) we find:
$$\frac { d [ \text{cN} ] } { d t } = L _ { n } + \frac { v _ { n } ^ { r } \kappa _ { n } ^ { r } [ \text{HetR} ] ^ { 2 } } { 1 + \kappa _ { n } ^ { r } [ \text{HetR} ] ^ { 2 } } - \delta _ { n } [ \text{cN} ]$$
where
$$\text{wnere} \\ L _ { n } = \frac { 1 } { \delta _ { n } } \left ( v _ { n } L _ { N i } + \delta _ { N i } L _ { n } ^ { \prime } \right ), \quad v _ { n } ^ { r } = \frac { v _ { n } ^ { \prime } } { \delta _ { n } ^ { \prime } } v _ { N i } ^ { r }, \quad \delta _ { n } = \delta _ { N i }, \quad \kappa _ { n } ^ { r } = \kappa _ { N i } ^ { r }.$$
To get a closed system of equations, we shall investigate the relation between cN and 2-OG. Both are related by means of the GS/GOGAT cycle (Fig. 2). Assuming the cycle is in equilibrium and reactions are grounded on the law of mass action, the following two conditions must be satisfied:
$$[ \text{glutamate} ] = \kappa _ { \leftarrow } [ \text{glutamine} ] [ 2 \text{-OG} ], \quad [ \text{glutamine} ] = \kappa _ { \rightarrow } [ \text{glutamate} ] [ \text{cN} ],$$
that lead to the relation:
$$[ 2 \text{-OG} ] = \frac { 1 } { \kappa _ { \leftarrow } \kappa _ { \rightarrow } [ c N ] }.$$
Table 1. Parameters for Eq. (11) that reproduce heterocyst formation under noisy conditions and pattern formation when PatS and cN diffuse along a strain of cyanobacteria.
| Constants | Constants | Constants | Constants |
|---------------|-----------------|----------------|----------------|
| l a = 0 . 2 | l r = 0 . 01 | l s = 0 . 0001 | l n = 0 |
| d a = 0 . 7 | d s = 0 . 05 | d n = 0 . 01 | β a a = 4 |
| β r a = 4 | β ar a = 8 | β a r = 1 | β r r = 1 |
| β ar r = 3 | β r s = 0 . 385 | β r n = 0 . 06 | γ a a = 3 |
| γ r a = 2 . 4 | γ r s = 1 . 2 | γ r n = 2 . 75 | γ r n = 2 . 75 |
However, this expression does not behave properly for small concentrations of cN, which are expected under cN deprivation: 2-OG levels would increase without limit. In fact, 2-OG production is controlled by some processes that are not considered in this work and so its value must be limited. We can effectively include such a limiting value by means of a translation on [cN] in Eq. (9)
$$[ 2 \text{-OG} ] = \frac { 1 } { \kappa _ { 2 \text{-OG} } + \kappa _ { \leftarrow } \kappa _ { \rightarrow } [ c N ] }$$
which reaches the maximum value [2-OG] max = 1 /κ 2-OG at [cN] = 0.
Finally we can introduce the differential equations governing cyanobacterial reaction to nitrogen deprivation. They represent the temporal variation of the most important factors of the genetic circuit, namely NtcA, HetR, PatS and cN. Using the production rates (26), (2), (3), (6) and introducing degradation rates constants, δ ∗ , we find:
$$\ n s t a n s, \delta _ { s }, \, \text{we find} \colon \\ \frac { d q _ { a } } { d \tau } & = l _ { a } + \frac { \beta _ { s } ^ { a } \gamma _ { a } ^ { a } q _ { a } ^ { 2 } + \beta _ { s } ^ { r } \gamma _ { a } ^ { r } q _ { r } ^ { 2 } ( 1 + q _ { n } ) + \beta _ { a } ^ { a r } \gamma _ { a } ^ { a } q _ { a } ^ { 2 } \gamma _ { a } ^ { r } q _ { r } ^ { 2 } } { ( 1 + q _ { n } + \gamma _ { a } ^ { a } q _ { a } ^ { 2 } ) ( 1 + \gamma _ { a } ^ { r } q _ { r } ^ { 2 } ) } - d _ { a } q _ { a }, \\ \frac { d q _ { r } } { d \tau } & = l _ { r } + \frac { \beta _ { r } ^ { a } q _ { a } ^ { 2 } ( 1 + q _ { s } ) + \beta _ { r } ^ { r } q _ { r } ^ { 2 } ( 1 + q _ { n } ) + \beta _ { r } ^ { a r } q _ { a } ^ { 2 } q _ { r } ^ { 2 } } { ( 1 + q _ { n } + q _ { a } ^ { 2 } ) ( 1 + q _ { s } + q _ { r } ^ { 2 } ) } - q _ { r }, \\ \frac { d q _ { s } } { d \tau } & = l _ { s } + \frac { \beta _ { s } ^ { r } \gamma _ { s } ^ { r } q _ { r } ^ { 2 } } { 1 + \gamma _ { a } ^ { r } q _ { a } ^ { 2 } } - d _ { s } q _ { s }, \\ \frac { d q _ { n } } { d \tau } & = l _ { n } + \frac { \beta _ { n } ^ { r } \gamma _ { n } ^ { r } q _ { r } ^ { 2 } } { 1 + \gamma _ { n } ^ { r } q _ { a } ^ { 2 } } - d _ { n } q _ { n }, \\ \text{we have introduced the adimensional variables} \colon$$
where we have introduced the adimensional variables:
$$q _ { a } = \underbrace { \sqrt { \frac { \kappa _ { r } ^ { a } } { \kappa _ { 2 } \text{-OG} } } } _ { \phi _ { a } } [ N t c A ], \quad q _ { r } = \underbrace { \sqrt { \kappa _ { r } ^ { r } } } _ { \phi _ { r } } [ \text{HetR} ], \quad q _ { s } = \underbrace { \kappa _ { r } ^ { s } } _ { \phi _ { s } } [ \text{PatS} ], \quad q _ { n } = \underbrace { \frac { \kappa _ { c } \kappa _ { \rightarrow } } { \kappa _ { 2 } \text{-OG} } } _ { \phi _ { n } } [ \text{cN} ], \quad \tau = \delta _ { r } t, \quad ( 1 2 )$$
and the constants
$$\text{samsung as} l _ { * } = \frac { L _ { * } \phi _ { * } } { \delta _ { r } }, \quad \beta _ { * } ^ { * } = \frac { v _ { * } ^ { * } \phi _ { * } } { \delta _ { r } }, \quad \gamma _ { * } ^ { * } = \frac { \kappa _ { * } ^ { * } } { \kappa _ { r } ^ { * } }, \quad d _ { * } = \frac { \delta _ { * } } { \delta _ { r } }.$$
Let us finally stress that this is a deterministic model for a single cyanobacterium. The study of the cyanobacterial strain is left to the final section. We advance that the main modification will be adding diffusion processes for the inhibitors PatS and cN through the chain. An important ingredient in pattern formation, noise, will be also added to the equations.
Figure 4. Adiabatic elimination of the fast variables q r and q a . Due to the fast dynamics that HetR and NtcA exhibit, we can approach the treatment of the system by adopting a point of view that follows the slower variables q s and q n . From this viewpoint, the time-evolution of the pair ( q s ( ) t , q n ( )) t is considered by assuming that q r and q a instantaneously relax to an equilibrium, which corresponds to a sink ( q ∗ r , q ∗ a ) for the fixed pair ( q s ( ), t q n ( )). t Depending on the region of the ( q , q s n )-plane, there are three fixed points (two sinks corresponding to the highest and the lowest concentrations respectively and a saddle in the middle) or one (a sink) for q r and q a (I and II). There are two one-sink regions that are separated from the two-sink region by saddle-node bifurcations (A-F). Sinks and saddles are represented by filled and unfilled circles respectively and arrows indicate the flow of the dynamics. We can then imagine the dynamics of q s and q n as evolving either in the bottom or in the top branch of I. In the two-sink region, both branches are plausible and the history of the dynamics determine the solution (hysteresis effect): a dynamics in a branch will continue in it until experiencing a bifurcation in the ( q , q r a ) plane (see Fig. 5 for examples).
## Unicellular dynamics
In this section we analyze the dynamical system (11) for a set of constants (Table 1) that exhibit both the dynamical and the structural properties of heterocyst differentiation. Following the usual procedure in the theory of dynamical systems, we study the basic properties of equations (11), such as fixed points and linear stability analysis, so to analyze the key features leading to heterocyst differentiation.
Taking into account the difference between the relaxation times of the constituents of the model, given by the inverses of d ∗ (see Table 1), we can think it is composed of two temporally separated systems: a rapid one, formed by HetR and NtcA, showing fast dynamics that relaxes to its steady state almost instantaneously and a slow one, composed of PatS and cN, whose evolution is dictated by the values of HetR and NctA in their instantaneous equilibrium. This corresponds to an adiabatic elimination technique [45] that helps in understanding the behavior of the system since it reduces the complexity of the dynamical system by splitting it into two simpler interdependent subsystems.
First, we look for the fixed points of the fast variables q a and q r for each pair of values of q s and q n
$$f _ { a } ( q _ { s }, q _ { n } ) = \frac { d q _ { a } } { d \tau } = 0, \quad f _ { r } ( q _ { s }, q _ { n } ) = \frac { d q _ { r } } { d \tau } = 0.$$
The numerical solution to this problem is sketched in Fig. 4. We find three different branches of solutions
## qcn) PHASE PORTRAIT
Figure 5. States of a cyanobacterium when subjected to different conditions of nitrogen and diffusion. When the cell is provided of cN ( l n = 0 03), there is only one stable fixed point (A) in the bottom . branch, which corresponds to a state in which the production of both HetR and PatS is minimum (vegetative-like state). When subjected to nitrogen deprivation ( l n = 0), there are two stable fixed points (B and C) each one in a different branch. The first point (B) is a vegetative-like state in which there exists an equilibrium between a small production of HetR, PatS and cN. The same kind of equilibrium is present in the second fixed point (C) but in this case the production of all TFs and cN is high (heterocyst-like steady state). When the cell is exposed to nitrogen stress its trajectory evolves from A to the steady state B and thus it remains vegetative. Assuming some diffusion of cN and PatS from the cell ( l s = -0 2 and . l n = -0 002), the only stable state (D) corresponds to a heterocyst-like . state with high levels of production of HetR, cN and PatS, being the latter transported to the surroundings of the cell.
that coexist in some regions. The fixed points on the lower and upper branches are always stable (blue region in Fig. 4I) and those lying on the middle branch (red region) are saddles. Transitions between the regions with one and three fixed points correspond to saddle-node bifurcations in which the middle branch of solutions coalesce with the lower and the upper one respectively. The basins of attraction of both stable fixed points are separated by the stable manifold of the saddle point (Fig. 4 II).
In the bistable region the system behaves as a switch that can be either OFF in a vegetative-like state (lower branch, with a small production of HetR and NtcA) or ON in a heterocyst-like one (upper branch, with a high production of HetR and NtcA). A sufficiently large perturbation may result in the system crossing the manifold of the saddle and falling into the other stable branch of solutions. The distance between the saddle and the nodes determines the size of the perturbation needed to activate or inactivate the system.
Once (14) is solved, we can apply the solution to calculate the effective field sensed by the ( q , q s n ) pair. In the regions showing bistability the field takes two very different forms, one corresponding to the values of the lower branch and another corresponding to those of the upper one (Fig. 5). We expect a hysteresis effect : if initially the dynamics lies on a particular branch it will remain on it unless a fluctuation or a bifurcation makes the system jump to the other branch.
When supplied of cN (Fig. 5A) we find only one stable fixed point that corresponds to a vegetative-like state, since it belongs to the lower branch. The upper branch is completely unstable: all the dynamics lying on it will fall down to the lower branch and eventually be attracted to the vegetative-like sink. The steady state is very robust against perturbations since it is far from the bifurcation region and there is a significant distance to the saddle in the q r -q a plane.
By reducing the flow of cN from the exterior of the cell ( l n = 0) we find that a stable fixed point appears in the upper branch, a heterocyst-like state, and the vegetative-like state gets closer to the bifurcation region becoming more susceptible to perturbations that can make the system jump to the upper branch. When cN is eliminated from the media the cyanobacterium would evolve from state A to state B in the lower branch until a perturbation pushes it to the upper branch, eventually becoming an heterocyst due to the field acting on that branch (Fig. 5B and C).
Diffusion protects cells in the neighborhood of the newly formed heterocyst to initiate the differentiation: as heterocysts are producers of cN and PatS, the vegetative fixed point of the cells in the neighborhood will move towards an A-like state, thus becoming more stable to perturbations. The heterocyst-like fixed point also becomes more stable due to diffusion, since its production of inhibitors is distributed among other cells (see Fig. 5D).
## Strains of cyanobacteria. Heterocyst patterns
In the previous section, we introduced a single cell model for the cyanobacterial reaction to nitrogenlimiting conditions. There we have shown that, for a specific range of parameters, the model exhibits features that would lead to heterocyst development under noisy conditions. Nevertheless, the model should be extended to cyanobacteria chains to account for heterocyst development since, as said above, isolated cyanobacteria do not become heterocysts by themselves; the action of the chain is needed to generate heterocysts.
In this section, we extend the previous results and consider a chain of vegetative cells facing nitrogen deprivation. The main modification is the introduction of diffusion of PatS and cN along the cyanobacteria chain. For this purpose we add to Equation (12) the discrete version of the diffusion equation:
$$\frac { d C _ { i } } { d t } = D _ { C } \left ( C _ { i + 1 } + C _ { i - 1 } - 2 C _ { i } \right ).$$
where D C is called the diffusion constant of the element C. Now, it is straightforward to introduce PatS and cN diffusion into the equations. The dynamics of cell i is characterized by the following set of equations:
$$\text{and} \, \text{in} \, \text{into} \, \text{to} \, \text{columns} \, \text{ } \text{ } \text{$1$ the dynamics of cell } \text{ is characterized by the following set of} \\ \text{quad} \, \text{$d_{i,a}$} & = l _ { a } + \frac { \beta ^ { a } \gamma ^ { a } _ { a } a ^ { 2 } _ { a } + \beta ^ { r } \gamma ^ { r } _ { a } a ^ { 2 } _ { a } \gamma ^ { 2 } _ { a } \gamma ^ { 2 } _ { a } } { ( 1 + q _ { n } + \gamma ^ { a } _ { a } a ^ { 2 } _ { a } ) ( 1 + \gamma ^ { r } _ { a } a ^ { 2 } _ { a } ) _ { a } } - d _ { a } q _ { i, a } + G _ { i, a } ( t ), \\ & \quad \, \text{$d_{i,r}$} & = l _ { r } + \frac { \beta ^ { a } _ { r } q _ { a } ^ { 2 } ( 1 + q _ { s } ) + \beta ^ { r } _ { r } q _ { r } ^ { 2 } ( 1 + q _ { n } ) + \beta ^ { a r } _ { r } q _ { a } ^ { 2 } q _ { r } ^ { 2 } } { ( 1 + q _ { n } + q _ { a } ^ { 2 } ) ( 1 + q _ { s } + q _ { r } ^ { 2 } ) } - q _ { i, r } + G _ { i, r } ( t ), \\ & \quad \, \text{$d_{i,s}$} & = l _ { s } + \frac { \beta ^ { r } _ { s } \gamma ^ { r } _ { s } q _ { r } ^ { 2 } } { 1 + K ^ { r } _ { r } q _ { a } ^ { 2 } _ { r } } - d _ { s } q _ { i, s } + D _ { s } \, ( q _ { i + 1, s } + q _ { i - 1, s } - 2 q _ { i, s } ) + G _ { i, s } ( t ), \\ & \quad \, \text{$d_{i,n}$} & = l _ { n } + \frac { \beta ^ { r } _ { n } \gamma ^ { r } _ { n } q _ { r } ^ { 2 } } { 1 + \gamma ^ { r } _ { n } q _ { r } ^ { 2 } _ { r } } - d _ { n } q _ { i, n } + D _ { n } \, ( q _ { i + 1, n } + q _ { i - 1, n } - 2 q _ { i, n } ) + G _ { i, n } ( t ), \\ \text{which constitutes the model for a cyanobacteria strain. }
. - + < > </ | " ( , ) [ ]$$
which constitutes the model for a cyanobacteria strain. To account for environment variability we add white noise, G i, ∗ ( ), t of the same amplitude, 〈 G i, ∗ ( ) t G i, ∗ ( t ′ ) 〉 = ξδ t ( -t ′ ), for all the components of the system. Based on these equations, we investigate the conditions that lead to a heterocyst pattern. It is easy to notice that they correspond to an activator-inhibitor system of cells coupled in a reaction-diffusion scheme [46]. This kind of systems produce regular pattern formation [47, 48]. Turing (linear stability) analysis of equations (16) (see SI) provides insight on the periodicity of patterns. It is interesting to show that the minimum periodicity observed in such analysis is larger than 1, which means that a single bacteria is unable to differentiate.
We performed the direct integration of equations (16) for chains of 200 cyanobacteria. We used a Runge-Kutta method for the numerical integration of stochastic differential equations (see Methods) [49]. The level of noise that best reproduces heterocyst pattern is ξ = 0 001 for the set of parameters of Table .
Figure 6. Time evolution of the main components of the differentiation in heterocysts (green) and vegetative cells (blue). Averages along the strain are also presented (black). Heterocysts, due to the early diffusion, evolve toward steady states of the type D of Fig 5 characterized by high levels of HetR and NtcA while vegetative cells present very low concentrations of them (A and B). The levels of PatS and cN in vegetative cells depend on their distance to close heterocysts: C and D show the concentrations of PatS and cN in a heterocyst and in its first two neihbouring vegetative cells, which clearly highlight the effect of diffusion along the strain.
1. Importantly, isolated cells do not initiate differentiation with this level of noise, in agreement with the results from the linear stability analysis. On the other hand, diffusion constants have been set to D s = 0 1 . and D n = 0 2. . Heterocysts patterns develop for different levels of noise and diffusion constants, but the model parameters, which characterize cell response to nitrogen deprivation, should change accordingly. This correlation between noise, diffusion and model parameters supports the idea that cyanobacteria have evolved towards the better response to the normal levels of noise in their environment.
In Fig. 6 we show the dynamics that the 4 variables exhibit when the chain is deprived of cN. We observe that the chain relaxes to the constant protein levels of the vegetative-like state we showed in the previous chapter. Then, due to the coupled action of noise and diffusion, some cells start to differentiate. As new forming heterocysts appear, their production and exportation of inhibitors to the surrounding cells make the latter more stable to perturbations stopping their differentiation. The model reproduces very well the initial peak that both NtcA and HetR present experimentally [33,50]. PatS increases more slowly to its steady value reducing the levels of NtcA and HetR and, finally, cN is generated by heterocysts stabilizing the pattern.
Fig. 7 shows the evolution of the profile for a 200 cells chain of cyanobacteria. We observe that heterocysts progressively appear in those regions in which others heterocysts do not have effect ( i.e. those vegetative cells that are not supplied of sufficient cN and PatS). Finally, a semiregular pattern is generated. PatS and cN diffuse along the strain exhibiting smooth variations between vegetative cells and heterocysts, while HetR and NtcA present very abrupt variations between cell types.
Figure 7. Time-evolution of the pattern of heterocysts (A) and of the probability distribution of the distance between consecutive heterocysts (B). Green and blue curves represent the concentration profiles of HetR and PatS (NtcA and cN are not presented since their behavior along the strain is comparable to the one of HetR and PatS, see Fig. 6 to see the similarities). Small perturbations along the strain of vegetative cells (initially in the steady state B of Fig. 5) are amplified due to diffusion processes in a demonstration of Turing's theory [47]. New heterocysts appear in regions that are not dominated by the action of other heterocysts. Finally, the competition between nearby differentiating cells ceases the differentiation of some of them, as observed in B: consecutive heterocysts, which are created by strong perturbations, finally dissapear due to the aforementioned competition. The final pattern presents localized levels of HetR (heterocysts) and a diffusive-like behavior of PatS, as it was expected.
Finally, in Fig. 7.B we show the time-evolution of the histogram for the distance between two consecutive heterocysts. It should be stressed that although initially some close heterocysts appear, they are eliminated by the non-linear action of the system during the differentiation process. Close heterocysts compete for the same region of action (the same vegetative cells that consume their PatS and cN) and then they cannot reach the optimal heterocyst-like state, which is stable to slight perturbations. Finally, one of them falls down from the upper branch becoming a vegetative cell. This behavior is typically observed experimentally [4,9]. The final histogram can be nicely fitted by a Γ-distribution.
## Discussion
The study of cell differentiation and its underlying mechanisms constitutes one of most intriguing problems in biology. This phenomena is the basis for multicellular organism and pattern formation. The approach we have chosen in this paper is to deal with a simple system, the heterocyst formation in cyanobacteria strains, yet complex enough to capture the main ingredients of some of the mechanism for cell differentiation and pattern formation under external driving. The knowledge of the basic regulatory genes and their corresponding interactions allows for a detailed description of cell dynamics. We have proposed the dynamical equations for these regulatory genes based in the statistical mechanics of the regulation process. This allow us to obtain a detailed description of the continuous time dynamics of the main regulatory protein, in contrast to other discrete approximations based in boolean dynamics [13]. This kind of analysis have been proved successfully in other time dependent phenomena in the cyanobacterial word like circadian cycles [51]. The analysis of the 'unicellular dynamics' has revealed that the two cellular stable states, vegetative and heterocyst cells, appear as attractors of the non-linear dynamics of the regulation equations. However, the study of many coupled cells is needed as cyanobacteria do not
differenciate when isolated.
The model is thus completed by coupling a number of cells in a one-dimensional array so that combined nitrogen and PatS can diffuse along the chain of cells. One importaint ingredient affecting the dynamical behavior of the chain is noise, which appears to have a key role in the transition from the initial chain of vegetative cells to the steady state in which heterocysts coexist with vegetative cyanobacteria. Thus, the appearance of differentiation is, in our model, a pure stochastic event. The cooperative character of the strain is clear from the amount of noise needed to start the differenciation process which appears significantly smaller than that needed in isolated cells. The source of noise as well as its biological consequences is, nowadays a current topic of research [52]. In fact, at its initial state, differentiation of cells appear randomly along the strain, but shortly after its onset a characteristic distribution of heterocyst emerges. This distribution can be compared with the experimental one with a fairly agreement [39].
Although the model presented here integrates both internal cell dynamics and the coupling between cells via diffusion, there exist other ingredients that can also be incorporated. One issue that have not been considered in this work is the possibility of replication of vegetative cells. This effect has been taken into account in [12]. Although this improvement is relevant, it only affects, in our approach, to the mean separation between heterocysts, by opening a gap in the Γ-function shaped of Fig. 6B and thus approaching better to the experimental distribution.
Other improvements to the approach presented here will come from the availability of more experimental data. Unlike other approaches [11] where comparison is done (globally) with heterocyst distributions, our work would allow for a qualitative comparison of each component involved in the differentiation (see Fig. 6). Unfortunately, experimental data are not enough to make a detailed fit and then extract reliable parameters. The availability of such data is extremely important both for having better set of model parameters and to validate new models. A complete understanding of the mechanism that derive in phenotypic differentiation is the first step for a modular comprehension of the whole cell [53].
## Methods
To reproduce the dynamics of Eq. (16) we make use of the integration scheme proposed in [49]. Eq. (16) is a set of stochastic differential equations (SDE) so its numerical integration requires generating a statistical representative trajectory for a discrete set of time-values. A SDE of the form
$$\dot { x } & = f ( x ) + G ( t ),$$
where G t ( ) is a Gaussian white noise with
$$\langle G ( t ) \rangle = 0, \text{ and, } \langle G ( t ) G ( t ^ { \prime } ) \rangle = \xi \delta ( t ^ { \prime } - t ),$$
can be integrated through a Runge-Kutta integration algorithm by adding a particular Gaussian signal at each stage of the scheme. This algorithm coincides with the usual Runge-Kutta scheme for ξ = 0. In this work we have employed a 3 O S 4 2 G algorithm, which is correct up to 3 th order, is developed in 4 stages and uses 2 independent Gaussian random variables.
## Acknowledgments
We acknowledge financial support from the Spanish MINECO under projects FIS2011-25167 and FIS201238266-C02-01. J.G.G. is supported by the Spanish MINECO through the Ramon y Cajal program. We are grateful to S. Ares and J. Mu˜oz-Garc´ ıa for sharing ideas and useful discussions. n We also acknowledge the 'Genetic Regulation and Physiology of Cyanobacteria' group at University of Zaragoza for sharing insight on the genetic of heterocyst formation.
## References
- 1. Koch A, Meinhardt H (1994) Biological pattern formation: from basic mechanisms to complex structures. Rev Mod Phys 66: 1481-1511.
- 2. S¨ uel GM, Garcia-Ojalvo J, Liberman LM, Elowitz MB (2006) An excitable gene regulatory circuit induces transient cellular differentiation. Nature 440: 545-550.
- 3. Wolk C (1998) Heterocyst formation. Annu Rev Genet 30: 59-78.
- 4. Flores E, Herrero A (2010) Compartmentalized function through cell differentiation in filamentous cyanobacteria. Nature Rev 8: 39-50.
- 5. Schirrmeister B, Antonelli A, Bagheri H (2011) The origin of multicellularity in cyanobacteria. BMC Evol Biol 11: 45.
- 6. Shi Y, Zhao W, Zhang W, Ye Z, Zhao J (2006) Regulation of intracellular free calcium concentration during heterocyst differenciation by HetR and NtcA in Anabaena sp. PCC7120. Proc Nat Acad Sci (USA) 103: 11334-11339.
- 7. Toepel J, Welsh E, Summerfield T, Pakrasi H, Sherman L (2008) Differential Transcriptional Analysis of the Cyanobacterium Cyanothece sp. Strain ATCC 51142 during Light-Dark and ContinuousLight Growth. J Bacteriol 190: 3904-3913.
- 8. Yoon H (1998) Heterocyst Pattern Formation Controlled by a Diffusible Peptide. Science 282: 935-938.
- 9. Zhang CC, Laurent S, Sakr S, Peng L, B´ edu S (2006) Heterocyst differentiation and pattern formation in cyanobacteria: a chorus of signals. Mol microbiol 59: 367-375.
- 10. Kumar K, Mella-Herrera R, Golden J (2010) Cyanobacterial heterocists. Cold Spring Harb Perspect Biol 2: a000315.
- 11. Allard J, Hill A, Rutenberg A (2007) Heterocyst patterns without patterning proteins in cyanobacterial filaments. Dev Biol 312: 427-434.
- 12. Brown A, Rutenberg A (2012) Reconciling cyanobacterial fixed-nitrogen distributions and transport experiments with quantitative modelling. Phys Biol 9: 016007.
- 13. Gedtzen Z, Salgado J, Osses A, Asenjo J, Rapaport I, et al. (2009) Modeling heterocyst pattern formation in cyanobacteria. BMC Bioinf 10: S16.
- 14. Laurent S, Chen H, B´ edu S, Ziarelli F, Peng L, et al. (2005) Nonmetabolizable analogue of 2oxoglutarate elicits heterocyst differentiation under repressive conditions in Anabaena sp. PCC 7120. P Natl Acad Sci USA 102: 9907-9912.
- 15. Muro-Pastor MI, Reyes JC, Florencio FJ (2001) Cyanobacteria perceive nitrogen status by sensing intracellular 2-oxoglutarate levels. J Biol Chem 276: 38320-38328.
- 16. V´ azquez-Berm´dez MF, Herrero A, Flores E (2003) Carbon supply and 2-oxoglutarate effects on u expression of nitrate reductase and nitrogen-regulated genes in Synechococcus sp. strain PCC 7942. FEMS Microbiol Lett 221: 155-159.
- 17. Muro-Pastor MI, Reyes JC, Florencio FJ (2005) Ammonium assimilation in cyanobacteria. Photosynth Res 83: 135-150.
- 18. Wei TF, Ramasubramanian TS, Golden JW (1994) Anabaena sp. strain PCC 7120 ntcA gene required for growth on nitrate and heterocyst development. J Bacteriol 176: 4473-4482.
- 19. Fr´ ıas JE, Flores E, Herrero A (1994) Requirement of the regulatory protein NtcA for the expression of nitrogen assimilation and heterocyst development genes in the cyanobacterium Anabaena sp. PCC 7120. Mol microbiol 14(4): 823-832.
- 20. V´ azquez-Berm´dez MF, Herrero A, Flores E (2002) 2-Oxoglutarate increases the binding affinity u of the NtcA (nitrogen control) transcription factor for the Synechococcus glnA promoter. FEBS lett 512: 71-74.
- 21. Tanigawa R, Shirokane M, Maeda Si Si, Omata T, Tanaka K, et al. (2002) Transcriptional activation of NtcA-dependent promoters of Synechococcus sp. PCC 7942 by 2-oxoglutarate in vitro. P Natl Acad Sci USA 99: 4251-4255.
- 22. Ramasubramanian TS, Wei TF, Oldham aK, Golden JW (1996) Transcription of the Anabaena sp. strain PCC 7120 ntcA gene: multiple transcripts and NtcA binding. J Bacteriol 178: 922-926.
- 23. Ramasubramanian TS, Wei TF, Golden JW (1994) Two Anabaena sp. strain PCC 7120 DNAbinding factors interact with vegetative cell- and heterocyst-specific genes. J Bacteriol 176: 12141223.
- 24. Buikema WJ, Haselkorn R (1991) Isolation and complementation of nitrogen fixation mutants of the cyanobacterium Anabaena sp. strain PCC 7120. J Bacteriol 173: 1879-1885.
- 25. Buikema WJ, Haselkorn R (1991) Characterization of a gene controlling heterocyst differentiation in the cyanobacterium Anabaena 7120. Gene Dev 5: 321-330.
- 26. Ehira S, Ohmori M (2006) NrrA directly regulates expression of hetR during heterocyst differentiation in the cyanobacterium Anabaena sp. strain PCC 7120. J Bacteriol 188: 8520-8525.
- 27. Wis´ en S, Jiang F, Bergman B, Mannervik B (1999) Expression and purification of the transcription factor NtcA from the cyanobacterium Anabaena PCC 7120. Protein expres purif 17: 351-357.
- 28. Alfonso M, Kirilovsky D (2001) Redox Control of ntcA Gene Expression in Synechocystis sp . PCC 6803 . Nitrogen Availability and NtcA Protein 1. Plant Physiol 125: 969-981.
- 29. Li JH (2003) An increase in the level of 2-oxoglutarate promotes heterocyst development in the cyanobacterium Anabaena sp. strain PCC 7120. Microbiology 149: 3257-3263.
- 30. Buikema WJ, Haselkorn R (2001) Expression of the Anabaena hetR gene from a copper-regulated promoter leads to heterocyst differentiation under repressing conditions. P Natl Acad Sci USA 98: 2729-2734.
- 31. Khudyakov IY, Golden JW (2004) Different functions of HetR, a master regulator of heterocyst differentiation in Anabaena sp. PCC 7120, can be separated by mutation. P Natl Acad Sci USA 101: 16040-16045.
- 32. Zhou R, Wei X, Jiang N, Li H, Dong Y, et al. (1998) Evidence that HetR protein is an unusual serine-type protease. P Natl Acad Sci USA 95: 4959-4963.
- 33. Huang X, Dong Y, Zhao J (2004) HetR homodimer is a DNA-binding protein required for heterocyst differentiation, and the DNA-binding activity is inhibited by PatS. P Natl Acad Sci USA 101: 4848-4853.
- 34. Muro-Pastor AM, Valladares A, Flores E, Herrero A (2002) Mutual dependence of the expression of the cell differentiation regulatory protein HetR and the global nitrogen regulator NtcA during heterocyst development. Mol microbiol 44: 1377-1385.
- 35. Valladares A, Muro-pastor AM, Herrero A, Flores E (2004) The NtcA-Dependent P 1 Promoter Is Utilized for glnA Expression in N 2 -Fixing Heterocysts of Anabaena sp . Strain PCC 7120. J Bacteriol 186: 7337-7343.
- 36. Fiedler G, Muro-Pastor A (2001) NtcA-Dependent Expression of the devBCAOperon, Encoding a Heterocyst-Specific ATP-Binding Cassette Transporter in Anabaena spp. J Bacteriol 183: 37953799.
- 37. Hebbar PB, Curtis SE (2000) Characterization of devH, a gene encoding a putative DNA binding protein required for heterocyst function in Anabaena sp. strain PCC 7120. J Bacteriol 182: 35723581.
- 38. Wu X, Liu D, Lee MH, James W, Golden JW (2004) patS Minigenes Inhibit Heterocyst Development of Anabaena sp . Strain PCC 7120 patS Minigenes Inhibit Heterocyst Development of Anabaena sp . Strain PCC 7120. J Bacteriol 186: 6422-6429.
- 39. Yoon Hs, Golden JW (2001) PatS and Products of Nitrogen Fixation Control Heterocyst Pattern PatS and Products of Nitrogen Fixation Control Heterocyst Pattern. J Bacteriol 183: 2605-2613.
- 40. Fay P (1992) Oxygen relations of nitrogen fixation in cyanobacteria. Microbiol Rev 56: 340-373.
- 41. Wolk CP, Austin SAMM, Bortins J, Galonsky A (1974) Autoradiographic localization of 13 N after fixation of 13 N-labeled nitrogen gas by a heterocyst-forming blue-gree alga. J Cell Biol 61: 440-453.
- 42. Buchler NE, Gerland U, Hwa T (2003) On schemes of combinatorial transcription logic. Proc Natl Acad Sci 100: 5136- 5141.
- 43. Bintu L, Buchler N, Garcia H, Gerland U, Hwa T, et al. (2005) Transcriptional regulation by the numbers: models. Current opinion in genetics & development 15: 116-124.
- 44. Phillips R, Kondev J, Theriot J, Garcia H (2012) Physical Biology of the Cell. New York: Garland Science, second edition.
- 45. Manneville P (1990) Dissipative structures and weak turbulence. London: Academic Press.
- 46. Morelli LG, Uriu K, Ares S, Oates AC (2013) Computational Approaches to Developmental Patterning. Science 336: 187-191.
- 47. Turing AM (1952) The Chemical Basis of Morphogenesis. Philos Trans R Soc London Ser B 237: 37-62.
- 48. Meinhardt H (1982) Models of Biological Pattern Formation. London: Academic Press.
- 49. Greenside, HS and Helfand, E (1981) Numerical-Integration of Stochastic Differential Equations 2. Bell System Technical Journal 60: 1927-1940.
- 50. Jang J, Shi L, Tan H, Janicki A, Zhang CC (2009) Mutual Regulation of ntcA and hetR during Heterocyst Differentiation Requires Two Similar PP2C-Type Protein Phosphatases, PrpJ1 and PrpJ2, in Anabaena sp Strain PCC 7120. J Bacteriol 191: 6059-6066.
- 51. Teng SW, Mukherji S, Moffitt JR, de Buyl S, OShea EK (2013) Robust circadian oscillations in growing cyanobacteria require transcriptional feedback. Science 340: 737-740.
- 52. Suel GM, Kulkarni RP, Dworkin J, Garcia-Ojalvo J, Elowitz MB (2007) Tunability and noise dependence in differentiation dynamics. Science 315: 1716-1719.
- 53. Karr J, Sanghvi J, Macklin D, Gutschow M, Jacobs J, et al. (2012) A whole-cell computational model predicts phenotype from genotype. Cell 150: 389 - 401.
## Supporting Information
## Regulatory equations: a statistical mechanics approach
To derive the equations regulating transcription processes during heterocyst differentiation, we follow an approach similar to the detailed in [42-44]. This is a thermodynamic approach to transcription in which binding sites are considered two-states systems, either empty or containing a binding protein. The probability that a given transcription factor (TF) is bound to its binding site is given by the Arrhenius formula
$$p _ { T F } = \frac { [ T F ] K _ { T F } } { 1 + [ T F ] K _ { T F } } = \frac { q _ { T F } } { 1 + q _ { T F } }$$
where [TF] is the concentration of the TF, K TF is the inverse of the effective dissociation constant, which represents the concentration of half-maximal occupation, and q TF = [TF] K TF is called the binding affinity. The denominator of Eq. (18) is nothing but the canonical partition function of the promoter Z = 1 + [TF] K TF , representing the Boltzmann-weighted sum of possible states of the binding site. Transcription starts with the binding of RNAp, which in the absence of interactions with TFs follows the probability law of Eq. (18):
$$p _ { R N A p } = \frac { q _ { R N A p } } { 1 + q _ { R N A p } }$$
Let us examine the case in which RNAp interacts with a TF within the promoter. In this case the partition function is
$$Z _ { R N A p, T F } = 1 + [ T F ] K _ { T F } + [ R N A p ] K _ { R N A p } + [ R N A p ] [ T F ] K _ { R N A p, T F }$$
with K RNAp,TF the inverse dissociation constant of the complex RNAp&TF that can be higher than K TF K RNAp if the interaction of the two proteins within the promoter is attractive or smaller if the interaction is repulsive. In the first case we say that the TF is an inhibitor while in the second case we say that the TF is an activator. We are interested in the probability that RNAp is bound to its binding site. This probability is
$$p _ { R N A P } & = \frac { [ R N A p ] K _ { R N A p } + [ R N A p ] [ T F ] K _ { R N A p, T F } } { 1 + [ T F ] K _ { T F } + [ R N A p ] K _ { R N A p } + [ R N A p ] [ T F ] K _ { R N A p, T F } } = \\ & = \frac { q _ { R N A p } \left ( 1 + q _ { T F } \omega _ { R N A p, T F } \right ) } { 1 + q _ { T F } + q _ { R N A p } \left ( 1 + q _ { T F } \omega _ { R N A p, T F } \right ) },$$
where we have used the definitions q TF = [TF] K TF , q RNAp = [RNAp] K RNAp and ω RNAp,TF = K RNAp,TF / K ( TF K RNAp ). Now, assuming that [RNAp] does not vary during the transcription process (or that it is not the limiting factor of transcription) and that transcription (which we remind is nothing but the production of mRNA) takes place at a given velocity ν when RNAp is bound to the promoter, we can find the relation between q TF and the transcription velocity v mRNA :
$$v _ { \text{mRNA} } = L ^ { A } + v _ { \text{TF} } \frac { q _ { \text{TF} } \kappa _ { \text{TF} } ^ { A } } { 1 + q _ { \text{TF} } \kappa _ { \text{TF} } ^ { A } },$$
for the case in which the TF is an activator ( ω RNAp,TF > 1) or
$$v _ { \text{mRNA} } = L ^ { I } + v _ { \text{TF} } \frac { 1 } { 1 + q _ { \text{TF} } \kappa ^ { I } },$$
in the case of an inhibitor ( ω RNAp,TF < 1). The remaining constants are
$$\ n i m o w o w \, ( \omega _ { N A p, T } & \prec 1 ). \, \text{ if remaining consumes are} \\ \nu _ { T } & = \nu \frac { q _ { R N A p } | \omega _ { R N A p, T } - 1 | } { ( 1 + q _ { R N A p } ) ( 1 + q _ { R N A p } \omega _ { R N A p, T } ) }, \\ \kappa ^ { A } & = \frac { 1 } { \kappa ^ { 1 } } & = \frac { 1 + \omega _ { R N A p, T F } q _ { R N A p } } { 1 + q _ { R N A p } }, \\ L ^ { A } _ { m R N A } & = \nu \frac { q _ { R N A p } } { 1 + q _ { R N A p } }, \\ L ^ { 1 } _ { m R N A } & = \nu \frac { q _ { R N A p } \omega _ { R N A p, T F } } { 1 + q _ { R N A p } \omega _ { R N A p, T F } }. \\ \text{s are of the form of the Michaelis-Menten equations of reaction kinetics with a leak term.}$$
These equations are of the form of the Michaelis-Menten equations of reaction kinetics with a leak term, represented by L A or L I depending on the case, that stands for the production of mRNA in the absence of regulation. This example shows the main features of the statistical mechanics approach to transcription. More complex transcription processes can be dealt with in a similar way by computing their corresponding partition function and counting the RNAp-active states.
For all processes in the article we assume that TFs do not interact within the promoter but rather that they cooperatively affect the velocity at which transcription takes place. This is the simplest way in which we can consider the interaction between different TFs and, on the other hand, it is rich enough to represent the main features of the transcription processes we need to account for.
Let us sketch, for instance, the regulation of ntcA in heterocyst development (see main text). ntcA is regulated partly by NtcA (in its dimer configuration) and 2-OG and also by HetR. The partition function in this case is:
$$\begin{array} {$$
Both NtcA and HetR acts as activators, and we finally arrive to the transcription velocity
$$v _ { a } = L _ { a } + \frac { v _ { a } ^ {$$
where v a a , v r a and v ar a represent the effective transcription velocity when NtcA, HetR or both are bound to DNA respectively. The constants κ are obtained from K by eliminating q RNAp from the equations following a procedure similar to that of Eq. (22) and can be thought as the inverse of effective dissociation constants associated with the binding of the different compounds.
Finally, we consider translation (the process by which mRNA is transformed into the corresponding protein) is produced at a constant rate η per mole of mRNA. Therefore, the concentration of a given TF is given by:
$$\frac { d [ \text{TF} ] } { d t } = \eta _ { \$$
with δ TF the inverse of the mean lifetime of the TF. On the other hand, the dynamics for concentration of mRNA is
$$\frac { d [ m R N A _ { T F } ] } { d t } = v$$
with δ mRNA TF the inverse of the mean lifetime of the mRNA. Usually, δ mRNA TF >> δ TF and, as we are interested in [TF], we can assume that from the viewpoint of the characteristic dynamics of the TF, [mRNA] relaxes instantaneously to its equilibrium value:
$$[ m R N A _ { T F } ] = \frac { v _ { m R N A$$
so
$$\frac { d [ \text{TF} ] } { d t } = \eta \frac { v _ { \text{mRNA} _ { \text{TF} } } } { \delta _ { \text{mRNA} _ { \text{TF} } } } - \delta _ { \text{TF} } [ \text{TF} ].$$
In an abuse of notation, we redefine the variables v · ∗ and l ∗ as v η /δ · ∗ ∗ mRNA ∗ and v η /δ · ∗ ∗ mRNA ∗ respectively to express the effective constants in Eq. (30). For instance, we find
$$\frac { d [ N t c A ] } { d t } = L _ { a } + \frac { v _ { a } ^ { a } \kappa _ { a } ^ { a } [ 2 \text{-} \text{-} \text{-} \text{-} } { } \frac { ( 1 + \kappa _ { a } ^ { a } [ 2 \text{-} \text{-} \text{-} } { } ^ { 2 } + v _ { a } ^ { r } \kappa _ { a } ^ { r } [ \text{Het} ] ^ { 2 } + v _ { a } ^ { a r } \kappa _ { a } ^ { a } \kappa _ { a } ^ { r } [ 2 \text{-} \text{-} \text{-} \text{-} } { } ( 1 + \kappa _ { a } ^ { a } [ 2 \text{-} \text{-} \text{-} \text{-} \text{-} } ) ( 1 + \kappa _ { a } ^ { r } [ \text{Het} ] ^ { 2 } ) ^ { 2 } - \delta _ { a } [ N t c A ] \ \ ( 3 1 )$$
## Turing linear stability analysis
Here we discuss the Turing stability analysis of our system [ ? ? , ]. Let us first discuss the general theory. We consider a chain composed of n -dimensional dynamical systems, each one i = 1 , . . . , n characterized by ˙ q i = f q ( i ). If we add diffusion along the chain, the behavior of cell i is characterized by:
$$\frac { d q _ { i } } { d \tau } = f ( q _ { i } ) + \widetilde { D } ( q _ { i + 1 } + q _ { i - 1 } - 2 q _ { i } ),$$
where ˜ D is the diffusion tensor . Now we consider a (stable) fixed point, q 0 , of the dynamical system, i.e. f q ( 0 ) = 0 . It also constitutes a fixed point for the entire chain since diffusion terms cancel ( q i = q j for all i and j ). We want to analyze the effect of a small perturbation, ∆ , around the steady state of the chain. Introducing the variables
$$q _ { i } = q _ { 0 } + \Delta _ { i },$$
into Equation 32, and expanding up to first order in ∆ one gets:
$$\frac { d \Delta _ { i } } { d \tau } = \nabla f ( q _ { 0 } ) \cdot \Delta _ { i } + \widetilde { D } ( \Delta _ { i + 1 } + \Delta _ { i - 1 } - 2 \Delta _ { i } ),$$
where ∇ f q ( 0 ) is the Jacobian matrix of the field evaluated in the point q 0 . Furthermore, we can decompose the perturbation in terms of plane waves:
$$\Delta _ { i } ( \tau ) & = \sum _ { k } \Delta _ { i, k }, \\ \Delta _ { i, k } ( \tau ) & = A e ^ { \omega _ { k } \tau } \cos ( k i ).$$
The admissible values of the wavevector k depend on the length of the chain and on the boundary conditions. For instance, k = nπ/L with n = 1 2 3 , , , . . . , L for Von Neumann (zero flux) boundary conditions. Introducing 35 into 34 we find:
$$\omega _ { k } A = \nabla f ( q _ { 0 } ) A + 2 D A ( \cos k - 1 ),$$
�
�
�
�
�
�
�
�
�
Figure 8. Results for ω max extracted from Equation 37 for the cyanobacterial system with D s = 0 1 . and D n = 0 2 (Blue). The black dashed line indicates the value at equilibrium in the absence of diffusion .
which has nontrivial solutions if
$$\det ( \nabla f ( q _ { 0 } ) + 2 \dot { D } ( \cos k - 1 ) - \omega _ { k } ) = 0.$$
Therefore, the k mode is related to some possible frequencies ω k . If those frequencies satisfy Re( ω k ) < 0, it is expected that the perturbation ∆ will be damped (negative exponential) and the system will recover its initial equilibrium. Nevertheless, if Re( ω k ) > 0 perturbations will be amplified (positive exponential) and thus it is expected that structures of wavelength 2 π/k will develop. The important value that determines if the system is stable to perturbations of some wavevector k is the largest real part of the ω k from 37.
We can now turn to our cyanobacteria chain characterized by Eq. (16) of the main text. We analyze the effect of perturbations around the vegetative-like state for D s = 0 1 and . D n = 0 2 (Figure 8). . We find that the system is unstable against perturbations of intermediate wavevectors, with an upper bound (representing a minimum length at which patterns can be formed) and a lower bound (maximum length). The value of the minimum length (measured in cell number) l min ≈ 8 / 7 > 1 so that a single cell is unable to differentiate.
|
10.1371/journal.pcbi.1004129
|
[
"Alejandro Torres-Sanchez",
"Jesus Gomez-Gardenes",
"Fernando Falo"
] |
2014-08-05T10:43:26+00:00
|
2014-08-05T10:43:26+00:00
|
[
"q-bio.CB",
"q-bio.MN"
] |
An integrative approach for modeling and simulation of Heterocyst pattern formation in Cyanobacteria strands
|
A comprehensive approach to cellular differentiation in cyanobacteria is
developed. To this aim, the process of heterocyst cell formation is studied
under a systems biology point of view. By relying on statistical physics
techniques, we translate the essential ingredients and mechanisms of the
genetic circuit into a set of differential equations that describes the
continuous time evolution of combined nitrogen, PatS, HetR and NtcA
concentrations. The detailed analysis of these equations gives insight into the
single cell dynamics. On the other hand, the inclusion of diffusion and noisy
conditions allows simulating the formation of heterocysts patterns in
cyanobacteria strains. The time evolution of relevant component concentrations
are calculated allowing for a comparison with experiments. Finally, we discuss
the validity and the possible improvements of the model.
|
1408.1086v1
|
## Discrete nature of thermodynamics in confined ideal Fermi gases ∗
Alhun Aydin and Altug Sisman † Istanbul Technical University, Energy Institute, 34469, Istanbul, Turkey (Dated: August 7, 2014)
Intrinsic discrete nature in thermodynamic properties of Fermi gases appears under strongly confined and degenerate conditions. For a rectangular confinement domain, thermodynamic properties of an ideal Fermi gas are expressed in their exact summation forms. For 1D, 2D and 3D nano domains, variations of both number of particles and internal energy per particle with chemical potential are examined. It is shown that their relation with chemical potential exhibits a discrete nature which allows them to take only some definite values. Furthermore, quasi-irregular oscillatory-like sharp peaks are observed in heat capacity. New nano devices can be developed based on these behaviors.
## I. INTRODUCTION
Nano scale thermodynamics may be regarded as a relatively new research area. Nevertheless, considering leap forwards in nanoscience and nanotechnologies in recent years, it is fair to say that understanding thermodynamics of nano systems became an inevitable necessity [1-13]. When dealing with thermodynamics of nano systems, it is crucial to understand in which ways they differ from macro systems.
Thermodynamic state functions are written as infinite summations, in their exact forms. In classical thermodynamics, sums are replaced by integrals, under continuum approximation [14]. However, this approximation is no longer valid at nano scale and thereby, summations and integrals give considerably different results. Therefore, either summation forms of state functions have to be calculated numerically, or derived analytically by using more precise methods like Poisson summation formula etc., in order to examine thermodynamic behaviors of systems at nano scale.
gas pressure, non-additivity of global thermodynamic properties, quantum forces and thermosize effects which are similar to thermoelectric effects [5, 6, 8, 9, 11-13]. Third term of the PSF is the discrete correction term to the integral representation of summations. While the contributions of both the second and third terms are negligible in macro scale, they become substantial in strongly confined nano structures. The contribution of third term becomes crucial only if at least one of the domain sizes is smaller than the de Broglie wavelength of particles. When analytical solution for the third term is not obtainable, it is preferred to calculate the summation itself numerically instead of calculating the whole terms of PSF.
To evaluate summations analytically, Poisson summation formula (PSF) for even functions is used and it is given as follows
$$\overbrace { \underset { \underset { i = 1 } { \underset { 0 } { \infty } } } { \overset { \dots } { \underset { i = 1 } { \underset { i = 1 } { \infty } } } } } ^ { \overset { \infty } { \infty } } \int _ { i = 1 } ^ { \infty } f ( i ) = \int _ { 0 } ^ { \infty } f ( i ) d i - \frac { f ( 0 ) } { 2 } + 2 \sum _ { s = 1 } ^ { \infty } \int _ { 0 } ^ { \infty } f ( i ) c o s ( 2 \pi s i ) d i \quad \text{are a} \\ \overbrace { \underset { \underset { \underset { i = 1 } { \infty } } } { \underset { \underset { \underset { \underset { \underset { \underset { \infty } { \infty } } { \infty } } } } } } ^ { \text{are a} } \text{and } d \\ \text{and } d \\ \text{aref term of $\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} } \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb$s} } \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{t}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$ \mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$} \text{$\mathbb{s}$}$$
First term of PSF is the conventional integral term and the second term represents the zero correction which is considered as quantum size effects in literature. For Maxwell-Boltzmann gases, the second term is responsible for quantum boundary layer, anisotropic
∗ NOTICE: This is the authors version of a work that was accepted for publication in Physics Letters A. Changes resulting from the publishing process, such as peer review, editing, corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. A definitive version was subsequently published in Phys. Lett. A, [378, 30-31, 6/13/2014] DOI: 10.1016/j.physleta.2014.05.044
† Corresponding author, [email protected]
In this paper, by considering the exact forms of thermodynamic state functions based on infinite sums, it is shown that the intrinsic discrete nature appears in thermodynamic quantities such as chemical potential, internal energy per particle and heat capacity per particle at constant volume for strongly confined and degenerate ideal Fermi gases. Besides the numerical results, some analytical expressions, representing the discrete nature of chemical potential and internal energy per particle, are derived and gaps between their discrete values are analytically expressed. Furthermore, oscillatory nature of heat capacity varying with particle number and domain size is observed. These behaviors are in agreement with the results of [15, 16] in which chemical potential and specific heat have been just numerically calculated for ultra-cold Fermi gases confined in harmonic and quartic traps. Surface density dependency of heat capacity for 2D Fermi gases has been considered as the manifestation of QSE in [17]. In addition, temperature dependence of specific heat capacity of non-interacting fermions with multifractal energy spectra has been numerically analyzed in literature [18, 19]. On the other hand, in this study, there is no restriction for temperature and analytical expressions representing the discrete nature of thermodynamic properties are given. Variation of heat capacity with domain size is also discussed as a manifestation of quantum size effects. Furthermore, oscillations and sharp peaks in heat capacity are explained by defining the discrete Fermi point, line and surface.
Discrete and oscillatory behaviors do not appear in Bose gases and they disappear even in Fermi gas at high temperature limit or for large domain sizes in comparison with the de Broglie wavelength of particles. In other words, these behaviors appear only when quantum degeneracy and size effects are strong enough to make the Pauli exclusion principle and quantum character of particles dominant in behavior of the system.
## II. INTRINSIC DISCRETE NATURE OF THERMODYNAMIC PROPERTIES IN AN IDEAL FERMI GAS CONFINED AT NANO SCALE
Considering mono particles (like monatomic gas, electron gas etc.) confined in a rectangular domain with dimensions L , L 1 2 and L 3 , energy eigenvalues of particles from Schr¨dinger equation are o
$$\varepsilon = \frac { h ^ { 2 } } { 8 m } \sum _ { n = 1 } ^ { 3 } \left ( \frac { i _ { n } } { L _ { n } } \right ) ^ { 2 }, \text{ with } i _ { n } = 1, 2, 3, \dots, \quad ( 2 ) \quad \text{$\text{$contin$}$} \quad \text{for $d$} \\ \text{goes}$$
where h is the Planck's constant, m is the mass of particle and n = 1 2 3 representing the spatial , , directions. In order to discuss confinement magnitudes of different domains, confinement parameters can be defined as α n = L /L c n where L c is a length scale based on de Broglie wavelength of particles that is expressed as L c = h/ √ 8 mk T b , where k B is the Boltzmann's constant and T is the temperature of the gas. Then, dimensionless energy eigenvalues in α -notation is
$$\frac { \varepsilon } { k _ { B } T } = \tilde { \varepsilon } = [ ( \alpha _ { 1 } i _ { 1 } ) ^ { 2 } + ( \alpha _ { 2 } i _ { 2 } ) ^ { 2 } + ( \alpha _ { 3 } i _ { 3 } ) ^ { 2 } ] \quad ( 3 ) \quad \max _ { \text{a re} } ^ { \text{com} }$$
Note that, confinement parameter α is inversely proportional to the domain size, so as the name implies, increasing α means increasing confinement and when α becomes greater than unity, domain size is smaller than L c .
By neglecting spin degrees of freedom g s (which is just a factor of 2 for a free electron gas or He 3 ), Fermi-Dirac distribution function is
$$f = \frac { 1 } { \exp ( \tilde { \varepsilon } - \Lambda ) + 1 } \quad \quad ( 4 ) \quad \ \ 1$$
where Λ is dimensionless chemical potential, defined as Λ = µ/k B T . Summation of the distribution function over momentum states in all directions gives the number of particles in the system
$$N = \sum _ { i _ { n } = 1 } ^ { \infty } \frac { 1 } { \exp ( \tilde { \varepsilon } - \Lambda ) + 1 } \quad \begin{array} { c c } ( 5 ) & \text{api} \\ \alpha _ { 1 } \\ \text{wh}. \end{array}$$
where i n = i 1 , i 2 , i 3 , so it is a triple sum.
In this letter, we consider an ideal Fermi gas confined in three different rectangular domains. Temperature is kept constant in all cases considered here. In the first case, domain is strongly confined in two directions (domain sizes are much smaller than L c in those directions) and relatively weakly confined in the other direction. In strong confinements, values of momentum component of that direction are always equal to the value of the ground state. Thus, no summations have to be done over that momentum components. In relatively weakly confined directions, however, particles can become excited so that value of momentum components in that directions can take the values of excited states and summations have to be done.
## A. 1D Fermi Gas
Although thermodynamic quantities are calculated based on infinite sums, after Fermi level ( i F = √ Λ /α ), contributions to the summation rapidly decrease and it is quite sufficient to sum up to i max = 2 i F . After i max , contributions become completely negligible. Moreover, for directions with α values much larger than unity, i F goes to zero for a limited range of Λ and this completely removes the summation process for that direction.
Let's consider an anisometric domain with confinement parameters α 1 = 1, α 2 = 40 and α 3 = 40. Note that, 40 is a realistic value for α , since for an electron confined in a single atom carbon layer (graphene) with 0 3 nm . thickness, α = 43 at 20K [20]. In this system, we can replace triple summations by a single sum as long as Λ < Λ 1 = ( α 1 ) 2 + ( α 2 ) 2 + (2 α 3 ) 2 . After the value of Λ , particles can occupy the excited states of momentum 1 components even in strongly confined directions and they make considerable contributions to the summation. As a result, for a definite range of Λ, triple sum can be represented by a single sum, since for strongly confined directions only ground state of momentum components contributes.
In case of α 1 << 1, gas is still 1D Fermi gas since { α , α 2 3 } >> 1 and it is enough to use first two terms of PSF to calculate the summation. In that case, total number of particles, Eq. (5), can be approximated for a 1D Fermi gas as
$$\i \, \text{$\text{$rem gas as}$} \\ ( 4 ) \quad & N _ { 1 D } \approx - \frac { \sqrt { \pi } } { 2 \alpha _ { 1 } } L i _ { 1 / 2 } [ - \exp ( \Lambda ^ { \prime } ) ] + \frac { 1 } { 2 } L i _ { 0 } [ - \exp ( \Lambda ^ { \prime } ) ] \quad ( 6 a ) \\ \i \, \text{$\text{$i$ as}$} \\ \text{$i$ as} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in} \\ \text{$i$ in}$$
where Λ ′ = Λ -α 2 2 -α 2 3 and Li denotes the polylogarithm function. Eq. (6b) is obtained by using the asymptotic approximations of polylogarithm functions for Λ >> 1 (strongly degenerate Fermi gas). On the other hand, if α 1 is not much less than unity, namely in nano scale, the whole terms of PSF have to be considered to evaluate the summation properly and analytically. Therefore, by using the whole terms of PSF and asymptotic forms of polylogarithm functions, after some mathematical
operations, 1D number of particles is obtained as
$$N _ { 1 D } ^ { P S F } \cong \frac { \sqrt { \Lambda ^ { \prime } } } { \alpha _ { 1 } } - \frac { 1 } { 2 } + \frac { 1 } { \pi } \arctan \left [ \cot \left ( \frac { \pi \sqrt { \Lambda ^ { \prime } } } { \alpha _ { 1 } } \right ) \right ] \quad ( 7 )$$
where PSF superscript indicates that the whole terms of PSF are used to obtain the expression. Variation of number of particles with dimensionless chemical potential Λ for anisometric 1D Fermi gas is given in Figure 1.
FIG. 1. Number of particles (left axis) and dimensionless internal energy per particle (right axis) varying with dimensionless chemical potential for 1D Fermi gas with α 1 = 1 , α 2 = 40 , α 3 = 40.
In Fig. 1, the variation of number of particles and dimensionless internal energy per particle with dimensionless chemical potential is given in left and right axes respectively. Considering the left axis first; the solid curve represents Eq. (5), that is the exact expression for number of particles confined in a 1D nano domain. Bold dots indicate the Λ values which correspond to the integer particle numbers and they are solved numerically from Eq. (5). Dashed curve represents Eq. (7), which gives almost the same result of Eq. (5) for 1D, except the sharpness of the edges of steps. Therefore, Eq. (7) correctly predicts that chemical potential is a discrete quantity as long as number of particles is integer and this discreteness of chemical potential becomes important in a degenerate Fermi gas confined in a nano domain. Dotted curve is obtained from Eq. (6b) and it passes exactly from the points where the particle number is integer even though it cannot represent the true discrete relation between number of particles and chemical potential.
For 1D Fermi gases that are strongly confined in second and third directions in a rectangular domain, a formula that gives the discrete Λ values corresponding to the integer number of particles can be defined. Either the third term of Eq. (7) is equalized to zero to find its roots, or more easily Λ can analytically be solved from the first two terms of Eq. (7) since third term is exactly zero when number of particles is integer. Then, for integer N >> 1
$$\Lambda ^ { \prime } = \alpha _ { 1 } ^ { 2 } \left ( N + \frac { 1 } { 2 } \right ) ^ { 2 }$$
As it is seen, Eq. (8) gives the possibility to write thermodynamic properties of 1D Fermi gases in terms of particle number instead of chemical potential. Besides, from Eq. (8), expression for dimensionless chemical potential intervals can easily be obtained
$$\Delta \Lambda ^ { \prime } = 2 \alpha _ { 1 } ^ { 2 } ( N + 1 ) = 2 \alpha _ { 1 } \sqrt { \Lambda ^ { \prime } } + \alpha _ { 1 } ^ { 2 } \quad \ \ ( 9 )$$
The other thermodynamic quantity that represents stepwise nature is the internal energy. Dimensionless internal energy is written in its exact form as
$$\frac { U } { k _ { B } T } = \widetilde { U } = \sum _ { i _ { n } = 1 } ^ { \infty } \frac { \tilde { \varepsilon } } { \exp ( \tilde { \varepsilon } - \Lambda ) + 1 } \quad \quad ( 1 0 )$$
Using the first two terms of PSF and asymptotic approximations of polylogarithm functions, dimensionless internal energy for 1D Fermi gases with integer number of particles is obtained as
$$\widetilde { U } _ { 1 D } = \frac { ( \Lambda ^ { \prime } ) ^ { 3 / 2 } } { 3 \alpha _ { 1 } } + \left ( \frac { \sqrt { \Lambda ^ { \prime } } } { \alpha _ { 1 } } - \frac { 1 } { 2 } \right ) ( \alpha _ { 2 } ^ { 2 } + \alpha _ { 3 } ^ { 2 } ) \quad ( 1 1 )$$
As long as particle number is integer, there are no longer continuous Λ values, and since all specific (per particle) thermodynamic state functions depend on Λ, specific internal energy and heat capacity have also discrete values corresponding to the integer particle number.
Considering the right axis of Fig. 1, discreteness of internal energy per particle is shown for 1D Fermi gas for particle numbers from 50 to 55 where the functional behavior is exactly the same as the N -Λ relation. This time, solid curve is obtained by dividing Eq. (10) to Eq. (5). Dotted line represents the result of Eq. (11) divided by Eq. (6b) and it matches with exact solution for integer N . If N is large enough, even for the unit value of α , Λ ′ becomes very large, so that by using Eq. (8) and Eq. (11) dimensionless internal energy is written in terms of N and α as
$$\widetilde { U } _ { 1 D } \cong \frac { \alpha _ { 1 } ^ { 2 } } { 3 } \left ( N + \frac { 1 } { 2 } \right ) ^ { 3 } + N ( \alpha _ { 2 } ^ { 2 } + \alpha _ { 3 } ^ { 2 } ) \quad ( 1 2 )$$
Using Eq. (12), dimensionless internal energy interval per particle for 1D Fermi gas, can be derived as
$$\Delta \left ( \frac { \widetilde { U } } { N } \right ) \cong \frac { 2 } { 3 } \alpha _ { 1 } ^ { 2 } \left ( N + \frac { 5 } { 4 } \right ) \text{ \quad \ \ } ( 1 3 )$$
From Eq. (13), it can be deduced that when α goes to zero (as in macro systems), discrete nature and
stepwise behavior practically disappears although Eq. (10) preserves its intrinsic discrete nature. Conversely, for large α values it become apparent. In addition, discreteness in N -Λ or ˜ U -Λ relations depend on confinement parameter α instead of temperature alone. So, even at T = 0K, steepness of the function in Fig. 1 does not get sharper as long as α is kept constant. On the other hand, both dimensionless chemical potential and internal energy intervals change with the value of α which depends on temperature, Eq. (9) and Eq. (13). When temperature goes to zero, α values and so the discreteness interval of Λ go to infinity.
## 1. Fermi Point and Zero Heat Capacity in confined 1D Fermi gas
Now let's examine the heat capacity of 1D Fermi gas at nano scale. Dimensionless heat capacity at constant volume is written in its exact form as
$$\frac { C _ { V } } { k _ { B } } = \widetilde { C } _ { V } = \sum _ { i _ { n } } ( \tilde { \varepsilon } ) ^ { 2 } f ( 1 - f ) - \frac { [ \sum _ { i _ { n } } \tilde { \varepsilon } f ( 1 - f ) ] ^ { 2 } } { \sum _ { i _ { n } } f ( 1 - f ) } \ \ ( 1 4 ) \quad \text{see $F$} \ \text{state.} \ \text{state.}$$
Derivation of heat capacity is done by differentiating internal energy U with respect to temperature T . In the derivation of exact expression of heat capacity, to be able to interchange derivative operator with sum operator, uniform convergency of the function is checked by comparing the analytical derivative with numerical one. It is seen that results perfectly match with each other.
As it's seen from Eq. (14), heat capacity expression contains the variance of Fermi-Dirac distribution function V ar f ( ) = f (1 -f ), which is equal to the derivative of distribution function with respect to Λ, df/d Λ = f (1 -f ). It is not possible to obtain an analytical expression for heat capacity by using the whole terms of PSF even for large values of Λ. For this reason, heat capacity is numerically calculated based on its summation form given by Eq. (14).
In degenerate Fermi gas (Λ >> 1), electronic contribution to heat capacity only comes from the momentum states around Fermi point, Fermi line and Fermi surface for 1D, 2D and 3D Fermi gases respectively. The state corresponding to the Fermi point is the one for which the distribution function is equal to 1 / 2 and the variance reaches its maximum value of 1 / 4, as it is seen in Fig 2. Full band width at half maximum (FWHM) of variance for 1D distribution function can be found by equating f (1 -f ) to 1 / 8
$$\delta _ { 1 D } = \frac { \arcosh ( 1 7 ) } { 2 \alpha _ { 1 } \sqrt { \Lambda ^ { \prime } } } \quad \quad ( 1 5 )$$
It is easy to see from Eq. (15) that if the confinement ( α ) or degeneracy (Λ) increases, FWHM decreases, so the peak of variance becomes too sharp that integer values of
FIG. 2. Distribution function (left axis) and its variance (right axis) around Fermi point of 1D Fermi gas with α 1 = 1 , α 2 = 40 , α 3 = 40 and N = 50.
momentum state variable, i 1 , correspond to the vanished values of the variance function, f (1 -f ). In other words, Fermi point corresponds to the half integer value of i 1 , see Fig. 2, which is not a valid value for a momentum state. In that case, contributions to heat capacity from integer values of i 1 are almost zero. As a result, heat capacity of 1D Fermi gas is nearly zero. When 0D system is considered, it is clear from Eq. (14) that summations die out and C V becomes exactly zero since there is only ground state in 0D system. Hence it is worth to notice that contributions from momentum states to C V always come from D -1 dimensional momentum space. For instance, in degenerate 2D and 3D Fermi gases, contributions come from the states on Fermi line and Fermi surface respectively.
## B. 2D and 3D Fermi Gases
Two different domains having 2D and 3D momentum spaces are considered in this section. Their α values are α 1 = 3, α 2 = 3, α 3 = 40 and α 1 = 3, α 2 = 3, α 3 = 3 respectively. These two cases are considered under the same section due to their similar nature. In 2D case, domain is relatively weakly confined in first and second directions whereas strongly confined in third one. For this kind of domain, there is no need to make summation over the states of momentum component in third direction since the only contribution comes from its ground state. By using Eq. (5) and the first two terms of PSF, an analytical expression for number of particles can be obtained as
$$N _ { 2 D } \approx \frac { \pi \Lambda ^ { \prime \prime } } { 4 \alpha _ { 1 } \alpha _ { 2 } } - \frac { \sqrt { \Lambda ^ { \prime \prime } } } { 2 } \left ( \frac { 1 } { \alpha _ { 1 } } + \frac { 1 } { \alpha _ { 2 } } \right ) + \frac { 1 } { 4 } \quad ( 1 6 )$$
where Λ ′′ = Λ -α 2 3 . By using the first two terms of PSF, internal energy expression for 2D Fermi gas is obtained
FIG. 3. Number of particles (left axis) and dimensionless internal energy per particle (right axis) varying with dimensionless chemical potential for 2D Fermi gas with α 1 = 3 , α 2 = 3 , α 3 = 40.
as
$$\widetilde { U } _ { 2 D } \approx \frac { \pi \Lambda ^ { \prime \prime } ^ { 2 } } { 8 \alpha _ { 1 } \alpha _ { 2 } } - \frac { \Lambda ^ { \prime \prime 3 / 2 } } { 6 } \left ( \frac { 1 } { \alpha _ { 1 } } + \frac { 1 } { \alpha _ { 2 } } \right ) + \alpha _ { 3 } ^ { 2 } N _ { 2 D } \quad ( 1 7 ) \quad \text{the t} \\.$$
As it's seen from Fig. 3, quasi-irregular stepwise relation of number of particles and internal energy changing with chemical potential is observed in 2D Fermi gases, unlike the regular stepwise nature in 1D. In Fig. 3, for the left axis, Eq. (16) represented by dashed line does not perfectly matches with the exact solution based on Eq. (5) represented by solid line. On the right axis of Fig. 3, quasi-irregular stepwise behavior is observed for dimensionless internal energy per particle. Solid curve represents Eq. (10) divided by Eq. (5) and the dashed curve is the result of Eq. (17) divided by Eq. (16). Bold points mark the values where the particle number is integer.
Similar quasi-irregular discrete nature of 2D appears also in 3D Fermi gas with α 1 = α 2 = α 3 = 3. In a similar way used for 2D Fermi gas, approximate analytical expressions for number of particles and internal energy of 3D Fermi gas are obtained as follows
$$N _ { 3 D } \approx \frac { \pi \Lambda ^ { 3 / 2 } } { 6 \alpha _ { 1 } \alpha _ { 2 } \alpha _ { 3 } } - \frac { \pi \Lambda } { 8 } \left ( \frac { 1 } { \alpha _ { 1 } \alpha _ { 2 } } + \frac { 1 } { \alpha _ { 2 } \alpha _ { 3 } } + \frac { 1 } { \alpha _ { 1 } \alpha _ { 3 } } \right ) \quad \text{$a\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\c
\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha_{\alpha \alpha_{\alpha_{\alpha_{\alpha_{\alpha \alpha_{\alpha \alpha_{\alpha \alpha_{\alpha \alpha_{\alpha \alpha_{\alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \alpha \beta \alpha \alpha \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta \beta$$
$$\widetilde { U } _ { 3 D } \approx \frac { \pi \Lambda ^ { 5 / 2 } } { 1 0 \alpha _ { 1 } \alpha _ { 2 } \alpha _ { 3 } } - \frac { \pi \Lambda ^ { 2 } } { 1 6 } \left ( \frac { 1 } { \alpha _ { 1 } \alpha _ { 2 } } + \frac { 1 } { \alpha _ { 2 } \alpha _ { 3 } } + \frac { 1 } { \alpha _ { 1 } \alpha _ { 3 } } \right ) \quad \text{
} \end{
\quad + \frac { \Lambda ^ { 3 / 2 } } { 1 2 } \left ( \frac { 1 } { \alpha _ { 1 } } + \frac { 1 } { \alpha _ { 2 } } + \frac { 1 } { \alpha _ { 3 } } \right ) - \frac { 1 } { 8 } \quad \text{
\quad \text{ }
- \frac { \pi \Lambda ^ { 3 / 2 } } { 1 2 } \left ( \frac { 1 } { \alpha _ { 1 } } + \frac { 1 } { \alpha _ { 2 } } + \frac { 1 } { \alpha _ { 3 } } \right ) - \frac { 1 } { 8 } \quad \text{
\quad \text{
\quad \text{ }
} \end{ }
} } } \end{ } } }$$
As it is seen in Fig. 4, Eqs. (18) and (19) represent
FIG. 4. Number of particles (left axis) and dimensionless internal energy per particle (right axis) varying with dimensionless chemical potential for 3D Fermi gas with α 1 = 3 , α 2 = 3 , α 3 = 3.
the trends of exact solutions quite well. Solid curve in Fig. 4 represent the results from the exact expressions based on Eqs. (5) and (10) for number of particles and dimensionless internal energy per particle respectively. Bold points indicate the values of Λ corresponding to the integer number of particles, N = { 50 , ..., 80 . } Dashed line represent the approximate analytical solutions, Eq. (18) and Eq. (19)/(18), based on the first two terms of PSF.
Interestingly, behaviors of 2D and 3D Fermi gases are quite different than that of 1D Fermi gases. To understand the reason, it is necessary to examine the nature of summations in detail. All thermodynamic properties are represented by summations over momentum state variables { i 1 , i 2 , i 3 } , which constitute a state matrix. Each element of the state matrix represented by { i 1 , i 2 , i 3 } , contributes to the summation differently. This state matrix can be decomposed into diagonal { i 1 = i 2 = i 3 } and non-diagonal matrices, except for 1D Fermi gas in which state matrix becomes a state vector. Quasi-irregular behaviors observed in 2D and 3D Fermi gases are the consequences of the contribution of the elements of non-diagonal state matrix. Unlike diagonal matrix, in non-diagonal matrix, there are mostly more than one element (momentum state) that make the same contribution to the summation, which is commonly called as degeneracy of energy levels. Since the degeneracy of energy levels changes due to the different combinatoric nature of double or triple summations, quasi-irregular patterns in Figs. 3 and 4 are different for 2D and 3D Fermi gases. On the other hand, regular stepwise behaviors instead of quasi-irregular one appear in 1D Fermi gas since there is no such degeneracy in energy levels in 1D case.
FIG. 5. Idealized and discrete Fermi lines and Fermi shell for 2D Fermi gas with α 1 = 3 , α 2 = 3 , α 3 = 40 and N = 50.
## 1. Discrete Fermi Line and Heat Capacity Oscillations in confined 2D Fermi gas
In 2D Fermi gas, contribution to heat capacity comes from the particles on the Fermi line. In other words, contributions from the particles outside of the Fermi line are totally negligible, in similar to outside of Fermi point in 1D Fermi gas. Therefore, it is enough to make summations just over the states on Fermi line to calculate the heat capacity of a 2D Fermi gas. To identify the Fermi line, it is useful to analytically define Round, Ceiling and Floor functions. Analytical expression of round of x is simply given by
$$R o u n d ( x ) = x + \frac { 1 } { \pi } \arctan \left [ \cot \left ( \pi \left ( x + \frac { 1 } { 2 } \right ) \right ) \right ] \ \ ( 2 0 ) \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
One can then easily obtain the Ceiling and the Floor functions by adding 1 / 2 to and subtracting 1 / 2 from x in Eq. (20) respectively. Hence, the states on Fermi line are represented by i 1 from 1 to i 1 max and i 2 from i 2 min to i 2 max where i 1 max = √ Λ ′′ /α 1 + 1 / 2, i 2 min = Ceiling( √ Λ ′′ -[ α 1 ( i 1 +1 2)] / 2 /α 2 -1 2) / and i 2 max = Floor( √ Λ ′′ -[ α 1 ( i 1 -1 / 2)] 2 /α 2 + 1 2). / Idealized (continuous) and exact (discrete) Fermi lines are shown in Figure 5 by dotted curve and solid stepwise function respectively. Contribution to heat capacity comes only from the states corresponding to the couples of integer i 1 and i 2 , which are represented by bold dots in Fig. 5. Two dashed curves enclosing the discrete Fermi line are ± 1 / 2 neighborhoods of the dotted (idealized) Fermi line. These two dashed curves define a Fermi shell originated from the Heisenberg uncertainty principle. Exact Fermi line always stays within Fermi shell. The closer a state inside the Fermi shell to the idealized Fermi line, the larger its contribution to heat capacity and vice versa.
In Figure 6, peakwise nature is seen in the variation of dimensionless specific heat capacity of 2D Fermi gas with number of particles.
When N is increased, Fermi line and consequently the Fermi shell extends while its thickness remains constant. Since the combination of the proximities of the states in Fermi shell to the idealized Fermi line changes, the magnitude of their contributions also changes and causes the peaks in heat capacity for certain number of particles. For example, in Fig. 6, there is a peak for 50 particles and it goes almost down to zero for 52 particles. As it is seen from Fig. 5, there are several states (bold points) which are close to the idealized Fermi line for 50 particles, so their contribution are large enough to generate a peak in heat capacity. Conversely, in Figure 7, for 52 particles, states in Fermi shell are mostly far from the idealized Fermi line, so heat capacity goes down to almost zero.
When N or Λ increases, Fermi line extends and Fermi shell starts to enclose more and more states. Even if the number of states in Fermi shell increases, FWHM of Fermi line decreases and most of the contributions of states in Fermi shell to heat capacity nearly vanish. Hence, despite the oscillations, when N →∞ or α →∞ , C V goes to zero.
In Figure 8, unusual effect of confinement on heat capacity is shown. When α (confinement) increases, quasi-irregular, non-periodic oscillations appear on specific heat capacity of 2D Fermi gases. This behavior again rises due to discrete nature of Fermi line. When the domain size is anisotropically changed, distribution and contributions of states in Fermi shell also change and cause oscillations seen in Fig. 8. On the other hand, when α values for all directions are increased at the same rate, oscillations almost disappear and heat capacity changes smoothly. The reason is that the distance of state points to the idealized Fermi line does not change while the sharpness of the variance function, Eq.(15), changes at the same rate for all the states in Fermi shell.
FIG. 6. Dimensionless heat capacity per particle vs number of particles for 2D Fermi gas with α 1 = 3 , α 2 = 3 , α 3 = 40.
FIG. 7. Idealized and discrete Fermi lines and Fermi shell for 2D Fermi gas with α 1 = 3 , α 2 = 3 , α 3 = 40 and N = 52.
## 2. Discrete Fermi Surface and Heat Capacity Oscillations in Confined 3D Fermi Gas
In 3D Fermi gas, contribution to heat capacity comes from the states on Fermi surface. Similar to 2D case, idealized Fermi surface (solid dark gray surface) as well as its ± 1 / 2 neighborhoods (solid light gray surfaces) and the exact or discrete Fermi surface (meshed surface) are shown in Fig. 9. Again, due to the nature of Fermi-Dirac variance function, contribution to heat capacity comes from the states within Fermi shell. Therefore, instead of making summations over { i 1 , i 2 , i 3 } from 1 to ∞ , it is enough to sum over i 1 from 1 to i 1 max = √ Λ /α 1 + 1 2, / i 2 from i 2 min = Ceiling( √ Λ -[ α 1 ( i 1 +1 2)] / 2 -[ α 3 ( i 3 +1 2)] / 2 /α 2 -1 2 ) to i 2 max = Floor( √ Λ -[ α 1 ( i 1 -1 / 2)] 2 -[ α 3 ( i 3 -1 / 2)] 2 /α 2 +1 2) and / i 3 from 1 to i 3 max = √ Λ /α 3 +1 2. /
FIG. 8. Dimensionless heat capacity per particle vs confinement parameter α 1 for 2D Fermi gas with α 2 = 3, α 3 = 40 and N = 50.
FIG. 9. Idealized and discrete Fermi surfaces and Fermi shell for 3D Fermi gas with α 1 = 3 , α 2 = 3 , α 3 = 3 and N = 50.
Similar to 2D case, there are significant sharp peaks in dimensionless heat capacity per particle vs number of particles graph, Figure 10.
Anomalies seen in the variation of heat capacity per particle with number of particles can be interpreted based on all the explanations that are used for 2D case.
If one or two directions are confined, non-periodic oscillations in specific heat capacity appear. In Fig. 11, oscillations in heat capacity per particle due to the variation of confinement rate in the first direction ( α 1 ) are shown for a 3D Fermi gas. On the other hand, similar to 2D case, when the sizes of an isometric system is reduced in all directions at the same rate, C V -α relation becomes a smooth decaying function instead of frequently oscillatory one.
FIG. 10. Dimensionless heat capacity per particle vs number of particles for 3D Fermi gas with α 1 = 3 , α 2 = 3 , α 3 = 3.
FIG. 11. Dimensionless heat capacity per particle vs confinement parameter α 1 for 3D Fermi gas with α 2 = 3, α 3 = 3 and N = 50.
## III. POSSIBILITY OF EXPERIMENTAL VERIFICATION AND EXCESS THERMAL ENERGY STORAGE AT NANO SCALE
Advances in nano technologies may provide an experimental verification for heat capacity anomalies (peaks and oscillations) in Fermi gases at nano scale. Consider an isometric 3D InSb (Indium antimonide) nanobox with domain sizes L 1 = L 2 = L 3 = 10 nm and conduction electron density 9 35x10 . 25 m -3 at T = 5K [21]. By changing the domain in the first direction from 10 nm to 12 nm, while keeping the conduction electron density between the range of 9 31x10 . 25 m -3 and 9 40x10 . 25 m -3 (by adjusting the number of electrons during the size reduction), one may observe the oscillations in the electronic contribution to the specific heat. For this system, particle number changes between 94 and 112 which is far from fluctuations. Due to possible experimental or computational errors, there may be inconsistencies in the comparison of experimental and theoretical results. To avoid this, one can repeat the experiment for the T = 7K and compare the ratio of the results at two temperatures with our theoretical estimate.
In Fig 12, solid curve is obtained by using the exact heat capacity expression in Eq. (14), while the dashed line is achieved by using the heat capacity per particle equation in continuum limit (without size effects), which is expressed as ˜ C /N V = π / 2 2Λ. It is seen that there are considerable and probably measurable deviations from the classical behavior. When domain size becomes comparable to L c , the wave functions of individual particles and the whole system is greatly affected by the boundaries of the domain. When the system is too small, any tiny changes in size greatly affects the system. So, specific heat capacity of Fermi gases at nano scale is strongly depend on anisometric size changes. This may lead us to make excess thermal energy storage devices at nano scale. By making small changes in domain size for one or two direction properly, one can increase or
1
FIG. 12. Ratio of heat capacity at T = 5K and T = 7K varying with domain size in the first direction, L 1 (nm), for 3D electron gas confined in InSb with N = 95 { , ..., 112 } , L 2 = 10 nm, L 3 = 10 nm and conduction electron density around n = 9 35x10 . 25 m -3 .
H
L
decrease the heat capacity of the substance. Thus, by adjusting domain sizes, high capacity thermal energy storage systems can be developed in nano scale.
## IV. CONCLUSION
In this paper, it is shown that thermodynamic properties of an ideal Fermi gas can have discrete nature in nano scale. Appearance of discrete nature is the combined consequence of internal features of the system such as Pauli exclusion principle and the quadratic energy-momentum dispersion relation, and the external parameters like quantum degeneracy and quantum confinement. Therefore, discrete nature of thermodynamic properties disappears in Bose and Maxwell gases or Fermi gas in thermodynamic limit (infinite domain size). It is seen that, as macroscopic properties of system, chemical potential and internal energy per particle can have discrete nature. QSE due to the strong confinement reveal themselves as stepwise behaviors in chemical potential and internal energy whereas peakwise and oscillatory behavior in heat capacity. The contributions from the states within Fermi shell centralized by idealized Fermi point, line and surface provide an explanation for unusual behaviors of heat capacity in 1D, 2D and 3D systems respectively. It is also shown that change in domain sizes strongly affects the behavior of heat capacity of Fermi gas at nano scale.
In nano systems with small number of particles, there is a need for the comparison of the order of fluctuations in particle number and energy with the order of discrete corrections based on QSE. The comparison of the order of fluctuations and QSE corrections has been discussed in literature [4]. For the cases considered here, it is seen that order of fluctuations in thermodynamic quantities are much less than the order of quantum
(discrete) corrections. Therefore, it may be possible to experimentally verify these results by measuring the heat capacities of nano scale structures at low temperatures since the electronic contribution becomes more significant than the lattice one at low temperature
- [1] M. I. Molina, Ideal gas in a finite container, Am. J. Phys. 64 503-505, (1996).
- [2] R. K. Pathria, An ideal quantum gas in a finite-sized container, Am. J. Phys. 66 1080-1085, (1998).
- [3] W. S. Dai and M. Xie, Quantum statistics of ideal gases in confined space, Phys. Lett. A 311 340-346, (2003).
- [4] W. S. Dai and M. Xie, Geometry effects in confined space, Phys. Rev. E 70 016103, (2004).
- [5] A. Sisman and I. M¨ller, The Casimir-like size effects in u ideal gases, Phys. Lett. A 320 360-366, (2004).
- [6] A. Sisman, Surface dependency in thermodynamics of ideal gases, J. Phys. A- Math. Gen. 37 11353-11361, (2004).
- [7] H. Pang, W. S. Dai and M. Xie, The difference of boundary effects between Bose and Fermi systems, J. Phys. A-Math. Gen. 39 2563-2571, (2006).
- [8] A. Sisman, Z. F. Ozturk and C. Firat, Quantum boundary layer: a non-uniform density distribution of an ideal gas in thermodynamic equilibrium, Phys. Lett. A 362 16-20, (2007).
- [9] C. Firat and A. Sisman, Universality of quantum boundary layer for a Maxwellian gas, Phys. Scripta 79 065002, (2009).
- [10] Z. F. Ozturk and A. Sisman, Quantum Size Effects on the Thermal and Potential Conductivities of Ideal Gases, Phys. Scripta 80 065402, (2009).
- [11] C. Firat, A. Sisman and Z. F. Ozturk, Thermodynamics of gases in nano cavities, Energy 35 , 814-819, (2010).
- [12] A. Sisman and G. Babac, Quantum Size Effects on Classical Thermosize Effects, J. Continuum Mechanics and Thermodynamics 24 , 339-346, (2012).
conditions. Verification of discrete nature can also be done by measuring Fermi surface of a strongly confined structures [22]. Possibility of creating new technologies like nano scale thermal energy storage devices based on these effects is proposed.
- [13] C. Firat and A. Sisman, Quantum forces of a gas confined in nano structures, Phys. Scr. 87 , 045008, (2013).
- [14] C. Kittel and H. Kroemer, Thermal Physics, 2nd Ed. (W. H. Freeman, 1980).
- [15] J. Schneider, H. Wallis, Mesoscopic Fermi gas in a harmonic trap, Phys. Rev. A 57 , 1253-1259, (1998).
- [16] L. Chen, G. Su, J. Chen, Ultracold finite-size Fermi gas in a quartic trap, Phys. Lett. A 376 , 2908-2910, (2012).
- [17] D. A. Tayurskii and A. R. Minnullin, Thermodynamic and magnetic properties of the confined neutral Fermi systems, Physica B 329333, 152153, (2003).
- [18] I. N. de Oliveira, M. L. Lyra, E. L. Albuquerque, Specific heat anomalies of non-interacting fermions with multifractal energy spectra, Physica A 343 , 424-432, (2004).
- [19] D. A. Moreira, E. L. Albuquerque, D.H.A.L. Anselmo, Specific heat spectra of non-interacting fermions in a quasiperiodic ladder sequence, Phys. Lett. A 372 , 5233-5238, (2008).
- [20] J. H. Warner, F. Schaffel, M. Rummeli, A. Bachmatiuk, Graphene: Fundamentals and emergent applications, 1st Ed. (Elsevier, 2012).
- [21] M. Gueron, Density of the conduction electrons at the nuclei in Indium Antimonide, Phys. Rev., 135 A1, (1964).
- [22] A. A. Kordyuk, S. V. Borisenko, M. S. Golden, S. Legner, K. A. Nenkov, M. Knupfer, J. Fink, H. Berger, L. Forro, R. Follath, Doping dependence of the Fermi surface in ( Bi, Pb ) 2 Sr CaCu O 2 2 8+ δ , Phys. Rev. B, 66 , 014502, (2002).
|
10.1016/j.physleta.2014.05.044
|
[
"Alhun Aydin",
"Altug Sisman"
] |
2014-08-05T11:29:54+00:00
|
2014-08-05T11:29:54+00:00
|
[
"cond-mat.quant-gas"
] |
Discrete nature of thermodynamics in confined ideal Fermi gases
|
Intrinsic discrete nature in thermodynamic properties of Fermi gases appears
under strongly confined and degenerate conditions. For a rectangular
confinement domain, thermodynamic properties of an ideal Fermi gas are
expressed in their exact summation forms. For 1D, 2D and 3D nano domains,
variations of both number of particles and internal energy per particle with
chemical potential are examined. It is shown that their relation with chemical
potential exhibits a discrete nature which allows them to take only some
definite values. Furthermore, quasi-irregular oscillatory-like sharp peaks are
observed in heat capacity. New nano devices can be developed based on these
behaviors.
|
1408.1087v1
|
## Competitive performance analysis of two evolutionary algorithms for routing optimization in graded network
## Kavitha Sooda
## T. R. Gopalakrishnan Nair
Advanced Networking Research Group (RIIC), DSI and Asst. Professor, Dept. of CSE Nitte Meenakshi Institute of Technology Bangalore - 560064, India e-mail: [email protected]
Abstract -In this paper we compare the two intelligent route generation system and its performance capability in graded networks using Artificial Bee Colony (ABC) algorithm and Genetic Algorithm (GA). Both ABC and GA have found its importance in optimization technique for determining optimal path while routing operations in the network. The paper shows how ABC approach has been utilized for determining the optimal path based on bandwidth availability of the links and determines better quality paths over GA. Here the nodes participating in the routing are evaluated for their QoS metric. The nodes which satisfy the minimum threshold value of the metric are chosen and enabled to participate in routing. A quadrant is synthesized on the source as the centre and depending on which quadrant the destination node belongs to, a search for optimal path is performed. The simulation results show that ABC speeds up local minimum search convergence by around 60% as compared to GA with respect to traffic intensity, and opens the possibility for cognitive routing in future intelligent networks.
Keywords- Graded Network; Evolutionary Algorithm; ABC; GA; Optimal path; Agent
## I. INTRODUCTION
In recent years Evolutionary algorithms has been applied for understanding the collective behavior of known population. Owing to the rapid increase in the end-users and new service demands in Internet, it has confronted with problems like scalability and performance issues along with increasing complexity as the network expands. There exist a requirement of Internet or future global network test-beds which support reliable end-to-end connectivity for federation and management of Internet. On this complex structure of Internet, there still exists a challenge of routing and imbibing intelligence among the routers. Current routing strategy whether it is wired or wireless routing protocols, all of them are based on some predetermined calculation or prior knowledge of the connection between the nodes. This information was simply based on the request packet sent. Generally for any intelligent action [1] to take place a collective behavior of the topology and elements, need to be studied prior to the path determination process. Also there exists a need for enabling a self organizing feature of the topology when changes occur. Such scenario where there is
ARAMCO Endowed Chair- Technology, PMU, KSA Advanced Networking Research Group VP, Research, Research and Industry Incubation Center (RIIC), Dayananda Sagar Institutions, Bangalore - 560078, India e-mail: [email protected], www.trgnair.org
a requirement for decision-taking process by router, evolutionary intelligence [2-3] has played a pre-dominant role in the recent past. It is a self- adaptive technique which can be applied for routing where the desired result is obtained based on the exchange of the information by the agents.
The remainder of the paper is organized as follows: Section 2 outlines the overview of Network Awareness and Graded network. Section 3 deals with ABC algorithm and GA. Section 4 present the details of the mathematical model. Section 5 consists of simulation results. Finally, the last section concludes the paper.
## II. NETWORK AWARENESS AND GRADED NETWORK
## A. Network Awareness
The awareness of a network helps in taking decision for routing, as it creates an understanding of the nodal-link setup and its pairing performance better. It supplements the knowledge of network characteristics when deciding on the forward channels to route. There exists a need for intelligent querying which is required by the nodes to participate in network routing. Existing intelligent routing technique works well for small scale systems. Thus creation of awareness of the environment will make the routing easy for largely distributed nodes as discussed in [4]. The exploitation of the environmental knowledge is aided by the topological setup and bandwidth availability. The nodes are queried for identification of better ones which is based on the observed facts. Here the observed facts include the network QoS metrics like node density (ND), resource allocation (RA), traffic congestion (TC), network lifetime (NL), delay and bandwidth. Querying here is to test whether a given QoS metric for the node is above or below a threshold value. This helps in identifying the nodes which are feasible for routing. The algorithm, initially assume that any given node is a feasible candidate for obtaining the optimal path. Decision is taken on nodes based on the query operator which will help to bifurcate the non-feasible nodes. Next the neighbor of the previous feasible node is considered and the process is repeated until the destination node is reached. Here the network aware algorithm is based on exploration of the neighbor nodes as illustrated in a simple way in Fig. 1.
Fig. 1. Network Awareness Routing
Thus the environment awareness [5] can extract the nodes which are feasible to propagate the data in an efficient manner. This leads to an improved performance due to reduction in the search space and provides feasible path for the data to reach the desired destination in an efficient manner.
Fig. 2. Network Awareness algorithm
The network awareness algorithm physical network where the nodes which participate in routing are available. Now a quadrant is designed where the quad is abstracted in vitrual circuit plane with source as the centre point as shown in Fig. 2. Here the nodes which belong to the destination quadrant will participate in routing. Thus the optimal path is determined assuming sufficient nodes are available in selected quadrant of interest. is applied on a
## B. Graded network
The concept of graded network has been discussed by the authors in [6-8]. Grade value estimation method and grade surface creation for implementing intelligent routing in the autonomic network is a core focus of research. Grade is like a coagulated index reflection for intelligent systems to work in network; it is made available everywhere and routing will much depend on it. It signifies the quality of the router, which is in fact the knowledge of the environment. The router must be an intelligent entity because it performs different operations based on the environment information. It depends on input, output, load and resource availability. The graded network is setup with six QoS parameters. There are considered at two levels; level-1 and level-2. applied region-wise and the goal is to achieve favorable routing based on selected attributes. The values obtained from Level-1 must be able to eliminate the non-production node. Non-production nodes are those that come in the pitfall of congestion and possess less resource availability. The grade approach must be able to exact the QoS satisfied nodes from the network setup. The selected nodes are assigned values that indicate how efficient they are. At this Level-1 is point, we are able to calibrate the nodes for the routing process region-wise. The algorithm used to determine the optimal path requires proactive decision making on the output obtained. This is because many paths exist to reach the destination, and we have to choose the most optimal path. Once the gradient value has been calculated, it can be made available as pervasive information packets to all the other nodes to obtain the optimal path constitution.
For the first level of node selection we apply the operation. The nodes are observation. Prioritizing is performed based on ontological reasoning as shown in Fig. 3: level-1 prioritized based on local
## ELSE
P=6;
Fig. 3. Priority Model
Based on the priority values, nodes which have value greater than or equal to three are selected for routing.
## Delay calculation
We assume that the node requirements γ ik , channel cost di(Ci),the constants D, µ and the flow ( λ i) are given and feasible. Thus the Delay T at the node, is given by (1). location, external traffic
The average rate of message flow equal to sum of the average message flow rate of all paths that traverse this channel. The traffic entering the network λ i , on the i th channel is
$$( 1 )$$
from the external sources forms a Poisson's process with a mean γ jk (messages per second) originating at node j and destined for node k. Therefore total external traffic entering and considered to be equal to λ i . for those messages leaving the network are
Now the graded function, i.e., calculated values as its input of function defines the route availability for the set of nodes considered. This shall be calculated for all the available paths leading towards the destination node. The mean value of the gradient in the grade function shows the success level of operation of the network, bandwidth. Calculation for traffic congestion and bandwidth availability are further explored in section 4. Level-2 , considers the level-1 . The output of this which is based on the
In recent years swarm intelligence has been the focus of interest in research. Swarm intelligence is about developing an algorithm or a distributed problem solving device inspired by the collective behavior of social insect colonies and other animals.
## A. Artificial Bee Colony(ABC) for network routing
Artificial Bee colony algorithm is developed by Karaboga in 2005 [10], which was motivated by the intelligent behavior of the bees. It is an optimization tool, where search is performed in a distributed environment. Here the artificial bee fly and find the food source based on experience or their nest mates and adjusts their position. Some artificial bees (scouts) choose the food source randomly without experience. If the nectar amount of the new food source is better, then the bees memorize this position and forget the previous food source position. Thus ABC performs a local search with onlooker and employed bees, with global search methods managed by the onlooker and scout bees, attempting to balance exploration and exploitation process.
The following shows the main steps in the algorithm:
Send scouts to the initial food source.
## REPEAT
Step 1: Employed bees are sent to the food sources to determine the nectar amount.
Step 2: Based on the preference of the onlooker bees calculate the probability value of the sources.
Step 3: Send onlooker bees to the food source and determine the nectar amount.
- Step 4: The sources which are exhausted with the nectar, are not exploited further.
- Step 5: Randomly send the scouts for discovering new food sources.
Step 6: Remember the best food source obtained so far.
UNTIL (requirement is met)
## B. Genetic algorithms
Genetic algorithms are a part of evolutionary computing which has been used as a powerful optimization technique for finding near optimum solution is large search space. It is also an efficient search method that has been used for path selection in networks [12].
The algorithm is as follows:
## begin PATHSELECTION\_GA
Create initial population of n nodes randomly.
while generation\_count < k do
/* k = max. no. of generations.*/
begin
Selection Fitness Function Modified crossover Mutation
Increment generation\_count.
## end ;
Output the optimal path by selecting the highest probability value chromosome on which data can be sent.
end PATHSELECTION\_GA.
## ABC reasons: i)
outperformance GA for the following three
- GA produces new solutions based on the present solution at hand whereas in ABC produces candidate solution from taking the difference between randomly determined part of the parent and a randomly chosen solution from its population. This helps in the convergence speed of search into a local minimum.
- ii) Diversity in solution is obtained by mutation operation in GA whereas in ABC a new one is randomly inserted by a scout. This helps is global search and prevents premature convergence.
- iii) In GA there are three control parameters, i.e, crossover rate, mutation rate and number of generation where is ABC there is only one control parameter, i.e, limit with fixed colony size and maximum cycle number.
## C. Fitness function
The simulation topology is generated randomly. Here relaxation is imposed on the number of onlooker bees. The number of onlooker bees is equal to the number of neighbors to the source node who belong to the quadrant of the destination node. Here the fitness value is calculated on nodes which belong to the quadrant. The nodes which satisfy the threshold bandwidth values are the ones which get selected and participate in routing which indicates the fitness value. Further the authors have planned to implement multiobjective parameters solvable by the fitness function.
## IV. MATHEMATICAL MODEL
The Fig. 4 shows how the network evolves itself to determine the optimal path. Here we have proposed a channel level stochastic Markov process for realizing the link capacity for random network model. Every region is considered to be a Jackson network [13] in this model. Here the set of n links, L={1,2,3, …,n} are assumed to have the same capacity, where l GLYPH<31> L.
Fig. 4. Network Setup
Let the flow arrival rate between any two nodes, say s and r be represented by λ sr . Along the route where packet flows we assume that each flow transmits packet at fixed unit rate with mean µ -1 .
The traffic intensity of the link can be calculated as,
$$\text{Traffic Intensity} = & \frac { P s. T l ( t ) } { A b } & & ( 4 ) \quad \text{is above} \\ & & \text{determ}$$
Here Ps indicates the packet size, Ab indicates the available bandwidth and the traffic load on the link l GLYPH<31> L at time t is denoted by Tl(t). This represents the total number of flows currently routed across it. Link l's flow arrival rate is denoted by γ l . This flow represents a part of the flow arrival between s and e that is routed to link l. Here the flow departure rate is proportional to the current link load T (t) l and is given by Tl(t) * µ. The bandwidth availability on the link and packet size for the simulation is fixed.
The updates on the link is done every 30*t sec, which would represent the delay involved in building a grade network. Here the model is built based on the given information on traffic load and flow rate. Here we assume that the traffic arrivals during [0, t] are routed based on the link states T (0) available at time 0. The link state T (t) tracks l l to the following differential equations which derives the total traffic arrival rate γ l to link l.
Tl(t) * µ)
$$\frac { d T l ( t ) } { d t } = ( \gamma _ { 1 }. T _ { 1 } ) \\ \frac { d t } { ( 5 ) } _ { \tau \cdot \cdot \cdot }$$
This is of the form,
ݔ݀ Therefore the Integrating Factor(IF) is,
$$\frac { d y } { d \tau } + R y = S$$
Y(IF) = S(IF)dt +c ∫ Hence eq 5 becomes,
$$T _ { 1 } ( t ) \, ^ { = } \, T _ { 1 } ( 0 ) \, e ^ { \mu t } + \gamma _ { 1 } \, / \, \mu [ 1 - e ^ { \mu t } ]$$
]
where Tl(0) is the link state information at time interval [0,t]. This value is subtracted from link capacity to obtain the remaining fraction of the bandwidth available.
$$B _ { a } = L _ { c } - T _ { 1 } ( 0 ) ; & & ( 7 ) & & \mathcal { A } \, \text{alg}$$
This output is considered by the level-2 operation for calculating the available bandwidth. The traffic intensity is thus determined from Eq.(4), using Eq. (6) and Eq. (7). Equation 7, helps in the calculation of whether congestion has occurred or not. Here we have TC determined for the priority model.
The model is designed based on nodes and link generated at random. To each node, the jobs arrive from outside which are Poisson in process with rate α >0. Each arrival is independently routed to node j with probability p j ≥ 0 and ܲ ൌ 1 ୀଵ . Here k represents the total number of packets which has arrived from node i.
Thus the determination of the link capacity helps to identify where congestion has occurred or not. The agent here is used for aiding the grade network with relevant information when required. For instance at the first level of grading the parameters which are looked for are network lifetime, congestion level check, delay, node density and resource allocation. All these five parameters are determined by the agent. For instance the network lifetime is randomly assigned by the system for simulation purpose. The agent finds the remaining lifetime of the node and check whether it is above a threshold value. The bandwidth availability is determined by Eq. (7) which in turn helps to check whether congestion has occurred or not. Delay is obtained by Eq. (4). Network node density is number of packets arriving to the node and resource allocation is randomly assigned.
We then apply local algorithm trying to achieve global optimization, which is referred to as constrained optimization. Based on agent information, we applying the probabilistic reasoning approach to reduce congestion. Here the agent tries to do the following:
Minimize
|ሺAct
ܶ
ሺjሻ െ Cur
ܶ
ሺjሻ ሻ|
Subject to ∑ Act ܶ ሺjሻ E ڱ ேሺሻ
ڱ
ேሺሻ
Over 0 ≤ Tl (j) ≤ 1 for all j
Here Act Tl(j) represents the actual bandwidth, Cur Tl(j) represents the current bandwidth. The summation of actual bandwidth must be ≥ envisaged bandwidth for a particular route in the direction of destination which is considered by the level-2 operation.
Upon this selected nodes the optimization algorithm are applied to compare the performance of GA and ABC for determining the optimal path.
## V. SIMULATION RESULTS
The initial configuration of simulation model consists of a randomly distributed set of nodes and links in a two dimensional search space as shown in Fig. 5. Now propagating message, demands user to enter the source and destination node for which a quadrant appears with source as the centre. Depending on the destination node quad, the grade values of the nodes are determined. As mentioned in Eq. 1 and Eq.7, the delay and bandwidth availability will be calculated and made available in the knowledge base. The quadrant structure reduces the search space nearly to onefourth depending on the source location. Now the optimal path determination is performed only in the quadrant where destination belongs to. The paths are obtained with ABC and GA algorithms with initial parameters as shown in table 1.
Fig. 5. Simulation Snapshot
The grade value shown remains same for execution of both the algorithms. The design of the nodes and link setup takes care that there is no much randomness exist in the search space.
Table. II shows the hop count derived from the path obtained after applying ABC and GA on the quadrant where destination belongs to. The values obtained here is one of the random run carried out.
| Initial Parameter | Value |
|-------------------------------------------|--------------------------------------------------------|
| Population Size/Colony size | 15-1024 |
| Number of Iteration/ Maximum Cycle Number | 30 |
| Selection, crossover and mutation in GA | Roulette wheel, two-point crossover with 0.1% mutation |
| Packet size | 200 Bytes |
| Maximum Bandwidth | 30 Mbps |
The column two of the table shows the reduction in the number of nodes participating in the optimal path determination with grade value.
Here the number of onlooker bees is chosen based on the number of neighbors the source possess.
TABLE II. HOP COUNT
| Total nodes | Number of nodes selected | Graded ABC | Graded GA |
|---------------|----------------------------|--------------------|--------------------|
| Total nodes | Number of nodes selected | Route length | Route length |
| 15 | 10 | 3 | 4 |
| 16 | 12 | 4 | 3 |
| 32 | 18 | 5 | 8 |
| 64 | 38 | 5 | 7 |
| 128 | 61 | 6 | 8 |
| 256 | 64 | Path not available | Path not available |
| 512 | 234 | 7 | 7 |
| 1024 | 432 | 6 | 7 |
Fig. 6. Graph determining the Traffic Intensity in ABC and GA.
The employed bees now search for the best nectar among the available neighbors for each one of them. This search is continued until the destination is reached. Now among the different paths which exist between the source and destination one complete path is selected based on the bandwidth estimation of the links. Since here the links selected have bandwidth above a threshold value, there is less likely scenario for congestion to occur.
Fig. 7. Graph determining the throughput in ABC and GA.
From Fig. 6, we observe that the convergence speed for ABC is 60% better than GA. The graph in Fig. 7, shows that most of the path obtained by ABC was of better grade value hence the link selected were of higher bandwidth. Since the work is based on the quadrant design we may not always get the best paths by ABC or GA. The technique applied here assures reduction in the search space and helps in convergence of path faster with quality. From Table. II, we saw that they may not exist fruitful path in the quadrant. We also observed that GA learnt better quality paths 10% of the time as compared to ABC and 10% equal quality paths.
## VI. CONCLUSION AND FUTURE WORK
The paper here presented the optimal path determination using the evolutionary artificial bee colony and genetic algorithm approaches in general network. ABC showed a good performance in terms of path convergence rate as the control parameter of the algorithm was only total number of cycle as against to GA which had three control parameters. A quality path with sufficient bandwidth was determined over the graded network. Here we have shown that without a pre-defined message path we are able to obtain the paths much optimally realizing a multi-hop connecting route. The routers are able to dynamically select the forward nodes and propagate the messages successfully to the destination. However the simulation has certain limitations as it is on progressive path. Currently only few hundreds of nodes were used to test and evaluate the principles of the model.
Future work involves expanding the simulation of higher number of nodes realizing near real life network scenarios. The simulation setup shall be modified to a zone based structure where there shall be a zone coordinator which will be responsible improving scalability and for achieving parallelism. The remaining two self organization properties shall be further implemented where parallel search techniques would be explored for intelligent state space update.
## REFERENCES
- [1] T. R. Gopalakrishnan Nair, Kavitha Sooda, Abhijith, 'Transformation of networks using cognitive approaches', JRI, Vol.1, No. 1, December, 2008, pp.7-14.
- [2] H.F. Wedde, M. Farooq, Y. Zhang, Beehive: an efficient fault-tolerant routing algorithm inspired by honeybee behavior, ant colony, optimization and swarm intelligence, in: 4th International Workshop, ANTS 2004, Brussels, Belgium, September 5-8.
- [3] J. Kennedy, R.C. Eberhart, 'Particle Swarm Optimization', IEEE International Conference on Neural Networks, vol. 4, 1995, pp. 19421948.
- [4] T R Gopalakrishnan Nair, M Jayalalitha, S Abhijith, 'Cognitive Routing with Stretched Network Awareness through Hidden Markov Model Learning at Router Level', IEEE Workshop on Machine learning in Cognitive networks, Hong Kong, 2008.
- [5] Yanif Ahmad Ugur Cetintemel 'Network-Aware Query Processing for Stream-based Applications', Proceedings of the 30th VLDB Conference, Toronto, Canada, 2004.
- [6] T. R. Gopalakrishnan Nair, Kavitha Sooda, 'Particle Swarm Optimization for Realizing Intelligent Routing in Networks with Quality Grading', WiCOM, 7 th IEEE International Conference on Wireless Communications, Networking and Computing, Wuhan, China, 23-25 th September, pp. 1-4, 2011.
- [7] T. R. Gopalakrishnan Nair, Kavitha Sooda,' Application of Genetic Algorithm on Quality Graded Networks for Intelligent Routing', World Congress on Information and Communication Technologies2011, December 2011, Mumbai, India, pp. 558-563, DOI: 10.1109/WICT.2011.6141306.
- [8] Kavitha Sooda, T. R. Gopalakrishnan Nair,'A Comparative Analysis for Determining the Optimal Path Using PSO and GA', International Journal of Computer Applications, Volume 32- No.4, October 2011, pp no. 8-12.
- [9] V. Tereshko, Reaction-diffusion model of a honeybee colony's foraging behaviour, in: M. Schoenauer (Ed.), Parallel Problem Solving from Nature VI, Lecture Notes in Computer Science, vol. 1917, Springer-Verlag, Berlin, 2000, pp. 807-816.
- [10] D. Karaboga, An idea based on honeybee swarm for numerical optimization, Technical Report TR06, Erciyes University, Engineering Faculty, Computer Engineering Department, 2005.
- [11] D. Karaboga, B. Basturk, On the performance of artificial bee colony (ABC) algorithm, Applied Soft Computing 8 (1) (2008), pp.687-697.
- [12] Shengxiang Yang, Hui Cheng and Fang Wang, ' Genetic Algorithms with Immigrants and memory schemes for dynamic shortest path routing problems in mobile AdHoc networks', IEEE Transactions on systems, man, and cybernetics - applications and reviews, vol. 40, no. 1, Jan 2010.
- [13] J. Medhi, Stochastic Models in Queueing theory , Elsevier, Academic press, 2006.
|
10.1109/IAdCC.2013.651430
|
[
"Kavitha Sooda",
"T. R. Gopalakrishnan Nair"
] |
2014-08-05T18:12:58+00:00
|
2014-08-05T18:12:58+00:00
|
[
"cs.NI"
] |
Competitive performance analysis of two evolutionary algorithms for routing optimization in graded network
|
In this paper we compare the two intelligent route generation system and its
performance capability in graded networks using Artificial Bee Colony (ABC)
algorithm and Genetic Algorithm (GA). Both ABC and GA have found its importance
in optimization technique for determining optimal path while routing operations
in the network. The paper shows how ABC approach has been utilized for
determining the optimal path based on bandwidth availability of the links and
determines better quality paths over GA. Here the nodes participating in the
routing are evaluated for their QoS metric. The nodes which satisfy the minimum
threshold value of the metric are chosen and enabled to participate in routing.
A quadrant is synthesized on the source as the centre and depending on which
quadrant the destination node belongs to, a search for optimal path is
performed. The simulation results show that ABC speeds up local minimum search
convergence by around 60% as compared to GA with respect to traffic intensity,
and opens the possibility for cognitive routing in future intelligent networks.
|
1408.1088v3
|
## Enumeration of monochromatic three term arithmetic progressions in two-colorings of any finite group
Erik Sj¨land o
November 11, 2014
## Abstract
There are many extremely challenging problems about existence of monochromatic arithmetic progressions in colorings of groups. Many theorems hold only for abelian groups as results on non-abelian groups are often much more difficult to obtain. In this research project we do not only determine existence, but study the more general problem of counting them. We formulate the enumeration problem as a problem in real algebraic geometry and then use state of the art computational methods in semidefinite programming and representation theory to derive lower bounds for the number of monochromatic arithmetic progressions in any finite group.
## Contents
| 1 | Introduction | 2 |
|-----|---------------------------------------------------|-----|
| 2 | Results | 2 |
| 3 | Polynomial optimization | 3 |
| 4 | Exploiting symmetries in semidefinite programming | 4 |
| 5 | Problem formulated as a semidefinite program | 6 |
| 6 | Proof of Theorem 2.1 | 9 |
| 7 | Methods for longer arithmetic progressions | 12 |
| 8 | Related problems | 12 |
## 1 Introduction
In Ramsey theory colors or density are commonly used to force structures. It follows from Szemer´ edi's theorem that there exist monochromatic arithmetic progressions of any length if we color the integers with a finite number of colors. Similarly there are arithmetic progressions of any length in any subset of the integers of positive density.
We can explore finite versions of these statements if we replace the integers by [ n ], or Z n . In this article we will focus on the cyclic group, as it is easier to analyze because of the symmetries. In most of the literature the posed question has been about how large n has to be for us to find a monochromatic progression of a desired length when Z n is colored by a fixed number of colors. A more difficult problem is to determine how many monochromatic progressions there are of desired length when Z n is colored by a fixed number of colors given n .
Many results on monochromatic progressions are first obtained for cyclic groups and then extended to results for other groups. Many results can be extended to abelian groups using the same machinery as for the cyclic group, whereas results about non-abelian groups are usually very difficult and require a completely different set of tools.
In this paper we reformulate the problem of counting monochromatic arithmetic progressions to a problem in real algebraic geometry that can be attacked by state of the art optimization theory. The methods allows us to find a lower bound to the amount of monochromatic progressions in any finite group, including non-abelian groups. One could find optimal lower bounds for small groups by explicitly counting them for all different colorings, but as this can be done only for very small groups it has not been considered in this article.
In the next section we present our results. In later sections we present our methods and proofs.
## 2 Results
The main theorem of this paper holds for any finite group G , including nonabelian groups for which very little is known about arithmetic progressions. The only information that is needed to get a lower bound for a specific group G is the number of elements of the different orders of G . The lower bound is sharp for some groups, for example Z p with p prime, but is not optimal for most groups. We have further included Table 1 where we have calculated the lower bound for some small symmetric groups.
Theorem 2.1. Let G be any finite group and let R (3 , G, 2) denote the minimal number of monochromatic 3 -term arithmetic progressions in any two-coloring of G . Let G k denote the set of elements of G of order k , N = | G | and N k = | G k | . Denote the Euler phi function φ k ( ) = |{ t ∈ { 1 , . . . , k } : t and k are coprime }| . Let K = { k ∈ { 5 , . . . , n } : φ k ( ) ≥ 3 k 4 } . For any G there are ∑ n k =4 N N · k 2 + N N · 3 24
Table 1: Calculation of ∑ k ∈ K N N · k 8 (1 -3 k -φ k ( ) φ k ( ) ) for small symmetric groups
| Group G | Number of 3-APs | Lower bound for R (3 , G, 2) |
|-----------|-------------------|--------------------------------|
| S 5 | 4540 | 90 |
| S 6 | 205440 | 3240 |
| S 7 | 11307660 | 306180 |
| S 8 | 774278400 | 16208640 |
arithmetic progressions of length 3. At least
$$R ( 3, G, 2 ) \geq \sum _ { k \in K } \frac { N \cdot N _ { k } } { 8 } ( 1 - 3 \frac { k - \phi ( k ) } { \phi ( k ) } )$$
of them are monochromatic in a 2-coloring of G .
## 3 Polynomial optimization
To obtain the main theorem we use methods from real algebraic geometry. It is important to note that the proof of the main theorem can be understood without this section, even though the methods played a vital role when finding the sum of squares based certificate in the proof. In this article we only give the elementary definitions relating to polynomial optimization that are needed to prove the results. For a more extensive review of the topic we refer to [8], and for implementation aspects we refer to [17].
Given polynomials f ( x , g ) 1 ( x , . . . , g ) m ( x ) we define a polynomial optimization problem to be a problem on the form
$$\rho _ { * } = \inf \quad f ( x ) \\ \text{subject to } \quad g _ { 1 } ( x ) \geq 0, \dots, g _ { m } ( x ) \geq 0, \\ x \in \mathbb { R } ^ { n },$$
The problem can be reformulated as follows
$$\begin{array} { c c } \rho ^ { * } = \sup & \lambda \\ \text{subject to } & f ( x ) - \lambda \geq 0, g _ { 1 } ( x ) \geq 0, \dots, g _ { m } ( x ) \geq 0 \\ & \lambda \in \mathbb { R }, x \in \mathbb { R } ^ { n } \end{array}$$
where f ( x , g ) 1 ( x , . . . , g ) m ( x ) are the same polynomials as before. We refer to the book by Lasserre [7] for an extensive discussion on the relationship between these problems. It is for example easy to see that ρ ∗ = ρ ∗ .
One of the most challenging problems in real algebraic geometry is to find the most useful relationships between nonnegative polynomials and sums of squares. These relationships are known as Positivstellens¨tze. a Let X be a
formal indeterminate and let us introduce the following optimization problem:
$$\sigma ^ { * } = \sup & \quad \lambda \\ \text{subject to} & \quad f ( X ) - \lambda = \sigma _ { 0 } + \sum _ { i = 1 } ^ { m } \sigma _ { i } g _ { i } \\ & \quad \sigma _ { i } \text{ is a sum of squares.}$$
One can easily see that σ ∗ ≤ ρ ∗ holds, and under some technical conditions (Archimedean) it was proven by Putinar that σ ∗ = ρ ∗ . The equality holds because of Putinar's Positivstellensatz, which is discussed further in [17]. The optimization problem can be relaxed by bounding the degrees of all involved monomials to d , lets call the solution σ ∗ d . It holds that σ ∗ d 1 ≤ σ ∗ d 2 for d 1 < d 2 , and it was proven by Lasserre [6] that if the Archimedean condition hold, then
$$\lim _ { d \to \infty } \sigma _ { d } ^ { * } = \sigma ^ { * } = \rho ^ { * } = \rho _ { * }.$$
The positivity condition is rewritten as a sum of squares condition because the latter is equivalent to a semidefinite condition: f ( X ) is a sum of squares of degree 2 d if and only if f ( X ) = v T d Qv d for some positive semidefinite matrix Q , where v d is the vector of all monomials up to degree d . This makes it possible to use semidefinite programming to find σ ∗ d as well as a sum of squares based certificate for the lower bound of our original polynomial, f ( X ) = σ ∗ d + σ 0 + ∑ m i =1 σ g i i ≥ σ ∗ d .
## 4 Exploiting symmetries in semidefinite programming
This section can be skipped by the reader who is just interested in the final proof of this article. The methods in this section were used and implemented to find parts of the numerical results that lead to the final proof, and is thus intended for the reader who wants to understand the full process from problem formulation to the end certificate, or solve similar problems with symmetries.
Let C and A , . . . , A 1 m be real symmetric matrices and let b 1 , . . . , b m be real numbers. To reduce the order of the matrices in the semidefinite programming problem
$$\max \{ \text{tr} ( C X ) \ | \ X \text{ positive semidefinite, tr} ( A _ { i } X ) = b _ { i } \text{ for } i = 1, \dots, m \}$$
when it is invariant under all the actions of a group is the goal of this section.
As in [5] and [4], we use a ∗ -representation to reduce the dimension of the problem. For the reader interested in ∗ -algebras we recommend the book by Takesaki [18]. A collection of several new methods, including the one we use, to solve invariant semidefinite programs can be found in [1]. Other important recent contributions in this area include [3, 2, 19, 10, 11, 14].
Definition 4.1. A matrix ∗ -algebra is a collection of matrices that is closed under addition, scalar multiplication, matrix multiplication and transposition.
Let G be a finite group that acts on a finite set Z . Define a homomorphism h : G → S Z , where S Z is the group of all permutations of Z . For every g ∈ G there is a permutation h g = h g ( ) of the elements of Z with the properties h gg ′ = h h g g ′ and h g -1 = h -1 g . For every permutation h g , there is a corresponding permutation matrix M g ∈ { 0 1 , } Z × Z , element-wise defined by
$$( M _ { g } ) _ { i, j } = \begin{cases} \ 1 & \text{if $h_{g}(i)=j$,} \\ 0 & \text{otherwise} \end{cases}$$
for all i, j ∈ Z . Let the span of these permutation matrices define the matrix ∗ -algebra
$$\mathcal { A } = \left \{ \sum _ { g \in G } \lambda _ { g } M _ { g } \ | \ \lambda _ { g } \in \mathbb { R } \right \}.$$
The matrices X satisfying XM g = M X g for all g ∈ G are the matrices invariant under the action of G . The ∗ -algebra consisting of the collection of all such matrices,
$$\mathcal { A } ^ { \prime } = \{ X \in \mathbb { R } ^ { n \times n } | X M = M X \text{ for all } M \in \mathcal { A } \},$$
is known as the commutant of A . We let d = dim A ′ denote the dimension of the commutant.
The commutant has a basis of { 0 1 , } -matrices, which we denote E , . . . , E 1 d , with the property that ∑ d i =1 E i = J , where J is the all-one Z × Z -matrix. We form a new basis for the commutant by normalizing the E i s
$$B _ { i } = \frac { 1 } { \sqrt { t r ( E _ { i } ^ { T } E _ { i } ) } } E _ { i }. \\ \text{pertry that } \text{tr} ( B _ { i } ^ { T } B _ { i } ) = \delta _ { i. i }, \text{v}$$
The new basis has the property that tr( B B T i j ) = δ i,j , where δ i,j is the Kronecker delta.
From the new basis we introduce multiplication parameters λ k i,j by
$$B _ { i } B _ { j } = \sum _ { k = 1 } ^ { d } \lambda _ { i, j } ^ { k } B _ { k }$$
for i, j, k = 1 , . . . , d .
We introduce a new set of matrices L , . . . , L 1 d by
$$( L _ { k } ) _ { i, j } = \lambda _ { k, j } ^ { i }$$
for k, i, j = 1 , . . . , d . The matrices L , . . . , L 1 d are d × d matrices that span the linear space
$$\mathcal { L } = \{ \sum _ { i = 1 } ^ { d } x _ { i } L _ { i } \colon x _ { 1 }, \dots, x _ { d } \in \mathbb { R } \}.$$
Theorem 4.2 ([5]) . The linear function φ : A → ′ R d × d defined by φ B ( i ) = L i for i = 1 , . . . , d is a bijection, which additionally satisfies φ XY ( ) = φ X φ Y ( ) ( ) and φ X ( T ) = φ X ( ) T for all X,Y ∈ A ′ .
Corollary 4.3 ([5]) . ∑ d i =1 x B i i is positive semidefinite if and only if ∑ d i =1 x L i i is positive semidefinite.
If it is possible to find a solution X ∈ A ′ , then Corollary 4.3 can be used to reduce the size of the matrix in the linear matrix inequality. Let us show that this is possible:
Lemma 4.4. There is a solution X ∈ A ′ to a G -invariant semidefinite program max tr( { CX ) | X positive semidefinite , tr( A X i ) = b i for i = 1 , . . . , m } .
Proof. Let C, A , . . . , A 1 m be Z × Z matrices commuting with M g for all g ∈ G . If X is an optimal solution to the optimization problem then the group average, X ′ = 1 | G | ∑ g ∈ G M XM g T g , is also an optimal solution: It is feasible since
$$i \stackrel { \sim } { \sim } g \stackrel { \sim } { \sim } g, \, \circ \, \cos \, \alpha \, \varphi \, \max \, \cos \, \nu \, \cos \, } \\ \text{tr} ( A _ { j } X ^ { \prime } ) \ \ & = \text{tr} ( A _ { j } \frac { 1 } { | G | } \sum _ { g \in G } M _ { g } X M _ { g } ^ { T } ) \\ \ \ & = \text{tr} ( \frac { 1 } { | G | } \sum _ { g \in G } M _ { g } A _ { j } X M _ { g } ^ { T } ) \\ \ \ & = \text{tr} ( \frac { 1 } { | G | } \sum _ { g \in G } A _ { j } X ) \\ \ \ & = \text{tr} ( A _ { j } X ), \\ \text{used that the well known fact that the trace is in}$$
where we have used that the well-known fact that the trace is invariant under change of basis. By the same argument tr( CX ′ ) = tr( CX ), which implies that X ′ is optimal. It is easy to see that X ′ ∈ A ′ .
The following theorem follows, which gives a tremendous computational advantage when d is significantly smaller than | Z | :
Theorem 4.5 ([5]) . The G -invariant semidefinite program
$$\max \{ \text{tr} ( C X ) \ | \ X \succeq 0, \text{tr} ( A _ { i } X ) = b _ { i } \text{ for } i = 1, \dots, m \}$$
has a solution X = ∑ d i =1 x B i i that can be obtained by
$$\max \{ \text{tr} ( C X ) \ | \ \sum _ { i = 1 } ^ { d } x _ { i } L _ { i } \succeq 0, \text{tr} ( A _ { i } \sum _ { j = 1 } ^ { d } B _ { j } x _ { j } ) = b _ { i } \text{ for } i = 1, \dots, m \}.$$
## 5 Problem formulated as a semidefinite program
We follow the convention that an arithmetic progression in G of length k is a set of k distinct element { a, b · a, . . . , b k -1 · a } where a ∈ G b , ∈ Z + . Also, as { 1 2 3 , , } , { 1 3 2 , , } , { 2 1 3 , , } , { 2 3 1 , , } , { 3 1 2 , , } and { 3 2 1 , , } denote the same set they are considered as the same arithmetic progression, thus when summing over all arithmetic progressions the set is only considered once.
Let χ : G →{-1 1 , } be a 2-coloring of the finite group G , and for simplicity let x g = χ g ( ) for all g ∈ G . Furthermore, let x G denote the vector of all variables x g . Let us introduce the polynomial
$$p ( x _ { a }, x _ { b }, x _ { c } ) \ & = \frac { ( x _ { a } + 1 ) ( x _ { b } + 1 ) ( x _ { c } + 1 ) - ( x _ { a } - 1 ) ( x _ { b } - 1 ) ( x _ { c } - 1 ) } { \quad } \quad & = \frac { x _ { a } x _ { b } + x _ { a } x _ { c } + x _ { b } x _ { c } + 1 } { 4 },$$
where a, b, c ∈ G , which has the property that
$$p ( x _ { a }, x _ { b }, x _ { c } ) = \begin{cases} \ 1 & \text{if $x_{a}=x_{b}=x_{c}$} \\ 0 & \text{otherwise.} \end{cases}$$
The polynomial p is one when a, b, c are the same color and zero otherwise. It follows that
$$R ( 3, G, 2 ) = \min _ { x _ { G } \in \{ - 1, 1 \} ^ { | G | } } \sum _ { \{ a, b, c \} \text{ is an A.P. in } G } p ( x _ { a }, x _ { b }, x _ { c } ).$$
The integer problem is relaxed to a problem on the hypercube to obtain a lower bound for R (3 , G, 2) . Since any solution of the integer program is also a solution to the hypercube problem we have
$$R ( 3, G, 2 ) & \geq \min _ { x _ { G } \in [ - 1, 1 ] | ^ { G | } } \sum _ { \{ a, b, c \} \text{ is an } A.P. \text{ in } G } p ( x _ { a }, x _ { b }, x _ { c } ) \\ & = \min _ { x _ { G } \in [ - 1, 1 ] | ^ { G | } } \sum _ { \{ a, b, c \} \text{ is an } A.P. \text{ in } G } x _ { a } x _ { b } + x _ { a } x _ { c } + x _ { b } x _ { c } + 1 \\ & = \min _ { x _ { G } \in [ - 1, 1 ] | ^ { G | } } \frac { p _ { G } } { 4 } + \sum _ { \{ a, b, c \} \text{ is an } A.P. \text{ in } G } \frac { 1 } { 4 } \$$
where
$$p _ { _ { G } } = \sum _ { \{ a, b, c \}$$
We immediately get a lower bound for R (3 , G, 2) by finding a lower bound for the homogeneous degree 2 polynomial p G . The coefficient of x x a b in p G equals the number of times the pair ( a, b ) occurs in a 3-arithmetic progression, which depends on the group G . After the coefficients are found the state-of-the-art methods surveyed in Sections 3 and 4 are used to find a lower bound for
$$\min _ { x _ { G } \in [ - 1, 1 ] ^ { | G |$$
Let us use the degree 3 relaxation of Putinar's Positivstellensatz, and let λ ∗ denote the maximal lower bound using this relaxation. Denote the elements of G by g , . . . , g 1 | G | and let v be the vector of all monomials of degree less or equal
than one in the formal indeterminates X g 1 , . . . , X g | G | ; v = [1 , X g 1 , . . . , X g | G | ] T . We get
$$\begin{array} {$$
For simplicity, let the Q -matrices entries be indexed by the set { 1, g 1 , . . . , g | G | } . Let A /multicloseright B denote the semidirect product of A and B , defined such that ( a, b ) ∈ A /multicloseright B takes i to a + bi (here + denotes the action of the group G , and bi denotes the repeated action of i on itself b times: bi = ∑ b j =1 i ). Arithmetic progressions are invariant under affine transformations: if ( a, b ) ∈ G /multicloseright Z + , and { i, j, k } is an arithmetic progression, then ( a, b ) · { i, j, k } = { a + bi, a + bj, a + bk } is also an arithmetic progressions. It follows that the semidefinite program is invariant under affine transformations, and thus that we can find an invariant solution by Lemma 4.4. This implies that the degree 3 relaxation has a solution for which Q 0 ( g , g i j ) = Q 0 ( a + bg , a i + bg j ) and Q 0 (1 , g i ) = Q 0 ( g , i 1) = Q 0 (1 , a + bg i ) for all ( a, b ) ∈ G /multicloseright Z + and g , g i j ∈ G . Also Q + g i ( g , g j k ) = Q + a + bg i ( a + bg , a j + bg k ), Q + g i (1 , g k ) = Q + g i ( g k , 1) = Q + a + bg i (1 , a + bg k ) and Q + g i (1 , 1) = Q + a + bg i (1 , 1). By the same argument we get similar equalities for the indices of the matrices Q -g i . From the equalities It is easy to see that the matrices Q + g i and Q -g i are obtained by simultaneously permuting the rows and columns of Q + g j and Q -g j respectively, and hence it is enough to require that Q + g 1 and Q -g 1 are positive semidefinite where g 1 ∈ G is the identity element.
In conclusion we see that the number of variables can be reduced significantly, and that only three | G | +1 ×| G | +1-matrices are required to be positive semidefinite instead of the 2 | G | +1 matrices required in the original formulation. The dimension of the commutant is small for some groups, and in those cases the size of these matrices can be reduced further using Theorem 4.5.
The certificates for the lower bounds we get from these methods are numerical, and to make them algebraic additional analysis of the numerical data, and possibly further restrictions, must be done. There is no general way to find algebraic certificates, and for many problem it might not even be possible. In this paper it is a vital step to find as the numerical certificates only gives certificates for one group at the time whereas we need a certificate for all groups.
It is in general very difficult to go from numerical patterns to algebraic certificates. For a specific group all information required to find a lower bound can be found in an eigenvalue decomposition of the involved matrices, but there is no general way of finding the optimal algebraic lower bound when the different eigenvalues and eigenvectors have decimal expansions that cannot trivially be translated into algebraic numbers. If one is interested in a rational approximation to the lower bound one can use methods by Parrilo and Peyrl [12]. These methods gives a good certificate for a specific group but does not help when one want to find certificates for an infinite family of groups.
To find a certificate for all groups, one of the tricks we used was to restrict the SDP above by requiring that some further entires equal one another, in order to at least get a lower bound for λ ∗ . There are also many other ways to restrict the SDP further, including setting elements to zero. Restricting the SDP either decreases the optimal value or reduces the number of solutions to the original problem. In the ideal scenario one can keep restricting the SDP until there is only one solution, without any change in the optimal value.
## 6 Proof of Theorem 2.1
To make the computations in the proofs that follow readable we introduce additional notation:
$$\sigma ( a ; b _ { 0 }, b _ { 1 }, \dots,$$
By elementary calculations we have the following equalities, which are needed in the proofs:
$$I _ { 1 } = \sum _ { i \in \mathbb { Z } _ { n$$
$$I _ { 2 } = \left ( \sum _ { i \in \mathbb { Z } _ { n } } X _ { i } \right ) ^ { 2 } = \sigma \left ( 0 ; 1, 2, \dots, 2 \right ),$$
$$I _ { 3 } ^ { j } = \frac { 1 } { 2 } \sum _ { i \in \mathbb { Z } _ { n } } ( X _ { i } - X _ { i + j } ) ^ { 2 } = \sigma \left ( 0 ; 1, 0, \dots, 0, - 1, 0, \dots, 0, - 1, 0, \dots, 0 \right ),$$
The first one is non-negative since we for the problem require that -1 ≤ x i ≤ 1 and the other polynomials are non-negative since they are sums of squares.
Proof of Theorem 2.1. Let us count all arithmetic progressions { a, b · a, b · b · a } , with ( a, b ) ∈ G × G k . There is a cyclic subgroup of G with elements { 1 , b, b 2 , . . . , b k -1 } . Let us write k = p e 1 1 p e 2 2 · · · p e s s for 1 < p 1 < · · · < p s distinct primes and e , . . . , e 1 s positive integers. The elements { b p i , b 2 p i , . . . , b ( k p i -1) p i } are of order less than k for all i whereas the elements U = { b t : t and k are coprime } are of order k . Note that φ k ( ) = | U | . If k < 3 there are no arithmetic progressions, so in all the following calculations we will always assume that k ≥ 3. If 3 /notbar k , then all triples { (1 , b i , b 2 i ) : b i ∈ U } will be distinct arithmetic progressions. Since there are N k φ k ( ) different cyclic subgroups of G of this type there are φ k ( ) N k φ k ( ) = N k arithmetic progressions of the form { 1 , b i , b 2 i } , where b i is an element of order k . Since { a, b · a, b · b · a } is an arithmetic progression if and only if { 1 , b, b · b } is an arithmetic progression it follows that there are N N · k 2 arithmetic progressions with ( a, b ) ∈ G × G k if 3 /notbar k . If 3 | k we have to be careful so that we do not count arithmetic progressions of the form { a, b k 3 · a, b 2 k 3 · a } three times. Since b k 3 will be of lower order than k if k > 3 it follows that this
kind of arithmetic progressions only occur for ( a, b ) ∈ G × G 3 . We conclude that there are N N · k 2 arithmetic progressions with ( a, b ) ∈ G × G k if k > 3, and N N · 3 6 if ( a, b ) ∈ G × G 3 .
The number of monochromatic arithmetic progressions is given by
$$R ( 3, G, 2 ) \ = \ \sum _ { \{ a, b, c \} \text{ is an } A. P. \text{ in } G } p ( x _ { a }, x _ { b }, x _ { c } ) \\ = \ \sum _ { \{ a, b, c \} \text{ is an } A. P. \text{ in } G } \frac { x _ { a } x _ { b } + x _ { a } x _ { c } + x _ { b } x _ { c } + 1 } { 4 }.$$
Let us rewrite this as
$$R ( 3, G, 2 ) = \frac { \sum _ { k = 4 } ^ { n } \frac { N \cdot N _ { k } } { 2 } + p _ { k } } { 4 } + \frac { N \cdot N _ { 3 } } { 2 4 } + \frac { p _ { 3 } } { 4 },$$
where p k is the sum of all polynomials x x s t s · + x x s t t s · · + x t b · x t t s · · where t is of order k . Let us also define the further reduced polynomial p ( a,b ) k for a fixed pair ( a, b ) ∈ G × G k by
$$p _ { k } ^ { ( a, b ) } = \sum _ { 0 \leq i < j \leq k - 1 } c _ { b ^ { i } \cdot a, b ^ { j } \cdot a } x _ { b ^ { i } \cdot a } x _ { b ^ { j } \cdot a }$$
/negationslash where c b i · a,b j · a denotes how many times the pair ( b i · a, b j · a ) is in an arithmetic progression { s, t · s, t · t · s } with t of order k . When a = 1 we will have a set { b i 1 , . . . , b i N k /φ k ( ) } of elements of order k such that b c 1 i j = b c 2 i k for all j, k ∈ { 1 , . . . , N k /φ k ( ) } and all c 1 , c 2 ∈ { 1 , . . . , k - } 1 , and can write
$$p _ { k } = \frac { 1 } { k } \sum _ { a \in G } \sum _ { j = 1 } ^ { N _ { k } / \phi ( k ) } p _ { k } ^ { ( a, b _ { i _ { j } } ) }.$$
It follows that if p (1 ,b ) k ≥ c for a b of order k , then p k ≥ N N k φ k k ( ) c .
/negationslash
Let k mod 2 = 1, k mod 3 = 0: It is fairly easy to see that if b i is of order k , then so is b -i and b 2 i . Also, if 2 | i then b i/ 2 is of order k , if 2 /notbar i then b ( n + ) i / 2 is of order k . Hence any b i is in both arithmetic progressions b -i , 1 , b i and 1 , b i , b 2 i , and also in either 1 , b i/ 2 , b i or 1 , b ( n + ) i / 2 , b i . These are all arithmetic progressions containing both 1 and b i . Note that the elements that are not of order k (apart from the identity element) are not in any of these arithmetic progressions, hence if k = p e 1 1 . . . p e s s , then
$$p _ { k } ^ { ( 1, b ) } ( X ) = \sum _ { i, j \in \mathbb { Z } _ { k } } b _ { j - i } X _ { i } X _ { j } = \sigma ( 0 ; b _ { 0 }, b _ { 1 }, \dots, b _ { k - 1 } )$$
where b tp i = 0 for t ∈ { 0 , . . . , k - } 1 and i ∈ { 1 , . . . , s } , and b t = 3 for all other t . In particular we get
$$p _ { k } ^ { ( 1, b ) } ( X ) = \frac { 3 } { 2 } I _ { 2 } + \frac { 3 } { 2 } \sum _ { t, i } I _ { 3 } ^ { t p _ { i } } + \frac { 3 } { 2 } ( 1 + k - \phi ( k ) - 1 ) I _ { 1 } - \frac { 3 } { 2 } ( k - \phi ( k ) ) k \geq - \frac { 3 } { 2 } ( k - \phi ( k ) ) k,$$
and it follows that
$$p _ { k } \geq - \frac { N } { k } \frac { N _ { k } } { \phi ( k ) } \frac { 3 } { 2 } ( k - \phi ( k ) ) k = - \frac { 3 } { 2 } \frac { N \cdot N _ { k } } { \phi ( k ) } ( k - \phi ( k ) ),$$
and furthermore that
$$\frac { N \cdot N _ { k } } { 2 } + p _ { k } \geq \frac { N \cdot N _ { k } } { 2 } ( 1 - 3 \frac { k - \phi ( k ) } { \phi ( k ) } ).$$
In cases when 1 -3 k -φ k ( ) φ k ( ) < 0, that is when φ k ( ) < 3 k 4 , we will instead use the trivial bound
$$\frac { N \cdot N _ { k } } { 2 } + p _ { k } \geq 0$$
k mod 2 = 0: b ∈ G k , and let us color all elements { 1 , b 2 , b 4 , . . . , b k -2 } blue and { b, b 3 , b 5 , . . . , b k -1 } red. In an arithmetic progression { a, c · a, c · c · a } with ( a, c ) ∈ { 0 , b, b 2 , . . . , b k -1 }×{ 0 , b, b 2 , . . . , b k -1 }∩ G k it holds that a and c · a are of different colors. The coloring can be extended too all pairs ( a, c ) ∈ G × G k , and thus there is a coloring without monochromatic arithmetic progressions. Hence we cannot hope to do better than the trivial bound using these methods:
$$\frac { N \cdot N _ { k } } { 2 } + p _ { k } \geq 0.$$
k mod 3 = 0: Let us color the elements { 0 , b, b 3 , b 4 , . . . , b k -3 , b k -2 } blue and { b 2 , b 5 , b 8 , . . . , b k -1 } red. As when k mod 2 = 0 we consider arithmetic progressions { a, c a, c c a · · · } with ( a, c ) ∈ { 0 , b, b 2 , . . . , b k -1 }×{ 0 , b, b 2 , . . . , b k -1 }∩ G k . It is easy to see that either a and c · a or a and c · c · a are of different colors. Again there is a coloring without monochromatic arithmetic progressions and so we cannot do better than the trivial bound:
$$\frac { N \cdot N _ { k } } { 2 } + p _ { k } \geq 0$$
for k > 3 and
$$\frac { N \cdot N _ { 3 } } { 6 } + p _ { 3 } \geq 0$$
for k = 3.
Let K = { k ∈ { 5 , . . . , n } : φ k ( ) ≥ 3 k 4 } . If 2 or 3 divide k then φ k ( ) < 3 k 4 , and hence none of those numbers are included in K . Summing up all cases we get
$$R ( 3, G, 2 ) \geq \frac { \sum _ { k \in K } \frac { N \cdot N _ { k } } { 2 } ( 1 - 3 \frac { k - \phi ( k ) } { \phi ( k ) } ) } { 4 }.$$
## 7 Methods for longer arithmetic progressions
Let χ : G → {-1 1 , } be a 2-coloring of the group G , and let x g = χ g ( ) for all g ∈ G . Let also x G denote the vector of all variables x g . For a, b, c ∈ G , let us introduce the polynomial
$$p ( x _ { a _ { 1 } }, \dots, x _ { a _ { k } } ) = \frac { ( 1 + x _ { a _ { 1 } } ) \cdots ( 1 + x _ { a _ { k } } ) + ( 1 - x _ { a _ { 1 } } ) \cdots ( 1 - x _ { a _ { k } } ) } { 2 ^ { k } },$$
which has the property that
$$p ( x _ { a _ { 1 } }, \dots, x _ { a _ { k } } ) = \begin{cases} \ 1 & \text{if $x_{a_{1}}=\dots=x_{a_{k}}$} \\ 0 & \text{otherwise.} \end{cases}$$
In other words, the polynomial p is one when { a , . . . , a 1 k } is a monochromatic arithmetic progression and zero otherwise. It follows that
$$R ( k, G, 2 ) = \min _ { x _ { G } \in \{ - 1, 1 \} ^ { | G | } } \sum _ { \{ a _ { 1 }, \dots, a _ { k } \} } \, \text{is an A.P. in $G$} } p ( x _ { a _ { 1 } }, \dots, x _ { a _ { k } } ).$$
To find a lower bound for R k, G, ( 2) we relax the integer quadratic optimization problem to an optimization problem on the hypercube. Since it is a relaxation, i.e. any solution of the integer program is also a solution to the hypercube problem, we have
$$R ( k, G, 2 ) \geq \min _ { x _ { G } \in [ - 1, 1 ] ^ { | G | } } \sum _ { \{ a _ { 1 }, \dots, a _ { k } \} \text{ is an A.P. in } G } p ( x _ { a _ { 1 } }, \dots, x _ { a _ { k } } ),$$
an optimization problem that we can find lower bounds for using Putinar's Positivstellensatz and the Lasserre Hierarchy.
## 8 Related problems
The methods developed in this article may also be applied in a wide variety of similar problems.
Let R (3 , [ n , ] 2) denote the minimal number of monochromatic arithmetic progressions of length 3 in a 2-coloring of [ n ]. Asymptotic bounds for R (3 , [ n , ] 2) have been found for large n using other methods [13]:
$$\frac { 1 6 7 5 } { 3 2 7 6 8 } n ^ { 2 } ( 1 + o ( 1 ) ) \leq R ( 3, [ n ], 2 ) \leq \frac { 1 1 7 } { 2 1 9 2 } n ^ { 2 } ( 1 + o ( 1 ) ).$$
The author has in collaboration with Oscar Kivinen found numerical results suggesting that the lower bound can be improved to 0 052341 . . . . using a degree 3-relaxation ( 1675 32768 = 0 05111 . . . . and 117 2192 = 0 05337 . . . . ) . The major challenge is to use the numerical results to obtain an algebraic certificate because of the lack of symmetries in the problem. For large n the numerical information seems to converge towards a fractal, and to obtain the improved lower bound based on
the proposed methods one would need to fully understand the behavior of the fractal. Even if it is not possible to understand the fractal, one can probably do approximations that would improve on the lower bound. This is work under progress.
Let R (3 , Z n , 2) denote the minimal number of monochromatic arithmetic progressions of length 3 in a 2-coloring of the cyclic group Z n . Optimal, or a constant from optimal, lower bounds for R (3 , Z n , 2) have been found for all n [15]:
$$n ^ { 2 } / 8 - c _ { 1 } n + c _ { 2 } \leq R ( 3, \mathbb { Z } _ { n }, 2 ) \leq n ^ { 2 } / 8 - c _ { 1 } n + c _ { 3 },$$
where the constants depends on the modular arithmetic and are tabulated in the following table.
| n mod 24 | c 1 | c 2 | c 3 |
|------------------------------------|-------|-------|--------|
| 1 , 5 , 7 , 11 , 13 , 17 , 19 , 23 | 1 / 2 | 3 / 8 | 3 / 8 |
| 8 , 16 | 1 | 0 | 0 |
| 2 , 10 | 1 | 3 / 2 | 3 / 2 |
| 4 , 20 | 1 | 0 | 2 |
| 14 , 22 | 1 | 3 / 2 | 3 / 2 |
| 3 , 9 , 15 , 21 | 7 / 6 | 3 / 8 | 27 / 8 |
| 0 | 5 / 3 | 0 | 0 |
| 12 | 5 / 3 | 0 | 18 |
| 6 , 18 | 5 / 3 | 1 / 2 | 27 / 2 |
A corollary is that we can find an optimal, or a constant from optimal, lower bound for the number of monochromatic arithmetic progressions for the dihedral group D 2 n for any n :
$$R ( 3, D _ { 2 n }, 2 ) = 2 R ( 3, \mathbb { Z } _ { n }, 2 ).$$
In particular
$$n ^ { 2 } / 4 - 2 c _ { 1 } n + 2 c _ { 2 } \leq R ( D _ { 2 n } ; 3 ) \leq n ^ { 2 } / 4 - 2 c _ { 1 } n + 2 c _ { 3 }$$
where the constants can be found in the table above.
Let R (4 , Z n , 2) denote the minimal number of monochromatic arithmetic progressions of length 4 in a 2-coloring of the cyclic group Z n . Asymptotic bounds for R (4 , Z n , 2) have been found for large n in [20], and the bounds have since then been improved in [9, Theorem 1.1, 1.2, 1.3] to:
$$\frac { 7 } { 1 9 2 } p ^ { 2 } ( 1 + o ( 1 ) ) \leq R ( 4, \mathbb { Z } _ { p }, 2 ) \leq \frac { 1 7 } { 3 0 0 } p ^ { 2 } ( 1 + o ( 1 ) )$$
when p is prime, and for other n
$$c _ { 1 } n ^ { 2 } ( 1 + o ( 1 ) ) \leq R ( 4, \mathbb { Z } _ { n }, 2 ) \leq c _ { 2 } n ^ { 2 } ( 1 + o ( 1 ) )$$
where the constants depends on the modular arithmetic on n in accordance with the following table
$$\begin{smallmatrix} n \mod 4 & c _ { 1 } & c _ { 2 } \\ 1, 3 & 7 / 1 9 2 & 1 7 / 3 0 0 \\ 0 & 2 / 6 6 & 8 5 4 3 / 1 4 5 2 0 0 0 \\ 2 & 7 / 1 9 2 & 8 5 4 3 / 1 4 5 2 0 0 0 \end{smallmatrix}$$
Furthermore [9, Theorem 1.5]
$$\varliminf _ { n \rightarrow \infty } R ( 4, \mathbb { Z } _ { n }, 2 ) \leq \frac { 1 } { 2 4 },$$
and it is conjectured that [9, Conjecture 1.1]:
$$\inf _ { n } \{ R ( 4, \mathbb { Z } _ { n }, 2 ) \} = \frac { 1 } { 2 4 }.$$
The author has tried to improve the bounds using a degree 4-relaxations with further restriction to simplify the problem, but as the author obtained algebraic bounds that are far from the current best bounds it seems like one needs to omit the restrictions and possibly use a higher degree relaxation to get relevant bounds. One of the main challenges is that the solution depends heavily on the modular arithmetic of n , and so one needs to numerically find solutions for fairly high values of n , which is not feasible for a high degree relaxation. Doing a full degree 4-relaxation is possible but tedious, and since one can only get numerical results for fairly small n it is difficult to find general patterns. One cannot exclude the possibility that this would be enough to improve the bounds, but what is more likely is that it would require additional tricks or techniques after one has found numerical certificates for small n .
Let R (4 , [ n , ] 2) denote the minimal number of monochromatic arithmetic progressions of length 4 in a 2-coloring of [ n ]. Let furthermore R (5 , Z n , 2) and R (5 , [ n , ] 2) denote the minimal number of monochromatic arithmetic progressions of length 5 in Z n and [ n ] respectively. Upper bounds have been found for R (4 , [ n , ] 2), R (5 , Z n , 2) and R (5 , [ n , ] 2). For n large enough [9, Equation (12)]:
$$R ( 4, [ n ], 2 ) \leq \frac { 1 } { 7 2 } n ^ { 2 } ( 1 + o ( 1 ) ).$$
When n is odd and large enough we have [9, Theorem 1.4]:
$$R ( 5, \mathbb { Z } _ { n }, 2 ) \leq \frac { 3 6 2 9 } { 1 3 1 4 2 4 } n ^ { 2 } ( 1 + o ( 1 ) ).$$
When n is even and large enough we have [9, Theorem 1.4]:
$$R ( 5, \mathbb { Z } _ { n }, 2 ) \leq \frac { 3 6 4 7 } { 1 3 1 4 2 4 } n ^ { 2 } ( 1 + o ( 1 ) ).$$
Furthermore [9, Theorem 1.5]:
$$\varliminf _ { n \rightarrow \infty } R ( 5, \mathbb { Z } _ { n }, 2 ) \leq \frac { 1 } { 7 2 }.$$
Finally when n is large enough we have [9, Equation (13)]:
$$R ( 5, [ n ], 2 ) \leq \frac { 1 } { 3 0 4 } n ^ { 2 } ( 1 + o ( 1 ) ).$$
These are all upper bounds that have been obtained through good colorings. Lower bounds to all these problems can possibly be found using the methods developed in this article using a relaxation of high enough order. As it is numerically very difficult to find solutions to relaxations of high orders, one should be aware that this is not guaranteed to work in practice.
Another type of problems that one can use the methods developed in this article to solve are enumeration problems in fixed density sets. As in this article it is of interest to count arithmetic progressions. Let W k,S,δ ( ) denote the minimal number of arithmetic progressions of length k in any subset of S of cardinality | S δ | . One can for example let S be a group or [ n ]. Note that if one could find strict lower bounds for W k, n , δ ( [ ] ) for all n , k and δ this would imply optimal quantitative bounds for Szemer´di's theorem. Although it might e be too ambitious to try to find strict lower bounds it might still be possible to find bounds good enough to generalize Szemer´ edi's theorem. This kind of results can be obtained by methods very similar to the ones used in this article as discussed in [16].
## Acknowledgements
I would like to thank Alexander Engstr¨m for o our long discussions on this topic and for all his valuable feedback. I would also want to thank Markus Schweighofer and Cynthia Vinzant for their comments, which have been incorporated in the article.
## References
- [1] Christine Bachoc, Dion C. Gijswijt, Alexander Schrijver and Frank Vallentin. Invariant semidefinite programs. Chapter of 'Handbook on Semidefinite, Conic and Polynomial Optimization'. Springer, Berlin, 2012. 219269.
- [2] Karin Gatermann and Pablo A. Parrilo. Symmetry groups, semidefinite programs, and sums of squares. J. Pure Appl. Algebra 192 (2004), no. 1, 95-128.
- [3] Yoshihiro Kanno, Makoto Ohsaki, Kazuo Murota and Naoki Katoh. Group symmetry in interior-point methods for semidefinite program. Optim. Eng. 2 (2001), no. 3, 293-320.
- [4] Etienne de Klerk, Cristian Dobre and Dmitrii V. Pasechnik. Numerical block diagonalization of matrix ∗ -algebras with application to semidefinite programming. Math. Program. 129 (2011), no. 1, 91-111.
- [5] Etienne de Klerk, Dmitrii V. Pasechnik and Alexander Schrijver. Reduction of symmetric semidefinite programs using the regular ∗ -representation. Math. Program. 109 (2007), no. 2-3, Ser. B, 613-624.
| [6] | Jean Bernard Lasserre. Global optimization with polynomials and the prob- lem of moments. SIAM J. Optim. 11 (2001), no. 3, 796-817. |
|-------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [7] | Jean Bernard Lasserre. Moments, positive polynomials and their applica- tions. Imperial College Press Optimization Series, 1. Imperial College Press, London, 2010. 361 pp. |
| [8] | Monique Laurent. Sums of squares, moment matrices and optimization over polynomials. Chapter of 'Emerging applications of algebraic geometry'. Springer, New York, 2009. 157-270. |
| [9] | Linyuan Lu and Xing Peng. Monochromatic 4-term arithmetic progressions in 2-colorings of Z n . J. Combin. Theory 119 (2012), Ser. A, no. 5, 1048- 1065. |
| [10] | Takanori Maehara and Kazuo Murota. A numerical algorithm for block- diagonal decomposition of matrix ∗ -algebras with general irreducible com- ponents. Jpn. J. Ind. Appl. Math. 27 (2010), no. 2, 263-293. |
| [11] | Kazuo Murota, Yoshihiro Kanno, Masakazu Kojima and Sadayoshi Ko- jima. A numerical algorithm for block-diagonal decomposition of matrix ∗ - algebras with application to semidefinite programming. Jpn. J. Ind. Appl. Math. 27 (2010), no. 1, 125-160. |
| [12] | Pablo A. Parrilo and Helfried Peyrl. Computing sum of squares decompo- sitions with rational coefficients. Theoret. Comput. Sci. 409 (2008), no. 2, 269-281. |
| [13] | Pablo A. Parrilo, Aaron Robertson, and Dan Saracino. On the asymp- totic minimum number of monochromatic 3-term arithmetic progressions. J. Combin. Theory 115 (2008), no. 1, Ser. A,185-192. |
| [14] | Cordian Riener, Thorsten Theobald, Lina Jansson Andr´n e and Jean B. Lasserre. Exploiting symmetries in SDP-relaxations for polynomial opti- mization. Math. Oper. Res. 38 , (2013), no. 1, 122-141. |
| [15] | Erik Sj¨land. o Enumeration of monochromatic three term arithmetic progressions in two-colorings of cyclic groups. Preprint available at http://arxiv.org/abs/1408.1058. |
| [16] | Erik Sj¨land. o Enumeration of three term arithmetic progressions in fixed density sets. Preprint available at http://arxiv.org/abs/1408.1063. |
| [17] | Erik Sj¨land. o Using real algebraic geometry to solve combinatorial problems with symmetries. Preprint available at http://arxiv.org/abs/1408.1065. |
| [18] | Masamichi Takesaki. Theory of operator algebras I . Encyclopaedia of Math- ematical Sciences, 124. Operator Algebras and Non-commutative Geome- |
- [19] Frank Vallentin. Symmetry in semidefinite programs. Linear Algebra Appl. 43 (2009), no. 1, 360-369.
- [20] Julia Wolf. The minimum number of monochromatic 4-term progressions in Z p . J. Comb. 1 (2010), no. 1, 53-68.
| null |
[
"Erik Sjöland"
] |
2014-08-05T18:34:50+00:00
|
2014-11-10T18:57:16+00:00
|
[
"math.CO"
] |
Enumeration of monochromatic three term arithmetic progressions in two-colorings of any finite group
|
There are many extremely challenging problems about existence of
monochromatic arithmetic progressions in colorings of groups. Many theorems
hold only for abelian groups as results on non-abelian groups are often much
more difficult to obtain. In this research project we do not only determine
existence, but study the more general problem of counting them. We formulate
the enumeration problem as a problem in real algebraic geometry and then use
state of the art computational methods in semidefinite programming and
representation theory to derive lower bounds for the number of monochromatic
arithmetic progressions in any finite group.
|
1408.1089v1
|
## A luminous, blue progenitor system for a type-Iax supernova
Curtis McCully , Saurabh W. Jha , Ryan J. Foley 1 1 2 3 , , Lars Bildsten 4 5 , , Wen-fai Fong , Robert P. Kirshner , G. H. Marion 6 6 7 6 , , Adam G. Riess 8 9 , , & Maximilian D. Stritzinger 10
August 7, 2014
authors' version, accepted to Nature ; final version available at http://dx.doi.org/10.1038/nature13615
1 Department of Physics and Astronomy, Rutgers, the State University of New Jersey, 136 Frelinghuysen Road, Piscataway, New Jersey 08854, USA. 2 Astronomy Department, University of Illinois at UrbanaChampaign, 1002 W. Green Street, Urbana, Illinois 61801, USA. 3 Department of Physics, University of Illinois at Urbana-Champaign, 1110 W. Green Street, Urbana, Illinois 61801, USA. 4 Department of Physics, University of California, Santa Barbara, California 93106, USA. 5 Kavli Institute for Theoretical Physics, University of California, Santa Barbara, California 93106, USA. 6 Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, Massachusetts 02138, USA. 7 Department of Astronomy, University of Texas at Austin, Austin, Texas 78712, USA. 8 Department of Physics and Astronomy, Johns Hopkins University, 3400 North Charles Street, Baltimore, Maryland 21218, USA. 9 Space Telescope Science Institute, 3700 San Martin Drive, Baltimore, Maryland 21218, USA. 10 Department of Physics and Astronomy, Aarhus University, Ny Munkegade 120, DK-8000 Aarhus C, Denmark.
SN 2012Z was discovered 16 in the Lick Observatory Supernova Search on UT 2012-Jan-29.15. It had an optical spectrum similar to the type-Iax (previously called SN 2002cx-like) SN 2005hk 3-5 (see Extended Data Fig. 1). The similarities between SN Iax and normal SN Ia make understanding SN Iax progenitors important, especially because no normal SN Ia progenitor has been identified. Like core-collapse SN (but also slowly-declining, luminous SN Ia), SN Iax are found preferentially in young, star-forming galaxies 17,18 . A single SN Iax, SN 2008ge, was in a relatively old (S0) galaxy with no indication of current star formation to deep limits 19 . Non-detection of the progenitor of SN 2008ge in Hubble Space Telescope ( HST ) pre-explosion imaging restricts its initial mass glyph[lessorsimilar] 12 M glyph[circledot] , and combined with the lack of hydrogen or helium in the SN 2008ge spectrum, favours a white dwarf progenitor 19 .
Type-Iax supernovae (SN Iax) are stellar explosions that are spectroscopically similar to some type-Ia supernovae (SN Ia) at maximum light, except with lower ejecta velocities 1, 2 . They are also distinguished by lower luminosities. At late times, their spectroscopic properties diverge from other SN 3-6 , but their composition (dominated by iron-group and intermediate-mass elements 1, 7 ) suggests a physical connection to normal SN Ia. These are not rare; SN Iax occur at a rate between 5 and 30% of the normal SN Ia rate . 1 The leading models for SN Iax are thermonuclear explosions of accreting carbon-oxygen white dwarfs (C/O WD) that do not completely unbind the star 8-10 , implying they are 'less successful' cousins of normal SN Ia, where complete disruption is observed. Here we report the detection of the luminous, blue progenitor system of the type-Iax SN 2012Z in deep pre-explosion imaging. Its luminosity, colors, environment, and similarity to the progenitor of the Galactic helium nova V445 Puppis 11-13 , suggest that SN 2012Z was the explosion of a WD accreting from a helium-star companion. Observations in the next few years, after SN 2012Z has faded, could test this hypothesis, or alternatively show that this supernova was actually the explosive death of a massive star 14,15 .
Deep observations of NGC 1309, the host galaxy of SN 2012Z, were obtained with HST in 2005-2006 and 2010, serendipitously including the location of the supernova before its explosion. To pinpoint the position of SN 2012Z with high precision, we obtained follow-up HST data in 2013. Colourcomposite images made from these observations before and after the supernova are shown in Fig. 1, and photometry of stellar sources in the pre-explosion images near the supernova location is reported in Extended Data Table 1. We detect a source, called S1, coincident with the supernova at a formal separation of 0 . ′′ 0082 ± 0 0103 (equal to 1.3 . ′′ ± 1.6 pc at 33 Mpc, the distance to NGC 1309 20,21 ). The pre-explosion data reach a 3 σ limiting magnitude of M V ≈ -3.5, quite deep for typical extragalactic SN progenitor searches 22 , but certainly the possibility exists that the progenitor system of SN 2012Z was of lower luminosity and would be undetected in our data (as has been the case for all normal SN Ia progenitor searches to date 23 ). However, the locations of SN 2012Z and S1 are identical to within 0.8 σ , and we estimate only a 0.24% (2.1%) probability that a random position near SN 2012Z would be within 1 σ (3 σ ) of any detected star, making a chance alignment unlikely (see Methods, and Extended Data Fig. 2). We also observe evidence for vari-
ability in S1 (plausible for a pre-supernova system; Extended Data Table 2), at a level exhibited by only 4% of objects of similar brightness. We thus conclude there is a high likelihood that S1 is the progenitor system of SN 2012Z.
The color-magnitude diagram (CMD) presented in Fig. 2 shows S1 to be luminous and blue, yet in an odd place for a star about to explode. If its light is dominated by a single star, S1 is moderately consistent with a ≈ 18.5 M glyph[circledot] main-sequence star 24 , an ≈ 11 M glyph[circledot] blue supergiant early in its evolution off the main sequence, or perhaps a ≈ 7.5 M glyph[circledot] (initial mass) blue supergiant later in its evolution (with core helium-burning in a blue loop, where models are quite sensitive to metallicity and rotation 25 ). None of these stars are expected to explode in standard stellar evolution theory, particularly without any signature of hydrogen in the supernova 22 .
The SN 2012Z progenitor system S1 is in a similar region in the CMD to some Wolf-Rayet stars 26 , highly evolved, massive stars, that are expected to undergo core collapse and may produce a supernova. If S1 were a single Wolf-Rayet star, its photometry is most consistent with the WN subtype and an initial mass ≈ 30-40 M glyph[circledot] , thought perhaps to explode with a heliumdominated outer layer as a SN Ib 27 , and unlikely to produce the structure and composition of ejecta seen in SN Iax 1, 3, 6, 7 . Moreover, isochrones 24 fit to the neighbouring stars (Extended Data Fig. 3) yield an age range of ≈ 10-42 Myr, longer than the 5-8 Myr lifetime of such a massive Wolf-Rayet star.
S1 may be dominated by accretion luminosity; its brightness in B and V is not far from the predicted thermal emission of an Eddington-luminosity Chandrasekhar mass WD (a supersoft source; SSS; Fig. 2). However, its V-I and V-H colours are too red for a SSS model. A composite scenario, with accretion power dominating the blue flux, and another source providing the redder light (perhaps a fainter, red donor star) may be plausible.
The leading models of SN Iax 8-10 are based on C/O WD explosions, so S1 may be the companion star to an accreting WD. Although there are a variety of potential progenitor systems (including main-sequence and red giant donors, which are inconsistent with S1 if they dominate the system's luminosity), in standard scenarios no companion star can have an initial mass greater than ≈ 7 M glyph[circledot] ; otherwise, there would not be enough time to form the primary C/O WD that explodes. Thus, the photometry of S1 suggests that if it is the companion to a C/O WD, recent binary mass transfer must have played a role in its evolution. One model for a luminous, blue companion star is a relatively massive ( ≈ 2 M glyph[circledot] when observed) helium star 11,28,29 , formed after binary mass transfer and a common envelope phase (e.g., a close binary with initial masses ≈ 7 and 4 M glyph[circledot] ). Although the model parameter space has not been fully explored, the predicted region for helium star donors in a binary system with a 1.2 M glyph[circledot] initial-mass accreting C/O WD 29 in the CMD is shown in Fig. 2, and S1 is consistent with being in this region. The evolutionary timescale for such a model is also well-matched to the ages of nearby stars (Extended Data Fig. 3).
SN 2012Z and the star S1 have an interesting analogue in our own Milky Way Galaxy: the helium nova V445 Puppis 11-13 , thought to be the surface explosion of a near-Chandrasekhar mass helium-accreting WD. Though S1 is somewhat brighter than the pre-explosion observations of V445 Pup, their consistent colours, similar variability amplitude 13 , and the physical connection between V445 Pup and likely SN Iax progenitors 8-10 is highly suggestive. Indeed, two SN Iax (though not SN 2012Z itself) have shown evidence for helium in the system 1, 17 . In this model, a low helium accretion rate could lead to a helium nova (like V445 Pup), whereas a higher mass-transfer rate could result in stable helium burning on the C/O WD, allowing it to grow in mass before the supernova. The accretion is expected to begin as the helium star starts to evolve and grow in radius; indeed, the S1 photometry is consistent with the evolutionary track of a helium star with a mass (after losing its hydrogen envelope) of ≈ 2 M glyph[circledot] , on its way to becoming a red giant 11 .
Though the scenario of a helium-star donor to an exploding carbon-oxygen white dwarf is a promising model for the progenitor and supernova observations, we cannot yet rule out the possibility that S1 is a single star that itself exploded. Fortunately, by late 2015, SN 2012Z will have faded below the brightness of S1, and HST imaging will allow us to distinguish these models. Our favoured interpretation of S1 as the companion star predicts that it will still be detected (though perhaps modified by the impact of its exploding neighbour, a reduction in accretion luminosity, or a cessation of variability). On the other hand, if S1 has completely disappeared, it will be a strong challenge to models of SN Iax, and perhaps significantly blur the line between thermonuclear white-dwarf supernovae and massive-star core-collapse supernovae, with important impacts to our understanding of stellar evolution and chemical enrichment.
- 1. Foley, R. J. et al. Type Iax Supernovae: A New Class of Stellar Explosion. Astrophys. J. 767 , 57 (2013).
- 2. Li, W. et al. SN 2002cx: The Most Peculiar Known Type Ia Supernova. Publ. Astron. Soc. Pacif. 115 , 453-473 (2003).
- 3. Jha, S. et al. Late-Time Spectroscopy of SN 2002cx: The Prototype of a New Subclass of Type Ia Supernovae. Astron. J. 132 , 189-196 (2006).
- 4. Phillips, M. M. et al. The Peculiar SN 2005hk: Do Some Type Ia Supernovae Explode as Deflagrations? Publ. Astron. Soc. Pacif. 119 , 360-387 (2007).
- 5. Sahu, D. K. et al. The Evolution of the Peculiar Type Ia Supernova SN 2005hk over 400 Days. Astrophys. J. 680 , 580-592 (2008).
- 6. McCully, C. et al. Hubble Space Telescope and Ground-based Observations of the Type Iax Supernovae SN 2005hk and SN 2008A. Astrophys. J. 786 , 134 (2014).
- 7. Stritzinger, M. D. et al. Optical and near-IR observations of the faint and fast 2008ha-like supernova 2010ae. Astron. Astrophys. 561 , A146 (2014).
- 8. Jordan, G. C., IV, Perets, H. B., Fisher, R. T. & van Rossum, D. R. Failed-detonation Supernovae: Subluminous Low-velocity Ia Supernovae and their Kicked Remnant White Dwarfs with Iron-rich Cores. Astrophys. J. 761 , L23 (2012).
- 9. Kromer, M. et al. 3D deflagration simulations leaving bound remnants: a model for 2002cx-like Type Ia supernovae. Mon. Not. R. Astron. Soc. 429 , 2287-2297 (2013).
- 10. Fink, M. et al. Three-dimensional pure deflagration models with nucleosynthesis and synthetic observables for Type Ia supernovae. Mon. Not. R. Astron. Soc. 438 , 1762-1783 (2014).
- 11. Kato, M., Hachisu, I., Kiyota, S. & Saio, H. Helium Nova on a Very Massive White Dwarf: A Revised Light-Curve Model of V445 Puppis (2000). Astrophys. J. 684 , 1366-1373 (2008).
- 12. Woudt, P. A. et al. The Expanding Bipolar Shell of the Helium Nova V445 Puppis. Astrophys. J. 706 , 738-746 (2009).
- 13. Goranskij, V., Shugarov, S., Zharova, A., Kroll, P. & Barsukova, E. A. The Progenitor and Remnant of the Helium Nova V445 Puppis. Peremennye Zvezdy 30 , 4 (2010).
- 14. Valenti, S. et al. A low-energy core-collapse supernova without a hydrogen envelope. Nature 459 , 674-677 (2009).
- 15. Moriya, T. et al. Fallback Supernovae: A Possible Origin of Peculiar Supernovae with Extremely Low Explosion Energies. Astrophys. J. 719 , 1445-1453 (2010).
- 16. Cenko, S. B. et al. Supernova 2012Z in NGC 1309 = Psn J032205351523156. Central Bureau Electronic Telegrams 3014 , 1 (2012).
- 17. Foley, R. J. et al. SN 2008ha: An Extremely Low Luminosity and Exceptionally Low Energy Supernova. Astron. J. 138 , 376-391 (2009).
- 18. Lyman, J. D. et al. Environment-derived constraints on the progenitors of low-luminosity Type I supernovae. Mon. Not. R. Astron. Soc. 434 , 527-541 (2013).
- 19. Foley, R. J. et al. On the Progenitor and Supernova of the SN 2002cxlike Supernova 2008ge. Astron. J. 140 , 1321-1328 (2010).
- 20. Riess, A. G. et al. Cepheid Calibrations of Modern Type Ia Supernovae: Implications for the Hubble Constant. Astrophys. J. Suppl. Ser. 183 , 109-141 (2009).
- 21. Riess, A. G. et al. A 3% Solution: Determination of the Hubble Constant with the Hubble Space Telescope and Wide Field Camera 3. Astrophys. J. 730 , 119 (2011).
- 22. Smartt, S. J. Progenitors of Core-Collapse Supernovae. Annu. Rev. Astron. Astrophys. 47 , 63-106 (2009).
- 23. Li, W. et al. Exclusion of a luminous red giant as a companion star to the progenitor of supernova SN 2011fe. Nature 480 , 348-350 (2011).
- 24. Bertelli, G., Nasi, E., Girardi, L. & Marigo, P. Scaled solar tracks and isochrones in a large region of the Z-Y plane. II. From 2.5 to 20 M glyph[circledot] stars. Astron. Astrophys. 508 , 355-369 (2009).
- 25. Georgy, C. et al. Populations of rotating stars. I. Models from 1.7 to 15 M at Z = 0.014, 0.006, and 0.002 with Ω Ω / crit between 0 and 1. Astron. Astrophys. 553 , A24 (2013).
- 26. Shara, M. M. et al. The Vast Population of Wolf-Rayet and Red Supergiant Stars in M101. I. Motivation and First Results. Astron. J. 146 , 162 (2013).
- 27. Groh, J. H., Meynet, G., Georgy, C. & Ekstr¨ om, S. Fundamental properties of core-collapse supernova and GRB progenitors: predicting the look of massive stars before death. Astron. Astrophys. 558 , A131 (2013).
- 28. Iben, I., Jr. & Tutukov, A. V. Helium star cataclysmics. Astrophys. J. 370 , 615-629 (1991).
- 29. Liu, W.-M., Chen, W.-C., Wang, B. & Han, Z. W. Helium-star evolutionary channel to super-Chandrasekhar mass type Ia supernovae. Astron. Astrophys. 523 , A3 (2010).
Acknowledgments Supernova 2012Z was discovered just weeks after the passing of our dear friend and colleague, Weidong Li, whose work on the Lick Observatory Supernova Search, SN 2002cx-like supernovae, and Hubble Space Telescope observations of SN progenitors, continues to inspire us. That those three foci of Weidong's research converge here in this paper makes our hearts glad, and we dedicate this paper to his memory.
We thank the SH 0 ES team for assistance with data from HST program GO-12880, E. Bertin for the development of the STIFF software to produce color images, and A. Dolphin for software and guidance in photometry. This research at Rutgers University was supported through NASA/ HST grant GO-12913.01, and National Science Foundation (NSF) CAREER award AST-0847157 to S.W.J. NASA/ HST grant GO-12999.01 to R.J.F. supported this work at the University of Illinois. At UC Santa Barbara, this work was supported by NSF grants PHY 11-25915 and AST 11-09174 to L.B. The Danish Agency for Science, Technology, and Innovation supported M.D.S. through a Sapere Aude Level 2 grant.
Support for HST programs GO-12913 and GO-12999 was provided by NASA through a grant from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Incorporated, under NASA contract NAS5-26555.
Author Contributions C.M., S.W.J., and R.J.F. performed the data analysis and were chiefly responsible for preparing the manuscript and figures. W.F., R.P.K., G.H.M., and A.G.R. assisted in developing the proposal to obtain HST observations, including acquiring supporting ground-based photometry and spectroscopy. L.B. provided insight into models for progenitor systems. M.D.S. analyzed groundbased photometry and spectroscopy of the supernova, used as input for this paper. All authors contributed to extensive discussions about, and edits to, the paper draft.
Competing Interests The authors declare that they have no competing financial interests.
Correspondence Correspondence and requests for materials should be addressed to S.W.J. ([email protected]).
## Methods
Observations and reduction. SN 2012Z provides a unique opportunity to search for a SN Iax progenitor, because of deep, pre-explosion HST observations. Its face-on spiral host galaxy NGC 1309 was also the site of the nearby, normal type-Ia SN 2002fk, and as such was targeted in 2005 and 2006 with the HST Advanced Camera for Surveys (ACS) and in 2010 with the HST Wide-Field Camera 3 (WFC3) optical (UVIS) and infrared (IR) channels, to observe Cepheid variable stars in order to anchor the SN Ia distance scale ( HST programs GO-10497, GO-10802, and GO-11570, PI: A. Riess; GO/DD-10711, PI: K. Noll). To measure the location of SN 2012Z with high precision, we used HST WFC3/UVIS images of NGC 1309 (fortuitously including SN 2012Z) taken on UT 2013-Jan-04 (program GO-12880; PI: A. Riess), as well as targeted HST WFC3/UVIS images of SN 2012Z taken on UT 2013-Jun-30 (GO-12913; PI: S. Jha).
The 2005-2006 ACS images include 2 visits totaling 9600s of exposure time in the F435W filter (similar to Johnson B ), 14 visits for a total exposure of 61760 sec in F555W (close to Johnson V ) and 5 visits for 24000s in F814W (analogous to Cousins I ). These are showing the blue, green, and red channels, respectively of panels a-c in Fig. 1. We re-reduced the archived data, combining the multiple exposures (including sub-sampling and cosmic ray rejection) using the AstroDrizzle software from the DrizzlePac package 30 , with the results shown in panels b and c of Fig. 1. The bottom right panels d and e of Fig. 1 show our combined HST WFC3/UVIS F555W (blue; 1836 sec) + F625W (green; 562 sec) + F814W (red; 1836 sec) images from January and June 2013, with SN 2012Z visible.
We used the DrizzlePac TweakReg routine to register all of the individual flatfielded (' flt ') frames to the WFC3/UVIS F555Wimage taken on UT 2013-Jan-04. The typical root-meansquare (rms) residual of individual stars from the relative astrometric solution was 0 . ′′ 009, corresponding to 0.18 pixels in ACS and 0.23 pixels in WFC3/UVIS. We drizzle the ACS images to the native scale of UVIS, 0 . ′′ 04 per pixel (20% smaller than the native 0 . ′′ 05 ACS pixels) and subsample the ACS point-spread function (psf) correspondingly with a pixel fraction parameter of 0.8.
Photometry and astrometry. In Extended Data Table 1, we present photometry of sources in the region based on the HST ACS images from 2005-2006 (F435W, F555W, and F814W), as well as WFC3/IR F160W data (6991 sec of total exposure time) from 2010. The stars are sorted by their proximity to the SN position, and their astrometry is referenced to SDSS images of the field, with an absolute astrometric uncertainty of 0 . ′′ 080 (but this is irrelevant for the much more precise relative astrometry of SN 2012Z and S1). We photometered the HST images using the psf-fitting software DolPhot, an extension of HSTPhot 31 . We combined individual flt frames taken during the same HST visit at the same position, and then used DolPhot to measure photometry using recommended parameters for ACS and WFC3.
The WFC3/UVIS images of SN 2012Z from January and June 2013 provide a precise position for the supernova of R.A. = 3 h 22 m 05 . s 39641, Decl. = -15 23 14 ◦ ′ . ′′ 9390 (J2000) with a registration uncertainty of 0 . ′′ 0090 (plus an absolute astrometric uncertainty of 0 . ′′ 08, irrelevant to the relative astrometry). As shown in Fig. 1 we detect a stellar source (called S1) in the preexplosion images coincident with the position of the SN with a formal separation, including centroid uncertainties, of 0 . ′′ 0082 ± 0 0103, indicating an excellent match and strong evidence for . ′′ S1 being the progenitor system of SN 2012Z.
Chance alignment probability. To estimate the probability of a chance alignment, we use the observed density of sources detected with signal-to-noise ratio S/N > 3.0 (in any filter) and S/N > 3.5 (in all bands combined, via DolPhot) in a 200 × 200 pixel (8 ′′ × 8 ′′ ) box centered on SN 2012Z (452 sources), and find only a 0.24% (2.1%) chance that a random position would be consistent with a detected star at 1 σ (3 σ ). Moreover, only 171 of these stars are as bright as S1, so a posteriori there was only a 0.09% (0.80%) chance of a 1 σ (3 σ ) alignment with such a bright object. As shown in Extended Data Fig. 2, these results are not especially sensitive to the size of the region used to estimate the density of sources, at least down to 50 × 50 pixels (2 ′′ × 2 ′′ ) around SN 2012Z. Nearer than this, the density of sources increases by a factor of a few, though with substantially larger uncertainty given the low number of sources (including S1 itself). We base our fiducial chance alignment probability on the larger region where the density of sources stabilizes with good statistics, but our qualitative results do not depend on this choice.
Given the surface brightness of NGC 1309, we crudely estimate ≈ 160 M glyph[circledot] in stars projected within an area corresponding to our 1 σ error circle ( ≈ 8 pc 2 ). These should be roughly uniformly distributed throughout this small region, so our chance alignment probability should accurately quantify the probability that SN 2012Z originated from an undetected progenitor that was only coincidentally near a detected source like S1.
The pre-explosion data for SN 2012Z are the deepest ever for a SN Iax, reaching M V ≈ -3.5; the next best limits come from SN 2008ge 19 , which yielded no progenitor detection down to M V ≈ -7. The star S1, at M V ≈ -5.3, would not have been detected in any previous search for SN Iax progenitors. This implies that our chance alignment probability calculation can be taken at face value; there have not been previous, unsuccessful 'trials' that would reduce the unlikelihood of a chance coincidence. Viewed in the context of progenitor searches for normal SN Ia, in only two cases: SN 2011fe 23 and SN 2014J 32 , would a star of the luminosity of S1 have been clearly detected in pre-explosion data. Other normal SN Ia, like SN 2006dd 33 , have progenitor detection limits in pre-explosion observations just near, or above, the luminosity of S1.
Variability of S1. NGC 1309 was imaged over 14 epochs in F555W with ACS before SN 2012Z exploded. Examining these individually, we find some evidence for variability in S1; the photometry is presented in Extended Data Table 2. Formally, these data rule out the null hypothesis of no variability at 99.95% (3.5 σ ), with χ 2 = 36.658 in 13 degrees of freedom. However, most of the signal is driven by one data point (MJD 53600.0; a 4.2 σ outlier); excluding this data point (though we find no independent reason to do so) reduces the significance of the vari-
ability to just 91.1% (1.7 σ ). To empirically assess the likelihood of variability, we looked to see how often stars with the same brightness as S1 (within 0.5 mag) doubled in brightness relative to their median like S1 in one or more of the 14 epochs; we find that just 4% do so (4 of the nearest 100, 9 of the nearest 200, and 20 of nearest 500 such stars). As variability might be expected in a SN progenitor before explosion (from non-uniform accretion, for example), combining this with the chance alignment probability strengthens the identification of S1 as the progenitor system of SN 2012Z. It also disfavours the possibility that S1 is a compact, unresolved star cluster.
Properties of S1 and nearby stars. In Fig. 2 we present colourmagnitude diagrams of S1 and other objects for comparison. In the figure, S1 has been corrected for Milky Way reddening (E(BV) MW = 0.035 mag, corresponding to A F435W = 0.14 mag, A F555W = 0.11 mag, A F814W = 0.06 mag), and host reddening (E(B-V) host = 0.07 mag; A F435W = 0.28 mag, A F555W = 0.22 mag, A F814W = 0.12 mag) based on narrow, interstellar absorption lines in high-resolution spectroscopy of SN 2012Z itself 34 . This low extinction is consistent with the photometry and spectroscopy of SN 2012Z, as well as its location in the outskirts of a face-on spiral host. For the potential Galactic analogue V445 Puppis progenitor, we correct its photometry for Galactic and circumstellar reddening 12 .
Fig. 2 also shows stellar evolution tracks 24 for stars with initial masses of 7, 8, 9, 10, and 11 M glyph[circledot] , adopting a metallicity of 0.87 solar, based on the H II region metallicity gradient 20 for NGC 1309 interpolated to the SN radial location. The Eddington-luminosity accreting Chandrasekhar-mass white dwarfs are shown as the large purple dots, with each subsequent dot representing a change in temperature of 1000 K. These 'super-soft' sources are fainter in F555W for higher temperatures (and bluer F435W -F555W colours) as the fixed (Eddington-limited) bolometric luminosity emerges in the ultraviolet for hotter systems. The shaded blue region represents the range of helium-star donors for C/O WD SN progenitor models starting with a 1.2 M glyph[circledot] white dwarf 29 . We converted the model temperatures and luminosities to our observed bands assuming a blackbody spectrum. The expected temperature and luminosity for this class of models is expected to vary with white dwarf mass, and therefore, we regard this region as approximate; its shape, size, and location is subject to change.
S1 is inconsistent with all confirmed progenitors of core-collapse SN (exclusively SN II), which are mostly red supergiants 22 . The blue supergiant progenitor of, e.g., SN 1987A, was significantly more luminous and likely more massive than S1 35 . However, we caution that our theoretical expectations for massive stars could be modified if S1 is in a close binary system where mass transfer has occurred.
The stars detected in the vicinity of S1 provide clues to the nature of recent star formation in the region. They include red supergiants (like S2 and S3 from Table 1) as well as objects bluer and more luminous than S1 (like S5). We show CMDs including these stars in Extended Data Fig. 3; the stars plotted have a signal-to-noise ratio S/N > 3.5 in the displayed filters, and were required to be no closer than 3 pixels to a brighter source to avoid photometric uncertainties from crowding. Model isochrones 24 imply that these stars span an age range of ≈ 10-42 Myr. These tracks favour an initial mass for S1 of ≈ 7-8 M glyph[circledot] (neglecting mass transfer) if it is roughly coeval with its neighbours. In other words, if S1 were a 30-40 M glyph[circledot] initial-mass Wolf-Rayet star with a predicted lifetime of only 5-8 Myr 27 , it would be the youngest star in the region.
- 30. Gonzaga, S., Hack, W., Fruchter, A. & Mack, J. The DrizzlePac Handbook (STScI, Baltimore, 2012).
- 31. Dolphin, A. E. WFPC2 Stellar Photometry with HSTPHOT. Publ. Astron. Soc. Pacif. 112 , 1383-1396 (2000).
- 32. Kelly, P. L. et al. Constraints on the Progenitor System of the Type Ia Supernova 2014J from Pre-explosion Hubble Space Telescope Imaging. Astrophys. J. 790 , 3 (2014).
- 33. Maoz, D. & Mannucci, F. A search for the progenitors of two Type Ia Supernovae in NGC 1316. Mon. Not. R. Astron. Soc. 388 , 421-428 (2008).
- 34. Stritzinger, M. D. et al. Comprehensive Observations of the Bright and Energetic Type Iax SN 2012Z: Interpretation as a Chandrasekhar Mass White Dwarf Explosion. Astron. Astrophys. (2014, submitted).
- 35. Arnett, W. D., Bahcall, J. N., Kirshner, R. P. & Woosley, S. E. Supernova 1987A. Annu. Rev. Astron. Astrophys. 27 , 629-700 (1989).
Figure 1 | Hubble Space Telescope colour images before and after supernova 2012Z. Panel a shows the Hubble Heritage image of NGC 1309 ( http://heritage.stsci.edu/2006/07 ), with panels b and c zooming in on the progenitor system S1 in the deep, pre-explosion data. Lower panels d and e show the shallower post-explosion images of SN 2012Z on the same scale as the panels above.
Figure 2 | Colour-magnitude diagrams of the SN 2012Z progenitor system S1 and comparison models. Panel a presents the F435W-F555W colour (roughly B-V), while panel b shows F555W-F814W (V-I), both plotted against the F555W (V) absolute magnitude. The black and brown crosses represent the progenitor systems for SN 2012Z and V445 Puppis 12 , respectively, with 1 σ photometric uncertainties. Other comparisons plotted include evolutionary tracks 24 for single stars (coloured dotted curves), thermal models for Eddington-luminosity super-soft sources (SSS; purple dots), candidate Wolf-Rayet stars 26 (blue-gray stars), and models for helium-star donors to 1.2 M glyph[circledot] initial mass C/O WDs 29 (shaded blue regions).
Extended Data Figure 1 | Spectra of SN Iax near maximum light. SN 2012Z 1 is similar to SN 2005hk 4 and SN 2008A 6 , all classified as SN Iax. The SN 2012Z spectrum was taken on UT 2012-02-16 with the Whipple Observatory 1.5m telescope (+FAST spectrograph) with a total exposure time of 1800 sec. Each spectrum is labelled by its rest-frame phase past B maximum light. Prominent features due to intermediate-mass and iron group elements are indicated; these features are also observed in luminous, slowly-declining SN Ia spectra at maximum light, though with higher expansion velocities.
Extended Data Figure 2 | Chance alignment probability calculation. Panel a shows the chance alignment probability between a random position and a detected source within a square box of the given width centered at the position of SN 2012Z. This calculation is based on a 1 σ position coincidence (0 . ′′ 0103 = 0.2575 WFC3/UVIS pixels); a 3 σ position coincidence gives probabilities approximately 9 times higher. The offset between SN 2012Z and S1 is 0.8 σ . The shaded regions show a naive Poisson-like uncertainty estimate with a fractional error given by 1 / √ N sources . Panel b shows the number of detected sources (including S1) as a function of the box size. The lines show results for all stellar sources ( > σ 3 detection in any band; blue) and just those as bright as S1 (red).
Extended Data Figure 3 | Stars in the neighbourhood of the SN 2012Z progenitor. The SN 2012Z progenitor system S1 (blue) is shown along with nearby stars (all with 1 σ photometric uncertainties), with three isochrones 24 with ages of 10.5 (dotted), 17.0 (dashed), and 41.7 Myr (solid). Panel a presents the F435W-F555W colour (roughly B-V), while panel b shows F555W-F814W (V-I), both plotted against the F555W (V) absolute magnitude. The large, filled circles correspond to stars within 10 WFC3/UVIS pixels (0 . ′′ 4) of the SN location, small filled circles are within 20 pixels (0 . ′′ 8), and the small open circles are within 30 pixels (1 . ′′ 2). Objects in the gray shaded regions would not be detected given the depth of the combined images.
Extended Data Table 1 | Pre-explosion photometry of SN 2012Z progenitor system S1 and nearby stars. The 1 σ photometric uncertainty is given in parentheses and non-detections are listed with 3 σ upper-limits. No corrections for Galactic or host extinction are made here.
| Star | R.A. (J2000) | Decl. (J2000) | F435W (mag) | F555W (mag) | F814W (mag) | F160W (mag) |
|--------|-----------------------|--------------------------|----------------|----------------|----------------|----------------|
| S1 | 3 h 22 m 05 . s 39591 | - 15 ◦ 23 ′ 14 . ′′ 9350 | 27.589 (0.122) | 27.622 (0.060) | 27.532 (0.135) | 26.443 (0.321) |
| S2 | 3 h 22 m 05 . s 40280 | - 15 ◦ 23 ′ 14 . ′′ 9402 | > 29.221 | 28.551 (0.116) | 27.463 (0.093) | 26.032 (0.238) |
| S3 | 3 h 22 m 05 . s 39483 | - 15 ◦ 23 ′ 14 . ′′ 8102 | 28.466 (0.258) | 27.221 (0.041) | 25.435 (0.022) | 23.699 (0.027) |
| S4 | 3 h 22 m 05 . s 38460 | - 15 ◦ 23 ′ 15 . ′′ 0314 | 28.258 (0.211) | 27.308 (0.046) | 25.717 (0.028) | 23.887 (0.033) |
| S5 | 3 h 22 m 05 . s 41013 | - 15 ◦ 23 ′ 14 . ′′ 9362 | 25.778 (0.029) | 25.918 (0.014) | 25.888 (0.030) | 25.435 (0.136) |
| S6 | 3 h 22 m 05 . s 37564 | - 15 ◦ 23 ′ 15 . ′′ 0702 | > 28.807 | 28.386 (0.116) | 26.553 (0.055) | 24.986 (0.085) |
| S7 | 3 h 22 m 05 . s 41970 | - 15 ◦ 23 ′ 14 . ′′ 8454 | 28.539 (0.286) | 27.942 (0.078) | 27.703 (0.146) | 26.436 (0.322) |
| S8 | 3 h 22 m 05 . s 37049 | - 15 ◦ 23 ′ 14 . ′′ 6690 | > 28.847 | 28.763 (0.164) | 27.279 (0.101) | 25.032 (0.092) |
| S9 | 3 h 22 m 05 . s 36477 | - 15 ◦ 23 ′ 15 . ′′ 0634 | 27.683 (0.127) | 27.470 (0.050) | 26.266 (0.042) | 24.753 (0.069) |
| S10 | 3 h 22 m 05 . s 39605 | - 15 ◦ 23 ′ 15 . ′′ 4062 | > 28.912 | 29.305 (0.273) | 26.969 (0.077) | 25.214 (0.102) |
| S11 | 3 h 22 m 05 . s 37904 | - 15 ◦ 23 ′ 14 . ′′ 5218 | 27.196 (0.083) | 27.329 (0.045) | 26.949 (0.076) | > 26.614 |
| S12 | 3 h 22 m 05 . s 37556 | - 15 ◦ 23 ′ 14 . ′′ 4202 | > 28.897 | 28.302 (0.109) | 27.095 (0.085) | 25.381 (0.131) |
| S13 | 3 h 22 m 05 . s 36527 | - 15 ◦ 23 ′ 14 . ′′ 5054 | 28.085 (0.181) | 27.560 (0.055) | 26.852 (0.069) | 25.740 (0.171) |
| S14 | 3 h 22 m 05 . s 40432 | - 15 ◦ 23 ′ 15 . ′′ 5778 | > 29.360 | 28.651 (0.113) | 26.078 (0.026) | 24.180 (0.042) |
| S15 | 3 h 22 m 05 . s 43040 | - 15 ◦ 23 ′ 14 . ′′ 4706 | 27.937 (0.164) | 28.465 (0.122) | 28.078 (0.208) | > 26.512 |
| S16 | 3 h 22 m 05 . s 42376 | - 15 ◦ 23 ′ 14 . ′′ 3726 | > 28.765 | 28.163 (0.092) | 26.905 (0.072) | 25.126 (0.094) |
| S17 | 3 h 22 m 05 . s 38194 | - 15 ◦ 23 ′ 14 . ′′ 2494 | > 28.926 | 29.007 (0.207) | 26.957 (0.077) | 26.230 (0.260) |
| S18 | 3 h 22 m 05 . s 35086 | - 15 ◦ 23 ′ 15 . ′′ 2402 | 27.831 (0.143) | 27.819 (0.069) | 26.969 (0.077) | 25.581 (0.145) |
| S19 | 3 h 22 m 05 . s 39387 | - 15 ◦ 23 ′ 14 . ′′ 2018 | 28.670 (0.302) | 28.417 (0.118) | 26.616 (0.056) | 25.072 (0.094) |
| S20 | 3 h 22 m 05 . s 37179 | - 15 ◦ 23 ′ 15 . ′′ 6186 | > 28.814 | 28.643 (0.147) | 26.660 (0.059) | 24.854 (0.075) |
| S21 | 3 h 22 m 05 . s 34566 | - 15 ◦ 23 ′ 15 . ′′ 4618 | > 28.813 | > 29.632 | 27.291 (0.102) | 25.572 (0.138) |
| S22 | 3 h 22 m 05 . s 39179 | - 15 ◦ 23 ′ 14 . ′′ 0298 | > 28.809 | 29.135 (0.235) | 27.219 (0.100) | 25.685 (0.161) |
| S23 | 3 h 22 m 05 . s 43718 | - 15 ◦ 23 ′ 15 . ′′ 6194 | 28.340 (0.241) | 29.474 (0.331) | > 28.665 | > 26.633 |
| S24 | 3 h 22 m 05 . s 35072 | - 15 ◦ 23 ′ 15 . ′′ 6022 | > 28.835 | 28.097 (0.089) | 26.975 (0.078) | 25.936 (0.202) |
| S25 | 3 h 22 m 05 . s 38396 | - 15 ◦ 23 ′ 13 . ′′ 9958 | 28.367 (0.230) | 28.382 (0.119) | 26.046 (0.035) | 24.549 (0.057) |
| S26 | 3 h 22 m 05 . s 33025 | - 15 ◦ 23 ′ 15 . ′′ 0438 | 27.518 (0.109) | 27.466 (0.055) | 26.284 (0.048) | 25.134 (0.094) |
| S27 | 3 h 22 m 05 . s 46232 | - 15 ◦ 23 ′ 14 . ′′ 9126 | > 28.841 | 28.581 (0.137) | 27.093 (0.084) | 25.445 (0.126) |
| S28 | 3 h 22 m 05 . s 39871 | - 15 ◦ 23 ′ 13 . ′′ 9614 | > 28.808 | 28.522 (0.132) | 27.052 (0.084) | 25.443 (0.137) |
| S29 | 3 h 22 m 05 . s 39210 | - 15 ◦ 23 ′ 13 . ′′ 9558 | > 28.908 | 28.386 (0.117) | 28.395 (0.288) | 26.075 (0.241) |
| S30 | 3 h 22 m 05 . s 40792 | - 15 ◦ 23 ′ 13 . ′′ 8970 | > 28.952 | 28.692 (0.154) | 27.315 (0.102) | 25.459 (0.135) |
| S31 | 3 h 22 m 05 . s 32555 | - 15 ◦ 23 ′ 14 . ′′ 6310 | > 28.961 | > 29.657 | 28.039 (0.201) | 25.475 (0.127) |
| S32 | 3 h 22 m 05 . s 32419 | - 15 ◦ 23 ′ 15 . ′′ 2534 | 27.612 (0.122) | 27.707 (0.062) | 27.701 (0.150) | > 26.591 |
| S33 | 3 h 22 m 05 . s 37008 | - 15 ◦ 23 ′ 15 . ′′ 9726 | 28.700 (0.320) | 28.055 (0.088) | 28.035 (0.206) | > 26.599 |
| S34 | 3 h 22 m 05 . s 42642 | - 15 ◦ 23 ′ 15 . ′′ 9482 | 27.164 (0.086) | 27.302 (0.045) | 26.856 (0.072) | > 26.520 |
| S35 | 3 h 22 m 05 . s 38581 | - 15 ◦ 23 ′ 16 . ′′ 0310 | 28.710 (0.329) | 27.696 (0.062) | 28.152 (0.227) | > 26.624 |
| S36 | 3 h 22 m 05 . s 34751 | - 15 ◦ 23 ′ 15 . ′′ 8186 | > 28.843 | > 29.672 | 27.447 (0.120) | 25.444 (0.127) |
| S37 | 3 h 22 m 05 . s 47012 | - 15 ◦ 23 ′ 15 . ′′ 2990 | 26.530 (0.051) | 26.607 (0.024) | 26.564 (0.053) | 25.154 (0.095) |
| S38 | 3 h 22 m 05 . s 35077 | - 15 ◦ 23 ′ 13 . ′′ 9774 | > 28.916 | > 29.672 | 27.501 (0.123) | 25.776 (0.173) |
| S39 | 3 h 22 m 05 . s 38399 | - 15 ◦ 23 ′ 13 . ′′ 7882 | > 28.831 | 27.847 (0.069) | 27.555 (0.138) | 25.455 (0.123) |
| S40 | 3 h 22 m 05 . s 42445 | - 15 ◦ 23 ′ 16 . ′′ 0202 | > 28.775 | 27.545 (0.057) | 26.175 (0.040) | 24.407 (0.050) |
| S41 | 3 h 22 m 05 . s 33343 | - 15 ◦ 23 ′ 14 . ′′ 1682 | 26.522 (0.049) | 26.491 (0.022) | 26.209 (0.041) | 25.470 (0.128) |
| S42 | 3 h 22 m 05 . s 31449 | - 15 ◦ 23 ′ 15 . ′′ 0958 | 28.411 (0.246) | 29.178 (0.236) | > 28.665 | > 26.583 |
Extended Data Table 2 | Photometric variability of the SN 2012Z progenitor system S1. The 1 σ photometric uncertainties are given in parentheses. No corrections for Galactic or host extinction are made here.
| Date (UT) | MJD | Exposure (sec) | Counts ( e - ) | F555W (mag) |
|-------------|---------|------------------|------------------|---------------|
| 2005-08-06 | 53588.7 | 2400 | 171 (71) | 28.36 (0.45) |
| 2005-08-17 | 53600 | 2400 | 656 (75) | 26.94 (0.13) |
| 2005-08-24 | 53606.8 | 2400 | 282 (67) | 27.79 (0.26) |
| 2005-08-27 | 53610 | 2400 | 392 (64) | 27.49 (0.18) |
| 2005-09-02 | 53615.6 | 2400 | 389 (69) | 27.48 (0.19) |
| 2005-09-03 | 53617 | 2400 | 308 (66) | 27.77 (0.23) |
| 2005-09-05 | 53618.6 | 2400 | 392 (71) | 27.49 (0.20) |
| 2005-09-07 | 53621 | 2400 | 329 (66) | 27.66 (0.22) |
| 2005-09-11 | 53624 | 2400 | 429 (68) | 27.37 (0.17) |
| 2005-09-16 | 53629.8 | 2400 | 242 (67) | 27.98 (0.30) |
| 2005-09-20 | 53633.4 | 2400 | 327 (64) | 27.70 (0.21) |
| 2005-09-27 | 53640.5 | 2400 | 193 (66) | 28.25 (0.37) |
| 2006-10-24 | 54032.6 | 2080 | 388 (76) | 27.26 (0.21) |
| 2006-10-07 | 54015.3 | 2080 | 264 (75) | 27.62 (0.31) |
|
10.1038/nature13615
|
[
"Curtis McCully",
"Saurabh W. Jha",
"Ryan J. Foley",
"Lars Bildsten",
"Wen-fai Fong",
"Robert P. Kirshner",
"G. H. Marion",
"Adam G. Riess",
"Maximilian D. Stritzinger"
] |
2014-08-05T20:00:00+00:00
|
2014-08-05T20:00:00+00:00
|
[
"astro-ph.SR",
"astro-ph.HE"
] |
A luminous, blue progenitor system for a type-Iax supernova
|
Type-Iax supernovae (SN Iax) are stellar explosions that are
spectroscopically similar to some type-Ia supernovae (SN Ia) at maximum light,
except with lower ejecta velocities. They are also distinguished by lower
luminosities. At late times, their spectroscopic properties diverge from other
SN, but their composition (dominated by iron-group and intermediate-mass
elements) suggests a physical connection to normal SN Ia. These are not rare;
SN Iax occur at a rate between 5 and 30% of the normal SN Ia rate. The leading
models for SN Iax are thermonuclear explosions of accreting carbon-oxygen white
dwarfs that do not completely unbind the star, implying they are "less
successful" cousins of normal SN Ia, where complete disruption is observed.
Here we report the detection of the luminous, blue progenitor system of the
type-Iax SN 2012Z in deep pre-explosion imaging. Its luminosity, colors,
environment, and similarity to the progenitor of the Galactic helium nova V445
Puppis, suggest that SN 2012Z was the explosion of a white dwarf accreting from
a helium-star companion. Observations in the next few years, after SN 2012Z has
faded, could test this hypothesis, or alternatively show that this supernova
was actually the explosive death of a massive star.
|
1408.1090v1
|
Mon. Not. R. Astron. Soc. 000 , 1-12 (2014)
Printed 7 August 2014 (MN LT E X style file v2.2) A
## Inter-comparison of Radio-Loudness Criteria for Type 1 AGNs in the XMM-COSMOS Survey
Heng Hao 1 2 , /star , Mark T. Sargent 3 4 5 , , , Martin Elvis , Eva Schinnerer , Gianni 2 5 Zamorani , Luis C. Ho 6 7 8 9 , , , Jennifer L. Donley 10 , Francesca Civano 2 11 12 , , , Vernesa Smolˇ ci´ c 13 , Annalisa Celotti 1 14 15 , , , Joanna Kuraszkiewicz , Mara Salvato 2 16 17 , , Marcella Brusa 6 18 19 , , , Peter Capak 20 , Chris L. Carilli 21 , Andrea Comastri , 6 Chris D. Impey 22 , Knud Jahnke , Anton M. Koekemoer 5 23 , Kevin Schawinski 24 , Jonathan R. Trump 25 , C. Megan Urry 11 , Cristian Vignali 6 18 , , Min Yun 26
- 1 SISSA, Via Bonomea 265, I-34136 Trieste, Italy
- 2 Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA 02138, USA
- 3 Astronomy Center, Department of Physics and Astronomy, University of Sussex, Brighton BN1 9QH, UK
- 4 CEA-Saclay, Service d'Astrophysique, Orme des Merisiers, Bat. 709, 91191 Gif-sur-Yvette, France
- 5 Max-Planck-Institut f¨r Astronomie, K¨nigstuhl 17, Heidelberg, D-69117, Germany u o
- 6 INAF - Osservatorio Astronomico di Bologna, via Ranzani 1, I-40127 Bologna, Italy
- 7 Kavli Institute for Astronomy and Astrophysics, Peking University, Beijing 100871, China
- 8
- Department of Astronomy, Peking University, Beijing, China
- 9 The Observatories of the Carnegie Institute for Science, Santa Barbara Street, Pasadena, CA 91101, USA 10
- Los Alamos National Laboratory, Los Alamos, NM 87544, USA
- 11 Yale Center for Astronomy and Astrophysics, 260 Whitney ave, New Haven, CT 06520, USA
- 12 Dartmouth College, Department of Physics and Astronomy, 6127 Wilder Lab, Hanover, NH 03755
- 13 Physics Department, University of Zagreb, Bijeniˇka cesta 32, 10002 Zagreb, Croatia c
- 14 INAF - Osservatorio Astronomico di Brera, via E. Bianchi 46, I-23807 Merate, Italy
- 15 INFN - Sezione di Trieste, via Valerio 2, 34127, Trieste, Italy
- 16 IPP - Max-Planck-Institute for Plasma Physics, Boltzmann Strasse 2, D-85748, Garching bei M¨nchen, Germany u 17
- Excellence Cluster, Boltzmann Strasse 2, D-85748, Garching bei M¨nchen, Germany u
- 18 Dipartimento di Fisica e Astronomia, Universit` degli studi di Bologna, viale Berti Pichat 6/2 40127 Bologna, Italy a
- 19 Max Planck Institute f¨r Extraterrestrische Physik, Postfach 1312, 85741, Garching bei M¨nchen, Germany u u
- 20 Spitzer Science Center, MS 314-6, California Institute of Technology, Pasadena, CA 91125, USA
- 21 National Radio Astronomy Observatory, P.O. Box 0, Socorro, NM 87801, USA
- 22 Steward Observatory, University of Arizona, 933 North Cherry Avenue, Tucson, AZ 85721, USA
- 23 Space Telescope Science Institute, 3700 San Martin Drive, Baltimore, MD 21218, USA
- 24 Institute for Astronomy, Department of Physics, ETH Zurich,Wolfgang-Pauli-Strasse 16, CH-8093 Zurich, Switzerland
- 25 Department of Astronomy and Astrophysics, Pennsylvania State University, University Park, PA 16802, USA
- 26 Department of Astronomy, University of Massachusetts, Amherst, MA 01003, USA
Version Jun 1st, 2014.
## ABSTRACT
Limited studies have been performed on the radio-loud fraction in X-ray selected type 1 AGN samples. The consistency between various radio-loudness definitions also needs to be checked. We measure the radio-loudness of the 407 type 1 AGNs in the XMM-COSMOS quasar sample using nine criteria from the literature (six defined in the rest-frame and three defined in the observed frame): R L = log( L 5 GHz /L B ), q 24 = log( L 24 µm /L 1 4 . GHz ), R uv = log( L 5 GHz /L 2500 ˚ ), A R i = log( L 1 4 . GHz /L i ), R X = log( νL ν (5 GHz /L ) X ), P 5 GHz = log( P 5 GHz ( W/Hz/Sr )), R L,obs = log( f 1 4 . GHz /f B ) (observed frame), R i,obs = log( f 1 4 . GHz /f i ) (observed frame), and q 24 ,obs = log( f 24 µm /f 1 4 . GHz ) (observed frame). Using any single criterion defined in the rest-frame, we find a low radio-loud fraction of /lessorsimilar 5% in the XMMCOSMOS type 1 AGN sample, except for R uv . Requiring that any two criteria agree reduces the radio-loud fraction to /lessorsimilar 2% for about 3/4 of the cases. The low radio-loud fraction cannot be simply explained by the contribution of the host galaxy luminosity and reddening. The P 5 GHz = log( P 5 GHz ( W/Hz/Sr )) gives the smallest radio-loud fraction. Two of the three radio-loud fractions from the criteria defined in the observed frame without k-correction ( R L,obs and R i,obs ) are much larger than the radio-loud fractions from other criteria.
Key words: galaxies: evolution; quasars: general; surveys
## 1 INTRODUCTION
Quasars are often classified into radio-loud (RL) and radioquiet (RQ), based on the presence or absence of strong radio emission. Radio-loud quasars are generally some three orders of magnitude more powerful at GHz radio frequencies relative to their optical or infrared fluxes than RQ quasars, while in the rest of their spectral energy distributions (SEDs), from mid-infrared to X-ray, there are only subtle differences between them (e.g., Elvis et al., 1994, E94 hereinafter). The strong radio emission is a result of RL quasars having a relativistic jet that generates synchrotron radiation in the radio (see review by Harris & Krawczynski 2006).
However, even RQ quasars can be detected as radio sources (Kellermann et al. 1989). This has led to two opposing views of the radio-loudness distribution which have long been debated. The first is that the radio-loudness distribution is bimodal (e.g. Kellermann et al. 1989; Miller et al. 1990; Visnovsky et al. 1992; Ivezi´ et al. 2002). The other c is that the distribution is continuous with no clear dividing line (Cirasuolo et al. 2003).
Typically, ∼ 10% of all quasars in optically selected samples are RL (e.g. Kellermann et al. 1989; Urry & Padovani 1995; Ivezi´ et al. 2002). Here we examine the RL fraction c of the 413 type 1 AGNs in the XMM-COSMOS sample to check both the RL fraction in X-ray selected samples and the consistency among various radio-loudness criteria. In this paper, we adopt the WMAP 5-year cosmology (Komatsu et al., 2009), with H 0 =71 km s -1 Mpc -1 , Ω M = 0.26 and Ω Λ = 0.74.
## 2 COSMOS TYPE 1 AGN PHOTOMETRY
The XMM-COSMOS survey (Hasinger et al. 2007) detected 1848 point sources down to ∼ 10 -15 erg cm -2 s -1 . Using a likelihood ratio technique, Brusa et al (2007, 2010) identified unique counterparts of 1577 (85%) XMM-COSMOS sources in the optical photometric catalog (Capak et al. 2007). A total of 886 XMM-COSMOS sources ( ∼ 50%) have well-determined spectroscopic redshifts from optical spectra (Trump et al. 2009a, Schneider et al. 2007, Lilly et al. 2007, 2009). From these spectra, 413 are identified as type 1 AGN, with emission line FWHM > 2000 km s -1 , forming the XMM-COSMOS type 1 AGN sample (XC413, Elvis et al. 2012, hereafter Paper I). The XC413 sample has 43 photometry bands extending from radio to X-ray, and spans a large redshift range (0 1 . /lessorequalslant z /lessorequalslant 4 3, with median 1.6), as well . as both large apparent magnitude (16 8 . /lessorequalslant i AB /lessorequalslant 25 0 with . median 21.2) and intrinsic luminosity (44 3 . /lessorequalslant log L bol /lessorequalslant 47 4 . with median 45.7) ranges. We now briefly review the optical, infrared (IR) and radio flux measurements which are crucial to the analysis.
All 413 quasars in XC413 have B band and J band detections (Paper I). In the mid-infrared range, 385 detections were obtained in the S-COSMOS MIPS 24 µm imaging (Sanders et al. 2007; Le Floc'h et al. 2009) by searching for counterparts within 2 ′′ of the optical counterpart. With the exception of one source (located outside the 24 µm survey area), we derived 3 σ upper flux density limits ( ∼ 54 µ Jy on average) for the remaining 27 unmatched quasars using a coverage-based rms map.
In the radio, the VLA-COSMOS 1.4 GHz survey detected 2865 sources at S/N=5 (rms=8-12 µ Jy, depending on the position in the field; Schinnerer et al. 2010). Radio counterparts to the XMM X-ray sources were determined by cross-correlating the optical quasar positions to source positions in the VLA-COSMOS joint catalog (Schinnerer et al. 2010) within a radius of 1 ′′ . This resulted in 61 (15% of 413) successful matches with the XC413 sample.
For the unmatched XMM-COSMOS quasars, the AIPS/MAXFIT peak finding algorithm was used to search for additional radio detections within a 2.5 ′′ × 2.5 ′′ box centered on the optical coordinates. The box size is chosen because the resolution of the radio beam is 2.5 ′′ . This yielded 78 additional detections in the 3 σ -5 σ range. In all, we have 139 sources with larger than 3 σ radio detection. We computed their total flux assuming that they are unresolved at 1.4GHz (beam FWHM 2.5 ′′ ). For lower significance peaks (286 out of 413) we adopted 3 σ upper flux density limits based on the local rms noise (calculated within a 17.5 ′′ × 17.5 ′′ box) at the position of the radio source. This box size was chosen based on the tests we made to obtain the most accurate map for the VLA-COSMOS Deep Project mosaic.
In summary, out of the 413 XMM-COSMOS quasars, 407 have either > 3 σ VLA detections or upper limits. We have no radio flux information about the remaining 6 AGNs as they lie outside the VLA-COSMOS 1.4 GHz coverage area, and in one case also outside the MIPS-COSMOS 24 µ m coverage area. We will only discuss the radio loudness for these 407 type 1 AGNs (XC407 sample) in this paper.
## 3 RADIO-LOUDNESS DEFINITIONS
Several criteria have been used to classify quasars as RL or RQ. As noted, mere radio detection is not enough. Radio power, either alone or relative to some other band is typically used. We determined the radio-loudness of the XC407 for nine definitions currently in use in the literature:
- (i) R L : the luminosity ratio of radio to optical emission R L = log( L 5 GHz /L B ) (Wilkes & Elvis 1987, Kellermann et al. 1989), with R L > 1 defining a RL source. This logarithmic (base ten, as for all following definitions) radio-tooptical luminosity ratio is the most widespread criterion for RL. COSMOS does not have 5 GHz coverage, but, as most of the sample sources are at redshift 1-2 (Paper I), the observed 1.4 GHz VLA band is close to the emitted 5 GHz frequency. We converted the observed 1.4 GHz luminosity to a rest-frame 5 GHz luminosity by assuming f ν ∝ ν -0 5 . (e.g., Ivezi´ 2004). c For most of the quasars in the XC407 sample at redshift z ∼ 2, the observed 1.4 GHz is at restframe ∼ 4 2 GHz, the residual k-correction is only 9%. The . B band luminosity is the luminosity at rest-frame B band ( λ eff = 4483 ˚ ) retrieved from the rest-frame SED for each A quasar, that is the linear interpolation of the adjacent observed photometry after moving to the rest-frame.
(ii) R L,obs : We also calculated R L,obs = log( f 1 4 . GHz /f B ) in the observed frame without k-correction for comparison, with R L,obs > 1 defined as radio-loud. This criteria is typically used when redshift information is not available.
As most of the XMM-COSMOS AGNs lie at redshift 1-2,
Figure 1. Distribution of radio-loudness measures: R L , R i , q 24 , R uv , R X , P 5 GHz , R L,obs , R i,obs , and q 24 ,obs . The black solid line show the distribution for quasars with radio detections; the blue dashed line show the distribution for quasars with upper limits in radio; the red dotted line show the distribution for all the 407 quasars with radio detection or upper limits.
the 1.4 GHz band lies close to rest-frame 5 GHz and the Jband lies close to rest-frame B band. Hence, for a consistency check, we can adopt R J = log( f 1 4 . GHz /f J ) in the observed frame as an alternative definition to RL that does not involve any assumptions about the k-correction.
(iii) R i : Balokovi´ et al. (2012) defined the radio-loudness c as the radio to i band luminosity ratio: R i = log( L 5 GHz /L i ). Here we calculate R i using the same k-correction in the radio as in R L and retrieve the rest-frame i band ( λ eff = 7523 ˚ ) A luminosity from the rest-frame SED by interpolation. We define R > i 1 as radio-loud.
as R i,obs = log( f 1 4 . GHz /f i ) in the observed frame without k-correction and R i,obs > 1 was defined as radio-loud. Considering most of the XMM-COSMOS AGNs lie at redshift 1-2, the 1.4 GHz band lies close to rest-frame 5 GHz and the K-band lies close to rest-frame i band. Similarly, we can adopt R K = log( f 1 4 . GHz /f K ) in the observed frame without k-correction as an alternative definition to radio-loudness.
- (iv) R i,obs : Ivezi´ et al. (2002) defined the radio-loudness c
- (v) q 24 : Appleton et al. (2004) introduced a new definition of radio-loudness using the Spitzer/MIPS 24 µ m flux: q 24 = log( L 24 µm /L 1 4 . GHz ), with q 24 < 0 defined as RL. We calculated q 24 in the rest-frame assuming the same power law in the radio as for the R L definition and the rest frame
24 µ m retrieved from the rest-frame SED of each quasar. Kuraszkiewicz et al. (in preparation) studied the correlation between q 24 and R L for a sample with limited reddening or host contamination, and find that, on average, R L > 1 corresponds to q 24 < . 0 24, rather than q 24 < 0. We use q 24 ,K to represent the criterion using this alternative dividing line.
(vi) q 24 ,obs : For easy comparison with R L,obs , which is defined in the observed frame, we also check q 24 in the observed frame and define it as q 24 ,obs = log( f 24 µm /f 1 4 . GHz ).
Out of the 407 quasars, 25 have upper flux density limits both at 1.4 GHz and 24 µ m. For these sources, q 24 and q 24 ,obs cannot be determined and we thus excluded these 25 sources from the discussion of the results for q 24 and q 24 ,obs .
(vii) R uv : Stocke et al. (1992) used the ratio between the 2500 ˚ A UV luminosity and radio luminosity: R uv = log( L 5 GHz /L 2500 ˚ ) A (see also Jiang et al. 2007). This criterion has the advantage of using the peak of the SED to define radio-loudness, but is strongly affected even by modest reddening. For example, E B ( -V ) > 0 3 decreases the 2500 ˚ . A UV luminosity by a factor > 7. The effect is to make more objects appear RL in X-ray selected samples, such as the XC407 sample, as these include a large number of sources with significant optical reddening compared to opticallyselected samples (Paper I, Hao et al. 2013, 2014). This criterion is therefore less useful for X-ray selected samples. In the Einstein Extended Medium Sensitivity Survey (EMSS), radio-loudness is defined with α ro = R uv / . 5 38 (Della Ceca et al. 1994, Zamorani et al. 1981). Sources with α ro > 0 35 . are defined as RL, that is equivalent to R uv > . 1 88. We use R uv,D to represent the criterion.
(viii) R X : Terashima & Wilson (2003) proposed a criterion based on the ratio between luminosity at 5 GHz and X-ray luminosity in the 2 -10 keV band. Sources with R X = log( νL ν (5 GHz /L ) X ) > -3 are RL (Pierce et al. 2011). This criterion is working both for heavily obscured AGN, N H /lessorsimilar 10 23 cm -2 and of being free of host galaxy contamination.
(ix) P 5 GHz : Goldschmidt et al. (1999) proposed a criterion based solely on radio power, where sources with P 5 GHz = log[ P 5 GHz ( W/Hz/Sr )] > 24 are considered to be RL. Given that Goldschmidt et al. (1999) assume a Hubble parameter H =50 km s 0 -1 Mpc -1 , this criterion corresponds to P 5 GHz = log[ P 5 GHz ( W/Hz/Sr )] > 23 7 for the cosmol-. ogy used in this paper.
The distributions of the nine radio loudness measures are shown in Figure 1. From these plots we could only see continuous distributions with long tail on the radio loud side and there is no clear sign of bimodality. The size of the sample is still too small to give statistical significant check on the bimodality of the radio loudness measures.
## 4 RADIO-LOUD FRACTION
## 4.1 Single Criterion
The numbers and fractions of RL quasars in the sample using the nine different RL selection criteria are summarized in Table 1. The RL-fraction is calculated (a) using KaplanMeier product limit estimator (Kaplan & Meier 1958) if the
Table 1. Radio-Loud Quasars by Different Criteria
| criterion | N(RL) | N(RQ) | N(amb) ∗ | Fraction(RL) |
|---------------------------|-----------|---------|------------|----------------------------------------------------|
| R L > 1 | 18 | 367 | 22 | 4.50% +2 . 73% - 1 . 52% |
| R i > 1 | 16 | 364 | 27 | 3.98% +2 . 67% - 1 . 38% |
| q 24 < 0 q 24 ,K < 0 . 24 | 13 a 25 a | 357 307 | 12 50 | 3.43% +1 . 67% - 1 . 11% 7.22% +2 . 43% - 2 . 70% |
| R uv > 1 R uv,D > 1 . 88 | 45 9 | 283 398 | 79 0 | 11.75% +4 . 02% - 2 . 60% 2.21% +2 . 09% - 1 . 00% |
| R X > - 3 | 10 | 397 | 0 | 2.46% +1 . 88% - 1 . |
| P 5 GHz > 23 . 7 | 8 | 399 | 0 | 23% 1.97% +1 . 83% - 1 . 03% |
| R L,obs > 1 | 67 | | 120 | 19.23% +4 . 92% - 3 . 58% +2 . 71% |
| R J > 1 | 16 | 220 360 | 31 | 4.02% - 1 . 39% |
| R i,obs > 1 R K > 1 | 38 10 | 300 392 | 69 5 | 10.03% +3 . 79% - 2 . 43% 2.46% +2 . 23% - 1 . 04% |
| q 24 ,obs < 0 | 8 | 370 | 4 b | 2.55% +1 . - 1 . 06% |
| | | | | 49% |
∗ Number of ambiguous sources that with upper/lower limits in the radio-loud region. See § 4 for details.
a,b The 2 sources (XMM ID: 320 and 5315) with VLA detection and MIPS upper limits have upper limits on q 24 and q 24 ,obs . Therefore: (a) they are already located in the RL region with upper limits, so they are RL by the q 24 criterion; (b) they are located in the RQ region with upper limits, so they can still be RL, so they are ambiguous for the q 24 ,obs criterion.
sample is singly-censored or (b) following the iterative procedure in Schmitt (1985) if both upper and lower limits on the loudness diagnostic are present ( q 24 and q 24 ,obs ). We also list the number of ambiguous sources in the table, which are those with upper/lower limits lying in the RL region.
The fraction of radio-loud quasars spans a wide range from ∼ 2% to ∼ 20%.
For the criteria defined in the rest-frame, R uv classifies the largest number ( ∼ 12%) of COSMOS AGNs as RL, where most of these quasars are RQ by all other criteria defined in the rest frame. The SEDs of the quasars classified as RL by R uv but not the other criteria generally do not show a 'big blue bump' feature in their SEDs that is characteristic of unobscured quasars (Paper I). We plot the sources which have either a direct measurement or upper limit of R uv that exceeds the selection threshold for radio loudness ( R uv = 1), but which are RQ according to all alternative definitions defined in the rest-frame, in the Hao et al. (2013) mixing diagram (Figure 2). These quasars are mainly located in the high reddening ( E B ( -V ) > 0 2) or . high galaxy fraction ( f g > 0 4) regions and well away from . the E94 SED region (red circle in Figure 2). This suggests that they are mostly reddened or galaxy dominated sources and their apparent radio-loudness by R uv is due to these contaminating factors. Note that if we change the RQ and RL dividing line, i.e. if we use the R uv,D criterion instead, the number of ambiguous sources would be much lower and the radio-loud fraction will drop to ∼ 2%. This criterion is more appropriate for X-ray selected samples.
Table 2. Radio-Loud Quasars with Two Criteria
| criterion | N(RL) | N(RQ) | N(amb) ∗ | N(amb) ∗ | N(amb) ∗ | Fraction(RL) |
|--------------------------------------|----------|----------|---------------|-------------|------------|----------------|
| | | | 1 | 2 | 2 | |
| R L > 1 q 24 < 0 | 8 | 339 | 10 5 a | 8 8 | 4 | 2.1% |
| R L > 1 R uv > 1(1 . 88) ∗∗ | 18(8) | 283(366) | 0(10) 27(1) | 0(22) 57(0) | 22(0) | 4.4%(2.0%) |
| R L > 1 R i > 1 | 14 | 355 | 4 2 | 5 10 | 17 | 3.4% |
| R L > 1 R X > - 3 | 7 | 364 | 11 3 | 22 0 | 0 | 1.7% |
| R L > 1 P 5 GHz > 23 . 7 | 6 | 365 | 12 2 | 22 0 | 0 | 1.5% |
| q 24 < 0 R uv > 1(1 . 88) ∗∗ | 13 b (7) | 270(355) | 0(6 a ) 32(2) | 5(12) 55(0) | 7(0) | 3.4%(1.8%) |
| q 24 < 0 R i > 1 | 8 | 342 | 5 a 8 | 7 7 | 5 | 2.1% |
| q 24 < 0 R X > - 3 | 7 | 354 | 6 a 3 | 12 0 | 0 | 1.8% |
| q 24 < 0 P 5 GHz > 23 . 7 | 6 | 355 | 7 a 2 | 12 0 | 0 | 1.6% |
| R uv > 1(1 . 88) ∗∗ R i > 1 | 15(7) | 277(362) | 30(2) 1(9) | 57(0) 5(27) | 22(0) | 3.7%(1.7%) |
| R uv > 1(1 . 88) ∗∗ R X > - 3 | 8(6) | 281(394) | 37(3) 2(4) | 79(0) 0(0) | 0(0) | 2.0%(1.5%) |
| R uv > 1(1 . 88) ∗∗ P 5 GHz > 23 . 7 | 6(5) | 281(395) | 39(4) 2(3) | 79(0) 0(0) | 0(0) | 1.5%(1.2%) |
| R i > 1 R X > - 3 | 9 | 363 | 7 1 | 27 0 | 0 | 2.2% |
| R i > 1 P 5 GHz > 23 . 7 | 6 | 362 | 10 2 | 27 0 | 0 | 1.5% |
| R X > - 3 P 5 GHz > 23 . 7 | 7 | 396 | 3 1 | 0 0 | 0 | 1.7% |
| R L,obs > 1 q 24 ,obs < 0 | 8 | 218 | 57 0 | 95 2 c | 2 | 2.1% |
| R i,obs > 1 q 24 ,obs < 0 | 8 | 293 | 28 0 | 49 2 c | 2 | 2.1% |
| R L,obs > 1 R i,obs < 0 | 38 | 220 | 29 0 | 51 0 | 69 | 9.3% |
∗ Number of ambiguous sources: 1) they are RL with one criterion, but RQ with the other; 2) the upper/lower limits locate in the radio-loud region for one or both criteria, that they could be RQ. See § 4 for details. ∗∗ The numbers in the parenthesis are for the R uv,D criterion.
a,b,c The 2 sources (XMM ID: 320 and 5315) with VLA detection and MIPS upper limits have upper limits on q 24 and q 24 ,obs . Therefore: (a) they are already located in the RL region with upper limits, so they are RL by the q 24 criterion. But they are not RL in the other criterion, so they are still ambiguous sources; (b) they are already located in the RL region with upper limits, so they are RL by the q 24 criterion. And they are RL by the R uv criterion, so they are not ambiguous sources; (c) they are located in the RQ region with the upper limits, so they can still be RL, so they are ambiguous for the q 24 ,obs criterion.
Figure 2. Mixing diagram (Hao et al. 2012) for RL quasars according to the R uv criterion, but classified as RQ by all other selections. Black points are used for quasars with radio detections, blue points for 'ambiguous' quasars (upper limit on R uv exceeding threshold R uv =1). The mixing diagram is the plot of the SED slopes α NIR (3 µm to 1 µm ) versus α OPT (1 µm to 0 3 . µm ). The E94 RQ mean SED is shown as the red cross. The red circle shows the dispersion of E94. The green triangles show galaxy templates from the SWIRE template library (Polletta et al. 2007). The purple arrow represents the reddening vector for the E94 RQ mean SED. The cyan lines connect the galaxy templates and the E94 mean SED are mixing curves showing templates with different galaxy fractions.
Therefore, for the criteria defined in the rest-frame, the RL fraction is 2.0%-4.5% using any single criterion (for R uv considering only the R uv,D ). This is small compared to the ∼ 10% seen in typical optically selected AGN samples (e.g. Peterson et al. 1997). To reach 10% would require an additional 22-32 AGNs to be classified as RL. The P 5 GHz ( ∼ 2.0%) is the most restrictive definition of radio-loudness, while R L ( ∼ 4.5%) gives somewhat larger RL samples.
For the criteria defined in the observed frame, we note that no k-correction is included, so for quasars at different redshift, the ratio is actually calculated at a different frequency. The radio loudness definition with radio-to-optical ratio all yield a high radio-loud fraction ( > 10%) in contrast to all the other criteria. This is because the quasars in the sample have a large redshift range (0 1 . /lessorequalslant z /lessorequalslant 4 3). . If we consider that most of the quasars in the XMM-COSMOS sample are at redshift 1-2, the rest-frame B band, i band and 5GHz would be at J band, K band and 1.4GHz in the observed frame, then we can define R J and R K instead of R L,obs and R i,obs , the radio-loud fraction reduce to ∼ 4% and ∼ 2 5% respectively. Although . q 24 ,obs also does not include k-correction, it still yields a low radio-loud fraction (2.55%).
Table 3. Radio-Loud Quasar Properties
| XID | z | log( L bol ) ∗ | log( M BH /M ⊙ ) ∗∗ | λ ∗∗∗ Edd |
|-------|-------|------------------|-----------------------|-------------|
| 40 | 0.971 | 45.24 | · · · | · · · |
| 2282 | 1.541 | 45.38 | · · · | · · · |
| 5230 | 1.317 | 46.44 | 8.21 | 1.337 |
| 5257 | 1.403 | 46.07 | 9.06 | 0.080 |
| 5395 | 1.472 | 45.68 | · · · | · · · |
| 5517 | 2.132 | 46.24 | 8.70 | 0.277 |
∗ log( L bol ) is calculated by integrating the rest-frame SED from 1 µm to 40 keV.
∗∗
The black hole mass estimates are from Trump et al. (2009b)
and Merloni et al. (2010).
∗∗∗
Eddington ratio
$$\lambda _ { E d d } = \frac { L _ { b o l } } { L _ { E d d } } = \frac { L _ { b o l } } { \frac { 4 \pi G c m _ { p } } { \sigma _ { e } } M _ { B H } } = \frac { L _ { b o l } } { 1. 2 6 \times 1 0 ^ { 3 8 } ( M _ { B H } / M _ { \odot } ) }$$
## 4.2 Pairs of Criteria
We also consider 18 pairs of RL criteria (Table 2). To avoid confusion with k-correction effects, criteria in the observed frame are not compared with criteria in the rest-frame. Thus, these 18 pairs include all the possible combinations of criteria pairs. The RL fraction is determined by the number of objects that lie in the RL-selection region for both measures of radio-loudness divided by the total number of sources in the sample (407; when q 24 or q 24 ,obs included, it is 382). The 'ambiguous' sources include: 1. sources identified as RL by one criterion (detections only) but RQ with the other; 2. sources with only upper/lower limits which lie in the RL region.
If we require any two criteria to agree, then the RL fractions are even smaller (1.5%-4.4%, Table 2) except for R L,obs and R i,obs . For about 3/4 of all possible combinations, the RL fraction are smaller or around 2%.
## 4.3 Solidly Radio Loud Quasars
Finally, if we require all the above criteria to be satisfied simultaneously, just six sources remain 1 . The according RL fraction is thus 1.5% (6 out of 407), only marginally lower than that obtained for a classification with the P 5 GHz criterion alone. We plot the SEDs of these quasars in Figure 3. The general properties of these quasars are listed in Table 3. For the three of the six, black hole estimates are available and all exceed 10 8 M ⊙ .
## 4.4 Inter-comparison of Criteria
Figure 4 & 5 show the distributions of the sources with respect to the different RL diagnostics to illustrate the RL fraction and compare the agreement among the above radio loudness definitions. On the side of each axis, we also plot the cumulative fraction of the sources as function of the corresponding radio-loudness measurements using the survival analysis methods (Kaplan & Meier 1958; Schmitt 1985) which we previously employed to compute the RL fractions.
1 XID=40, 2282, 5230, 5275, 5517, 5395. Note if we use the R uv,D criterion, the quasar XID=5275 is classified as RQ.
Figure 3. The spectral energy distribution (SED) of the six quasars which are radio-loud in all criteria. The red dashed line is the E94 RQ mean SED. The green dotted line is the E94 RL mean SED. The data points in the SED are color-coded as in Elvis et al. (2012). From low to high frequency, the black data points are: 1.4GHz, 24 µ m, 8 µ m, 5.7 µ m, 4.5 µ m, 3.6 µ m, K-band, H-band, J-band, NUV, FUV and 2keV. The blue data points are the Subaru broad bands (B J , g, r, i, z). The green data points are the (CFHT) K-band, and the (CFHT) u band and i band. The purple data points are the 6 Subaru intermediate bands for season 1 (2006) (IA427, IA464, IA505, IA574, IA709, IA827). The cyan data points are the 5 Subaru intermediate bands for season 2 (2007) (IA484, IA527, IA624, IA679, IA738, IA767).
We marked the 6 quasars in the sample that are classified as RL according to all nine criteria with green circles. Their SEDs are plotted in Figure 3.
for all RL criteria. In compiling these numbers, we use the criterion R uv,D instead of R uv , as it is subject to reddening and host contamination issues ( 4). §
For the criteria defined in the rest-frame, the radioloudness defined by the ratios of radio to optical luminosity ( R L , R uv and R i ) correlate well with each other (Figure 4 & 5). The R X and P 5 GHz criteria agree the best with the fewest 'ambiguous' quasars.
## 5 DISCUSSION
## 5.1 Low Radio-Loud Fraction
We have found a low RL fraction of 1.5%-4.5% (using six different criteria defined in the rest-frame) in the XMMCOSMOS type 1 AGN sample, as compared with about 10% in optically selected samples. For example, in the BQS, Kellermann et al. (1989), find 15% using the R L criterion; in LBQS, Hooper et al. (1996) find 9% using the rest-frame 8.4 GHz luminosity log L 8 4 . > 25 and 9% using the flux ratio between the rest-frame 8.4 GHz and B band R 8 4 . > 1. The difference between XC407 and these other RL fractions is significant at a confidence level of > 99% and is observed
Similarly low RL fractions have been reported or inferred in a few other samples, all of which include infrared selection. For example, Richards et al. (2006) reported only 8 RL quasars among a Spitzer-Sloan Digital Sky Survey (SDSS) quasar sample of 259 sources, giving a similarly small RL fraction of 3%, using the criterion of radio luminosity L rad > 10 33 erg s -1 Hz -1 , that is log[ L rad ( W/Hz )] > 26, which is stricter than the typical P 5 GHz criterion we applied in the present analysis.
Donley et al. (2007) found a RL fraction of 3% when they applied the q 24 criterion to a sample of 62 X-ray selected power-law AGNs in the Chandra Deep Field North (CDFN) whose Spitzer IRAC SEDs exhibit the characteristic power-law emission expected for luminous AGNs. Donley et al. (2012) attribute this low RL fraction to the fact that IRAC color-color selection is biased against sources with particularly bright hosts and radio-loud AGN tend to be hosted by bright elliptical galaxies.
Of the criterion defined in the observed frame, the criteria R L,obs and R i,obs yield a high radio-loud fraction
Figure 4. Radio-loudness measures: R L , q 24 , R uv , R i , R X , and P 5 GHz . Black crosses = radio detections; blue arrows = upper/lower limits; green circles = RL in all criteria. Solid lines = limits assumed for the radio-loudness definition. Dashed line perpendicular to q 24 axis = Kuraszkiewicz et al. (in preparation) adjusted q 24 criterion. Dashed line perpendicular to R uv axis = Della Ceca et al. (1994) adjusted R uv criterion. Along the upper/right-hand edge we show the cumulative distribution functions which include a correction for upper/lower limits. Dash-dotted lines indicate the RQ fraction.
( > 10%). This number is comparable to previous studies. For example, in the Sloan Digital Sky Survey -DR2/FIRST, Ivezi´ et al. (2002) find 8% c ± 1% using the flux ratio between the 1.4 GHz and i band in the observed frame R i,obs > 1. However, we note that no k-correction is included in these two criteria and the XMM-COSMOS sample has a large redshift range. Considering most of the quasars in the XMMCOSMOS sample are at redshift 1-2, we define R J and R K instead to ensure a more meaningful comparison with the observed frame criteria applied to low redshift samples. The radio-loud fraction then drops to about 4% and 2 5%, re-.
spectively. Even though q 24 ,obs also involves no k-correction, the corresponding RL fraction is nevertheless low (2.55%). Therefore, for samples with large redshift range, the criteria defined in the observed frame not including k-correction could give different radio-loud fraction compared to other criteria defined in the rest-frame.
X-ray samples often select more obscured or host dominated AGNs than optically selected samples (Hao et al. 2014, Kuraszkiewicz et al. 2003), because X-ray emission is ubiquitous in AGNs which makes X-ray surveys the most complete census of AGNs of any single band (Risaliti & Elvis
Figure 5. Radio-loudness measures: R uv , R i , R X , P 5 GHz , R L,obs , q 24 ,obs and R i,obs . Black crosses = radio detections; blue arrows = upper/lower limits; green circles = RL in all criteria. Solid lines = limits assumed for the radio-loudness definition. Dashed line = Della Ceca et al. (1994) adjusted R uv criterion. Along the upper/right-hand edge we show the cumulative distribution functions which include a correction for upper/lower limits. Dash-dotted lines indicate the RQ fraction.
2004). Reddening will increase the RL fraction, as we confirm in the present XMM-COSMOS sample based on the R uv criterion. A large host galaxy contribution could artificially boost the apparent optical flux, thereby fortuitously reducing the RL fraction for most of these criteria (e.g. R L ), while the combination effect of host and reddening still results in higher values of R uv . The host contribution and reddening will effect the criteria defined with an optical luminosity the most. However, both the q 24 and P 5 GHz criteria are insensitive to reddening and host contribution, and we find both q 24 and P 5 GHz to agree with R L for this sample (Figure 4). This suggests that neither reddening nor host- contamination are the main reasons for the low RL fraction in this sample.
To ascertain in more detail that reddening and hostcontamination do not play a major part in producing the low radio-loud fraction in the XMM-COSMOS sample, we use the mixing diagram (Hao et al. 2013) to estimate the host galaxy fraction at 1 µm ( f g ) for each quasar. Here we choose the mixing curve connecting Ell5 (SWIRE galaxy template of elliptical galaxy with age of 5 Gyr, Polletta et al. 2007) and E94 (Elvis et al. 1994 mean quasar SED). As shown in Hao et al. (2013), the difference in galaxy fraction is negligible regardless of the mixing curves cho-
Figure 6. The radio-loudness after correction for the host contribution and reddening ( R L,corr ) compared to the observed radioloudness ( R L ). The plot is color-coded as in Figure 4&5.
sen. The AGN contribution of the luminosity at 1 µm is thus L AGN, µm 1 = L 1 µm (1 -f g ). We assume that the unreddened intrinsic quasar SED has the shape of the E94 SED, which is a reasonable assumption (Hao et al. 2014). We normalize the E94 SED to the pure AGN luminosity at 1 µm , and recover the AGN rest-frame B band luminosity from L B,corr = L 1 µm (1 -f g ) -L E 94 1 , µm + L E 94 ,B . As the radio emission mainly comes from the AGN, the new R L,corr = log( L 5 GHz /L B,corr ) is the radio-loudness corrected for the host galaxy contribution and reddening. Figure 7 shows the comparison between the radio-loudness before and after the correction. The radio-loud fraction for this corrected radio-loudness is 6 08% . +3 18% . -1 75% . . We can see that it is slightly higher than the radio-loud fraction before the correction, which means the host contribution and reddening in combination decrease the radio-loud fraction. This is consistent with the case of nearby Seyferts (Ho & Peng 2001). However, they are not the most significant factor leading to the low radio-loud fraction found in the XMM-COSMOS sample.
RL quasars usually reside in very massive galaxies and typically have a lower optical or X-ray output at given stellar mass (i.e. lower L/L Edd at given L , Sikora et al. 2007) compared to RQ quasars. This means that a L X -limited sample will have a lower RL quasar fraction, compared to a masslimited sample. COSMOS is a deep survey in a limited area, thus it will sample a few high mass ( L > 5 L ∗ ) galaxies at any redshift, compared to wide area shallow surveys (e.g. SDSS), similar to the CDFN sample (Donley et al. 2007) which also has a low radio-loud fraction.
We also note that XC407 is not a complete X-ray selected sample, but a sub-sample with spectroscopic coverage. The optical-near infrared magnitudes of the spectroscopic sub-sample thus tend to be brighter than the complete sample, which in principle produces a bias toward low radio- loudness in the radio-loudness distribution of the complete sample.
## 5.2 Correlation of Radio-loudness with Other Properties
In the EMSS, Della Ceca et al. (1994) found a 10.6% radioloud fraction and a trend of lower RL fractions for absolute magnitudes fainter than M B = - 24. Note that this sample is also X-ray selected but is not restricted to type 1 AGNs. The XMM-COSMOS B band absolute magnitude distribution is similar to that of the EMSS sample (Elvis et al. 2012), with 191 (46% of the whole sample) having M B > -24. We plot R L versus M B for XC407 in the left panel of Figure 7. Note that we have changed to the same cosmology (H 0 =50 km s -1 Mpc -1 ) as Della Ceca et al. (1994) for M B . No trend of reduced RL fraction at M B < -24 is seen in our sample.
Radio-loudness is seen to increase with decreasing Eddington ratio ( L/L Edd ), in particular R L ∝ ( L/L Edd ) -1 at L/L Edd > . 0 001 (Sikora et al. 2007, Ho 2002). However, the XMM-COSMOS sources have relatively high Eddington ratios ( > . 0 01, with a median of 0.2, Hao et al. 2014). This is similar to the values for PG quasars (Kellerman et al., 1987, Sikora et al. 2007). We plot R L versus log( L/L Edd ) in the middle panel of Figure 7. No trend of reduced RL fraction for large Eddington ratio is visible in our sample. In fact, all the R L > 1 AGNs have log( L/L Edd ) > -1.
In E94, the mean SED of the RL and RQ sample show differences in the X-rays (see Figure 3 red and green curve), with RL quasars tending to have relatively brighter X-ray luminosity (about 0.5 dex higher). We checked the R L versus the X-ray luminosity at 2keV ( νL 2 keV ) in the right panel of Figure 7. No obvious trend toward reduced RL fraction at low X-ray luminosity is seen. If we divided the sample at the median νL 2 keV (43.96 erg/s), slightly more radio loud quasars are above the median, which is 10 radio-loud quasars with radio detections has X-ray luminosity above median compared to 8 radio-loud quasars with radio detections has X-ray luminosity below the median. This is consistent with the mean SED get in E94.
## 6 SUMMARY
We have used nine different radio-loudness criteria to study the radio loud fraction of the XMM-COSMOS type 1 AGN sample, in which six were defined in the rest-frame (radio loud fraction ranging from 1.5%-4.5%) and three were defined in the observed frame without k-correction. The poor statistics on the RL sources does not allow to infer statistically significant results on a dichotomy between RQ and RL AGN, of which we do not see any sign.
The criteria defined in the rest-frame generally agree with each other and gives similar radio-loud fractions. The criterion R X and P 5 GHz agree the best with the smallest number of ambiguous sources. Radio-loudness defined via a radio-to-optical luminosity ratio ( R L , R uv and R i ) display the strongest correlation. The radio power ( P 5 GHz ) gives the strongest restriction, that yields the lowest radio-loud fraction for any single criteria. If we require all the criteria to be satisfied at the same time, the radio-loud fraction is
Figure 7. R L versus M B (left), log( L/L Edd ) (middle) and νL 2 keV (right) for the XC407 sample, respectively. Here M B is calculated using H 0 =50 km s -1 Mpc -1 . The Eddington ratio is λ Edd = L bol /L Edd . The νL 2 keV is the 2keV X-ray luminosity in units of erg/s. Note that in the middle plot, we only show 204 among the 407 quasars in the XC407 that have black hole mass estimates (Trump et al. 2009b; Merloni et al. 2010). For the rest of the sample, the broad emission lines are locate at the edge of the spectrum, such that reliable estimates of the black hole mass are not possible (Elvis et al. 2012). The black crosses show the quasars with > 3 σ detections in the radio and the blue arrows show the quasars with upper limits in the radio. The green circles indicate the 6 sources classified as RL by all criteria. The red dashed line in the center panel shows R L = -log( L/L Edd ) -1 which represents the general trend found in Sikora et al. (2007).
marginally smaller than the radio-loud fraction we get from the P 5 GHz criterion only.
gratefully acknowledges support from Swiss National Science Foundation Grant PP00P2 138979/1.
Two of the criteria defined in the observed frame without k-correction give a much higher radio-loud fraction, but if we take the redshift distribution of the sample into consideration, the radio-loud fraction is greatly reduced and becomes consistent with the results from the criteria defined in the rest-frame. Thus, we need to be careful when citing the radio-loud fraction using the criteria defined in the observed frame without k-correction.
If we corrected the host galaxy contribution and reddening, the radio-loud fraction in the XC407 sample will rise to 6 08% . +3 18% . -1 75% . , still a bit smaller than the typical value of 10% in optical selected samples, which might be caused by the selection of the sample being L X -limited and biasing towards bright optical quasars. No correlation of the radio-loudness with M B , L/L Edd or L X is seen in the XC407 sample.
The combination of newly approved deep 3GHz EVLA observations (P.I.: Smolcic) with the completion of Chandra coverage of the whole COSMOS field in the COSMOS Legacy survey (P.I.: Civano) will help us understand the origin of the low RL fraction of type 1 AGNs.
## ACKNOWLEDGMENTS
HH thank Belinda Wilkes for useful discussion. This work was supported in part by NASA Chandra grant number G07-8136A (HH, ME, CV). LCH acknowledges support from the Kavli Foundation, Peking University, and the Chinese Academy of Science through grant No. XDB09030100 (Emergence of Cosmological Structures) from the Strategic Priority Research Program. Support from the Italian Space Agency (ASI) under the contracts ASI-INAF I/088/06/0 and I/009/10/0 is acknowledged (AC and CV). MS acknowledges support by the German Deutsche Forschungsgemeinschaft, DFG Leibniz Prize (FKZ HA 1850/28-1). KS
## REFERENCES
Appleton, P. N., et al., 2004, ApJS, 154, 147 Balokovi´, M., Smol˘i´, V., Ivezi´, ˘ ., Zamorani, G., Schin- c c c c Z nerer, E., and Kelly, B. C., ApJ, 2012, 759, 30 Brusa, M., et al., 2005, A&A, 2005, 421, 69 Brusa, M., et al., 2007, ApJS, 2007, 172, 353 Brusa, M., et al., 2010, ApJ, 716, 348 Capak, P., et al., 2007, ApJS, 172, 99 Capak, P., et al., 2010, in preparation Cirasuolo, M., et al., 2007, MNRAS, 380, 585 Cirasuolo, M., et al., 2003, MNRAS, 341, 993 Della Ceca, R., Zamorani, G., Maccacaro, T., Wolter, A., Griffiths, R., Stocke, J. T., & Setti, G., 1994, ApJ, 430, 533 Donley, J. L., Rieke, G. H., P´rez-Gonz´lez, P. G., Rigby, e a J. R., Alonso-Herrero, A., 2007, ApJ, 660, 167 Donley, J. L., et al., 2012, ApJ, 748, 142 Elvis, M. et al., 1994, ApJS, 95, 1 Elvis, M. et al., 2012, ApJ, 759, 6 Feigelson, E. D. & Nelson, P. I., 1985, ApJ, 293, 192 Goldschmidt, P., Kukula, M. J., Miller, L., & Dunlop, J. S., 1999, ApJ, 511, 612 Hao, H. et al. 2014, MNRAS, 438, 1288 Hao, H. et al. 2013, MNRAS, 434, 3104 Harris, D. E., & Krawczynski, H., 2006, ARA&A, 44, 463 Hasinger, G., et al., 2007, ApJS, 172, 29 Ho, L. C., 2002, ApJ, 564, 120 Ho, L. C., & Peng, C. Y., 2001, ApJ, 555, 650 Hooper, E. J., et al., 1996, ApJ, 473, 746 ˇ .,
Ivezi´, c Z 2004, in Multiwavelength AGN Surveys, ed. R.M´ ujica & R. Maiolino (Singapore: World Scientific), 53 Ivezi´, Z., et al., 2002, AJ, 124, 2364 c ˇ
## 12 Heng Hao et al.
| Jiang, L., Fan, X., Ivezi´,Z., c ˇ Richards, G. T., Schneider, D. P., Strauss, M. A., & Kelly, B. C., 2007, ApJ, 656, 680 |
|-----------------------------------------------------------------------------------------------------------------------------|
| Kaplan, E. L., & Meier, P. 1958, J. Am. Statistical Associ- ation, 53, 457 |
| Kellermann, K. I., et al., 1989, AJ, 98, 1195 |
| Komatsu, E., et al., 2009, ApJS, 180, 330 |
| Kuraszkiewicz, J. K., et al. 2003, ApJ, 590, 128 |
| Le Floc'h, E. et al., 2009, ApJ, 703, 222L |
| Lilly, S. J., et al., 2007, ApJS 172, 70 |
| Lilly, S. J., et al., 2009, ApJS 184, 218 |
| Merloni, A., et al. 2010, ApJ 708, 137 |
| Miller, R. G., Jr. 1981, Survival Analysis (New York; Wi- ley) |
| Miller, L., et al., 1990, MNRAS, 244, 207 |
| Peterson, B. M. 1997, An Introduction to Active Galactic Nuclei (Cambridge: Cambridge University Press) |
| Pierce, C. M., Ballantyne, D. R., & Ivison, R. J., 2011, ApJ, 742, 45 Polletta, M. et al. 2007, ApJ, 663, 81 |
| Richards, G. T., et al., 2006, ApJ, 166, 470 Risaliti, G. & Elvis, M., 2004, in ASSL Vol |
| 308, Supermas- sive Black Holes in the Distant Universe, ed. A. J. Barger (Dordrecht: Kluwer), 187 |
| Schinnerer, E., et al., 2010, ApJS, 188, 384. |
| Schmitt, J. H. M. M. 1985, ApJ, 293, 178 |
| Scoville, N. Z., et al., 2007, ApJS, 172, 1 |
| Sikora, M., Stawarz, L. & Lasota, J. P., 2007, ApJ, 658, 815 |
| Scoville, N. Z., et al., 2007, ApJS, 172, 38 |
| Spergel, D. N., et al., 2007, ApJS, 170,377 |
| Stocke, J. T., Morris, S. L., Weymann, R. J. & Foltz, C. B., 1992, ApJ, 396, 487 |
| Terashima, Y. & Wilson, A. S. 2003, ApJ, 583, 145 |
| Trump, J. R., et al., 2007, ApJS, 172, 383 Trump, J. R., et al., 2009a ApJ, 696, 1195 |
| Trump, J. R., et al., 2009b ApJ, 700, 49 |
| Urry, C. M. & Padovani, P., 1995, PASP, 107, 803 Visnovsky, K. L., et al., 1992, ApJ, 391, 560 |
| Wilkes, B. J. & Elvis, M., 1987, ApJ, 323, 243 |
| Zamorani, G., et al., 1981, ApJ, 245, 357 |
| null |
[
"Heng Hao",
"Mark T. Sargent",
"Martin Elvis",
"Eva Schinnerer",
"Gianni Zamorani",
"Luis C. Ho",
"Jennifer L. Donley",
"Francesca Civano",
"Vernesa Smolcic",
"Annalisa Celotti",
"Joanna Kuraszkiewicz",
"Mara Salvato",
"Marcella Brusa",
"Peter Capak",
"Chris L. Carilli",
"Andrea Comastri",
"Chris D. Impey",
"Knud Jahnke",
"Anton M. Koekemoer",
"Kevin Schawinski",
"Jonathan R. Trump",
"C. Megan Urry",
"Cristian Vignali",
"Min Yun"
] |
2014-08-05T20:00:01+00:00
|
2014-08-05T20:00:01+00:00
|
[
"astro-ph.GA"
] |
Inter-comparison of Radio-Loudness Criteria for Type 1 AGNs in the XMM-COSMOS Survey
|
Limited studies have been performed on the radio-loud fraction in X-ray
selected type 1 AGN samples. The consistency between various radio-loudness
definitions also needs to be checked. We measure the radio-loudness of the 407
type 1 AGNs in the XMM-COSMOS quasar sample using nine criteria from the
literature (six defined in the rest-frame and three defined in the observed
frame): $R_L=\log(L_{5GHz}/L_B)$, $q_{24}=\log(L_{24\mu m}/L_{1.4GHz})$,
$R_{uv}=\log(L_{5GHz}/L_{2500\AA})$, $R_{i}=\log(L_{1.4GHz}/L_i)$,
$R_X=\log(\nu L_{\nu}(5GHz)/L_X)$, $P_{5GHz}=\log(P_{5GHz}(W/Hz/Sr))$,
$R_{L,obs}=\log(f_{1.4GHz}/f_B)$ (observed frame),
$R_{i,obs}=\log(f_{1.4GHz}/f_i)$ (observed frame), and $q_{24,
obs}=\log(f_{24\mu m}/f_{1.4GHz})$ (observed frame). Using any single criterion
defined in the rest-frame, we find a low radio-loud fraction of $\lesssim 5\%$
in the XMM-COSMOS type 1 AGN sample, except for $R_{uv}$. Requiring that any
two criteria agree reduces the radio-loud fraction to $\lesssim 2\%$ for about
3/4 of the cases. The low radio-loud fraction cannot be simply explained by the
contribution of the host galaxy luminosity and reddening. The
$P_{5GHz}=\log(P_{5GHz}(W/Hz/Sr))$ gives the smallest radio-loud fraction. Two
of the three radio-loud fractions from the criteria defined in the observed
frame without k-correction ($R_{L,obs}$ and $R_{i,obs}$) are much larger than
the radio-loud fractions from other criteria.
|
1408.1091v1
|
## POSSIBLE DETECTION OF THE STELLAR DONOR OR REMNANT FOR THE TYPE Iax SUPERNOVA 2008ha
Ryan J. Foley 1,2 , Curtis McCully 3 , Saurabh W. Jha 3 , Lars Bildsten 5,4 , Wen-fai Fong 6 , Gautham Narayan 7 , Armin Rest 7 , and Maximilian D. Stritzinger 9
Draft version August 7, 2014
## ABSTRACT
Type Iax supernovae (SNe Iax) are thermonuclear explosions that are related to SNe Ia, but are physically distinct. The most important differences are that SNe Iax have significantly lower luminosity (1%-50% that of typical SNe Ia), lower ejecta mass ( ∼ 0.1 - 0.5 M /circledot ), and may leave a bound remnant. The most extreme SN Iax is SN 2008ha, which peaked at M V = -14 2 mag, about 5 mag below that of . typical SNe Ia. Here, we present Hubble Space Telescope ( HST ) images of UGC 12682, the host galaxy of SN 2008ha, taken 4.1 years after the peak brightness of SN 2008ha. In these deep, high-resolution images, we detect a source coincident (0.86 HST pixels; 0.043 ′′ ; 1.1 σ ) with the position of SN 2008ha with M F814W = -5 4 mag. . We determine that this source is unlikely to be a chance coincidence, but that scenario cannot be completely ruled out. If this source is directly related to SN 2008ha, it is either the luminous bound remnant of the progenitor white dwarf or its companion star. The source is consistent with being an evolved > 3 M /circledot initial mass star, and is significantly redder than the SN Iax 2012Z progenitor system, the first detected progenitor system for a thermonuclear SN. If this source is the companion star for SN 2008ha, there is a diversity in SN Iax progenitor systems, perhaps related to the diversity in SN Iax explosions. If the source is the bound remnant of the white dwarf, it must have expanded significantly. Regardless of the nature of this source, we constrain the progenitor system of SN 2008ha to have an age of < 80 Myr.
Subject headings: galaxies-individual(UGC 12682), supernovae-general, supernovae-individual (SN 2008ha)
## 1. INTRODUCTION
Type Iax supernovae (SNe Iax) are a class of thermonuclear stellar explosions physically distinct from SNe Ia, but sharing several characteristics (see Jha et al. 2006 and Foley et al. 2013 for reviews). This class is the most common of all types of peculiar SNe by both number and rate, occurring at roughly 30% the rate of SNe Ia (Foley et al. 2013). All observational properties indicate that these explosions are less energetic than SNe Ia (e.g., Li et al. 2003; Phillips et al. 2007; Foley et al. 2013).
While the progenitor systems of SNe Ia are still uncertain (see Maoz et al. 2013 for a review), a reasonable progenitor model and explosion mechanism that explains most SN Iax observations has emerged. As the number of observations has increased (Foley et al. 2009, 2010a,b, 2013; Valenti et al.
- 1 Astronomy Department, University of Illinois at UrbanaChampaign, 1002 W. Green Street, Urbana, IL 61801, USA
2009; Maund et al. 2010; McClelland et al. 2010; Milne et al. 2010; Narayan et al. 2011; Kromer et al. 2013; McCully et al. 2014a; Stritzinger et al. 2014), the best model for these SNe continues to be a deflagration of a C/O white dwarf (WD) that does not burn through the entire WD as first proposed by Foley et al. (2009). Because of He seen in the spectra of two SNe Iax and their implied delay time of /lessorsimilar 500 Myr, Foley et al. (2013) expanded on this initial model by employing a He-star companion, similar to a previously suggested progenitor system for SNe Ia (Iben & Tutukov 1991; Liu et al. 2010). Observations suggest, and this model implies, that at least some SNe Iax do not unbind their progenitor star, leaving behind a remnant (Foley et al. 2009, 2010a, 2013; Jordan et al. 2012; Kromer et al. 2013; McCully et al. 2014a). This model can explain all observational properties of SNe Iax except perhaps the low polarization seen in a single SN Iax (Chornock et al. 2006; Maund et al. 2010).
- 2 Department of Physics, University of Illinois UrbanaChampaign, 1110 W. Green Street, Urbana, IL 61801, USA
- 5 Kavli Institute for Theoretical Physics, Santa Barbara, CA 93106, USA
- 3 Department of Physics and Astronomy, Rutgers, The State University of New Jersey, 136 Frelinghuysen Road, Piscataway, NJ 08854, USA
- 4 Department of Physics, University of California, Santa Barbara, CA 93106, USA
- 7 National Optical Astronomy Observatory, 950 North Cherry Avenue, Tucson, AZ 85719-4933, USA
- 6 Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA 02138, USA
- 8 Space Telescope Science Institute, 3700 San Martin Drive, Baltimore, MD 21218, USA
- 9 Department of Physics and Astronomy, Aarhus University, Ny Munkegade, DK-8000 Aarhus C, Denmark
This model has been directly tested by pre-explosion Hubble Space Telescope ( HST ) images of SN 2008ge (Foley et al. 2010b) and SN 2012Z (McCully et al., in press; hereafter M14). No progenitor was detected in the relatively shallow images for SN 2008ge, but the data excluded most massive stars as the progenitor (as suggested by Valenti et al. 2009 and Moriya et al. 2010). Deep limits on the star formation for the host galaxy of SN 2008ge also indicated that SN 2008ge did not have a massive progenitor star (Foley et al. 2010b).
M14detected a star in pre-explosion HST images at the position of SN 2012Z. This was the first direct detection of a thermonuclear SN progenitor system. This relatively
Fig. 1.False-color HST image of UGC 12682, the host galaxy of SN 2008ha. The red, green, and blue channels correspond to F814W, F625W, and F435W, respectively. Left : 75 ′′ × 75 ′′ image. The red box is 15 ′′ × 15 ′′ and centered on the position of SN 2008ha. Top-Right : 15 ′′ × 15 ′′ image corresponding to the red box in the left panel. The red box is 2 ′′ × 2 ′′ . Bottom-Right : 2 ′′ × 2 ′′ image corresponding to the red box in the top-right panel. A 3 σ ellipse is centered on the position of SN 2008ha. There is a single source near the center of the ellipse, which we label 'S1.' The bright source to the south-east of S1 is an H II region with a measured metallicity of 12 + log[O / H] = 8 16 . (Foley et al. 2009).
luminous star is most consistent with a massive ( M ≈ 2 M /circledot at the time of explosion) He star, as predicted by Foley et al. (2013).
discuss the implications of our findings and summarize our conclusions in Section 4.
In this manuscript, we present post-explosion HST images of the host galaxy of SN 2008ha, the least luminous SN Iax (Foley et al. 2009, 2010a; Valenti et al. 2009; Stritzinger et al. 2014). SN 2008ha was discovered on 2008 November 7.17 (UT dates are used throughout this paper; Puckett et al. 2008) in UGC 12682, an irregular galaxy with a recession velocity of 1393 km s -1 , corresponding to a Virgo-infall corrected distance modulus of µ = 31 64 . ± 0 15 mag and . D = 21 3 . ± 1 5 Mpc, assuming . H 0 = 73 ± 5 km s -1 Mpc -1 . Foley (2008) first classified SN 2008ha as a SN Iax, but also noted its extreme characteristics and the possibility that SN 2008ha may not have unbound its progenitor star.
In the post-explosion images, we detect a point source near the position of SN 2008ha. We outline the data acquisition and reduction in Section 2. In Section 3, we detail the properties of this source, and examine several possibilities for its nature including being a chance coincidence, SN emission, the stellar remnant, and the companion star from the progenitor system. We also constrain the properties of the progenitor system and remnant if the source is unrelated to SN 2008ha. We
## 2. OBSERVATIONS AND DATA REDUCTION
UGC 12682 was observed with HST /ACS on 2012 January 2.2 (Program GO-12999; PI Foley). The full observing sequence obtained 1164, 764, 840, and 1240 seconds in the F435W, F555W, F625W, and F814W filters (corresponding roughly to BVRI ).
We combined exposures (including cosmic ray rejection) using AstroDrizzle. We register the individual flatfielded (flt) frames using TweakReg from the DrizzlePac package. We registered the absolute astrometry to ground-based images of the same field using the F555W exposures. We then tied the relative astrometry of the other filters to this image. The typical rms of the relative astrometry solution was 0 005 . ′′ = 0 1 ACS pixels. . We drizzle the images to the native scale of ACS, 0.05 ′′ per pixel. A false-color image of UGC 12682 and zoomed panels of the region near the position of SN 2008ha are presented in Figure 1.
We photometered the HST images using the PSF photometry software DolPhot, an extension of HSTPhot (Dolphin 2000). DolPhot photometers individual flt frames and then combines the photometry. We use the
suggested parameters for ACS from the Dolphot manual. We only consider photometry with a 'flag' of 0 and an 'object type' of 1 (point sources) in our results.
Using an r -band image taken with the 1.0-m Swope telescope, we were able to determine the astrometric position of SN 2008ha in the HST images. The Swope image, from which the Stritzinger et al. (2014) SN 2008ha photometry was derived, had a template image subtracted for the region around SN 2008ha. This provides a precise measurement of the position of SN 2008ha relative to field stars without contamination from underlying emission (particularly from a nearby H II region).
The Swope and HST images had a total of 14 stars in common with well-measured positions. Following the procedures of Foley et al. (2010b, and references therein), we determined a geometric transformation between the two images. However, tests showed that inclusion of stars far from SN 2008ha adversely affected the accuracy of the transformation as determined by looking at positions of stars not included in the fitting. The poor accuracy is likely because of higher-order distortions on large spatial scales that could not be adequately removed with the limited number of stars. However, we found that using only the 8 stars closest to SN 2008ha removed all obvious bias near the location of SN 2008ha.
Using these 8 stars, the accuracy of the geometric transformation is robust near the SN position. However, the nominal uncertainties in the SN position ( ∼ 0.4 ACS pixels; ∼ 0.02 ′′ ) may be underestimated given the previously determined errors. To provide a better estimate of the positional uncertainty, we re-determined the position of SN 2008ha 100 times by selecting bootstrapped samples (with replacement) from the initial sample of 8 stars. We performed this analysis several times with varying orders for the polynomial transformations. Using outlier-resistant statistics, the average SN 2008ha position from the bootstrapped samples are consistent with our best estimate. We use the outlier-resistant scatter of the resulting positions, 0.44 and 0.50 ACS pixels in the R.A. and Dec. directions, respectively, which are larger than the nominal uncertainty from the nominal position, as the uncertainty in the position of SN 2008ha from the geometric transformation. To this, we add in quadrature the statistical uncertainty in the position of SN 2008ha in the Swope image, which is equivalent to 0.08 ACS pixels, and the estimated systematic uncertainty in determining the SN position, which is equivalent to 0.44 ACS pixels. The total uncertainty in measuring the position of SN 2008ha in the HST images is thus 0.71 and 0.75 ACS pixels (0.037 ′′ and 0.039 ′′ ) in the R.A. and Dec. directions, respectively.
Offset by 0.86 ACS pixels, corresponding to 0.043 ′′ , from the SN position, there is a source clearly detected in the F814W band (S N = 6 9). This source is marginally / . detected in the F625W band (S N = 1 8), and unde-/ . tected in the bluer bands. We label this source 'S1.' The uncertainty in the position of this source is roughly 0.4 ACS pixels. Adding this in quadrature to the uncertainty in the position of SN 2008ha, S1 is 1.1 σ from the position of SN 2008ha. Photometry of S1 yields magnitudes of < 27.7, < 27.3, 28 1 . ± 0 6, and 26 34 . . ± 0 16 mag . in the HST filters, from blue to red, respectively, with the limits being 3σ limits.
The F625W -F814W color of S1 is redder than almost all nearby stars. Of the 100 stars closest to the position of SN 2008ha (including S1), 84 are bluer, and 13 have non-detections in F625W, but are fainter in F814W such that they could be bluer than S1. That is, only 2 of the closest 100 stars are definitely redder than S1.
Assuming our nominal distance modulus and correcting for Milky Way reddening of E B ( -V ) = 0 07 mag . (Schlegel et al. 1998; Schlafly & Finkbeiner 2011), we can derive the luminosity of S1. Fitting a blackbody to the photometry of S1 (assuming no host-galaxy reddening), we find a best-fit temperature of 2100 ± 500 K and a radius of 1500 ± 800 R /circledot , and a bolometric luminosity of L bol = 1 5 . × 10 38 erg s -1 . Using this temperature and the F814W absolute magnitude, we place S1 on a Hertzsprung-Russell (HR) diagram (Figure 2).
Although it is difficult to measure host-galaxy reddening from SN Iax colors (Foley et al. 2013), SN 2008ha did not appear to have any host-galaxy reddening (Foley et al. 2009). The position of the SN in its host galaxy also makes a substantial amount of dust in the interstellar medium unlikely. However, it is possible that there is newly formed circumstellar dust.
V445 Pup, a helium nova that ejected ∼ 10 -4 M /circledot of material formed a significant amount of circumstellar dust 7.5 months after discovery(Kato et al. 2008). Under the right conditions, SN 2008ha could have generated a similar amount of dust, which could also account for the very red colors of S1. Unfortunately, we cannot constrain the extinction from newly formed circumstellar dust with our current observations. Ignoring the possibility of circumstellar dust, the colors and luminosity of S1 are consistent with an thermal-pulsing asymptotic giant branch (TP-AGB) star that had an initial mass of > 3 M /circledot (see Section 3.4).
## 3. ANALYSIS
Having identified a source that is spatially coincident with SN 2008ha, we now examine its physical properties in detail. There are four scenarios for this source: it is unrelated to SN 2008ha, it is emission from the SN itself, it is the luminous remnant of the WD progenitor, or it is the companion star to the WD.
Additionally, we examine constraints on the progenitor system and remnant if S1 is not associated with SN 2008ha.
## 3.1. Chance Coincidence
There are many bright, resolved stars in our images of UGC 12682. For certain positions, there will be a chance superposition with one of these stars. As noted above, S1 is 0.86 ACS pixels, equivalently 0.043 ′′ and 1.1 σ , from the position of SN 2008ha. We now calculate the probability of such a similar chance alignment.
Our analysis follows that of M14. Within a 40 × 320 pixel (2 ′′ × 16 ) box centered on SN 2008ha, we de-′′ tect 89 stars with S / N > 3 in at least one of our HST images. The dimensions of the box were chosen such that the major axis was roughly perpendicular to the stellar density gradient, and the density of stars within this box should be roughly representative of the stellar density at the position of SN 2008ha.
Given our astrometric uncertainty, there is a 0.90% (7.0%) chance that a random position near SN 2008ha would be within 1.1 σ (3 σ ) of any detected star (in any of
Fig. 2.Hertzsprung-Russell diagram for S1 (thick black cross; 2100 ± 500 K), the HST limits for a blackbody source of a given temperature (black line), the SN 2012Z progenitor system (thin black cross), and the V445 Pup progenitor system (brown cross; F814W magnitude estimated from temperature and V -band magnitude). Also plotted are stellar evolution tracks for stars with initial masses of 2, 3, 4, and 5 M /circledot (Bertelli et al. 2009) and the region predicted for one set of He-star progenitor models from Liu et al. (2010) (blue region). The limits shown are the single-band 3σ limits for black bodies. The grey band indicates the additional 1σ distance modulus uncertainty. The distance uncertainty is included in the uncertainties for S1.
the HST images), making a chance alignment somewhat unlikely. Additionally, there are only 53 stars within the same area that are brighter than S1; there is only a 0.54% (4.2%) chance of being within 1.1 σ (3 σ ) of a star at least as bright as S1. Although this latter measurement is an a posteriori statistic, it is helpful for determining the rarity of such stars and what we would have determined for slightly shallower images.
The likelihood of chance alignment is small, but not sufficiently small to completely rule out that scenario. Our above calculation assumes no spatial clustering of stars, which is incorrect. However, it is not possible to precisely determine the necessary correction with the current data. Nonetheless, if S1 is not part of the SN 2008ha progenitor system, but is located nearby because it was born in the same cluster, this is not a chance coincidence. In that scenario, S1 will have a similar age (and composition) as the SN 2008ha progenitor system. Future observations could possibly differentiate these scenarios.
Fig. 3.R -band light curve of SN 2008ha and photometry of S1 (converting our blackbody fit to an R band magnitude). The blue points are the R -band light curve of SN 2005hk (Phillips et al. 2007; McCully et al. 2014a), shifted to match the peak brightness of SN 2008ha. The red-dotted line is the expected light curve extrapolated from the last detection (and is thus an upper limit) if the SN is only powered by 56 Co decay. The black solid lines are the model SN Ia light curves from R¨pke et al. (2012) that include o additional radioactive elements. The two curves represent different explosion mechanisms and thus slightly different nucleosynthetic yields. S1 is > 2 mag brighter than the expected emission from SN 2008ha at the time of our HST images.
## 3.2. Supernova Emission
SN 2008ha was a very low luminosity event, likely the least luminous hydrogen-deficient SN yet observed. It peaked at V = 17 68 . mag, corresponding to M V = -14 17 . mag, and declined quickly with ∆ m 15 ( B ) = 2 17 . mag (Foley et al. 2009). SN 2008ha was last detected at V = 20 2 . mag 68 days after B maximum (Valenti et al. 2009). A later non-detection of R > 22 5 . mag at 230 days after B maximum was consistent with 56 Co decay extrapolated from the last detection (Foley et al. 2010a), but the true brightness of the SN may have been much fainter. We reproduce the SN 2008ha light curve of Foley et al. (2010a) in Figure 3, including our observation of S1. The last detection and the non-detection are 8.6 and 5.6 mag brighter than S1 (and roughly the same for our limits), respectively. However, our observation of S1 was obtained 1505 days after B maximum (in the rest frame), and extrapolating the 56 Co decay to that time, we expect the SN to be at least 6.2 mag fainter than S1.
It is expected that SN light curves should fade slower than the 56 Co decay rate at times /greaterorsimilar 1000 days after B maximum, when the 57 Co decay becomes the dominant contribution to the SN luminosity (Seitenzahl et al. 2009; R¨ opke et al. 2012). We place the R¨ opke et al. (2012) model SN Ia light curves (scaled in luminosity to match SN 2008ha) in Figure 3. The models invoke different explosion mechanisms (either violent merger or delayed detonation), which result in somewhat different amounts of particular radioactive nucleosynthetic products and thus slightly different luminosities at very late times.
The details of the nucleosynthesis may be particularly important for SN 2008ha, but both model light curves are much fainter (2.3 and 2.8 mag) than S1 at the time of our observation. In fact, S1 is brighter than the expected brightness of SN 2008ha at 1000 days, before the other radioactive products are expected to strongly affect the light curve. Since the radioactive decay is exponential, even a doubling of the abundance of these long-lived radioactive species will not dramatically affect the latetime luminosity.
Additionally, S1 is more luminous than SN Iax 2005hk at 630 days after maximum (McCully et al. 2014a), after shifting its peak brightness to match that of SN 2008ha. Unless the SN faded in an unexpected way (or rebrightened), S1 is inconsistent with being emission from the SN. For S1 to be SN emission, it must have faded by an average of < 1.4 mag yr -1 from the time of the nondetection.
Any circumstellar dust extinction will only increase the discrepancy between the expected SN luminosity and that of S1.
A final possibility is that S1 is a light echo of SN 2008ha. This possibility is highly unlikely; there is no indication of dust in front of the SN, S1 is not resolved, and light echoes are usually bluer than the SN.
## 3.3. Remnant
If the explosion fails to unbind the progenitor star, then there will be a gravitationally bound remnant. Such an object has never before been identified, but it is possible that it could be quite luminous.
From early-time spectra, we know that some of the SN 2008ha ejecta is traveling at speeds much greater than the escape velocity of the WD, and Foley et al. (2010a) estimated that there was 0.3 M /circledot of unbound material. From mass loss alone, we expect the radius of the WD to change, but not significantly. If the progenitor was a Chandrasekhar-mass WD, it would have lost ∼ 20% of its mass in the explosion and would have expanded to about twice its radius. At lower masses, the radius of a WD has a weak dependence on its mass, R ∝ M -1 / 3 ; a 0.45 M /circledot WD losing 0.3 M /circledot (67% of its mass), would only have its radius increase by 44%.
Probably more important to the radius and luminosity of any remnant is the additional energy input directly from the explosion and from radioactive material left in the star. If there is a surviving companion star, then the radius of the remnant is effectively limited by the Roche radius under normal conditions. Once the remnant grows to that radius, mass transfer to the companion will begin. However, the conditions 4 years after the explosion may not be close to an equilibrium; for instance, the remnant may drive a strong opaque wind, which would increase the effective radius beyond the Roche radius, creating a common envelope. This super-Eddington wind could explain the persistent photosphere with low-velocity PCygni profiles seen in the late-time spectra of SNe Iax (Jha et al. 2006; Foley et al. 2013; McCully et al. 2014a).
Blackbody fits to the S1 photometry yield a best-fit temperature of 2100 ± 500 K and a radius of 1500 ± 800 R /circledot , both consistent with a TP-AGB star, although the implied temperature is significantly lower and the implied radius is significantly large than most TP-AGB stars and red supergiants (although perhaps consistent during an extreme mass-loss event). We caution overinterpretation of these values; although S1 was strongly detected in one band, it is only marginally detected in a second band with limits in two other bands. Additionally, a blackbody spectrum may not be appropriate for this source. Furthermore, circumstellar dust may make S1 appear redder and fainter than it really is. However if S1 has minimal circumstellar reddening and is roughly a blackbody, it must be somewhat cool and large.
About 2 × 10 46 erg of energy is needed to expand a 1 M /circledot WD to ∼ 1000 R /circledot . There was ∼ 2 × 10 48 erg of kinetic energy coupled to the ejecta of SN 2008ha (Foley et al. 2009). Therefore, about 1% of the explosion energy needs to be injected into the star to expand it to our measured radius. Even less energy is needed if only a low-mass envelope expanded to those radii. If the star were to do this, the effective temperature would decrease, and the remnant may appear similar to an TPAGB star.
Jordan et al. (2012) and Kromer et al. (2013) presented single-plume deflagration explosion models for SNe Iax. These models, which do not burn through the entire WD, left a ∼ 1 M /circledot remnant that was polluted with nucleosynthetic ash, and particularly some 56 Ni. Assuming that there was full γ -ray trapping in the remnant, Kromer et al. (2013) indicated that the remnant may become more luminous than the SN at ∼ 100 days after B -band maximum. The luminosity of this remnant, however, is tied to the amount of heating from the explosion and the 56 Ni in the remnant, making the remnant luminosity only slightly larger than the SN luminosity at late times. The remnant luminosity should decrease slower than that of the SN, but this decay is still larger than observed for S1. The remnant was not resolved in their simulation, but Kromer et al. (2013) expects the remnant to expand and be 'puffed up,' qualitatively consistent with the large radius inferred for S1.
## 3.4. Companion Star
M14 discovered the progenitor system for SN 2012Z, another SN Iax, in pre-explosion HST images. For that case, it is impossible for the detected source to be the SN emission or a remnant. Although it is possible that this emission came from an accretion disk around the WD, the best interpretation is that it is the companion star to the WD that exploded. For SN 2008ha, with the deep HST images obtained after the SN, the interpretation of S1 is not as clear, but it could be a companion star.
Examining Figure 2, we see that S1 is near the AGB/red supergiant portion of the HR diagram, but is redder than model stellar evolution tracks. The red color of S1 may be caused by newly formed circumstellar dust, but we have no constraint on any circumstellar dust reddening. Using stellar evolution tracks (Bertelli et al. 2009) with a low metallicity (12 + log[O / H] = 8 16, . as measured from the nearby H II region; Foley et al. 2009) and ignoring possible circumstellar reddening, S1 has the temperature and luminosity of a TP-AGB star. Because of the scant photometric data for S1, we use the F814W luminosity to determine the possible initial mass range. Since this calculation does not incorporate any color information, we consider all models consistent to within five times the photometric uncertainty of the F814W luminosity of S1 to be consistent. This
limit includes the possibility of circumstellar reddening of E B ( -V ) ≤ 0 4 mag ( . A V ≤ 1 2 mag). . The data are consistent with a TP-AGB star that is in the midst of dramatic mass loss. The luminosity of S1 is consistent with TP-AGB stars with initial masses of M > 3 M /circledot (for Z = 0 004; Marigo & Girardi 2007). If there is even more . circumstellar reddening, the temperature could change dramatically, and specific conclusions about the nature of S1 would need to be adjusted.
This mass range is consistent with having a C/O WD companion, as is necessary for SN 2008ha (Foley et al. 2010a). Therefore, the TP-AGB explanation makes S1 a viable candidate for a companion star in the SN 2008ha progenitor system.
SNe Iax come from relatively young populations (Foley et al. 2009; Lyman et al. 2013), and thus the stars close to the position of SN 2008ha can constrain the age of the progenitor system. Fitting isochrones to nearby stars (Bertelli et al. 2009; Figure 4), we find a best-fit logarithmic age (in years) of 7 74 . ± 0 09, corresponding . to t = 55 +13 -10 Myr. However, if we exclude one of the four nearby stars, we determine a 1σ upper limit on the age of 80 Myr. This age is consistent with our mass range of TP-AGB stars for S1.
It is possible, although very unlikely, that SN 2008ha had a massive-star progenitor (Valenti et al. 2009; Moriya et al. 2010). For this scenario, the progenitor must have been a very massive star that had its hydrogen and a significant amount of helium stripped from its outer layers. Any such star would have too short a life to allow a 2 - 6 M /circledot companion to evolve to become a TPAGB star. Therefore, if S1 is the companion star to the SN 2008ha progenitor, it is one more reason that its progenitor could not be a massive star. Furthermore, this scenario requires that the progenitor star, which would necessarily had a very short lifetime ( < 10 Myr), would be much younger than nearby stars.
S1 is much redder than the likely companion star for SN 2012Z and the progenitor of V445 Puppis, a Galactic helium nova (Kato & Hachisu 2003; Woudt et al. 2009). Furthermore, S1 is inconsistent with specific models for a massive helium-burning star companion (Liu et al. 2010), which has been proposed as the progenitor systems for SNe Iax (Foley et al. 2013) and SN 2012Z in particular (M14). If S1 is intrinsically as blue as these other stars, but reddened by circumstellar dust, then S1 must also be much more luminous than these stars. It is unlikely that S1 is a reddened, but intrinsically blue, He star.
However, only a small portion of the parameter space for WD/He star progenitor systems has been probed. It is possible that S1 is a He star, but instead of a mainsequence He star, it would be a He red giant, similar in color and luminosity to R Cor Bor. Kato et al. (2008) examined the stellar evolution of He stars to the red giant branch. Their approach was similar to that of Paczy´ski n (1971) and others. Kato et al. (2008) found that single He stars with M > 0 8 M . /circledot would evolve to become much more luminous and redder, with luminosities consistent (although typically higher) with S1. However, the lowest model temperatures were roughly 4500 K, hotter than S1. Although this model should be investigated further, especially if it is found that S1 is reddened, the data are currently inconsistent with the models.
Although observations of S1 after the SN are consistent with being a TP-AGB star, the SN may have affected the observational properties of any companion star. For SNe Ia, it is expected that any nearby companion will have mass stripped and will be shock heated (e.g., Marietta et al. 2000; Pan et al. 2012; Shappee et al. 2013); as a result, their luminosity is expected to increase. Pan et al. (2012), Pan et al. (2013), and Liu et al. (2013) specifically looked at Hestar companions, the same companion star suggested for SN 2012Z (McCully et al. 2014b). These studies found that because of their size, the amount of mass stripping was much lower for He star companions than mainsequence or TP-AGB companions, but still suggest that the star will 'puff up' and become quite luminous.
None of these studies indicated that the companion star should expand to ∼ 1500 R /circledot or become as cool as S1. Moreover, SN 2008ha had an ejecta mass and kinetic energy one and three orders of magnitude less than that of a SN Ia, respectively. Any effect from the SN on the companion star should be less than that of a SN Ia explosion. We therefore think that it is unlikely that any companion star is significantly perturbed, and it is extremely unlikely that S1 is an impacted main-sequence He star, but detailed modeling is required to see the full effects of this scenario.
As seen in Figure 2, S1 has very different physical parameters from the SN 2012Z companion star. There are five scenarios that are consistent with the observations of SNe 2008ha and 2012Z: (1) S1 is not the companion star of in the SN 2008ha progenitor system and the blue He-star companion has a luminosity below our current limits, (2) S1 was a blue He star and the SN significantly altered its appearance (which we consider unlikely), (3) the blue emission from the SN 2012Z progenitor system was from the WD accretion disk (a super-soft source) and the companion star is red and lower luminosity like S1 (which McCully et al. 2014b consider unlikely), (4) S1 is a He red giant, simply a slightly more evolved version of the companion star for SN 2012Z, or (5) SNe Iax have a variety of progenitor systems.
## 4. DISCUSSION AND CONCLUSIONS
Using HST images obtained 4.1 years after the explosion, we detected a stellar source very close to the position of SN 2008ha, which we call 'S1.' We determined that this source is unlikely to be a chance superposition, but have not ruled out such a coincidence. We have also shown that the luminosity of S1 is much higher than the expected luminosity of the SN at the time of the observations, leaving three possibilities for S1: it is not directly related to SN 2008ha, it is the remnant of the WD that produced SN 2008ha, or it is the companion star to the WD that produced SN 2008ha.
Fitting isochrones to stars near the position of SN 2008ha, we have determined that the stars near SN 2008ha have a likely age of < 80 Myr. This is consistent with the average value determined for SNe Iax based on their locations relative to H α emission (30-50 Myr; Lyman et al. 2013).
The photometry of S1 suggests that it is a TP-AGB star with an initial mass of > 3 M /circledot (Marigo & Girardi 2007). This mass range is consistent with being the com-
Fig. 4.Color-magnitude diagrams for the region near SN 2008ha. The large filled (with error bars), small filled, and open circles represent stars within 15, 30, and 45 pixels (0.75, 1.5, and 2.25 ′′ ) of the position of SN 2008ha, respectively. The grey area represents magnitudes below our 3σ detection limit. The black curve is a 55 Myr isochrone.
panion of a C/O WD and inconsistent with being the companion of a massive star progenitor.
S1 appears to be too cool to be a He red giant. However, the previous arguments for a He companion star for SN Iax progenitor systems, the properties of the SN 2012Z progenitor system, and the obvious mass transfer triggering by stellar evolution make this option somewhat attractive. Binary evolution models and models of how a SN might change a He red giant star could improve our understanding of the plausibility of this option. Newly formed circumstellar dust may make such a star appear much redder than its intrinsic color.
Regardless, S1 has observational properties very different from the progenitor system for SN 2012Z (M14). It is possible that SNe 2008ha and 2012Z had similar progenitor systems and S1 has expanded and cooled as the result of the SN impact or that the emission in the SN 2012Z progenitor system is dominated by emission from the accretion disk. However, detailed calculations are required to test these possibilities. If S1 is the companion star for the SN 2008ha progenitor system and its observational properties have not changed significantly from before the SN explosion, then SNe Iax may have come from a variety of progenitor systems. Perhaps the significant diversity in the properties of SNe Iax (SN 2012Z was 40 times as luminous as SN 2008ha at peak), is related to the diversity of progenitor systems.
It is possible that S1 is an inflated remnant of the WD. SN 2008ha ejected only 0.3 M /circledot of material (Foley et al. 2010a) and therefore likely left a remnant. Models indicate that this remnant should be relatively luminous at late times and 'puffed up' (Jordan et al. 2012; Kromer et al. 2013; Fink et al. 2014). S1 has a blackbody radius of ∼ 1500 R /circledot , which is qualitatively consis- tent with these models; moreover, the energy required for this expansion is relatively small compared to the energy of the explosion.
Future observations should be able to further confirm if S1 is the remnant and may rule it out as a companion star. The remnant should vary on years timescales, while it is unlikely that the luminosity of the companion star will change significantly. Eventually, we will build large enough telescopes to obtain a spectrum of S1; if S1 is indeed the remnant, it will likely have a peculiar spectrum.
Deeper observations at bluer wavelengths will have a better chance of detecting a helium star (the preferred companion star for SN 2012Z) at a position more coincident with SN 2008ha than S1. New HST images could also improve our astrometry by both increasing the S/N of S1 and by measuring additional stars for the astrometric linking. In the future, obtaining deep highresolution images of SNe Iax will improve our ability to constrain progenitor systems and remnants associated with SNe Iax.
S1 represents potentially the second progenitor system of a thermonuclear SN detected, after SN 2012Z. If this is the correct interpretation, it would complicate the Hestar companion scenario of Foley et al. (2013) and require a variety of progenitor systems for SNe Iax. Instead if S1 is the remnant of SN 2008ha, it would be the first detection of such an object. Continued monitoring of S1 will improve our understanding of SN 2008ha and SNe Iax.
Facility: Hubble Space Telescope(ACS)
We thank the anonymous referee for helpful and in-
formed comments.
Based on observations made with the NASA/ESA Hubble Space Telescope , obtained at the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-26555. These observations are associated with program GO-12999.
through NASA/ HST grant GO-12913.01, and National Science Foundation (NSF) CAREER award AST0847157 to S.W.J. M.D.S. gratefully acknowledges generous support provided by the Danish Agency for Science and Technology and Innovation realized through a Sapere Aude Level 2 grant. This work was supported by the NSF under grants PHY 11-25915 and AST 11-09174.
This research at Rutgers University was supported
## REFERENCES
Bertelli, G., Nasi, E., Girardi, L., & Marigo, P. 2009, A&A, 508, 355
Marigo, P., & Girardi, L. 2007, A&A, 469, 239
- Chornock, R., Filippenko, A. V., Branch, D., Foley, R. J., Jha, S.,
- & Li, W. 2006, PASP, 118, 722
- Dolphin, A. E. 2000, PASP, 112, 1383
- Fink, M., et al. 2014, MNRAS, 438, 1762
- Foley, R. J. 2008, Central Bureau Electronic Telegrams, 1576, 2
Foley, R. J., Brown, P. J., Rest, A., Challis, P. J., Kirshner, R. P.,
& Wood-Vasey, W. M. 2010a, ApJ, 708, L61
Foley, R. J., et al. 2009, AJ, 138, 376
- --. 2010b, AJ, 140, 1321
- --. 2013, ApJ, 767, 57
- Iben, Jr., I., & Tutukov, A. V. 1991, ApJ, 370, 615
- Jha, S., Branch, D., Chornock, R., Foley, R. J., Li, W., Swift,
- B. J., Casebeer, D., & Filippenko, A. V. 2006, AJ, 132, 189
- Jordan, IV, G. C., Perets, H. B., Fisher, R. T., & van Rossum,
- D. R. 2012, ApJ, 761, L23
- Kato, M., & Hachisu, I. 2003, ApJ, 598, L107
- Kato, M., Hachisu, I., Kiyota, S., & Saio, H. 2008, ApJ, 684, 1366 Kromer, M., et al. 2013, MNRAS, 429, 2287
- Li, W., et al. 2003, PASP, 115, 453
- Liu, W.-M., Chen, W.-C., Wang, B., & Han, Z. W. 2010, A&A, 523, A3
- Liu, Z.-W., et al. 2013, ApJ, 774, 37
- Lyman, J. D., James, P. A., Perets, H. B., Anderson, J. P.,
- Gal-Yam, A., Mazzali, P., & Percival, S. M. 2013, MNRAS, 434, 527
- Maoz, D., Mannucci, F., & Nelemans, G. 2013, ArXiv e-prints, 1312.0628
- Marietta, E., Burrows, A., & Fryxell, B. 2000, ApJS, 128, 615
- Maund, J. R., et al. 2010, ApJ, 722, 1162 McClelland, C. M., et al. 2010, ApJ, 720, 704 McCully, C., et al. 2014a, ApJ, 786, 134 McCully, C., Jha, S. W., Foley, R. J., & et al. 2014b, Nature, in press. Milne, P. A., et al. 2010, ApJ, 721, 1627 Moriya, T., Tominaga, N., Tanaka, M., Nomoto, K., Sauer, D. N., Mazzali, P. A., Maeda, K., & Suzuki, T. 2010, ApJ, 719, 1445 Narayan, G., et al. 2011, ApJ, 731, L11+ Paczy´ski, B. 1971, Acta Astronomica, 21, 1 n Pan, K.-C., Ricker, P. M., & Taam, R. E. 2012, ApJ, 750, 151 --. 2013, ApJ, 773, 49 Phillips, M. M., et al. 2007, PASP, 119, 360 Puckett, T., Moore, C., Newton, J., & Orff, T. 2008, Central Bureau Electronic Telegrams, 1567, 1 R¨ opke, F. K., et al. 2012, ApJ, 750, L19 Schlafly, E. F., & Finkbeiner, D. P. 2011, ApJ, 737, 103 Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, ApJ, 500, 525 Seitenzahl, I. R., Taubenberger, S., & Sim, S. A. 2009, MNRAS, 400, 531 Shappee, B. J., Stanek, K. Z., Pogge, R. W., & Garnavich, P. M. 2013, ApJ, 762, L5 Stritzinger, M. D., et al. 2014, A&A, 561, A146 Valenti, S., et al. 2009, Nature, 459, 674
- Woudt, P. A., et al. 2009, ApJ, 706, 738
|
10.1088/0004-637X/792/1/29
|
[
"Ryan J. Foley",
"Curtis McCully",
"Saurabh W. Jha",
"Lars Bildsten",
"Wen-fai Fong",
"Gautham Narayan",
"Armin Rest",
"Maximilian D. Stritzinger"
] |
2014-08-05T20:00:04+00:00
|
2014-08-05T20:00:04+00:00
|
[
"astro-ph.HE",
"astro-ph.SR"
] |
Possible Detection of the Stellar Donor or Remnant for the Type Iax Supernova 2008ha
|
Type Iax supernovae (SNe Iax) are thermonuclear explosions that are related
to SNe Ia, but are physically distinct. The most important differences are that
SNe Iax have significantly lower luminosity (1% - 50% that of typical SNe Ia),
lower ejecta mass (~0.1 - 0.5 M_sun), and may leave a bound remnant. The most
extreme SN Iax is SN 2008ha, which peaked at M_V = -14.2 mag, about 5 mag below
that of typical SNe Ia. Here, we present Hubble Space Telescope (HST) images of
UGC 12682, the host galaxy of SN 2008ha, taken 4.1 years after the peak
brightness of SN 2008ha. In these deep, high-resolution images, we detect a
source coincident (0.86 HST pixels; 0.043"; 1.1 sigma) with the position of SN
2008ha with M_F814W = -5.4 mag. We determine that this source is unlikely to be
a chance coincidence, but that scenario cannot be completely ruled out. If this
source is directly related to SN 2008ha, it is either the luminous bound
remnant of the progenitor white dwarf or its companion star. The source is
consistent with being an evolved >3 M_sun initial mass star, and is
significantly redder than the SN Iax 2012Z progenitor system, the first
detected progenitor system for a thermonuclear SN. If this source is the
companion star for SN 2008ha, there is a diversity in SN Iax progenitor
systems, perhaps related to the diversity in SN Iax explosions. If the source
is the bound remnant of the white dwarf, it must have expanded significantly.
Regardless of the nature of this source, we constrain the progenitor system of
SN 2008ha to have an age of <80 Myr.
|
1408.1092v2
|
## HALF OF THE MOST LUMINOUS QUASARS MAY BE OBSCURED: INVESTIGATING THE NATURE OF WISE-SELECTED HOT, DUST-OBSCURED GALAXIES
R.J. Assef 1 , P.R.M. Eisenhardt 2 , D. Stern 2 , C.-W. Tsai 2,3 , J. Wu 4 , D. Wylezalek 5 , A.W. Blain 6 , C.R. Bridge 7 E. Donoso 8 , A. Gonzales 7,9,10 , R.L. Griffith 11 , T.H. Jarrett 12
Draft version February 20, 2015
## ABSTRACT
The WISE mission has unveiled a rare population of high-redshift ( z = 1 -4 6), . dusty, hyperluminous galaxies, with infrared luminosities L IR > 10 13 L /circledot , and sometimes exceeding 10 14 L /circledot . Previous work has shown that their dust temperatures and overall far-IR spectral energy distributions (SEDs) are significantly hotter than expected to be powered by star-formation. We present here an analysis of the rest-frame optical through mid-IR SEDs for a large sample of these so-called 'Hot, DustObscured Galaxies' (Hot DOGs). We find that the SEDs of Hot DOGs are generally well modeled by the combination of a luminous, yet obscured AGN that dominates the rest-frame emission at λ > 1 µ m and the bolometric luminosity output, and a less luminous host galaxy that is responsible for the bulk of the rest optical/UV emission. Even though the stellar mass of the host galaxies may be as large as 10 11 -10 12 M /circledot , the AGN emission, with a range of luminosities comparable to those of the most luminous QSOs known, require that either Hot DOGs have black hole masses significantly in excess of the local relations, or that they radiate significantly above the Eddington limit, at a level at least 10 times more efficiently than z ∼ 2 QSOs. We show that, while rare, the number density of Hot DOGs is comparable to that of equally luminous but unobscured (i.e., Type 1) QSOs. This may be at odds with the trend suggested at lower luminosities for the fraction of obscured AGN to decrease with increasing luminosity. That trend may, instead, reverse at higher luminosities. Alternatively, Hot DOGs may not be the torus-obscured counterparts of the known optically selected, largely unobscured HyperLuminous QSOs, and may represent a new component of the galaxy evolution paradigm. Finally, we discuss the environments of Hot DOGs and statistically show that these objects are in regions as dense as those of known high-redshift proto-clusters.
Subject headings: galaxies: active - galaxies: evolution - galaxies: high-redshift - quasars: general - infrared: galaxies
## 1. INTRODUCTION
Massive galaxies are thought to evolve from starforming disks into passive ellipticals through major mergers that trigger star-formation and intense episodes of accretion into their central super-massive black holes
- 1 N´ ucleo de Astronom´ ıa de la Facultad de Ingenier´ ıa, Universidad Diego Portales, Av. Ej´ ercito Libertador 441, Santiago, Chile. E-mail: [email protected]
- 2 Jet Propulsion Laboratory, California Institute of Technology, 4800 Oak Grove Drive, Pasadena, CA 91109, USA.
- 4 UCLA Astronomy, PO Box 951547, Los Angeles, CA 900951547, USA
- 3 NASA Postdoctoral Program Fellow
- 5 European Southern Observatory, Karl-Schwarzschildstr.2, D-85748 Garching bei M¨nchen, Germany u
(SMBHs) (e.g. Hopkins et al. 2008). Such a picture can explain several properties of galaxies, such as the tight correlations between the mass of their SMBH ( M BH ) and the mass, luminosity and velocity dispersion of the galaxy's spheroidal component (e.g. Magorrian et al. 1998; Ferrarese & Merritt 2000; Tremaine et al. 2002; Marconi & Hunt 2003; Bentz et al. 2009; G¨ltekin et al. u 2009), and the evolution of the galaxy luminosity density (e.g., Faber et al. 2007). In these scenarios, the host galaxy stellar mass is assembled through starformation ahead of the onset of the active galactic nuclei (AGN), which through a feedback mechanism heats up the gas and expels some of it, thereby quenching its starformation.
- 6 Physics & Astronomy, University of Leicester, 1 University Road, Leicester LE1 7RH, UK
- 8 Instituto de Ciencias Astron´micas, de la Tierra, y del Eso pacio (ICATE), 5400, San Juan, Argentina
- 7 Division of Physics, Math, and Astronomy, California Institute of Technology, Pasadena, CA 91125, USA
- 9 Department of Earth, Atmospheric and Planetary Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
- 11 Department of Astronomy and Astrophysics, The Pennsylvania State University, 525 Davey Lab, University Park, PA 16802, USA
- 10 Scripps College, 1030 Columbia Avenue, Claremont, CA 91711, USA
- 12 Astrophysics, Cosmology and Gravity Centre (ACGC), Astronomy Department, University of Cape Town, Private Bag X3, Rondebosch 7701, South Africa
An important characteristic of both the intense starformation and AGN episodes is the significant quantities of dust associated with them. In galaxies undergoing extreme star-formation and AGN activity, a large fraction of the luminous energy is absorbed by dust and then re-radiated at infrared/sub-mm wavelengths, as observed for populations such as Ultra-Luminous Infrared Galaxies (ULIRGs; Sanders & Mirabel 1996), Sub-mm Galaxies (SMGs; Blain et al. 2002; Casey et al. 2014) and Dust-Obscured Galaxies (DOGs; Dey et al. 2008). It follows then that studying the most luminous infrared galaxies in the Universe, which host the most intense star-formation and AGN activity, likely probe extreme scenarios within the galaxy evolution paradigm.
NASA's Wide-field Infrared Survey Explorer (WISE; Wright et al. 2010) was launched in December 2009, and surveyed the entire sky in four mid-IR bands centered at 3 4 . µ m (W1), 4 6 . µ m (W2), 12 µ m (W3) and 22 µ m (W4). One of the main goals of the WISE mission was to identify the most luminous infrared galaxies in the universe, and the mission has achieved considerable success towards this goal in the past few years (Eisenhardt et al. 2012; Wu et al. 2012; Bridge et al. 2013; Jones et al. 2014; Stern et al. 2014; Wu et al. 2014; Eisenhardt et al. in prep.; Lonsdale et al. in prep.; Tsai et al. 2014). One highly successful method to identify such sources is to target objects that are extremely red in the WISE bands, with faint or no detections in the more sensitive W1 and W2bands, but well detected in W3 and W4. These selection criteria were presented by Eisenhardt et al. (2012) and have been shown to successfully identify a population of luminous galaxies with z /greaterorsimilar 1 (for details on the selection criteria, see Eisenhardt et al. 2012, and § 2.1). Using sub-mm observations, Wu et al. (2012) showed that these objects are indeed extremely luminous, with bolometric luminosities exceeding 10 13 L /circledot , and sometimes exceeding 10 14 L /circledot . These observations also showed that the dust in these objects is at temperatures significantly higher than in other luminous infrared populations, such as ULIRGs and SMGs, peaking at rest-frame wavelengths λ /lessorsimilar 20 µ m. Such high dust temperatures are consistent with AGN heating, suggesting the bulk of the luminosity is produced by accretion onto the galaxy's central SMBH rather than by star-formation. Indeed, Eisenhardt et al. (2012) in a detailed study of one of these sources, WISE J181417.29+341224.9 (W1814+3412 hereafter), determined that its spectral energy distribution (SED) is consistent with a heavily obscured ( A V ∼ 50 mag) AGN producing the bulk of the luminosity. Due to their high dust temperatures, and yet similar optical-to-mid-IR colors to DOGs, we adopt the terminology of Wu et al. (2012) and refer to these objects as Hot, Dust-Obscured Galaxies or Hot DOGs.
Follow-up studies have provided additional interesting aspects of this population. The observedframe optical spectra of these objects (Eisenhardt et al. 2012; Wu et al. 2012; Stern et al. 2014; Eisenhardt et al. in prep.) have diverse properties. While many of these objects show narrow emission lines commonly associated with obscured AGN activity present in the IR, some show features more closely associated with starformation, with mostly absorption lines and only Ly α emission (e.g., W1814+3412 Eisenhardt et al. 2012), and some even show red continua with a lack of emission lines (e.g.,WISE J092625.44+423251.9; Wu et al. 2012). Imaging obtained with the Hubble Space Telescope /WFC3 and through adaptive optics with Keck/NIRC2 have shown that these objects typically are not gravitationally lensed, implying their luminosities are intrinsic (Eisenhardt et al. 2012; Wu et al. 2014; Tsai et al. 2014). Recently, Jones et al. (2014) reported observations at 850 µ m with SCUBA-2 at the James Clerk Maxwell Telescope (JCMT) of a subsample of 10 Hot DOGs, which suggest these objects are located in arcmin-scale overdensities of luminous dusty galaxies (see § 6) and confirm the hot dust temperatures determined by Wu et al. (2012). Jones et al. (2014) constrain the contribution of ULIRG-type star-formation to less than 30% of the IR luminosity, and of spiral-type star-formation to less than 3%. Similar conclusions are reached by Wu et al. (2014), who studied two Hot DOGs at submm and mm wavelengths at higher spatial resolution using the Submillimeter Array (SMA) and the Combined Array for Research in Millimeter-wave Astronomy (CARMA). Wu et al. (2014) was able to constrain their cold dust masses to amounts comparable to those of quasars with comparable luminosities. On the other end of the electromagnetic spectrum, Stern et al. (2014) studied the AGN nature of three Hot DOGs using Xray observations obtained with the X-ray Multi-Mirror Mission ( XMM-Newton ) and the Nuclear Spectroscopic Telescope Array ( NuSTAR ), finding that the AGN emission is heavily absorbed, possibly Compton-thick. Using a similar sample to that defined by the Hot DOG selection criteria, Bridge et al. (2013) determined that a significant fraction of such objects show extended Ly α emission on 30 - 100 kpc scales, and pointed out this could be consistent with the presence of intense quasar feedback. Bridge et al. (2013) also further constrained the high temperature of these objects by using Herschel /PACS and SPIRE observations to map the full shape of their far-IR SEDs. Finally, Lonsdale et al. (in prep.) present a study based on ALMA Cycle 0 observations of the farIR SEDs of radio-selected, red WISE objects that are possibly the radio-loud counterparts of Hot DOGs. Although these objects are located at somewhat lower redshifts (0.47-2.85), they share some of the same characteristics, including the overall high dust temperatures and an even more dramatic apparent overdensity of nearby luminous dusty galaxies (Jones et al. in prep.).
In this work we study the physical properties of Hot DOGs by analyzing their SEDs, number densities and environments. In a companion paper, Tsai et al. (2014) presents a detailed study of the most luminous Hot DOGs, those with bolometric luminosities in excess of 10 14 L /circledot . The article is structured as follows. In Section 2 we discuss the sample selection and the follow-up photometric and spectroscopic observations. In Section 3 we present our SED modeling methodology, while in Section 4 we apply it to model our sample of Hot DOGs and discuss their inferred physical properties. In Section 5 we compare the number density of Hot DOGs to that of comparably luminous QSOs. Finally, in Section 6 we study the density of the environments in which Hot DOGs are found using follow-up Warm Spitzer /IRAC imaging. We discuss how the environments compare to known clusters at similar redshifts, and how this constrains the stellar masses of Hot DOGs. Throughout this work we assume a flat ΛCDM cosmology with H 0 = 73 km s -1 , Ω M = 0 3, . and Ω Λ = 0 7. . We refer to all magnitudes in the Vega photometric system. For convenience, the different samples used throughout this work are summarized in Table 1.
## 2. SAMPLE SELECTION AND MULTI-WAVELENGTH FOLLOW-UP OBSERVATIONS
## 2.1. WISE and the W12drop Selection
The WISE mission observed the full sky in four midIR photometric bands with a FWHM of 6 ′′ in W1-3 and 12 ′′ in W4. We use the WISE All-Sky data release, which includes all observations obtained during the fully cryogenic mission. WISE surveyed the sky in a polar or-
Table 1 Sample Definition
| Sample | Description | All | All | IRAC Detected | IRAC Detected |
|---------------------------------------------|-----------------------------------|-------|-------|-----------------|-----------------|
| | | N | N z | N | N z |
| Full Sample W12drop | Eqn. (1) | 934 | 155 | 711 | 115 |
| Full Sample Hot DOG | z > 1 and W12drop | · · · | 122 | · · · | 96 |
| Core Sample W12drop | W4 < 7.2 and W12drop | 252 | 95 | 103 | 58 |
| Core Sample Hot DOG | W4 < 7.2 and Hot DOG | · · · | 77 | · · · | 52 |
| Hot DOGs for number density analysis ( § 5) | 2 < z < 4 and Core Sample Hot DOG | · · · | · · · | · · · | 42 |
Note . -All results presented in this article are based on objects detected by Warm Spitzer /IRAC (see § 2.2.1). The definition of the Hot DOG population presented in this table reflects the definition adopted for this work but is somewhat more restrictive than those used elsewhere (e.g., as in Wu et al. 2012; see § 4). Note that for samples where z is not required for their selection, N is the total number of objects regardless of whether they have been observed spectroscopically, while Nz shows the number of objects with a measured redshift.
bit with respect to the ecliptic, simultaneously obtaining images in all four bands. Hence, the number of observations in a field increases with its ecliptic latitude. While fields near the ecliptic were typically observed 12 times, the number can grow to thousands near the ecliptic poles (e.g., Jarrett et al. 2011). The median coverage across the sky is approximately 15 frames per passband. Detailed accounts of the mission are presented by Wright et al. (2010) and in the WISE All-Sky data release explanatory supplement 13 .
to limit contamination by Galactic objects. All objects are required to be free of artifacts flagged by the WISE pipeline and to not be associated with either known asteroids or those discovered by WISE (Mainzer et al. 2011).
As discussed earlier, our canonical picture of galaxy evolution suggests the existence of key stages where massive galaxies experience extremely luminous but heavily dust-enshrouded star-formation and nuclear activity. For the most massive galaxies, these stages may reach infrared luminosities L IR > 10 13 L /circledot , and hence be classified as Hyper-Luminous Infrared Galaxies (HyLIRGs), but can be very faint in the optical bands due to obscuration. Wu et al. (2012) and Eisenhardt et al. (2012) presented a large sample of WISE-selected HyLIRGs, which are the main target of this study. The selection criteria used by Eisenhardt et al. (2012) and Wu et al. (2012) specifically target galaxies red enough to be well detected in the long wavelength WISE bands W3 and W4, but are poorly or undetected at the shortest wavelength, more sensitive W1 and W2 bands. Samples selected in this way are referred to as 'W1W2-dropouts' by Eisenhardt et al. (2012) and Wu et al. (2012), but for brevity we use 'W12drops' here as an equivalent term.
W12drop selection requires that W1 > 17 4 mag, and . that either
$$W 4 \, < \, 7. 7 \, \max \, \wedge \, W 2 - W 4 \, > \, 8. 2 \, \max, \quad ( 1 ) \quad \text{$\dots$}$$
or
$$W 3 \, < \, 1 0. 6 \, \max \, \wedge \, W 2 - W 3 \, > \, 5. 3 \, \max. \, \ ( 2 ) \quad _ { e } ^ { \, \text{to} \, b }$$
Furthermore, objects are required to be farther than 30 ◦ from the Galactic Center and 10 ◦ from the Galactic Plane
13 http://wise2.ipac.caltech.edu/docs/release/allsky/expsup/
Finally, we required candidates to pass a series of visual inspections of both individual exposures and coadded images for any given source. We focused such efforts on the brighter candidates with W4 < 7 2, . resulting in a sample of 252 objects over approximately 32 , 000 deg . 2 We refer to these objects as the 'core sample'. A search to W4 /lessorsimilar 7.7 using preliminary reductions covering 70% of the total area was also carried out, resulting in an additional sample of 682 W12drops which is somewhat less complete and well-characterized than the core sample. We refer to the total sample of 934 objects as the 'full sample'.
It is now known that WISE All-Sky profile-fitting derived fluxes of very faint sources are significantly biased due to excess sky subtraction during the data processing (see Lake et al. 2013). The effect is somewhat stochastic in nature, but can be well modeled as a constant underestimation of 9 29 . ± 0 04 . µ Jy and 10 38 . ± 0 07 . µ Jy in the W1 and W2 fluxes of the WISE All-Sky release catalog (S. Lake, private communication). Some of these issues have been corrected in the latest WISE data release, dubbed AllWISE, but because the All-Sky Catalog was used for the W12drop selection we use the WISE All-Sky fluxes and apply the corrections outlined above when modeling the selection function rather than trying to translate the W12drop selection function to the AllWISE data release. We note that the SED modeling discussed later is not affected by this issue, since we rely on deeper Warm Spitzer observations for those wavelengths (see § 2.2.1 and § 4 for details), but it will prove to be important when considering the W12drop selection function (see § 5 for details).
Additionally, Wright et al. (2010) found systematic differences in the W3 and W4 magnitudes of red and blue calibrators, with the red calibrators being 17% too faint in W3 and 9% too bright in W4. Given the red col-
ors of our sources, we have corrected their W3 and W4 magnitudes by adding -0.17 mag and 0.09 mag respectively to the quantities reported in the All-Sky release for modeling their SEDs (see § 3). Since these corrections are not considered by the selection criteria, we remove them when evaluating the W12drop selection function in § 5. Brown et al. (2014) have recently suggested a somewhat larger correction of 0.13 mag instead 0.09 mag for sources as red as those considered here. Our main results would not be qualitatively affected by using this larger correction.
## 2.2. Follow-up Observations
## 2.2.1. Warm Spitzer Observations
We obtained observations of Hot DOGs with the IRAC instrument (Fazio et al. 2004) onboard the Spitzer Space Telescope (Werner et al. 2004). In its non-cryogenically cooled state, known as Warm Spitzer , the IRAC camera obtains photometry in two broad-band channels centered at 3 6 . and 4 5 . µ m, referred to as [3.6] and [4.5]. The channels are similar to the WISE W1 and W2 bands, but because of its larger aperture, longer exposure time, and smaller PSF (FWHM of ≈ 1.7 ′′ in each band), IRAC provides significantly deeper images. We refer the reader to Griffith et al. (2012) for details of this program, as well as of the data reduction and photometric measurements. Of the 934 W12drops in the full sample, 712 were observed with Warm Spitzer in the [3.6] and [4.5] IRAC channels. All but one of these is well detected in both bands, with the one outlier object (W0149-8257) detected only in the [4.5] band. We limit our parent sample to those 711 objects detected in both Warm Spitzer /IRAC bands. The core sample (W4 < 7.2) is similarly reduced to 103 targets.
## 2.2.2. Ground-Based Near-IR Imaging
We obtained follow-up near-IR observations of our sample using the Wide-field IR Camera (WIRC; Wilson et al. 2003) on the Hale 200-inch telescope at Palomar Mountain, the WIYN High-Resolution Infrared Camera (WHIRC; Meixner et al. 2010) at the 3.5m WIYN telescope, the Ohio State InfraRed Imager/Spectrometer (OSIRIS; Depoy et al. 1993) at the 4m SOAR telescope, and the SAO Widefield InfraRed Camera (SWIRC; Brown et al. 2008) at the 6.5m MMT telescope. Table 2 provides more details about these observations.
All images were reduced following standard IRAF procedures using the XDIMSUM package 14 , and all fluxes were obtained in 4.0 ′′ diameter apertures. Each image was flux-calibrated using the 2MASS point source catalog (PSC, Skrutskie et al. 2006), using comparison stars within the field of view whenever possible, or by using the closest observation in time of a field containing 2MASS detected stars if conditions were photometric. The latter was only necessary for some of the OSIRIS observations, which have an 80 ′′ field of view. We add the dispersion of the zero-point calibration in quadrature to the photometric uncertainty of each source. Magnitudes are listed in Table 2 for each of the W12drops for which we obtained follow-up near-IR observations.
Table 2 Ground-Based NIR Follow-Up
| WISE ID | Band | Mag † | Unc. | Instrument ‡ |
|-------------------------|--------|-------------------------------------------|-------------------------------|----------------|
| WISEJ000431.34-192301.8 | H | 19.453 21.795 20.270 18.833 19.794 19.861 | 0.125 0.630 0.117 0.142 0.166 | A |
| WISEJ000709.03+730831.2 | J | | | B |
| WISEJ002659.24+201556.2 | J | | | B |
| | Ks | | | C |
| WISEJ002933.06+020505.4 | J | | | B |
| WISEJ012611.98-052909.6 | J | | 0.135 | B |
| WISEJ013400.59-260726.5 | J | 20.163 | 0.123 | A |
| WISEJ014747.59-092350.5 | J | > 23.000 | | B |
Note . - Table 2 is published in its entirety in the electronic edition of ApJ. A portion is shown here for guidance regarding its form and content.
- † 1 σ upper bounds shown for undetected sources.
- ‡ Telescope and instrument used for the observations, defined as A : SOAR/OSIRIS; B : WIYN 3m/WHIRC; C : Hale 200inch/WIRC; and D : MMT/SWIRC.
Of the 711 (103) objects in our full (core) W12drop sample with Warm Spitzer observations, 84 (52) have been observed in J -band, 23 (16) in H -band, and 37 (19) in K -band. Of these, only 1 (0) object has been observed in all three bands, and only 26 (16) have been observed in more than one band.
## 2.2.3. Optical Spectroscopy
Optical spectroscopy was performed for a large fraction of our sample using several facilities, with greater emphasis given to the core sample. These were obtained primarily using LRIS on the Keck-I telescope (see Wu et al. 2012, for examples) and GMOS-S on the Gemini-S telescope. Redshifts are generally based on multiple features and are therefore considered secure. Eisenhardt et al. (in prep.) will provide a comprehensive description of the optical spectroscopy. Figure 1 shows the redshift distribution of all 115 (58) objects in the full (core) W12drop sample with Warm Spitzer observations and successful redshift measurements. The full sample shows a clear minimum at 0 6 . < z < 2, suggesting the low- and high-redshift populations are distinct. The core sample distribution is also consistent with the presence of the bimodality, albeit at lower significance due to the smaller number of objects. Since we are only interested in very luminous objects, we focus on the sample of 96 (52) objects from the full (core) sample with z > 1. Eisenhardt et al. (in prep.) shows that approximately 70% of the objects targeted for spectroscopy yielded redshift measurements, but that objects with failed spectroscopic measurements are primarily due to optical faintness, suggesting they are typically located at high redshift and are bona-fide HyLIRGs. We note that during some of the spectroscopic observing runs, we biased against targets detected in both the B and R bands of the Digitized Sky Survey 15 to reduce the incidence of low redshift contaminants. This bias has only a minor effect for the z > 1 population, so we do not discuss it any further.
## 3. SED MODELING METHODOLOGY
Figure 1. Redshift distribution of the 115 W12drops with spectroscopic redshift and IRAC measurements. The shaded histogram shows the distribution for the 58 objects in the core sample (W4 < 7.2, see § 2.1).
In order to gain insight into the physical properties of Hot DOGs, we study their rest-frame optical through mid-IR multi-wavelength SEDs. Specifically, we study their rest-frame optical through mid-IR properties by combining the WISE data with optical spectroscopy, Warm Spitzer observations and ground-based near-IR photometry. We model the SEDs following the approach applied in Eisenhardt et al. (2012) to study Hot DOG W1814+3412. Namely, we use the low resolution AGN and galaxy SED templates and the respective modeling algorithm of Assef et al. (2010). Briefly, every object is modeled as a non-negative linear combination of three host galaxy SED templates (broadly resembling E, Sbc and Im types, see Assef et al. 2010, for details) and one AGN SED template. Eisenhardt et al. (2012) used the AGN template of Richards et al. (2006) because of its longer wavelength range. Here we use the Assef et al. (2010) AGN template instead, but there is no substantial difference between these AGN templates for the purpose of the SED modeling of Hot DOGs over the UV-throughmid-IR wavelength range studied in this work.
We also fit an AGN reddening component, which we parametrize by the color excess E B ( -V ), considering values from 0 to 10 1 5 . . The assumed reddening-law corresponds to an SMC-like extinction for λ < 3300 ˚ , and A a Galactic extinction curve at longer wavelengths. Additional details are provided in Assef et al. (2010). Although in some unresolved obscured sources it may be important to modify these reddening curves to account for the optical and UV photons scattered into the line-ofsight (see Kochanek et al. 2012), our sources are under too much obscuration for this to be a significant issue. In the next sections we show that this approach does, in general, perform a good job of modeling Hot DOGs, although possible shortcomings are discussed in detail.
One of the important quantities we want to study is the intrinsic bolometric luminosity of the underlying AGN. We estimate it using the scaling relation of Kaspi et al. (2000),
$$L _ { \text{AGN} } ^ { \text{Bol} } \, = \, 9 \, \lambda L _ { \lambda } ( 5 1 0 0 \AA ). \, \text{ \quad \ \ } ( 3 )$$
The continuum luminosity at 5100 ˚ is calculated by takA ing only the best-fit reddened AGN component, and removing the obscuration. While a more self-consistent bolometric luminosity can be obtained by integrating over the best-fit AGN component to the SED (see, e.g., Assef et al. 2010, 2013; Eisenhardt et al. 2012), this scaling is widely used, so adopting it simplifies the comparison with other results in the literature. For reference, the luminosity obtained integrating over the unreddened AGN template between 0.1 and 30 µ m is greater than the bolometric luminosity estimated using equation (3) by a factor of 1.3. By using a single AGN template we are implicitly assuming the torus covering fraction of the accretion disk in all Hot DOGs is equal to that implied by the template, which is likely about 50% given the results of Assef et al. (2013). For luminous Type 1 quasars Roseboom et al. (2013) found that the mean covering fraction was 39% with a dispersion of 18%. Since the amplitude of the AGN component in the best-fit SED is primarily anchored by the W3 and W4 fluxes, a larger covering fraction could reduce the deduced AGN continuum luminosity at 5100 ˚ . In principle, this luminosity could A be overestimated by up to a factor of 2 in the extreme scenario where the torus leaves no open line of sight towards the accretion disk and is composed of uniformly distributed hot dust clouds, as all dust must strongly radiate at λ /lessorsimilar 10 µ m to affect our results. This scenario is, however, unlikely, given the detection of AGN narrowemission lines in many of the rest-frame UV spectra (see, e.g. Wu et al. 2012).
We also wish to estimate the stellar mass ( M ∗ ) of each host galaxy. We estimate this quantity by multiplying the rest-frame luminosity of the host component in the K -band by the mass-to-light ratio ( M/L ) in that band. The value of M/L depends on many parameters, including the galaxy's star-formation history, metallicity, stellar initial mass function (IMF) and contribution from thermally pulsating asymptotic giant branch (TP-AGB) stars. Because we generally have only a single photometric band probing the host properties, we only aim to place meaningful bounds on M/L . Although the lower end of the M/L range is only loosely bound, the upper end is much better constrained, as it is primarily limited by the age of the Universe at the redshift of the object. Hence, for the purpose of this study, we will focus on estimating upper bounds on the stellar mass of each Hot DOG. We estimate these upper bounds for each object using the EzGal code of Mancone & Gonzalez (2012) with the stellar population models of Bruzual & Charlot (2003). We choose these stellar population models in favor of more recent ones available for EzGal (e.g., Maraston 2005; Conroy et al. 2009; Conroy & Gunn 2010) since they have the lowest contribution of TP-AGB stars to the composite SEDs, implying the highest M/L values. The M/L values in the rest-frame near-IR are rather insensitive to variations in the metallicity, with lower metallicities implying higher M/L values in K s . To be
conservative we consider the lowest metallicity available for EzGal, Z = 0 008 ( . ≡ 0 4 . Z /circledot ). For the star-formation history, we consider a simple stellar population (SSP) with a formation redshift z F = 15. Finally, for the IMF, we consider the results of Conroy et al. (2013), who have shown that in early-type galaxies, the M/L ratio in K -band can be up to twice that expected for the Milky Way. We use M/L values two times higher than those estimated assuming a Chabrier (2003) IMF. Our only assumption that would tend to underestimate the stellar mass is that of little host obscuration in the rest-frame K -band. We discuss this issue further in detail in § 4 and § 6, where evidence is shown that although Hot DOGs live in significantly dense regions, possibly clusters or dense filaments, higher stellar masses would predict even richer environments than observed.
Finally, we also attempt to estimate the central SMBH mass ( M BH ) for these objects. In many cases, however, we do not quote the values of M BH but of the Eddington luminosity defined as
$$L _ { \text{Edd} } \, = \, 3. 2 8 \times 1 0 ^ { 4 } \, \left ( \frac { M _ { \text{BH} } } { M _ { \odot } } \right ) \, L _ { \odot }, \quad ( 4 ) \quad \text{$\text{$\text{$wh}$} } { \text{for}$} \\ \text{ish a common to the minimum of $t$ which $h$} \, $h$} \, $ \text{that} \, $ \text{$\text{$the}$} \, $ \text{$h$} \, $ h $ }$$
which corresponds to the luminosity at which photon pressure inhibits isotropic accretion onto an isotropically radiating body. This is an interesting quantity to study for AGN since most energy is generated by accretion onto the SMBH. Furthermore, Kollmeier et al. (2006, also see Shen et al. 2008) have shown that luminous QSOs at similar redshifts as Hot DOGs tend to radiate in a limited range of Eddington ratios, defined as λ E = L Bol AGN /L Edd .
Direct estimates of M BH in AGN, and hence L Edd , based on single optical spectra are possible by combining the width of their broad emission lines and the luminosity of their accretion disks (e.g., Vestergaard & Peterson 2006), typically limiting such measurements to unobscured objects. In an upcoming article (Wu et al. in prep.) we explore such estimates for a handful of Hot DOGs where we observed broad H α in the near-IR, but such methods are certainly not applicable to the sample of objects we study here. Hence, we consider two alternative methods, and explore their consequences in depth in § 4. First, we estimate M BH through equation (4) by assuming that the AGN in Hot DOGs radiate at the typical λ E = 0.30 determined by Kollmeier et al. (2006) for QSOs at similar redshifts. Alternatively, we assume that M BH is related to the stellar mass in the host galaxy in the same way as found for local galaxies. Specifically, we use the relation between the spheroidal component mass ( M Sph ) and M BH of Bennert et al. (2011a),
$$\log \frac { M _ { \text{BH} } } { M _ { \odot } } \, = \, - 3. 3 4 \, + \, 1. 0 9 \log \frac { M _ { \text{Sph} } } { M _ { \odot } }. \quad ( 5 ) \quad \text{(} ^ { \text{$i$} } _ { \text{nie} 1$}$$
Given that the near-IR imaging reported here is either ground based or from Spitzer , we cannot attempt a morphological decomposition of the bulge or spheroidal component. Furthermore, due to the extreme nature of Hot DOGs, it is not even clear the definition of a spheroidal component would be sensible. Hence, we assume that all of the detected rest-frame K -band light belongs to the spheroidal component. Combined with the fact that our stellar mass estimates are also upper bounds, our estimates of M BH obtained in this manner should be con- sidered as generous upper limits. In the next section we discuss how both estimates lead to very different scenarios for Hot DOGs, highlighting that accurate estimates of M BH in Hot DOGs are a crucial element for understanding their nature, requiring more extensive optical/nearIR spectral coverage.
Due to the low number of degrees of freedom, there can be considerable uncertainty for the best-fit parameters of each galaxy. To account for this, we do a Monte Carlo approach. For each object we resample each available flux from a Gaussian distribution centered at the measured value with a standard deviation equal to the photometric uncertainty, and re-estimate all the parameters described above. We repeat this 1000 times per galaxy and re-estimate the results discussed in the next section, finding no significant difference.
Finally, we caution that the infrared emission might not be powered by AGN activity but might be powered by extreme starburst disks, such as proposed by Thompson et al. (2005), who found that, under the appropriate conditions, a disk of star-formation can form where gas and dust are supported primarily by starformation feedback and radiation pressure. The typical dust temperature can become quite high, comparable to the temperatures of ∼ 60 K typical of Hot DOGs (Wu et al. 2012). Extreme star-formation rates, on the order of ∼ 10 4 M /circledot yr -1 , would be needed to power the luminosities seen in Hot DOGs (see, e.g. Eisenhardt et al. 2012), making this scenario quite unlikely. Also, Thompson et al. (2005) find that it is inevitable to form a bright AGN at the center of such a starburst disk and that, furthermore, the star-formation driven rest-frame 10 µ memission still relates to the accretion disk emission in the same way as for a regular Type 1 QSO as judging from the templates of Elvis et al. (1994). Hence, even in this arguably unlikely physical scenario, the main results presented here are still valid. We do not refer to this scenario hereafter.
## 4. ANALYSIS
By design, W12drops are either undetected or marginally detected in W1 and W2. Because of this, we fit their SEDs using the Warm Spitzer [3.6] and [4.5] photometry, W3 and W4 photometry, and, whenever possible, the ground-based near-IR photometry. Figure 2 shows examples of SED fits to two high-redshift and two low-redshift W12drops. Objects with z < 1 have much lower luminosities and are much worse fit by the models than their higher-redshift counterparts, reinforcing the idea they compose two distinct populations. From now on, we will use interchangeably the terms 'Hot DOG' (Wu et al. 2012) and 'W12drop at z > 1' for convenience in this article.
Our SED fits show that the AGN component dominates the luminosity of Hot DOGs, accounting for > 97% of the total 0 1 . -30 µ m output in all objects. Figure 3 shows the residuals of the SED fits both as a ratio between the observed and model fluxes (top panel) and as the difference with respect to the measurement uncertainties (bottom panel). The ground-based near-IR photometry shows large discrepancies with respect to the modeled fluxes but only in absolute terms, as the deviations are in almost no case beyond 3 σ . The Warm Spitzer /IRAC and W3-4 fluxes show only small abso-
Figure 2. Examples of SED fits to two high-redshift (top panels) and two low-redshift W12drops (bottom panels). All four have χ 2 values close to the median values of their populations. Each panel shows the observed fluxes ( green filled circles ), as well as the expected fluxes from the best-fit SED model ( black open triangles ). The solid lack line shows the best-fit SED, composed of an AGN under E B ( -V ) of obscuration ( dashed blue line ) plus a quiescent ( dotted red ), intermediate ( long-dashed green line ) and strongly star-forming ( dot-dashed cyan line ) stellar populations (see § 3 for details). Most galaxies do not require all three stellar templates. Notice the much higher χ 2 values of the low-redshift galaxy fits.
lute discrepancies from the model, without a significant systematic component. The lack of systematic residuals also suggests it is unlikely there is a significant amount of unrecognized obscuration on the host galaxy that would bias our maximal stellar mass estimates (see § 3).
Figure 4 shows the distribution of the bolometric luminosity of the AGN component L Bol AGN for Hot DOGs in the full sample as well as in the core sample. Hot DOGs tend to have quite luminous AGN, with bolometric luminosities between 10 47 and 10 48 erg s -1 . In the same panel, we show the values of the characteristic quasar luminosity function (QLF) luminosity L ∗ at two different redshifts, z = 1 0 and 2.0, determined by converting ap-. propriately the values of M ∗ ,J determined by Assef et al. (2011). The AGN components in Hot DOGs are among the most luminous AGN at their redshifts. We discuss this further in § 5.
Figure 5 shows the distribution of dust obscuration towards the accretion disk, parametrized by E B ( -V ). Eisenhardt et al. (2012) studied Hot DOG W1814+3412 in detail using 8 bands of optical through mid-IR photometry, finding an obscuration of E B ( -V ) = 15 6 . ± 1 4 . with a very similar approach to that used here. With an updated processing of the WISE data and the corrections to W3 and W4 discussed in § 2.1 but a more limited multi-wavelength photometry set than that used by Eisenhardt et al. (2012), we find here a lower obscuration of E B ( -V ) = 11 7 . ± 1 2 for . W1814+3412. This obscuration is lower primarily due to the W3 and W4
Figure 3. Flux residuals of the SED models for the 89 out 96 Hot DOGs with χ 2 < 20.
Figure 4. Distribution of AGN bolometric luminosities derived from our SED modeling of Hot DOGs (see § 3 and § 4 for details). The solid hashed histogram shows the distribution of the core sample (i.e., W4 < 7.2), while the dashed open one shows the distribution of the full sample. The vertical dashed gray lines show the characteristic AGN bolometric luminosity derived from the J -band QLF of Assef et al. (2011), assuming the unobscured AGN SED template of Assef et al. (2010) and equation (3).
Figure 5. Distribution of AGN obscurations derived from our SED modeling of Hot DOGs (see § 3 and § 4 for details). The solid hashed histogram shows the distribution of the core sample (i.e., W4 < 7.2), while the dashed open one shows the distribution of the whole sample. The dashed gray line shows the obscuration derived for HyLIRG W1814+3412, studied by Eisenhardt et al. (2012). For reference, we note that for the reddening law used in this article, the attenuation factor at 1 µ m is given by A 1 µm = 1 24 . ∗ E B ( -V ).
band corrections applied here. If we do not consider these corrections, we find an obscuration of E B ( -V ) = 15 1 . ± 1 2 . for W1814+3412, consistent with the value found by Eisenhardt et al. (2012). Figure 5 shows that W1814+3412has significantly more obscuration than the average Hot DOG, for which 〈 E B ( -V ) 〉 = 6 8 in the . core sample and 〈 E B ( -V ) 〉 = 6 4 in the full sample. . For completeness, we note the median obscuration is E B ( -V ) = 6 0 in the core sample and . E B ( -V ) = 5 5 . in the full sample. The best-fit AGN obscuration for Hot DOGs ranges between 2 5 . < E B ( -V ) < 21 5. . Assuming the median gas-to-dust ratio of Maiolino et al. (2001), E B ( -V ) /N H = 1 5 . × 10 -23 cm 2 mag, the average obscuration corresponds to gas column densities of approximately of 4 × 10 23 cm -2 , and a range of 1 6 . × 10 23 < N H < 1 4 . × 10 24 cm -2 . This is over 10 times more absorption than the typical dividing line between Type 1 and Type 2 AGN ( N H = 10 22 cm -2 ; see, e.g., Ueda et al. 2003), to just being slightly below Compton thick ( N H > . 1 5 × 10 24 cm -2 ). Using X-ray observations of a sample of three Hot DOGS obtained with XMM-Newton and NuSTAR , Stern et al. (2014) inferred for each of them neutral hydrogen column densities of N H /greaterorsimilar 10 24 cm -2 , consistent with Compton-thick obscuration, and in good agreement with the values inferred above from dust obscuration.
Figure 6 shows our maximal stellar mass estimates. For reference, we also show the stellar mass of an L ∗ galaxy today ( ∼ 5 × 10 10 M ; e.g., Baldry et al. 2008). /circledot Even if we were to assume a mass-to-light ratio 10 times lower, as appropriate for extreme starbursts, Hot DOGs
Figure 6. Distribution of stellar mass upper bounds derived from our SED modeling of Hot DOGs (see § 3 and § 4 for details). The solid hashed histogram shows the distribution of the core sample (i.e., W4 < 7.2), while the dashed open one shows the distribution of the whole sample. The dashed gray line shows the stellar mass of an L ∗ galaxy today from Baldry et al. (2008).
would still have massive host galaxies. In § 6 we further discuss these host mass estimates and explore their environments in the context of such massive host galaxies. However, as massive as the host galaxies are, the AGN still dominates the emission by orders of magnitude, which has very interesting implications for the nature of Hot DOGs depending on the mass of their SMBH. As discussed earlier, however, we are not able to obtain unique estimates of M BH . Below we explore two scenarios for which we can estimate M BH based on indirect considerations as well as the implications of each scenario to the nature of Hot DOGs.
## 4.1. M BH Estimates based on a Fixed λ E
The first scenario estimates M BH assuming a fixed Eddington ratio of λ E = 0.30, the same Eddington ratio as that determined by Kollmeier et al. (2006) for z ∼ 2 QSOs. Since the AGN emission is so luminous, a very large M BH is needed to have an SMBH accreting at the same level of regular QSOs. In fact, the SMBH mass is much larger than expected for the local stellar bulge to M BH ratio (Bennert et al. 2011a) and the stellar mass of the host galaxy, even for the upper limits calculated above, as shown in Figure 7. Even if we assume that the stellar mass is equal to the upper bound and is dominated by an spheroidal component, Hot DOGs sit an order of magnitude above the local relation, but the discrepancy is likely much larger. If these galaxies are to evolve in such a way that at z = 0 they would fall in the local measured relations, the bulge will have to grow at a much faster rate than the accreting central SMBH, implying that SMBH growth precedes the host galaxy growth. In such a scenario, the stellar mass growth could
Figure 7. The position in the M BH -M Sph diagram of Hot DOGs ( red circles ) if we assume the AGN in Hot DOGs radiate at the typical λ E = 0 30 of . z ∼ 2 QSOs. For comparison, we also show the relation of local active galaxies determined by Bennert et al. (2011a), as well as the position in the diagram of different high redshift populations in the literature, as described in the top left labels. For the literature samples, the label shows their respective mean redshifts.
not be quenched by intense AGN activity, as is assumed in many evolution models of massive galaxies (as in, e.g., Hopkins et al. 2008), since the epoch of intense SMBH growth significantly precedes the end of the host galaxy assembly, but may be consistent with quenching by lowlevel AGN accretion through radio-mode feedback (as in, e.g., Croton et al. 2006).
While this discrepancy can be alleviated by considering a larger value for λ E , it is consistent with what has been found previously for some Type 1 QSOs. For example, Bennert et al. (2011b) studied the evolution of the M BH -M Sph relation with cosmic time using a sample of z ∼ 1 unobscured QSOs observed with the Hubble Space Telescope where M Sph was determined through SED and morphological modeling of the host. Bennert et al. (2011b) found a trend in the same direction suggested by our data, namely M BH /M Sph ∝ (1+ z ) 1 96 . ± 0 55 . , while combination with more heterogeneous measurements in the literature yielded M BH /M Sph ∝ (1 + z ) 1 16 . ± 0 15 . . Figure 7 shows the set of objects studied in detail by Bennert et al. (2011b), illustrating a displacement from the local relation comparable to that of Hot DOGs. Similar results have also been obtained for unobscured QSOs at z ∼ 6 by Wang et al. (2010, and references therein) and at z ∼ 2 by Coppin et al. (2008), where the host stellar masses come from dynamical estimates based on CO emission and the assumption of random orientation (i.e., i = 32 7 ). . o These findings support the assumption that the AGN in Hot DOGs radiate at the same λ E as similar redshift QSOs. For comparison, Figure 7 also shows the location in this diagram of z ∼ 2 SMGs, as determined by Borys et al. (2005), which are highly discrepant from the local relation but in the opposite direction to QSOs. Interestingly, Bongiorno et al. (2014) has recently shown that z ∼ 2 Type 1.8/1.9 QSOs may actually show better agreement with the local relation than Type 1 QSOs, with a discrepancy that may depend on M BH .
Adifferent picture to interpret these results is that proposed by Peng (2007, see also Jahnke & Macci` 2011), o who have argued that the local host galaxy -black hole mass correlations can naturally arise through galaxy mergers without co-evolution between them and no initial relation. The prediction for that scenario is that the dispersion of the M BH -L Bulge relation would increase substantially with redshift, possibly consistent with the large spread observed in Figure 7. Hot DOGs and QSOs could then be regarded as the tail of that distribution, namely exceedingly massive black holes in faint host galaxies.
## 4.2. M BH Estimates based on the Local M BH -M Sph Relation
In the second scenario, we assume that the local relationship between the spheroidal component stellar mass and M BH is appropriate for Hot DOGs (see § 3), and obtain an upper bound of the SMBH mass from the estimates of the maximal stellar mass. Combined with the AGN bolometric luminosity of Hot DOGs, this provides an estimate of the 'minimum' Eddington ratio ( λ E ) at which the AGN is radiating. Figure 8 shows the distribution of the implied Eddington ratio of Hot DOGs. The peak is about 2-3 times the Eddington limit, and almost all objects are accreting at λ E ≥ 1. For comparison, we show the typical Eddington ratio of ∼ 0 30 for QSOs at . a redshift of ∼ 2 determined by Kollmeier et al. (2006).
Although the scenario discussed in § 4.1 may seem more likely, Hot DOGs might indeed be radiating at several times their Eddington limit. Note the Eddington limit is not necessarily a hard, physical boundary in this case, since the accretion is most likely not isotropic. Due to photon pressure, it is possible that the AGN would be in the process of mechanically expelling material from the regions close to the central SMBH and, possibly, from the galaxy's potential well. This would be consistent with the high fraction of extended Ly α emission reported by Bridge et al. (2013), and may relate specifically to the key stage in the standard galaxy evolution paradigm where star-formation is being shut down, and thus transitioning to the QSO stage where it will remain until the accreting gas supply stops (see, e.g., Hopkins et al. 2008). Were we to assume the evolving M BH /M Sph relation of Bennert et al. (2011b, see § 4.1), our SMBH masses would be underestimated by a factor of 3.5-5 and the AGN in the Hot DOG population would still be radiating at, or close to the Eddington limit, well above the mean QSO Eddington ratio found by Kollmeier et al. (2006, see Fig. 8). A correction factor closer to 10 is needed to yield consistency with the mean λ E of Kollmeier et al. (2006). In § 5 we discuss the number density of these objects and further argue that Hot DOGs might indeed be radiating at high Eddington ratios.
The presence of extended Ly α emission may put interesting constraints on the lifetime of Hot DOGs if we assume it comes from winds launched by a super-Eddington
Figure 8. Distribution of the minimum Eddington ratio for the AGN in Hot DOGs, assuming that the local relation between M Bulge and M BH holds at high redshift. The solid hashed histogram shows the distribution of the core sample (i.e., W4 < 7.2), while the dashed open one shows the distribution of the whole sample. The dashed gray line shows the the peak Eddington ratio found by Kollmeier et al. (2006) for z ∼ 2 Type 1 QSOs.
accreting AGN. Bridge et al. (2013) find the emitting gas has projected velocities of up to several thousand km s -1 . Assuming a constant velocity of ∼ 10 000 km s , -1 , for the gas to have been launched from the central regions of the galaxy, this stage would have to last /greaterorsimilar 10 Myr to reach a distance of ∼ 100 kpc. This is consistent with the lifetimes of 1 Myr < t QSO < 20 Myr estimated by Trainor & Steidel (2013) for a sample of hyper-luminous QSOs (which they defined as L UV ∼ 10 14 L /circledot ). On the other hand, if the gas were simply responding to the emission of the AGN but was not associated to feedback, the light travel time puts a weak constraint on the lifetime of this phase to /greaterorsimilar 3 × 10 5 yr.
## 5. NUMBER DENSITIES OF HOT DOGS AND QSOS
Several lines of evidence imply that Hot DOGs are indeed powered by heavily obscured AGN (Wu et al. 2012; Eisenhardt et al. 2012, in prep.), and so it is natural to compare their properties to those of comparably luminous but unobscured Type 1 QSOs, which are much better studied. The simplest comparison is of their space densities. For this, we focus on the core sample of 252 W12drops described in § 2.1 (i.e., W4 < 7.2). We know that of the 103 core W12drops with Warm Spitzer /IRAC observations, 58 have reliable spectroscopic redshifts. Of these, 52 (90%) are at z ≥ 1 and can hence be considered Hot DOGs, while 43 (74%) are at z ≥ 2. Assuming that objects without spectroscopic redshift measurements have a redshift distribution approximately equal to that of the spectroscopic sample (Eisenhardt et al. in prep.), there are 187 Hot DOGs with z ≥ 2 and W4 < 7.2. Of them, Hot DOG W2246-0526 ( z = 4 6) .
is the only one at z > 4 and it seems to be somewhat of an outlier based on its photometric properties, so we will limit the study to the Hot DOGs in the redshift range 2 < z < 4.
We study the effects of the selection function using the 42 core sample Hot DOGs with reliable spectroscopic redshifts in the range 2 < z < 4 and IRAC observations. Because the Hot DOG selection is based upon the observed WISE magnitudes and colors, objects identified as core sample Hot DOGs may not have been recognized as such if located at a different redshift. This causes a significant sample incompleteness which we take into account by using the V/V Max method of Schmidt (1968). In short, we use the best-fit SED model of each individual source to evaluate the redshift range, and hence volume V , for which the object could have been detected and identified by WISE as a Hot DOG within the full volume V Max , corresponding to the redshift range 2 < z < 4. We then simply assume that the intrinsic distribution of sources is uniform across the given volume V Max such that the effective number surface density of Hot DOGs is then
$$N _ { \text{HD} } \, = \, \frac { 1 } { f _ { z } A _ { \text{Sky} } } \, \sum _ { i } \, \frac { 1 } { V _ { i } / V _ { \text{Max} } }, \quad \, ( 6 ) \\ \therefore f _ { \cdot } = 5 8 / 2 5 2 = 0 \, 2 3 \, \text{is the fraction of core samble}$$
where f z = 58 252 = 0 23 is the fraction of core sample / . Hot Dogs with IRAC observations and a reliable redshift measurement, and A Sky = 32 000 deg , 2 is the area of the sky surveyed for Hot DOGs. We estimate the error as
$$\delta N _ { \text{HD} } \, = \, \frac { 1 } { f _ { z } A _ { \text{Sky} } } \, \left ( \sum _ { i } \, \frac { 1 } { ( V _ { i } / V _ { \text{Max} } ) ^ { 2 } } \right ) ^ { 1 / 2 }. \quad ( 7 ) \\ \text{Two caveats to this process should be noted. First, we}$$
Two caveats to this process should be noted. First, we apply the corrections of Lake et al. (2013, and Lake, S. private comm.) to the model fluxes before evaluating the selection function to account for the flux bias of the WISE All-Sky release (see § 2.1 for details). And, second, we apply a small correction to the W3 and W4 model fluxes to perfectly match the observed values at the object's redshift. Even after applying the flux bias correction, we find that 10 of the 52 Hot DOGs in the core sample with IRAC observations and spectroscopic z ≥ 1 (6 of which are in the redshift range 2 < z < 4) were only selected as such because of the stochastic nature of the flux bias. The SED modeling of these objects, driven primarily by the IRAC fluxes, predicts W1 and W2 fluxes too bright to be selected as Hot DOGs. We do not use these objects to estimate N HD . Note the change to f z from eliminating these 7 objects from the sample is negligible, decreasing from 0.23 to 0.21.
Using equation (6) we find then that the effective number surface density of Hot DOGs is N HD = 0.032 ± 0.004 deg -2 , or approximately one for every 31 ± 4 deg . 2 Figure 9 shows the cumulative, volume-corrected surface density of Hot DOGs as a function of increasing luminosity of the obscured AGN component. We parametrize the AGN component luminosity by M AGN J , the absolute J -band magnitude it would have under no obscuration. A drawback of our V/V Max approach to correct for incompleteness is that it disregards evolution in the Hot DOG number density between 2 < z < 4. Furthermore, using such a large redshift range can lead to very large correction factors, which could be introducing significant noise
Figure 9. The top left panel shows the number density of Hot DOGs in the redshift range 2 < z < 4 ( solid blue line ) corrected for volume incompleteness through the V/V Max approach of Schmidt (1968, see § 5 for details). We also show the region marked by the error-bars ( solid blue region ). The dotted blue lines show the uncertainty region taking into account a nominal scatter in the AGN covering fractions, as discussed in the text. For comparison, we also show the expected number density of Type 1 QSOs as predicted by the QLFs of Richards et al. (2006, green) , Assef et al. (2011, cyan) and Ross et al. (2013, red) . The remaining three panels show the same but for three redshift slices within the full range. For comparison, the Hot DOG distribution of the 2 < z < 2 5 slice is shown . in the other two slices ( dashed gray line ).
to our estimates. To avoid these issues, Figure 9 also shows the effective number of Hot DOGs as a function of the AGN component luminosity for three redshift slices: 2 < z < 2 5, 2 5 . . < z < 3 and 3 < z < 4. Due to the low number of sources, the uncertainties are large in the last redshift bin. Interestingly, little evolution is observed between the first two redshifts slices, while some evolution may be present when compared to the 3 < z < 4 slice. The statistics are currently insufficient to be more definitive about this.
The estimates of M AGN J from our SED modeling are based on the assumption of a single AGN covering fraction ( c f ) for all Hot DOGs, as discussed in § 3. It is possible, however, that there is a significant scatter on the c f of Hot DOGs. To explore this we re-estimate the uncertainty regions now including the effects of a scatter in c f equal to that found by Roseboom et al. (2013) for Type 1 QSOs. The scatter in c f found by Roseboom et al. (2013) can be expressed as an uncertainty of 0.5 mag in the magnitude of any band dominated by the accretion disk emission for a fixed mid-IR flux. Although the restframe J -band luminosity is only partly powered by the accretion disk, we consider the full 0.5 mag uncertainty in our estimates of M AGN J for simplicity. Figure 9 shows the uncertainty region considering the effects of such a scatter. Including this effect increases the size of the uncertainty region, particularly at the bright end, but does not qualitatively change our results.
To obtain the number density of comparably luminous Type 1 AGN we use the functional form of the optically selected QLF of Richards et al. (2006) and Ross et al. (2013), derived from SDSS observations, and of the midIR and X-ray selected QLF of Assef et al. (2011), derived from the deeper but smaller area observations of the NDWFS Bo¨ otes field. The brighter end of the luminosity function is much better constrained by the SDSS observations used by Richards et al. (2006) and Ross et al. (2013) than by the much smaller Bo¨tes field used by o Assef et al. (2011), suggesting they are a better comparison to our HyLIRG sample. However, the mid-IR nature of our Hot DOG selection may be better represented by the selection function of Assef et al. (2011). We note that Richards et al. (2006) assumed a flattening of the QLF at z > 2 4, . motivated by the flatter 4 < z < 5 QLF of Fan et al. (2001), which has since been shown to be incorrect (McGreer et al. 2013) and an artifact of a rapidly evolving break luminosity (see also Assef et al. 2011). The QLF of Ross et al. (2013) is only defined over the redshift range 2 2 . < z < 3 5, so we extrapolate the . parameters using their functional form to the modestly larger 2 < z < 4 range of Hot DOGs.
Let Φ( M ,z J ) be the space density of QSOs at redshift z with absolute magnitude M J . The number surface density of QSOs in the sky that are brighter than a certain luminosity in the redshift range 2 < z < 4 is then given by
$$N _ { \text{QSO} } \, = \, \int _ { 2 } ^ { 4 } \int _ { - \infty } ^ { M _ { J } } \phi ( M _ { J }, z ) \, \frac { d V _ { C } } { d z } \, d M _ { J } d z, \quad ( 8 ) \\ \vdots \quad \tau \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdot \, \cdots$$
where V C is the comoving volume. The luminosity functions of Richards et al. (2006) and Ross et al. (2013) are parametrized as a function of M i ′ , the absolute magnitude in the i ′ -band, which corresponds to the SDSS i -band shifted to z = 2. We use the rest-frame AB color of a Type 1 QSO with no host-contamination of i ′ -J = -0 78 determined by Assef et al. (2011) to con-. vert between the absolute i ′ and J -band magnitudes. Figure 9 shows the cumulative surface density of Type 1 QSOs according to each of the QLF parametrizations. Each of the uncertainty regions shown in Figure 9 are estimated by calculating 500 realizations of the QLF created through re-sampling of each of its parameters according to their published 1 σ uncertainties and assuming Gaussian statistics. The shaded regions show the 68% confidence interval of the 500 realizations. Note that as the co-variances between parameters are neglected, the uncertainty regions could be somewhat overestimated.
Figure 9 shows that the counts we obtained for Hot DOGs are quite well matched at the bright end by those predicted by the QLF of Ross et al. (2013), suggesting that Hot DOGs are as common as QSOs of comparable luminosity. This is roughly consistent with the comparison to the Assef et al. (2011) QLF except in the highest redshift range, where error bars are quite large due to the small size of the NDWFS Bo¨tes field. o Unsurprisingly, there is a significant discrepancy with the Richards et al. (2006) QLF, consistent with the large discrepancy between the latter and the Ross et al. (2013) QLFs. Also note that the optical color selection function of Richards et al. (2006) is least effective at z ∼ 2 5 .
due to confusion with the colors of the stellar locus (Fan 1999). The difference in the faint-end of the slopes of the respective curves at 2 0 . < z < 2 5 may also suggest that . Hot DOGs follow a different luminosity function than QSOs, although this is not observed in the higher redshift bins.
Some studies of lower-luminosity QSOs have found that the fraction of Type 2 AGN decreases strongly as a function of increasing bolometric luminosity (see, e.g., Ueda et al. 2003; Hasinger 2004; Simpson 2005; Assef et al. 2013), implying there should be a very small number of obscured QSOs at the luminosities of Hot DOGs. This prediction is significantly at odds with the results shown in Figure 9, since Hot DOGs appear to be as common as comparably luminous Type-1 QSOs. This discrepancy could indicate a reversal in the trends found at lower luminosities, implying that the fraction of Type-2 QSOs increases with luminosity towards the upper end of the QLF. Such a reversal has also been suggested by Banerji et al. (2012), although with limited statistics, based on the obscuration fraction of NIR selected QSOs. Banerji et al. (2012) suggests that these reddened, high luminosity QSOs may probe an evolutionary phase rather than an orientation effect. Other studies, primarily at lower luminosities, have found that the obscuration fraction may be a shallow function of the bolometric luminosity (Lusso et al. 2013) or independent of it (e.g., Wang & Jiang 2006; H¨ onig et al. 2011; Lawrence & Elvis 2010; Lacy et al. 2013), so that a high fraction of obscured QSOs at high luminosities may not be surprising. The obscuration of the lower-luminosity QSOs is thought to come from dust primarily in the vicinity of the SMBH, namely the dust torus, and models that replicate the lowering fraction of obscured objects with increasing luminosity have been devised for these structures (see, e.g., Lawrence 1991; Simpson 2005). This, however, may not be the case for Hot DOGs, which may be obscured by a different dust structure, such as dust on significantly larger physical scales or with a significantly different geometry or covering fraction, although we caution the reader that a significantly different covering fraction/geometry than that implied by our AGN SED template could have a considerable effect over the estimated M J value (see also the discussion in § 3). This could naturally solve the tension with the obscuration trends found for the lower-luminosity QSOs, and would imply that Hot DOGs are not the torus-obscured counterparts of the known Type 1 QSOs of similar luminosities. Lonsdale et al. (in prep.) have come to a similar conclusion for objects that may be the radio-loud counterparts of Hot DOGs. Measuring the total dust content from a combination of mid-IR and ALMA sub-mm observations, Lonsdale et al. (in prep.) suggests the torii of these objects would have to be unrealistically large to explain the high luminosities observed.
Alternatively, we could consider the possibility discussed in the previous section that the AGN in Hot DOGs radiate at an Eddington ratio significantly above unity, in contrast to typical z ∼ 2 QSOs which radiate at λ E ∼ 0 30 (Kollmeier et al. 2006). . We can speculate then that Hot DOGs are objects going through a phase of their evolution in which, for a brief period of time, they radiate well above their Eddington limit. If so, we should compare their number density to that of lower luminos- ity QSOs instead. In this case, we would conclude that Hot DOGs only constitute a small fraction of SMBHs of moderate to high mass, rather than a large fraction of the tip of the SMBH mass function. The fraction of QSOs of the same SMBH mass that are in this 'Hot DOG phase' would constrain the duration of the latter relative to the QSO lifetime, and could be discussed in the context of the timescales estimated at the end of § 4.2. We, however, refrain from exploring this any further here since there are too many uncertainties for a meaningful discussion.
As noted in § 2.2.3, 30% of objects targeted for spectroscopy did not yield a successful redshift measurement, primarily due to optical faintness (Eisenhardt et al. in prep.). This means that while we considered 58 core sample W12drops with spectroscopic z and Warm Spitzer imaging, 83 were targeted. Their optical faintness implies the 25 additional objects have high-z, and here we have assumed they follow the same redshift distribution of Figure 1. The effects of a different redshift distribution can be approximated by modifying f z , but is unlikely to qualitatively modify our conclusions. For example, if all of these 25 objects had 2 < z < 4 and had the same magnitude distribution, the effects over the 2 < z < 4 number density distribution of Figure 9 (top left panel)can be approximated by making f z = ((58 + 25) / 252) × (42 / (42 + 25)) = 0 21 instead of . f z = 0 23. Alternatively, in the unlikely scenario that all . 25 objects had z < 2, we could approximate the effects by making fz = 83 252 = 0 33. / .
## 6. THE ENVIRONMENT OF HOT DOGS
Recently, Jones et al. (2014) found evidence of an overdensity of sub-mm neighbors to a small sample of Hot DOGs using 850 µ m observations with SCUBA-2 at JCMT, and also noted that the overdensity did not show an angular dependence around the Hot DOGs within 90 . Assessing the environments of these objects is, how-′′ ever, not trivial, as their high redshifts and corresponding faintness makes getting spectroscopic distances a daunting task. Hence, we must rely on a statistical approach.
We use the deep Warm Spitzer /IRAC imaging described in § 2.2.1 to count the number of galaxies neighboring Hot DOGs. Of the 96 Hot DOGs (see Table 1), we use the 90 that were observed as part of the same snapshot program (ID 70162; see Griffith et al. 2012, for details). Such imaging is ideal for this as the peak of the stellar emission is redshifted into the IRAC bands at the distance of Hot DOGs, allowing us to probe much lower stellar masses than what, for example, optical imaging would allow us to. We follow the approach of Wylezalek et al. (2013), and use the IRAC imaging to study the field density in the vicinity of Hot DOGs in comparison with the field density in two control samples: i) around random pointings in the Warm Spitzer UKIDSS Ultra Deep Survey (SpUDS, P.I.:J. Dunlop) representative of field galaxies, and ii) around radio-loud AGN in the Clusters Around RadioLoud AGN survey (CARLA; Wylezalek et al. 2013, 2014; and see Galametz et al. 2010, 2013 for spectroscopic confirmation of two of these clusters). CARLA identifies moderately massive clusters at high redshift.
We start by counting for each of the 90 Hot DOGs the number of red galaxies ([3.6]-[4.5] > 0.37, to select only z /greaterorsimilar 1 galaxies) found in a 1 ′ radius around it. We
Figure 10. The solid line shows the distribution of the host galaxy observed-frame [4.5] magnitudes, obtained from the SED modeling described in § 3. For comparison, the dotted line shows the observed [4.5] magnitudes, including the contribution of the AGN. The vertical dashed line shows the 10 µ Jy completeness limit of our IRAC imaging.
use a circle with a 1 ′ radius because such size is small enough to fit well within a Spitzer /IRAC image, but large enough to encompass a typical mid-IR selected cluster with log ( M 200 /M /circledot ) ∼ 14 at z > 1 (Brodwin et al. 2011; Wylezalek et al. 2013). We then repeat this around all 420 CARLA radio-loud AGN (RLAGN) and in 437 randomly selected pointings within the SpUDS survey. We determined that a grid of 19 × 23=437 apertures maximized the number of independent 1 ′ radii apertures extracted from the SpUDS survey region. The shallowest IRAC depth of the three samples is that of Hot DOGs, with a limiting [4.5] flux of 10 µ Jy, so we only consider objects down to that depth in all three samples. We note that the [4.5] magnitudes of the host galaxies in Hot DOGs, obtained from the SED modeling described in § 3, are typically ∼ 1 mag brighter than the 10 µ Jy depth of our IRAC imaging, as shown in Figure 10.
Figure 11 shows the results of this analysis. The density of galaxies in the Hot DOG imaging is greater than that in the SpUDS control sample, suggesting the environment of Hot DOGs is significantly more dense than that of field galaxies. The Hot DOGs show good agreement with the CARLA sources, suggesting that Hot DOGs live in environments similar to those of RLAGN. Formally, a K-S test shows that the probability of the Hot DOG surface density distribution being drawn from the same parent population as that of the random pointings on the SpUDS survey field is 1 2 . × 10 -10 , while the probability raises to 0.69 when comparing with CARLA. The agreement with the CARLA survey fields is unexpectedly good, and leads to the speculation of whether Hot DOGs could be the precursors of RLAGN. In a follow-up work we will study in depth the radio properties of Hot DOGs,
Figure 11. The solid red histogram shows the fraction of Hot DOGs with a given number density of objects brighter than 10 µ Jy at [4.5] and with [3.6]-[4.5] > 0.37 within 1 . ′ We only consider Hot DOGs with IRAC observations and a reliable spectroscopic redshift. For comparison, we also show the distribution for all objects in the CARLA survey ( dotted blue histogram ) and for a sample of randomly selected positions in the SpUDS survey ( dashed green histogram ). The Hot DOG fields are overly dense compared to the random field, and similar to the CARLA fields.
and explore this suggestion further (Tsai et al. in prep.). As discussed earlier, one of the key assumptions of our SED analysis is that little to no obscuration is present in Hot DOG host galaxies. Under significant obscuration, Hot DOG stellar masses could be considerably larger and affect the interpretation of the results presented in § 4. We note, however, that the stellar masses estimated by De Breuck et al. (2010) for RLAGN in CARLA clusters, likely the most massive members of the respective clusters, are not significantly above our upper bound estimates for Hot DOGs. If the upper bounds were underestimated due to unrecognized stellar obscuration, Hot DOGs would be expected to live in considerably denser environments.
To illustrate this point, we note that a study of the luminosity function of clusters in CARLA, has shown that at redshifts between 2 6 . < z < 3, their [4.5] luminosity function is consistent with a Schechter function with best-fit values of α = -1 28 . +0 15 . -0 10 . and m ∗ [4 5] . = 19 59 . +0 25 . -0 25 . (Wylezalek et al. 2014). Using this luminosity function, we can estimate a lower bound on the expected number of galaxies brighter than 10 µ Jy in a cluster containing a given Hot DOG by assuming the Hot DOG is the brightest cluster galaxy. When we consider all Hot DOGs with IRAC imaging, we find a marginal agreement between our predicted field densities and observed density distribution of the Hot DOG fields, with a K-S probability of 0.19 that both are drawn from the same parent population. Yet, if we assume that fluxes (hence stellar masses) are underestimated by a factor as small as 1.5, the K-S
Figure 12. Differential surface density distribution, Σ, of red galaxies surrounding Hot DOGs ( red ) and CARLA RLAGNs ( blue ) as a function of the distance. The black dashed vertical line shows the distance considered for estimating the surface densities shown in Figure 11. The points from the CARLA fields have been offset by 1 ′′ to the left for clarity. Compared to the CARLA fields, the environments around Hot DOGs are significantly less concentrated, despite the fact that the number density of galaxies within 1 ′ are similar. This is consistent with the low angular concentration in Hot DOG fields found by Jones et al. (2014) at 850 µ m.
probability decreases to 6 × 10 -7 .
We conclude from these results that while Hot DOGs may live in dense environments, their field densities are inconsistent with clusters that host galaxies more massive than the upper bound stellar masses we estimated for them in § 3. This reinforces our assumption of little to no host galaxy obscuration and shows that systematically underestimated stellar masses are not the drivers of the results discussed in § 4.1 and § 4.2.
Additionally, we can use the IRAC images to look at the concentration of galaxies in the vicinity of Hot DOGs. To do this we measure the angular surface density, Σ, of red galaxies (again selected as objects with [3.6]-[4.5] > 0.37 and f [4 5] . > 10 µ Jy), in the fields of Hot DOGs as a function of the distance to the given Hot DOG. The results are shown in Figure 12. For comparison, we repeat the process in the CARLA fields, centered on the RLAGN targeted by the survey. We note the redshift distribution of CARLA RLAGN is quite similar to that of Hot DOGs. Compared to the CARLA fields, the environments around Hot DOGs are significantly less concentrated, despite the fact that the number density of galaxies within 1 ′ are similar. Such a difference could imply that Hot DOGs live in very dense filaments, or possibly in clusters that are in an earlier state of virialization. Alternatively, such results could be interpreted as Hot DOGs living in clusters as dense or denser than those found by CARLA, but that the Hot DOGs are not the central galaxy of the cluster. This is, however, unlikely, in light of the argument presented earlier based on the cluster luminosity function. We note this low angular concentration is consistent with the results of Jones et al. (2014) at 850 µ m within a similar radial distance of Hot DOGs.
## 7. CONCLUSIONS
We have presented a detailed study of the observed near-IR through mid-IR SEDs of a large set of Hot DOGs identified by WISE, focusing on the subsample with W4 < 7.2 (the core sample). Using the SED templates of Assef et al. (2010), we find that Hot DOGs are generally well fit by a combination of a luminous and obscured AGN that dominates the emission at rest-frame wavelengths λ /greaterorsimilar 1 µ m, and a host galaxy that dominates the bluer emission. The AGN in Hot DOGs are among the most luminous AGN known and dominate the bolometric luminosity of these objects, accounting for > 97% of the total 0 1 . -30 µ m output in all objects.
Using these SED models, we find that the AGN in Hot DOGs display a large range of obscurations, with 2 5 . < E B ( -V ) < 21 5, and a mean of . 〈 E B ( -V ) 〉 = 6 8 . (6 4) . in the core (general) sample. Using the median dust-to-gas ratio in AGN of Maiolino et al. (2001), E B ( -V ) /N H = 1 5 . × 10 -23 cm 2 mag, these obscurations correspond to gas column densities of 1 7 . × 10 23 < N H < 1 4 . × 10 24 cm -2 , or over 10 times the 10 22 cm -2 column used to separate Type 1 and Type 2 AGN, reaching into the Compton-thick regime. While significant host galaxy obscuration is unlikely, such obscuration would make us underestimate the dust absorption towards the accretion disk.
We estimate upper bounds on the stellar mass of Hot DOGs using the rest-frame K -band luminosities of the modeled host-galaxy component. These range from 11 < log M /M ∗ /circledot < 12 5, implying Hot DOGs could be . some of the most massive galaxies at their redshifts. It is unlikely these upper bounds are underestimated, because the environmental densities of Hot DOGs are inconsistent with those needed to host more massive galaxies at their redshifts (see § 6).
We investigated two approaches to estimate M BH in Hot DOGs. If we assume the AGN in Hot DOGs radiate at the same λ E of similar redshift QSOs, then Hot DOGs must deviate significantly from the local M Sph -M BH relation. Such deviations are also observed for high redshift QSOs, and imply that the SMBH is assembled considerably before the stellar mass, constraining our current galaxy evolution models. Alternatively, we can estimate M BH by assuming Hot DOGs follow the local M Sph -M BH relation, and derive a minimum λ E for these objects. If this is the case, AGN in Hot DOGs must be radiating at significantly super-Eddington ratios. This could imply that Hot DOGs could be at the transition point where the AGN is possibly expelling gas from the galaxy and quenching their star-formation, a scenario that may be consistent with the high fraction of extended Ly α emission found by Bridge et al. (2013).
We show in § 5 that although very rare, the number density of Hot DOGs is comparable to that of equally luminous Type 1 AGN in the redshift range 2 < z < 4. This suggests that Hot DOGs may not be the torus-obscured counterpart of the equally luminous Type 1 AGN, as the dust torus obscuration fraction is expected to be exceedingly small at these luminosities (e.g., Lawrence 1991; Simpson 2005; Assef et al. 2013). Considering the large Eddington ratios we estimate for these objects, we speculate that Hot DOGs may host AGN with significantly less massive SMBHs than the equally luminous Type 1 AGN, but that they are going through a phase where their accretion rates are temporarily enhanced, possibly to the point of quenching star-formation. In this scenario, when radiating at
their 'normal' Eddington ratios, the AGN in Hot DOGs would constitute a small fraction of the Type 1 AGN with comparable mass SMBHs.
Finally, in § 6 we study the environments of Hot DOGs using follow-up IRAC imaging. We show that the number of galaxies within a 1 ′ radius is significantly above the number observed in random pointings implying Hot DOGs live in dense environments. Furthermore, we show that the environments are as dense as those of the clusters identified by the CARLA survey.
Further constraints on the host galaxy properties will allow us to better place these objects in the galaxy evolution context. The physical properties we have been able to determine here highlight how unusual Hot DOGs are among the general galaxy population or even other previously identified extreme populations (e.g., SMGs), and suggest that they may represent a pivotal transition in the galaxy evolution paradigm.
We are indebted to all WISE team members. We thank the anonymous referee for comments and suggestions that helped to improve this article. RJA was supported by Gemini-CONICYT grant number 32120009. This publication makes use of data products from the Widefield Infrared Survey Explorer, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration. This work is based in part on observations made with the Spitzer Space Telescope , which is operated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with NASA. Kitt Peak National Observatory and Cerro Tololo InterAmerican Observatory, National Optical Astronomy Observatory, are operated by the Association of Universities for Research in Astronomy (AURA) under cooperative agreement with the National Science Foundation. The WIYN Observatory is a joint facility of the University of Wisconsin-Madison, Indiana University, Yale University, and the National Optical Astronomy Observatory. The SOAR Telescope is a joint project of: Conselho Nacional de Pesquisas Cient´ ıficas e Tecnol´gicas CNPqo Brazil, The University of North Carolina at Chapel Hill, Michigan State University, and the National Optical Astronomy Observatory. Based partly on observations obtained at the Hale Telescope, Palomar Observatory as part of a continuing collaboration between the California Institute of Technology, NASA/JPL, NOAO, Oxford University, Stony Brook University, and the National Astronomical Observatories of China.
Facilities: WISE, Spitzer (IRAC), Hale (WIRC), MMT (SWIRC), SOAR (OSIRIS), WIYN (WHIRC)
## REFERENCES
Assef, R. J., et al. 2010, ApJ, 713, 970
- -. 2011, ApJ, 728, 56 -. 2013, ApJ, 772, 26 Baldry, I. K., Glazebrook, K., & Driver, S. P. 2008, MNRAS, 388, 945 Banerji, M., McMahon, R. G., Hewett, P. C., et al. 2012, MNRAS, 427, 2275 Bennert, V. N., Auger, M. W., Treu, T., Woo, J.-H., & Malkan, M. A. 2011a, ApJ, 726, 59 -. 2011b, ApJ, 742, 107
- Bentz, M. C., Peterson, B. M., Pogge, R. W., & Vestergaard, M. 2009, ApJ, 694, L166
- Blain, A. W., Smail, I., Ivison, R. J., Kneib, J.-P., & Frayer,
- D. T. 2002, Phys. Rep., 369, 111
- Bongiorno, A., Maiolino, R., Brusa, M., et al. 2014, MNRAS, 443, 2077
- Borys, C., Smail, I., Chapman, S. C., Blain, A. W., Alexander,
- D. M., & Ivison, R. J. 2005, ApJ, 635, 853
- Bridge, C. R., et al. 2013, ApJ, 769, 91
- Brodwin, M., et al. 2011, ApJ, 732, 33
- Brown, W. R., McLeod, B. A., Geary, J. C., & Bowsher, E. C. 2008, Proc. SPIE, 7014,
- Brown, M. J. I., Moustakas, J., Smith, J.-D. T., et al. 2014, ApJS, 212, 18
- Bruzual, G., & Charlot, S. 2003, MNRAS, 344, 1000
- Cardelli, J. A., Clayton, G. C., & Mathis, J. S. 1989, ApJ, 345, 245
- Casey, C. M., Narayanan, D., & Cooray, A. 2014, Phys. Rep., accepted (arXiv:1402.1456)
- Chabrier, G. 2003, PASP, 115, 763
- Conroy, C., Dutton, A. A., Graves, G. J., Mendel, J. T., & van Dokkum, P. G. 2013, ApJ, 776, L26
- Conroy, C., & Gunn, J. E. 2010, ApJ, 712, 833
- Conroy, C., Gunn, J. E., & White, M. 2009, ApJ, 699, 486
- Coppin, K. E. K., et al. 2008, MNRAS, 389, 45
- Croton, D. J., et al. 2006, MNRAS, 365, 11
- Depoy, D. L., Atwood, B., Byard, P. L., Frogel, J., & O'Brien,
T. P. 1993, in Society of Photo-Optical Instrumentation
Engineers (SPIE) Conference Series, Vol. 1946, Infrared
Detectors and Instrumentation, ed. A. M. Fowler, 667-672
De Breuck, C., Seymour, N., Stern, D., et al. 2010, ApJ, 725, 36
Dey, A., et al. 2008, ApJ, 677, 943
- Eisenhardt, P. R. M., et al. 2012, ApJ, 755, 173
- -. in prep.
- Elvis, M., et al. 1994, ApJS, 95, 1
- Faber, S. M., et al. 2007, ApJ, 665, 265
- Fan, X. 1999, AJ, 117, 2528
- Fan, X., et al. 2001, AJ, 121, 54
- Fazio, G. G., et al. 2004, ApJS, 154, 10
- Ferrarese, L., & Merritt, D. 2000, ApJ, 539, L9
- Galametz, A., Stern, D., Stanford, S. A., De Breuck, C., Vernet,
J., Griffith, R. L., & Harrison, F. A. 2010, A&A, 516, A101
- Galametz, A., et al. 2013, A&A, 559, A2
- Gordon, K. D., & Clayton, G. C. 1998, ApJ, 500, 816
- Griffith, R. L., et al. 2012, AJ, 144, 148
- G¨ ultekin, K., et al. 2009, ApJ, 698, 198
- Hasinger, G. 2004, Nuclear Physics B Proceedings Supplements, 132, 86
- H¨ onig, S. F., Leipski, C., Antonucci, R., & Haas, M. 2011, ApJ, 736, 26
- Hopkins, P. F., Hernquist, L., Cox, T. J., & Kereˇ, D. 2008, s ApJS, 175, 356
- Jahnke, K., & Macci`, A. V. 2011, ApJ, 734, 92 o
- Jarrett, T. H., et al. 2011, ApJ, 735, 112
- Jones, S., et al. in prep.
- Jones, S. F., et al. 2014, MNRAS, 443, 146
- Kaspi, S., Smith, P. S., Netzer, H., et al. 2000, ApJ, 533, 631
- Kochanek, C. S., Khan, R., & Dai, X. 2012, ApJ, 759, 20
- Kollmeier, J. A., et al. 2006, ApJ, 648, 128
Lacy, M., Ridgway, S. E., Gates, E. L., et al. 2013, ApJS, 208, 24
Lake, S. E., Wright, E. L., Cutri, R. M., & Eisenhardt, P. R. 2013, in American Astronomical Society Meeting Abstracts, Vol. 221,
American Astronomical Society Meeting Abstracts, #443.09
- Lawrence, A. 1991, MNRAS, 252, 586
- Lawrence, A., & Elvis, M. 2010, ApJ, 714, 561
- Lonsdale, C., et al. in prep.
- Lusso, E., Hennawi, J. F., Comastri, A., et al. 2013, ApJ, 777, 86
- Magorrian, J., et al. 1998, AJ, 115, 2285
- Mainzer, A., et al. 2011, ApJ, 731, 53
- Maiolino, R., Marconi, A., Salvati, M., Risaliti, G., Severgnini, P.,
- Oliva, E., La Franca, F., & Vanzi, L. 2001, A&A, 365, 28
- Mancone, C. L., & Gonzalez, A. H. 2012, PASP, 124, 606
- Maraston, C. 2005, MNRAS, 362, 799
- Marconi, A., & Hunt, L. K. 2003, ApJ, 589, L21
- McGreer, I. D., et al. 2013, ApJ, 768, 105
- Meixner, M., et al. 2010, PASP, 122, 451
- Peng, C. Y. 2007, ApJ, 671, 1098
Richards, G. T., et al. 2006, AJ, 131, 2766
- Roseboom, I. G., Lawrence, A., Elvis, M., et al. 2013, MNRAS,
- 429, 1494
- Ross, N. P., et al. 2013, ApJ, 773, 14
- Sanders, D. B., & Mirabel, I. F. 1996, ARA&A, 34, 749
- Schmidt, M. 1968, ApJ, 151, 393
- Shen, Y., Greene, J. E., Strauss, M. A., Richards, G. T., &
Schneider, D. P. 2008, ApJ, 680, 169
- Simpson, C. 2005, MNRAS, 360, 565
- Skrutskie, M. F., et al. 2006, AJ, 131, 1163
- Stern, D., et al. 2014, ApJ, 794, 102
- Thompson, T. A., Quataert, E., & Murray, N. 2005, ApJ, 630, 167
- Trainor, R., & Steidel, C. C. 2013, ApJ, 775, L3
- Tremaine, S., et al. 2002, ApJ, 574, 740
- Tsai, C.-W., et al. 2014, ApJ, submitted (arXiv:1410.1751)
- Tsai, C.-W., et al. in prep.
- Ueda, Y., Akiyama, M., Ohta, K., & Miyaji, T. 2003, ApJ, 598, 886
- Vestergaard, M., & Peterson, B. M. 2006, ApJ, 641, 689
- Wang, J. X., & Jiang, P. 2006, ApJ, 646, L103
- Wang, R., et al. 2010, ApJ, 714, 699
- Werner, M. W., et al. 2004, ApJS, 154, 1
- Wilson, J. C., et al. 2003, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 4841, Instrument Design and Performance for Optical/Infrared Ground-based Telescopes, ed. M. Iye & A. F. M. Moorwood, 451-458
- Wright, E. L., et al. 2010, AJ, 140, 1868
- Wu, J., et al. 2012, ApJ, 756, 96
- -. 2014, ApJ, accepted (arXiv:1405.1147)
- -. in prep.
- Wylezalek, D., et al. 2013, ApJ, 769, 79
- -. 2014, ApJ, 786, 17
York, D. G., et al. 2006, MNRAS, 367, 945
|
10.1088/0004-637X/804/1/27
|
[
"Roberto J. Assef",
"Peter R. M. Eisenhardt",
"Daniel Stern",
"Chao-Wei Tsai",
"Jingwen Wu",
"Dominika Wylezalek",
"Andrew W. Blain",
"Carrie R. Bridge",
"Emilio Donoso",
"Alexandria Gonzales",
"Roger L. Griffith",
"Thomas H. Jarrett"
] |
2014-08-05T20:00:08+00:00
|
2015-02-19T20:04:35+00:00
|
[
"astro-ph.GA"
] |
Half of the Most Luminous Quasars May Be Obscured: Investigating the Nature of WISE-Selected Hot, Dust-Obscured Galaxies
|
The WISE mission has unveiled a rare population of high-redshift ($z=1-4.6$),
dusty, hyper-luminous galaxies, with infrared luminosities $L_{\rm IR} >
10^{13}~L_{\odot}$, and sometimes exceeding $10^{14}~L_{\odot}$. Previous work
has shown that their dust temperatures and overall far-IR SEDs are
significantly hotter than expected for star-formation. We present here an
analysis of the rest-frame optical through mid-IR SEDs for a large sample of
these so-called "Hot, Dust-Obscured Galaxies" (Hot DOGs). We find that the SEDs
of Hot DOGs are generally well modeled by the combination of a luminous, yet
obscured AGN that dominates the rest-frame emission at $\lambda > 1\mu\rm m$
and the bolometric luminosity output, and a less luminous host galaxy that is
responsible for the bulk of the rest optical/UV emission. Even though the
stellar mass of the host galaxies may be as large as
$10^{11}-10^{12}~M_{\odot}$, the AGN emission, with luminosities comparable to
those of the most luminous QSOs known, require that either Hot DOGs have black
hole masses significantly in excess of the local relations, or that they
radiate significantly above the Eddington limit. We show that, while rare, the
number density of Hot DOGs is comparable to that of equally luminous but
unobscured (i.e., Type 1) QSOs. This is inconsistent with the trend of a
diminishing fraction of obscured objects with increasing luminosity found for
less luminous QSOs, possibly indicating a reversal in this relation at high
luminosity, and that Hot DOGs are not the torus-obscured counterparts of the
known optically selected, largely unobscured Hyper-Luminous QSOs. Hot DOGs may
represent a different type of galaxy and thus a new component of the galaxy
evolution paradigm. Finally, we discuss the environments of Hot DOGs and show
that these objects are in regions as dense as those of known high-redshift
proto-clusters.(Abridged)
|
1408.1093v1
|
## Comprehensive Observations of the Bright and Energetic Type Iax SN 2012Z: Interpretation as a Chandrasekhar Mass White Dwarf Explosion /star
M. D. Stritzinger 1 , S. Valenti 2 3 , , P. Hoeflich 4 , E. Baron 5 6 , , M. M. Phillips 7 , F. Taddia 8 , R. J. Foley 9 , E. Y. Hsiao 1 7 , , S. W. Jha 10 , C. McCully 10 , V. Pandya 10 , J. D. Simon 11 , S. Benetti 12 , P. J. Brown 13 , C. R. Burns 11 , A. Campillay , C. 7 Contreras 1 7 , , F. F¨rster o 14 , S. Holmbo 1 , G. H. Marion 15 , N. Morrell 7 , and G. Pignata 16
- 1 Department of Physics and Astronomy, Aarhus University, Ny Munkegade 120, DK-8000 Aarhus C, Denmark (e-mail: [email protected] )
- 2 Las Cumbres Observatory Global Telescope Network, Inc. Santa Barbara, CA 93117, USA
- 3 Department of Physics, University of California, Santa Barbara, Broida Hall, Mail Code 9530, Santa Barbara, CA 93106-9530, USA
- 4 Department of Physics, Florida State University, Tallahassee, FL 32306, USA
- 5 Homer L. Dodge Department of Physics and Astronomy, University of Oklahoma, 440 W. Brooks, Rm 100, Norman, OK 730192061, USA
- 6 Hamburger Sternwarte, Gojenbergsweg 112, 21029 Hamburg, Germany
- 7 Carnegie Observatories, Las Campanas Observatory, Casilla 601, La Serena, Chile
- 8 The Oskar Klein Centre, Department of Astronomy, Stockholm University, AlbaNova, 10691 Stockholm, Sweden
- 9 Astronomy Department, University of Illinois at Urbana-Champaign, 1002 W. Green Street, Urbana, IL 61801, USA
- 10 Department of Physics and Astronomy, Rutgers, The State University of New Jersey, Piscataway, NJ 08854, USA
- 11 Observatories of the Carnegie Institution for Science, 813 Santa Barbara St., Pasadena, CA 91101, USA
- 12 INAF-Osservatorio Astronomico di Padova, vicolo dell Osservatorio 5, 35122 Padova, Italy
- 13 George P. and Cynthia Woods Mitchell Institute for Fundamental Physics & Astronomy, Texas A. & M. University, Department of Physics, 4242 TAMU, College Station, TX 77843
- 14 Center for Mathematical Modelling, Universidad de Chile, Avenida Blanco Encalada 2120 Piso 7, Santiago, Chile
- 15 University of Texas at Austin, 1 University Station C1400, Austin, TX 78712-0259, USA
- 16 Departamento Ciencias Fosicas, Universidad Andres Bello, Av. Republica 252, Santiago, Chile
Received XX XXXX 2014 / Accepted XX XXXX 2014
## Abstract
We present ultraviolet through near-infrared (NIR) broad-band photometry, and visual-wavelength and NIR spectroscopy of Type Iax supernova (SN) 2012Z. The data set consists of both early and late-time observations, including the first late phase NIR spectrum obtained for a spectroscopically classified SN Iax. A detailed comparison is made to the well-observed Type Iax SN 2005hk, revealing a striking resemblance between the two objects. Simple model calculations of its bolometric light curve suggest SN 2012Z produced ∼ 0.3 M /circledot of 56 Ni, ejected about a Chandrasekhar mass of material, and had an explosion energy of ∼ 10 51 erg, making it one of the brightest ( M B = -18 3 . mag) and most energetic SN Iax yet observed. Early and late phase visual-wavelength spectra of SN 2012Z are found to resemble similar epoch spectra of SN 2005hk, with the main difference being the former exhibiting broader absorption features, which we attribute to variations in the distribution of 56 Ni . The late phase ( + 269d) NIR spectrum of SN 2012Z is found to broadly resemble similar epoch spectra of normal SNe Ia; however, like other SNe Iax, corresponding visual-wavelength spectra differ substantially compared to all supernova types. Constraints from the distribution of intermediate mass elements, e.g. silicon and magnesium, indicate that the outer ejecta did not experience significant mixing during or after burning, and the late phase NIR line profiles suggests most of the 56 Ni is produced during high density burning. The various observational properties of SN 2012Z are found to be consistent with the theoretical expectations of a Chandrasekhar mass white dwarf progenitor that experiences a pulsational delayed detonation, which produced several tenths of a solar mass of 56 Ni during the deflagration burning phase and little (or no) 56 Ni during the detonation phase. Within this scenario only a moderate amount of Rayleigh-Taylor mixing occurs both during the deflagration and fallback phase of the pulsation, and the layered structure of the intermediate mass elements is a product of the subsequent denotation phase. The fact that the SNe Iax population does not follow a tight brightness-decline relation similar to SNe Ia can then be understood in the framework of variable amounts of mixing during pulsational rebound and variable amounts of 56 Ni production during the early subsonic phase of expansion.
Key words. supernovae: general - supernovae: individual: SN 2005hk; supernovae: individual: SN 2012Z
/star Based on observations collected at the European Organization for Astronomical Research in the Southern Hemisphere, Chile (ESO Program 088.D-0222, 184.D-1152), the Magellan 6.5 m telescopes at Las Campanas Observatory, and the Nordic Optical Telescope, op-
erated by the Nordic Optical Telescope Scientific Association at the Observatorio del Roque de los Muchachos, La Palma, Spain, of the Instituto de Astrofisica de Canarias. Also partially based on observa-
## 1. Introduction
Over recent years dedicated non-targeted and targeted transient search programs have revealed the existence of a multitude of peculiar transients lurking in the nearby Universe. A particular group of objects that share a number of commonalities, warranting their own spectroscopic designation are Type Iax supernovae (see Foley et al. 2013, for a review). Supernova (SN) 2002cx serves as the prototypical example, of what was originally entitled by Li et al. (2003) as, 'The Most Peculiar Known Type Ia Supernova.' Since the discovery of SN 2002cx, it has become clear that they are not all that rare. They have been estimated to account for 5 -30 % of the overall SNe Ia rate (Li et al. 2011; Foley et al. 2013; White et al. 2014). SNe Iax are characterized by a number of observational peculiarities, including, among others: a significant range in peak luminosity ( M B ∼ -14 to -18 mag); spectral signatures characteristic of over luminous 1991T-like SNe Ia such as blue continua at early phases and iron lines observed at all epochs; expansion velocities that are about half (4000 to 9000 km s -1 ) of what is observed in normal SNe Ia (10,000 to 15,000 km s -1 ) at similar epochs; and late phase optical spectra which bear little to no resemblance to other types of supernovae.
Detailed studies of the light curves and spectra of wellobserved SNe Iax indicate that they generate between ∼ 0.0030.30 M /circledot of 56 Ni , and have explosion energies ranging from ∼ 10 49 -10 51 erg. The progenitors of the entire class of SNe Iax are unknown and are a matter of open debate. Recent efforts to model the bright SN Iax (e.g. SN 2005hk) suggest that their origins may lie with the partial thermonuclear incineration of a carbon-oxygen Chandrasekhar mass white dwarf (WD), which leaves a bound remnant (Jordan et al. 2012; Kromer et al. 2013; Fink et al. 2013). However, pure deflagration models are currently unable to account for the stratified distribution of intermediate mass elements (IMEs) that are inferred from the spectroscopic observations of some bright SNe Iax. All current deflagration models produce significant mixing throughout the explosion (Livne & Arnett 1993; Livne 1993; Khokhlov 1995; Gamezo et al. 2003; Reinecke, Hillebrandt & Niemeyer 2002; Gamezo, Khokhlov, Oran 2005; R¨ opke, Woosley & Hillebrandt 2007; Seitenzahl et al. 2009, 2013), and it is unclear whether additional physics can sufficiently suppress mixing in deflagrations (Hoeflich et al. 2013). This leaves open room for other possible progenitor scenarios, such as for example, the partial thermonuclear disruption of helium accreting sub -Chandrasekhar mass white dwarfs (Wang, Justham & Han 2013) or mergers (Piersanti et al. 2003). In this paper we discuss an alternative scenario that explains most of the features of the recently observed Type Iax SN 2012Z (Foley et al. 2013), as well as some other members of the SNe Iax class. However, it is important to keep in mind that SNe Iax are not a homogeneous class and several explosion scenarios may contribute to the entire observed population.
Recently we studied SNe 2008ha and 2010ae, which are the least luminous Type Iax objects yet observed (Stritzinger et al. 2014). Here, our attention is turned towards the bright end of the luminosity distribution of SNe Iax. Specifically we present ultraviolet (UV), optical and near-infrared (NIR) observations of SN 2012Z. The data reveal that this object bears a striking resemblance to the well-observed Type Iax SN 2005hk (Chornock et al. 2006; Phillips et al. 2007; Sahu et al. 2008;
tions made with the Southern African Large Telescope (SALT), and the W. M. Keck Observatory located on the summit of Mauna Kea.
Maund et al. 2010; McCully et al. 2014a), and further indicates that SN 2012Z is among the brightest and most energetic SN Iax yet observed. The majority of the data presented in this study were obtained as part of the Carnegie Supernova Project-II (CSP-II), which is a four year NSF-funded SN followup program aimed at obtaining optical and NIR observations of ∼ 100 SNe Ia located in the Hubble flow. The CSP-II is placing a particular emphasis on carrying out detailed NIR spectroscopy, and as a result, this paper presents the most comprehensive NIR time series of a SN Iax. We also note that revised optical and NIR photometry of SN 2005hk are presented in the Appendix. Since these data were first published by Phillips et al. (2007), the images of SN 2005hk have undergone proper host-galaxy template subtraction, and as well, the calibration of the local sequence has been improved.
In a companion paper, we report on the detection of a luminous and blue point source coincident with the location of SN 2012Z in pre-explosion Hubble Space Telescope (HST) images (McCully et al. 2014b). This direct detection is consistent with a helium-star donating material to an accreting white dwarf as the progenitor for SN 2012Z. Followup HST observations have the potential to determine if the progenitor system was correctly identified and consistent with such a model.
The observations of SN 2012Z presented here provide unique coverage of the light curve and time evolution of the spectra at both optical and NIR wavelengths for a SN Iax. In supernovae, we are able to peer into deeper layers of the ejecta over time, and therefore, our observations allow us to probe the entire envelope structure. The nature of SNe Iax is under discussion, as are many aspects of the explosion physics and the nature of the variations observed within SNe Ia. A comprehensive discussion of the latter point is beyond the scope of this paper and therefore, we refer the reader to Hoeflich (2006), Hoeflich et al. (2013), and references therein. In what will follow, we discuss the various constraints on the underlying progenitor provided by individual observational components, as well as compare SN 2012Z to other SNe Ia and SNe Iax. As a reference scenario to compare SN 2012Z with normal SNe Ia, we adopt one of a series of delayed-detonation models of the later because it is readily available, fits many observational properties of SNe Ia, and can be regarded as a limiting case for one of the models discussed below. The goal of this work is not to rule out a specific model, but rather to piece together the arguments within all scenarios discussed in the literature. By considering the individual observational constraints of SN 2012Z and confronting them with various scenarios, we suggest a unified picture that may explain the nature of at least some SNe Iax.
## 1.1. Supernova 2012Z
SN 2012Z was discovered with the KAIT telescope as a part of LOSS (see Filippenko et al. 2001; Filippenko 2005) on 2012 Jan. 29.15 UT, at an apparent unfiltered magnitude of 17.6 (Cenko et al. 2012). With J2000.0 coordinates of α = 03 22 h m 05 35 and . s δ = -15 23 15 6, the position of this super-◦ ′ . ′′ nova is 14 . ′′ 6 West and 42 . ′′ 1 North from the center of the SA(s)bc galaxy NGC 1309. Figure 1 contains an image of the host galaxy with SN 2012Z indicated. Spectroscopic confirmation obtained three days after the discovery indicated that this was a SN Iax on the rise (Cenko et al. 2012).
NEDlists the redshift of NGC 1309 as cz = 2136 km s -1 or z = 0 007125 . . NGC 1309 has a number of Cepheid-based distance measurements, and as well has hosted the normal Type Ia SN 2002fk (Riess et al. 2009; Cartier et al. 2014). A recent re-
calibration of the distance scale by Riess et al. (2011), provides an up-to-date Cepheid based distance to NGC 1309 of 33.0 ± 1.4 Mpc or µ = 32 59 . ± 0.09. Given the accuracy of this distance measurement, it is adopted in this paper to set the absolute flux scale of SN 2012Z.
## 2. Observations
## 2.1. Ultraviolet, optical, and NIR imaging
SN 2012Z was observed from space with the UVOT camera aboard the Swift X-ray telescope (Burrows et al. 2005; Roming et al. 2005). Up to 13 epochs of UV uvw 2 , uvm 2 , and uvw 1 -band imaging was obtained over the duration of a month. Template images of the host obtained during early 2013 were used to estimate the background emission at the position of the SN. Aperture photometry of the SN was then computed as described by Brown et al. (2009), using the calibration zero-points and sensitivity corrections from Breeveld et al. (2011). The final UV photometry is listed in Table 1.
Optical ( ugriBV ) imaging of SN 2012Z was initiated near the end of the first campaign of the CSP-II with the Swope 1.0 m ( + SITe3 direct CCD camera) telescope at the Las Campanas Observatory (LCO). In addition, NIR ( Y JH ) imaging was also obtained with the du Pont 2.5 m telescope ( + RetroCam). All CSP-II imaging was processed in the standard manner, following the procedures described by Hamuy et al. (2006) and Contreras et al. (2010).
Before computing photometry of the SN, galaxy subtraction was performed on all science images. High signal-to-noise optical and NIR template images of NGC 1309 were obtained under favorable seeing conditions with the du Pont telescope between the 10 th and the 17 th of October 2013. Multiple exposures were stacked to create master templates, which were then subtracted from the science images following the techniques described by Contreras et al. (2010).
PSF Photometry of SN 2012Z was computed differentially with respect to a local sequence of stars in the field of NGC 1309. The optical sequence consists of 17 stars calibrated with respect to Landolt (1992) ( BV ) and Smith et al. (2002) ( ugri ) photometric standard fields, which were observed over the course of multiple photometric nights. The NIR local sequence, on the other hand, contains 6 stars calibrated relative to Persson et al. (1998) standards also observed over multiple photometric nights. As the Persson et al. (1998) catalog of standards does not include Y -band photometry, the Y -band local sequence was calibrated relative to unpublished Y -band photometry of a subset of the Persson et al. standards. The standard star Y -band photometry was computed from multiple observations conducted at LCO as part of the CSP-I (Hamuy et al. 2006). Final absolute photometry of the optical and NIR local sequence in the natural system is listed in Table 2. The quoted uncertainty for each magnitude corresponds to the weighted average of the instrumental errors of the photometry during the nights photometric standard fields were observed.
With photometry of the local sequences and template subtracted science images in hand, PSF photometry of SN 2012Z was computed as prescribed by Contreras et al. (2010). The definitive optical and NIR photometry in the natural photometric system of the Swope and du Pont telescopes is listed in Tables 3 and 4, respectively. The uncertainties accompanying each photometric measurement correspond to the sum in quadrature of the instrumental error and the nightly zero-point error. The early phase UV, optical, and NIR light curves of SN 2012Z are plotted relative to the time of B -band maximum (hereafter T B ( ) max ) in Figure 2. We note that a single epoch of BVi -band optical photometry was also obtained at late phases. This epoch of data is used below to provide a lower limit on the UVOIR flux at + 266d (see below).
## 2.2. Optical and NIR spectroscopy
Five epochs of early phase ( + 2d to + 21d), low-resolution visual-wavelength spectroscopy of SN 2012Z was obtained by the CSP-II with the Nordic Optical Telescope (NOT + Alfosc), and the Magellan Clay telescope ( + LDSS3). In addition, a single high-resolution optical spectrum was obtained on -7d with the Magellan Inamori Kyocera Echelle (MIKE; Bernstein et al. 2003) spectrograph attached to the Magellan Clay telescope. All low-resolution spectra were reduced in the standard manner using IRAF 1 scripts, while the high-resolution spectrum was reduced using the MIKE pipeline (Kelson 2003) following the methods described by Simon et al. (2010). Complementing the early spectroscopy are 3 epochs of late phase ( + 194d to + 248d), low-resolution spectra, which were obtained with SALT ( + RSS), the du Pont ( + WFCCD) and Keck ( + DEIMOS) telescopes.
A summary of our spectroscopic observations and those of Foley et al. (2013) is provided in the spectroscopic journal given in Table 5. As seen in Figure 3, our early phase spectra nicely complement the previously published spectra of SN 2012Z. The three late phase visual-wavelength spectra are plotted in Figure 4.
Ten epochs of NIR spectroscopy were also obtained for SN 2012Z, and details concerning these observations are provided in Table 5. This includes 5 epochs taken with the Magellan Baade telescope ( + FIRE), 4 epochs with the Very Large Telescope (VLT + ISAAC), and finally, one epoch with the New Technology Telescope (NTT + Sofi). The FIRE spectra were reduced as described by Hsiao et al. (2013), while the VLT and NTT spectra were reduced as described by Smartt et al. (2013). The NIR spectroscopic sequence plotted in Figure 5 extends from -7d to + 269d, and consists of the most comprehensive NIR time-series yet obtained for a SN Iax, including the first late phase NIR spectrum.
## 3. Dust extinction
The Milky Way color excess in the direction of NGC 1309 is listed in NED to be E B ( -V ) MW = 0 035 . ± 0 012 . mag (Schlafly & Finkbeiner 2011), which when adopting a Fitzpatrick (1999) reddening law characterized by an R V = 3 1 . , corresponds to A V = 0 11 . mag.
Estimating the level of reddening that affects the emission of supernovae due to dust external to the Milky Way (MW) is a challenging prospect, particularly for peculiar types that do not have well-defined intrinsic colors. One must therefore find alternative methods to estimate host reddening. In the case of SN 2012Z, we first examine the host-galaxy color-excess value inferred from the broad-band optical and NIR light curves of the normal Type Ia SN 2002fk, which occurred in an inner region of NGC 1309. From a detailed analysis of its broad-band colors,
1 The image Reduction and Analysis Facility (IRAF) is maintained and distributed by the Association of Universities for Research in Astronomy, under the cooperative agreement with the National Science Foundation.
Cartier et al. (2014) estimate the color excess along the line-ofsight of SN 2002fk to be E B ( -V ) host = 0 057 . ± 0 036 . mag. As SN 2012Z is located in the outskirts of NGC 1309 one may expect that it suffered a similar level or even less reddening.
Using the high-resolution MIKE spectrum, we next explored if a more direct handle on the host reddening could be obtained via column density measurements of neutral sodium and potassium inferred from host absorption components of the Na i D λλ 5890, 5896 and K i λλ 7665, 7699 doublets. Line fits computed using the Voigt profile fitting program VPFIT provided column densities of log 10 N Na i = 12 79 . ± 0 09 . cm -2 and log 10 N K i = 11 30 . ± 0 09 . cm -2 . These values give a N Na i to N K i ratio that is perfectly normal for the MW (see Phillips et al. 2013, Figure 13). Adopting the MW relations between log 10 N Na i and log 10 N K i vs. A V , as derived by Phillips et al. (2013, see their Equations 4 and 5), implies visual extinction values of A V (Na ) = 0 45 i . ± 0 24 . mag and A V (K ) = 0 50 i . ± 0 36 . mag, respectively.
Phillips et al. (2013) also recently showed that the 5780 A ˚ diffuse interstellar band (DIB) can provide a robust means to estimate host reddening. Under close inspection of the highresolution spectrum of SN 2012Z this feature is not detectable. However from the signal-to-noise ratio of the MIKE spectrum at the expected position of this feature, a 3 σ upper limit on the equivalent width of EW(5780) ≤ 35 m ˚ A is estimated. Adopting the Phillips et al. (2013, see their Equation 6) MW relation that connects log 10 EW (5780) to A V then implies an A V ≤ 0 18 . ± 0 09 . mag. This value and those derived from Na i and K i doublets are generally consistent, and when combined via a weighted average imply an A V = 0 23 . ± 0 08 . mag, which when adopting an R V = 3 1 . , corresponds to a host galaxy color excess E B ( -V ) host = 0 07 . ± 0 03 . mag. Interestingly, this value is fully consistent with what is derived from the broad-band colors of SN 2002fk. In what follows we adopt this value for the host reddening of SN 2012Z, and when combining with the MW component, gives the total color excess E B ( -V ) tot = 0 11 . ± 0 03 . mag.
## 4. Results
## 4.1. UV, optical and NIR light curves
The photospheric phase optical light curves of SN 2012Z plotted in Figure 2 consist of dense photometric coverage, while the UV light curves are of considerably lower quality, and the NIR light curves contain only a handful of epochs, beginning well after maximum. The optical light curves reveal the characteristic bell-shape appearance that is typical of stripped-envelope supernovae, indicative of a rather low ejecta mass. Similar to other bright SNe Iax, the r - and i -band light curves appear broad and slowly evolving as compared to the bluer bands. In addition, the r - and i -band light curves show no evidence for a secondary maximum, which is a defining characteristic of normal SNe Ia (Hamuy et al. 1996). Interestingly, the uvw 2 - and uvm 2 -band light curve are found to be already declining 8 days before T B ( max ) , while the uvw 1 light curve appear to be at peak brightness. Clearly, the UV light curves reach maximum well before the optical light curves.
Light curve parameters were measured from Gaussian Process functional fits to the optical light curves using software contained within the light-curve fitter SNooPy (Burns et al. 2011). The fit parameters consist of: (i.) the time of maximum, (ii.) peak apparent magnitude and, (iii.) an estimate of the decline-rate parameter, ∆ m 15 2 . These fit parameters are listed in Table 6, along with those measured from the revised photometry of SN 2005hk (see Appendix), accompanied with robust uncertainties derived via Monte Carlo simulations. The light-curve fit parameters indicate that the bluer the light curve the earlier maximum is reached, with a ∆ t ≈ 13 day difference between the time of u - and i - band maxima, and the redder the light curve the smaller its decline-rate. The decline-rate ranges from as much as 1.92 ± 0.08 mag in the u band, all the way down to 0.54 ± 0.04 mag in the i band.
Figure 7 shows the ( B -V ) , ( V -r ) and ( V -i ) color evolution of SN 2012Z compared to that of the Type Iax SN 2005hk and the normal, un-reddened Type Ia SN 2006ax. The color curves of both SNe Iax have been corrected for MW and host reddening. Overall the colors and the color evolution of SNe 2005hk and 2012Z are very similar, with the former appearing marginally bluer. At the start of the observations the color curves of SN 2012Z are at their bluest. They then slowly evolve and reach a maximum value in the red around 20 days past T B ( ) max , at which point an inflection is observed followed by a linear evolution with a negative slope. The color evolution of the two SNe Iax also have a similar overall shape to normal SNe Ia, but with colors that are consistently redder as the supernovae evolve. It will be interesting to see in the future, as more well-observed, bright SNe Iax become available, if this family of objects displays intrinsic colors or color evolution which could allow for an accurate estimation of the host extinction.
## 4.2. Absolute magnitudes, UVOIR light curve, and light curve modeling
With estimates of peak apparent magnitude for the ugriBV bands in hand, peak absolute magnitudes were computed adopting an E B ( -V ) tot = 0 11 . ± 0 03 . mag and the direct distance measurement µ = 32 59 . ± 0 09 . mag. The resulting peak values are listed in Table 6 3 , with accompanying uncertainties that account for both the error in the fit to the time of maximum, and the error associated with the distance to the host galaxy. Reaching an absolute peak B -band magnitude ( M B ) ≈ -18 3 . ± 0.1 with ∆ m 15 ( B ) = 1 43 . ± 0 02 . mag, SN 2012Z is among the brightest and slowest declining SN Iax yet observed. This is demonstrated in Figure 6, where the location of SN 2012Z in the M B vs. ∆ m 15 diagram is shown compared to a handful of other well-observed SNe Iax, and an extended sample of SNe Ia observed by the CSP.
Armed with the broad-band photometry of SNe 2005hk and 2012Z, we proceeded to construct UVOIR bolometric light curves. For SN 2005hk we adopted the early phase photometry presented in Appendix A, as well as the late phase photometry published by Sahu et al. (2008). The early phase filtered light curves of both objects were fit with spline functions, which allowed us to fill in missing gaps in the data based on interpolation. For SN 2012Z the u -band light curve was extended so that its temporal coverage matched that of the other optical filters. To do so, we resorted to extrapolation by adopting an average
2 The light curve decline-rate parameter, ∆ m 15 , is defined as the magnitude change from the time of maximum brightness to 15 days later. For normal SNe Ia, ∆ m 15 correlates with the peak absolute luminosity in such a way that more luminous objects exhibit smaller ∆ m 15 values (Phillips 1993), however this is not the case for SNe Iax (see e.g., Figure 6).
3 To derive absolute magnitudes of SN 2005hk (and absolute luminosity, see Section 4.2), we have adopted a distance modulus of µ = 33 46 . ± 0 27 . mag (Phillips et al. 2007) and the total color excess value E B ( -V ) host = 0 11 . mag.
( u -B ) color derived from the last three epochs in which u -band photometry was obtained.
With the definitive observed light curves in hand, the photometry was corrected for extinction, and then converted to flux at the effective wavelength of each passband. This allowed us to construct SEDs spanning from ∼ 300 to 2500 nm. The full series of SEDs were then summed over wavelength using a trapezoidal integration technique, assuming zero flux beyond the limits of integration. Due to our limited NIR photometric coverage of SN 2012Z it was necessary to account for the NIR contribution of the flux. Given the close resemblance between the peak luminosity, light curve shape (see Section 6), and color evolution between SNe 2005hk and SN 2012Z, we resorted to adopting the fraction of NIR flux estimated from the full optical/NIR spectral energy distribution (SED) of SN 2005hk. The final UVOIR light curves of SNe 2005hk and SN 2012Z are plotted in Figure 8, with the associated error-bars accounting for uncertainties in the photometry, total reddening, and distance. Note that also over plotted in Figure 8 are a handful of flux points of SN 2012Z, which were computed using its optical and limited NIR photometry. These points are in full agreement with the UVOIR flux points of SN 2012Z that contain the contribution of NIR flux estimated from SN 2005hk.
To derive rough estimates of the explosion parameters for these SNe Iax, we turned to toy model fits of the UVOIR light curves based on Arnett's equations (Arnett 1982), following the implementation described by Valenti et al. (2008). The key parameters of interest are the amount of radioactive 56 Ni synthesized during the explosion, the ejecta mass ( M ej ), and the explosion energy ( E K ). The model calculations rely upon a number of underlying assumptions and therefore the results should be approached with caution. With this caveat in mind, the model assumptions include: complete energy deposition supplied from the radioactive decay chain 56 Ni → 56 Co → 56 Fe, spherical symmetry, homologous expansion of the ejecta, no appreciable mixing of 56 Ni, a constant optical opacity (i.e. κ opt = 0.1 cm 2 g -1 ), and the diffusion approximation for photons. A key input parameter in computing the model calculations is a measure of the expansion velocity ( v ph ) of the ejecta. From a detailed inspection of the optical spectral features of SN 2012Z (see Section 4.3), this parameter is estimated to range between 7 000 , /lessorsimilar v ph /lessorsimilar 9 000 , kms -1 . In the case of SN 2005hk, based on velocity measurements presented by Phillips et al. (2007) and McCully et al. (2014a), we infer the range 5 000 , /lessorsimilar v ph /lessorsimilar 7 000 , km s -1 . For both objects a grid of models was computed that encompasses the appropriate range of v ph . Final best-fit model calculations are over-plotted in Figure 8. The range of explosion parameters computed for SN 2005hk are M Ni = 0 15 . -0 20 . M /circledot , M ej = 1 5 . -2 0 . M /circledot and E k = 0 4 . -1 2 . × 10 51 erg; while those of SN 2012Z are M Ni = 0 25 . -0 29 . M /circledot , M ej = 1 4 . -2 6 . M /circledot and E k = 0 7 . -2 8 . × 10 51 erg.
These rough estimates of the explosion parameters are in essence based on Arnett's Rule and the the photospheric expansion velocity, v phot . It should be stressed that a portion of the large range in the estimated parameters is due to the degeneracy between the E k , M Ni and M ej , and the reliance on, in effect, a single bolometric light curve. Furthermore, we note that the estimated values of the various explosion parameters do not capture enough of the physical variations displayed by the various suite of SNe Ia explosion models found in the literature. Even for normal SNe Ia, Arnett's Rule is but a useful means of obtaining a qualitative understanding of the relation between the mass of 56 Ni produced and the peak brightness.
In particular, due to the strong temperature dependence of the opacity, opacities vary by two and a half orders of magnitude at each phase of the evolution of the supernova (Hoeflich et al. 1992). This opacity variation is required in order to reproduce the brightness decline-rate relation (Hoeflich et al. 1996; Nugent et al. 1995; Pinto & Eastman 2001; Maeda et al. 2003; Baron et al. 2012). Attempts to account for both gamma-rays escape and energy stored within the expanding ejecta due to variations in the diffusion time and non-constant opacities, have led to further parameterization to the standard form of Arnett's Rule. This is normally accomplished by the introduction of the parameter, α , which leads to L max bol ≈ αS t ˙ ( max ) . The parameter α is thought to lie somewhere in the range 0 7 . ∼ < α ∼ < 1 5 . (e.g. Branch 1992; Khokhlov, Mueller & Hoeflich 1993; Hoeflich & Khokhlov 1996). 4 With such a large range in α , attempts to use Arnett's Rule for quantitative analysis will only be accurate to about 50%.
SNe Iax light curves are characterized by a short rise time as compared with normal SNe Ia. The hydrodynamics is determined by the total ejected mass and the total energy produced by thermonuclear fusion; however, the light curves are determined not by the total mass, but rather by the optical depths, that is the product of the column density and the opacities. As the opacities vary by several orders of magnitude over time as well as varying with abundances and ionization stage, a change of the opacity by a factor of two is equivalent to a variation in the mass by ∼ √ 2 . Choosing a constant opacity to fit with the risetime is equivalent to fixing the mass scale and the opacity choice of Arnett's Rule is tuned to the choice of a Chandrasekhar mass.
In Section 5 we attempt to gain an additional handle on the explosion parameters of SN 2012Z through comparison of the observations to modern explosion models.
## 4.3. Early phase spectroscopy
The sequence of early phase visual-wavelength spectra plotted in Figure 3 reveals all of the quintessential features of a SN Iax located at the bright end of the luminosity distribution. At the earliest epochs, SN 2012Z exhibits a 'hot' continuum dotted with prevalent iron features. These characteristics are reminiscent of early spectra of luminous 1991T-like SNe Ia, however, they differ substantially at late phases. The blue portion of the -9d spectrum exhibits clear features produced by Ca ii H & K , Fe ii λ 4555 and Fe iii λλ 4404 5129 , , while red-wards of ∼ 5000 A ˚ the spectrum is relatively smooth, showing only subtle notches at the expected locations of the IMEs S ii λλ 5454 5640 , and Si ii λ 6355 . By maximum, the strengths of these IME features mildly increase, however soon thereafter, their presence is diluted by the emergence of numerous permitted lines of Fe ii and Co ii (see Branch et al. 2004, for detailed SYNOW fits to similar epoch spectra of SN 2002cx). Simultaneously, a conspicuous absorption feature attributed to the Ca ii NIR triplet emerges, and subsequently grows in strength over the duration of a fortnight. Finally, we note that the near maximum spectra of SN 2012Z shows a weak feature that has been attributed to C ii λ 6580 (see Foley et al. 2013, for a discussion).
The NIR spectral time-series of SN 2012Z presented in Figure 5 contains both the earliest and latest observed epochs of any SN Iax published to date, and therefore provides unique insight into the SN Iax class. Similar to normal SNe Ia observed at very early times, the earliest epoch ( -7 d) NIR spectrum of
4 In the reference model adopted in this paper α = 1 22 . .
SN 2012Z exhibits a rather blue, featureless continuum. Several days later, just prior to maximum the spectrum reveals a shallow feature at wavelengths expected for Mg ii λ . 1 0952 µ m. There is also weak evidence of C i lines at 0 9093 . , 0 9406 . , 1 0693 . µ m, which have been clearly identified in the subluminous Type Ia SN 1999by (Hoeflich et al. 2002). As the SN ejecta expands, the photosphere recedes into the inner regions, resulting in a reduction of the continuum flux, while line features largely associated with Fe-group elements emerge. The evolution of the spectrum and the emergence of Fe-group features tends to occur on a faster timescale in SNe Iax as compared to normal SNe Ia.
Close inspection of the post maximum spectra reveals that by + 22d, Co ii features dominate the NIR spectrum red-wards of 1.5 µ m. This is expected, owing to the fact that Co ii features are known to be ubiquitous to the NIR spectra of all other SNe Iax observed to date (see Stritzinger et al. 2014). Turning to the last two observed NIR spectra, minimal evolution of the overall line features is observed to occur between + 58d and + 269d. Finally, we note that the NIR spectra of SN 2012Z do not exhibit any compelling evidence for spectral features related to He i (White et al. 2014).
The time-evolution of the blue-shifts of a handful of ions present in the early spectra of SN 2012Z were measured in order to estimate the velocity distribution of the emitting regions of the ejecta. The results of which are shown in Figure 9, and include IME ions of Ca ii H & K , Si ii λ 6355 , C ii λ 6580 and Ca ii λ 8498 , as well as the iron lines Fe ii λ 4555 and Fe iii λ 4404 , λ 5129 . The expansion velocity measurements reveal that these ions are related to material with relatively low velocities, all consistently below 10,000 km s -1 . Interestingly, at + 2d, ions attributed to Ca ii H & K , Fe iii λ 4404 and Si ii λ 6355 exhibit consistent line velocities of ≈ 7,500 km s -1 . This value is bracketed by velocities of Fe ii λ 4555 and Fe iii λ 5129 , which are ∼ 1,000 km s -1 higher and lower, respectively. Inspection of Figure 9 also indicates that at a month past T B ( ) max the line velocities evolve rather slowly, with velocities ranging between 6,000 and 8,000 km s -1 . Taken together the line velocities inferred from the spectroscopic sequence of SN 2012Z suggests a rather low explosion energy compared to normal and overluminous SNe Ia.
To gain a handle on the radial structure of IMEs in the outer layers of the ejecta, we measured the blue wing velocities in addition to the absorption minima. Blue wings are important since they will not appear if the particular ion does not exist at that velocity. In particular we measured the velocities of the Si ii λ 6355 and Mg ii λ . 1 0952 µ m features. Metal lines are not appropriate since even primordial abundances of calcium and beyond can produce background lines. The Si ii blue wing is measured to be v ≈ -11 000 km s , -1 on -8d and v ≈ -10 000 km s , -1 on + 1.8d, while Mg ii is measured with v ≈ -13 500 km s , -1 on -0 4 . d. These measurements along with the flat evolution of features in velocity space suggests a layered distribution of the IMEs in the outer ejecta (Nomoto, Thielemann & Yokoi 1984). Similar abundance structures have been inferred from the spectral analyses of normal SNe Ia (Jeffery & Branch 1990; Hoeflich et al. 1995; Hoeflich & Khokhlov 1996; Baron et al. 2006; Dessart et al. 2014b), and provides robust constraints on the explosion physics (see below).
## 4.4. Late phase spectroscopy
The late-phase visual-wavelength spectra of SN 2012Z (see Figure 4) contains nearly a dozen broad emission features, with the majority of cooling occurring through the most prevalent fea- ture centered around ≈ 7400 ˚ . This dominant emission feature A peaks by a factor of 3 higher in flux than any of the other emission feature, and has a full-width-half-maximum (FWHM) velocity ∼ 10 000 km s , -1 . For comparison the weaker emission features exhibit FWHM values between 5,700 to 9,000 km s -1 . Based on similar epoch spectra of other SNe Iax that have lower ejecta velocities, we attribute the ≈ 7400 ˚ A feature to a blend of a handful of lines, including [Ca ii ] λλ 7291, 7324 and [Fe ii [ λ 7155. Foley et al. (2013) remarked that some SNe Iax exhibit [Ni ii ] λ 7378 and [ Fe ii ] λ 7453; unfortunately in SN 2012Z the significant blending of features in this wavelength range prevents simple line identifications and therefore, it not straightforward to determine if these lines are present. Other notable features in the late phase spectra located at ∼ 5200 A ˚ and 5400 A are likely ˚ produced by the blending of numerous permitted Fe ii lines, which have been identified in other SN Iax (Jha et al. 2006; Foley et al. 2013). The emission feature at ≈ 5900 ˚ is likely formed from A Na i D , while at ≈ 8300 ˚ we A find features which are more than likely formed by the Ca ii NIR triplet. The NIR triplet is commonly seen in SNe Ia and other supernovae, and is strong even at solar abundances. An alternative identification of the 5900 ˚ A feature could be [Co iii ] λ 5888 (Dessart et al. 2014a). Note that we find a blend of [Co iii ] λ 5890 , and contributions of [Co iii ] λ 5908 and [Fe ii ] λ 5872 in our reference model, but much weaker than observed (see Section 5.3).
Beyond these prevalent features, the late phase spectrum of SN2012Zcontain a multitude of low amplitude and low velocity (500-1000 km s -1 ) features. These features appear to be generic to the SNe Iax class, and have been linked to both forbidden and allowed transitions of iron (e.g. Jha et al. 2006; McCully et al. 2014a).
We now turn to the examination of the late phase NIR spectrum of SN 2012Z, which to date, is the first of its kind obtained for a SN Iax. Plotted in Figure 10 is the comparison between our late phase ( + 269d) NIR spectrum of SN 2012Z, and a similar epoch spectrum of the normal Type Ia SN 2012fr. Clearly, SN 2012Z exhibits many of the same broad features as is seen in the normal SN Ia, particularly at the wavelength regions that correspond to the J and H passbands. Overall, the similarity between the two spectra is quite striking considering the discrepancy between the visual-wavelength late phase spectra between SN 2012Z and other SNe Iax in general, compared to normal SNe Ia (see Figure 11). Close inspection of the late phase NIR spectrum, albeit of low signal-to-noise, indicates that the [Fe ii ] λ 1.64 µ m feature has a pot-bellied profile rather than a peaked profile. 5 Similar features have been previously observed in a handful of normal SNe Ia (e.g. Hoeflich et al. 2004; Motohara et al. 2006), and as we discuss below, is a signature of high density burning.
## 5. Bridging observations to model calculations
In the subsequent section, we use the various observational properties of SN 2012Z described in the previous section as a guide
5 There is confusion in community in the nomenclature to describe central line profiles. Line profiles from constant emissivity shells which are truly flat-topped have been called by the same name as those due to a nickel hole where energy input from gamma rays and positrons causes the emissivity to vary. While flat-topped has previously been used in the literature, we want to avoid the confusion between the description of truly flat-topped shell emission that is used to describe optically thin line profiles in a moving shell (Owocki & Cohen 2001) and the profiles we are discussing here, which we refer to as pot-bellied.
to argue that the progenitor is consistent with the theoretical expectations of a pulsating delayed detonation (PPD) explosion of a Chandrasekhar mass ( M Ch) WD. To do so, we rely upon existing explosion models and link together model predictions to the various observables including amongst others: (1) the light curve properties, (2) the layered structure of the ejecta and, (3) the pot-bellied [Fe ii ] 1.64 µ m profile. The various lines of reasoning are then connected and summarized in Section 7 to provide a progenitor scenario to explain at least some SNe Iax.
## 5.1. Interpreting the broad-band emission through model comparison
SNe Ia explosions produce a form of stellar amnesia (Hoeflich 2006) and consequently, most explosion scenarios produce very similar structures, ranging from classical deflagration models like W7 to the inner structures of merger (Nomoto, Thielemann & Yokoi 1984; Benz et al. 1990; Livio 1989; Khokhlov 1991; Yamaoka et al. 1992; Hoeflich & Khokhlov 1996; Lor´ en-Aguilar, Isern & Garc´ ıa-Berro 2009; Gamezo et al. 2 003; Gamezo, Khokhlov, Oran 2005; R¨ opke, Woosley & Hillebrandt 2007; Seitenzahl et al. 2009; Pakmor et al. 2011). The abundance distributions reflect the conditions under which burning takes place and therefore exhibit the imprint of the detailed physics that we discuss below. Delayed detonation models (DDT) give the overall chemical stratification where the burning propagates as a deflagration front and then transitions into a detonation (Khokhlov 1991; Yamaoka et al. 1992). Within the DDT scenarios there are two realizations: (1) the WD becomes unbound during the deflagration phase (these are the DD models) and, (2) the WD remains bound at the end of the deflagration phase, then undergoes a pulsation followed by a delayed detonation transition. This latter scenario is called a pulsational delayed detonation (PDD). These two realizations form a sequence which merge at the regime where the WD just becomes unbound. In what follows a classical DD model is used as reference, as well as a PDD model with a large amplitude pulsation to span the range (Hoeflich et al. 1995; Hoeflich & Khokhlov 1996). Beyond our own past work on PPD models, we note that Dessart et al. (2014b) have recently presented a suite of PDD models that cover the range of observed SNe Ia luminosities. Their results are similar to those of Hoeflich et al. (1995); Hoeflich & Khokhlov (1996), and show that PDDs can encompass a wide variety of SNe Ia conditions.
We identified a set of simulations based on the spherical explosion of a M Ch WD, which inherently suppresses mixing. One particular simulation within a series of DD models was identified to exhibit the approximately correct brightness of SN 2012Z. This particular simulation, designated 5p0z22.16 (Hoeflich et al. 2002) is a DD model in which a subsonic deflagration burning front turns into a supersonic detonation wave and mixing in the deflagration phase is suppressed. The 5p0z22.16 simulation gives [ M , B B ( -V ) max , ∆ m 15 ( B , ) ∆ m 15 ( V )] of [ -18 22 0 14 1 54 0 99] . , . , . , . , which compares well to [ -18 00 0 07 1 62 0 92] . , . , . , . and [ -18 27 0 01 1 43 0 89] . , . , . , . for SN 2005hk and SN 2012Z, respectively. In addition, the model has total and 56 Ni masses of 1 362 . and 0 268 . M /circledot , respectively, an explosion energy of 1 21 . × 10 51 ergs, and also produces light curves where the primary and secondary NIR maxima merge into a single maximum (Hoeflich et al. 1995, 2002).
From the observed light curves, M Ch models are consistent with SN 2012Z within the uncertainties of ≈ 10 % for dif- ferent progenitors and ≈ 20 % for different explosion scenarios (Hoeflich & Khokhlov 1996; Hoeflich et al. 2013, and references therein). However, in 5p0z22.16 the photosphere recedes from a velocity of about 13 000 km s , -1 on -10 d to about 9 000 km s , -1 by maximum light, which is too high compared to what is observed in SN 2012Z. Light curves capture the diffusion time scales of the energy production, but, as discussed below, it takes line profiles to probe the velocity distribution of the elements to understand the nature of SNe Iax.
## 5.2. Constraints from early phase spectroscopy
In what follows 5p0z22.16 is used as a reference model, but we must consider variations from it that can explain the features of SNe Iax in general, and SN 2012Z in particular. Namely, the model predictions of the photospheric velocities must match the observations. Although this model is a classical DD model, we can think of it as the first in a series of PDD models (Hoeflich et al. 1995), but with a shell mass of zero. Below we consider the effects of variations on the shell mass and how this allows us to subdue the photospheric velocities to match those observed in SN 2012Z. We will be guided in our considerations by the results shown in Figure 10 of Quimby et al. (2007). Due to the low density in the shell some of the original WD's carbon and oxygen (C + O) remains completely unburned in the outer layers. This is in distinction to the classical DD model where all of the material undergoes at least carbon burning.
Early-time spectra can be used to probe the decoupling region of the photons and give a unique measure of the velocity distribution of the elements. Thus, spectral observations probe the density under which burning took place. In velocity space, the 5p0z22.16 simulation (see Figure 3 in Hoeflich et al. 2002) shows products of explosive carbon burning, namely oxygen and magnesium, incomplete oxygen burning, that is silicon and sulphur, and complete burning close to nuclear statistical equilibrium (NSE), i.e. iron/cobalt/nickel (Fe/Co/Ni), in the velocity ranges of [13 , 000 -18 000] , , [7 , 000 -14 000] , and [0 -6 000] km s , -1 , respectively. An inner region of about 3 000 km s , -1 shows elements of stable isotopes of Fe/Co/Ni due to electron capture, a signature of high-density, M Ch models. Calcium is produced in the range [5 , 000 -13 000] km s , -1 , and is destroyed at lower velocities.
In the standard DD scenario for subluminous SNe Ia, specifically model 5p0z22.16, deflagration burning with mixing during the deflagration suppressed, only burns material up to velocities of about 4 300 km s , -1 , (see e.g., Fig. 1 in Hoeflich et al. 2002). SN 2012Z shows Fe-group elements up to, at least, 9 000 km s , -1 (Figure 9), which is some 3 000 km s , -1 higher than where 56 Ni is produced during the detonation phase of the reference model. Fe-group elements are formed during high density burning beyond 1-2 × 10 7 g cm -3 , strongly suggesting mixing in a M Ch scenarios, or an increase in production of 56 Ni during the detonation phase.
An increased production of 56 Ni within the classicaldetonation scenario would lead to an even more extended region of IMEs than 5p0z22.16, which is inconsistent with SN 2012Z. Detonations in smooth density structures will unavoidably produce extended distributions of IMEs in velocity space or if the density is high, hardly any production of IMEs. Because the specific energy production depends on the nuclear processes, the latter case would also produce high velocity 56 Ni (Khokhlov, Mueller & Hoeflich 1993; Hoeflich & Khokhlov 1996). We note that the same arguments would apply to subM Ch explosions triggered by the igni-
tion of an outer helium layer (Woosley, Weaver & Taam 1980; Nomoto 1982; Woosley & Weaver 1994; Hoeflich & Khokhlov 1996; Nugent et al. 1997; Sim et al. 2012).
Mixing may occur both during the deflagration phase of burning and during a pulsation phase of the WD prior to the initiation of the detonation, while the detonation phase leads to a layered structure. During the last decade, significant progress has been made in our understanding of the deflagration phase. On a microscopic scale, a deflagration propagates due to heat conduction by electrons. However, all current simulations have found burning fronts to be Rayleigh-Taylor (RT) unstable increasing the effective burning speed, predicting homogeneous mixing of products of incomplete and complete burning (Livne & Arnett 1993; Livne 1993; Khokhlov 1995; Gamezo et al. 2003; Reinecke, Hillebrandt & Niemeyer 2002; Gamezo, Khokhlov, Oran 2005; R¨ opke, Woosley & Hillebrandt 2007; Seitenzahl et al. 2009, 2013). As seen in Figure 9, mixing of IMEs and calcium is not observed in SN 2012Z at velocities less than ≈ 6 500 km s , -1 . Ca II is commonly seen in SNe Ia even in the outer layers, that at most undergo carbon burning at solar metallicities (see for example Lentz et al. 2000), arguing against mixing of the inner layers during a deflagration burning phase.
Both mixing and an upper limit for the distribution of burning products can be understood in terms of models that undergo a pulsational phase prior to the detonation (Khokhlov, Mueller & Hoeflich 1993; Hoeflich et al. 2004; Hoeflich & Khokhlov 1996; Livne 1999; Livne, Asida & Hoeflich 2005; Quimby et al. 2007; Dessart et al. 2014b). In this class of PDD models, a M Ch WD remains bound after the deflagration phase. The deflagration mostly produces Fe-peak isotopes; subsequently, an extended envelope is formed. As the explosion progresses, and the bound shell falls back a detonation is triggered that leads to the production of IMEs and a massive shell. The mass of the shell determines the upper velocity limit of the burning products (Khokhlov, Mueller & Hoeflich 1993; Hoeflich & Khokhlov 1996; Quimby et al. 2007). To be aligned with SN 2012Z's measured blue wing velocity of Si ii (see Section 4.3), a shell mass of ≈ 0 05 . -0 12 . M /circledot is required (see Figure 10 in Quimby et al. 2007). In addition the blue wing velocity of Mg ii is consistent with a shell mass of ∼ 0 08 M . /circledot . Early multi-dimensional calculations indicated that WDs become unbound during the deflagration phase (Gamezo et al. 2003; R¨ opke, Woosley & Hillebrandt 2007), but this conclusion depends sensitively on the assumption for the thermonuclear runaway (Livne, Asida & Hoeflich 2005). This strong dependency on the basically unknown initial conditions of the WD is an obstacle for fitting details of a specific supernova, but is consistent for the wide variety of properties displayed by SNe Iax.
It is well known that weak interaction/pulsation or the presence of a low-mass buffer around the exploding white dwarf can dramatically alter the outer ejecta layers of the SN Ia explosion (Hoeflich et al. 1995; Hoeflich & Khokhlov 1996; Dessart et al. 2014b). The specific properties of the pulsation can affect the appearance of C i , C ii , and O ii lines The presence or absence of C i , C ii and O ii is also sensitive to the details of the excitation. Some realizations within PDD and DD models do in fact show strong optical and NIR C i and C ii lines (Hoeflich et al. 2002; Dessart et al. 2014b). However, C ii λ 6580 lines at velocities as ∼ 7 000 km s , -1 (Fig. 9) is somewhat in contradiction with also weak C i λ . 1 0694 µ m in the NIR at higher velocity
( ∼ 10 000 km s , -1 ). Alternative identifications for the optical feature are due to lines of Co ii and Si ii .
We note that in PDDs there can be extra energy left over from shock heating in the unburned regions with may lead to higher temperatures than in classical DD models (Dessart et al. 2014b). Some energy from interaction my be also lead to higher temperatures and hotter spectra (Gerardy et al. 2004). Mixing of radioactive material through the C+O region is disfavored as we have already discussed above.
## 5.3. Late phase spectroscopy
Independent of the details of the explosion scenario the inner regions of SNe Ia ejecta are quite similar. Modulations due to variations in PDD models do not impact the bulk of the inner regions of the ejecta for models that produce the same amount of 56 Ni (Hoeflich et al. 1995; Hoeflich & Khokhlov 1996; Dessart et al. 2014b). As discussed above we can probe the 56 Ni distribution using our reference model.
## 5.3.1. Methods
Calculations of explosions, light curves, and spectra are performed using the HYDro RAdiation transport (HYDRA) code (Hoeflich 1990; Hoflich 1995a; Hoeflich 2003; Hoeflich 2009a). HYDRA solves the hydrodynamics using the explicit Piecewise Parabolic Method (PPM Colella & Woodward 1984), detailed nuclear and atomic networks (Cyburt et al. 2010; Kurucz 1994; Hoflich 1995a; Seaton 2005), radiation transport including γ -rays and positrons by variable Eddington Tensor solvers and Monte Carlo Methods (Mihalas & Mihalas 1984; Stone et al. 1992; Hoeflich 2003; Hoeflich et al. 1993; Hoeflich 2009b; Penney 2011). For this study, forbidden line transitions have been updated using the atomic data in De et al. (2010a,b), Friesen et al. (2014), and 'The Atomic Line List V2.05B18' at http://www.pa.uky.edu/˜ peter/newpage/ by Peter van Hoof. For the overall evolution, we use simplified atomic models limited to about 2,000 -10,000 discrete levels using 'level merging' or superlevels (Anderson 1989; Dreizler & Werner 1993; Hubeny & Lanz 1995; Schweitzer et al. 2000; Hillier 2003; Hoeflich 2003; Hoeflich 2009a). To compute NIR line profiles, we do not use superlevels but calculate the emission by postprocessing of the background model (Hoeflich et al. 2004; Penney 2011; Sadler et al. 2012; Hoeflich et al. 2013). Nevertheless, line ratios and excitation state depend sensitively on collisional de-excitation -uncertainties in the collisional cross sections have been and continue to be the main uncertainies in determining the level populations (Axelrod 1980; Bowers et al. 1997; Liu et al. 1997; Gerardy et al. 2007).
## 5.3.2. Spectral Analysis
All thermonuclear explosion models show centrally peaked density profiles, therefore at late times, due to simple geometrical dilution line formation should occur closer to the center at lower velocities yielding narrow line profiles. In fact, if we see to the center the line profile should peak at zero velocity, due to the central caustic. However, due to the mass-radius relation of WDs, lower mass progenitors will have significantly lower central densities. If the central density becomes very low, such as in a sub-Chandrasekhar detonation or merger, than calcium will not be burned at low velocity, as noted above.
The low velocity Fe-group material seen in the late phase visual-wavelength spectra are a natural result of the PDD scenario, but can also be explained by pure deflagration models that produce very low energy explosions (see, for example Reinecke, Hillebrandt & Niemeyer 2002), as well as by deflagrations leaving bound remnants (Kromer et al. 2013; Fink et al. 2013), and merger models (Piersanti et al. 2003; Pakmor et al. 2011).
Late phase spectra beyond 200 days are dominated by forbidden emission lines which are powered by positrons originating from the β + decay of 56 Co . At this phase, hard γ rays from the decay escape for all classes of explosion scenarios (Hoeflich et al. 1992). Up to about one year after the explosion, the mean free path of positrons is small, leading to local energy deposition and excitation of levels (Milne et al. 2002; Penney 2011; Sadler et al. 2012), which traces the initial 56 Ni distribution. Thus, emission line profiles at optical and NIR wavelengths traces the 56 Ni distribution in velocity space, though blends may broaden the features and affect the details of the line profiles, particularly in the optical.
NIR spectra at late epochs of normal SNe Ia are dominated by blended forbidden lines of [Fe i ] , [Fe ii ] , [Fe iii ] , [Co ii ] and [Co iii ] (Hoeflich et al. 2004; Motohara et al. 2006). As normal SN Ia evolve in time, the strengths of their NIR spectral features change due to the ionization state of the line forming regions, which is linked to the ever decreasing decay of 56 Co. During this time the 1.5-1.7 µ m region is dominated by a multitude of blended forbidden [Fe ii ] and [Co iii ] lines, with major contributions attributed to [Fe ii ] λλ 1.53, 1.64, 1.67 µ m, and [Co iii ] λ 1.55 and λ 1.76 µ m (Hoeflich et al. 2004). Eventually, the IR-catastrophe (Fransson, Houck, Kozma 1994) is expected to cause a redistribution of the energy deposition to the NIR and beyond, however light curve observations of SNe Ia constrain this possible phase to much later epochs than considered in this paper (Sollerman et al. 2004; Strizinger & Sollerman 2007; Leloudas et al. 2009; McCully et al. 2014a).
Fig. 12 shows the comparison between the +215 d optical (left panel) and +269 d NIR (right panel) spectra of SN 2012Z compared to the +220 d reference model. As discussed above, the spectra are dominated by single and doubly ionized lines of iron group elements, with almost all of the optical features formed from blends of various species. Many of the more prevalent observed features are reproduced by the synthetic spectrum, however some discrepancies are apparent. The prominent ∼ 7400 A feature is well-matched by the synthetic spectrum. ˚ However, the observed feature is slighly asymmetric on the blue wing, whereas the synthetic feature shows multiple components formed by [Ca ii ] λ 7325.9, [Fe ii ] λλ 7390.2, 7454.6 7639.6, with a significant contribution from [Ni ii ] λλ 7379.9, 7413.6. Here [Ca ii ] contributes significantly to the line width of the blue wing. The blue features are too strong in the synthetic spectrum, which is likely due to inaccuracies in the atomic data and the use of superlevels.
Other notable features are found to be formed by [Fe iii ] λλ 4659.5, 4703.0,4735.3, 4770.9, [Fe iii ] λλ 4932.0, 5012.8, [Fe ii ] λ 5113.1, [Fe ii ] λ 5160.2, [Fe ii ] λλ 5221.5, 5263.1, [Fe iii ] λ 5272.0, [Fe ii ] λλ 5263.1, 5298.3, [Fe ii ] λλ 5335.1, 5377.9, [Co iii ] λλ 5890.1, 5908.4 blended with [Fe ii ] λ 5872 , [Co iii ] λ 6578.1, [Ni ii ] λ 6668.6, and [Co iii ] λ 6855.4 blended with [Fe ii ] λλ 6875 7 6898 1 . , . .
In addition, we see [Fe ii ] λ 8619.3 and [Fe iii ] λλ 8840.6, 9036.0 though it appears slightly too low, most likely due to discrepancies in the atomic line data. [C ] i λ 8729.5 may contributes to the observed feature at ∼ 8600 ˚ ; A however, in the reference model, it forms at too high a velocity and is not the dominant transition. Another interesting feature present in both the models and SN 2012Z is the [O ] i λλ 6302.0, 6302.0, 6393.5 triplet, which appears rather weak. The weak oxygen excitation expected in both DD and PDD models is due to the separation of 56 Ni and the region of explosive oxygen burning (that is where silicon and sulfur are formed), if mixing in the PDD scenario remains limited to a region smaller than the Si/S shell, as it appears to be in SN 2012Z.
The NIR synthetic spectrum of Figure 12 also shows a host of [Fe ii ] and [Fe iii ] lines as seen in the optical. Specifically showing features at 0 95 . , 1 25 . , 1 5 . , 1 6 . , 1 8 . µ m, including the strong feature at 1 25 . µ m which is a blend of [Fe ii ] , [Fe iii ] and [Ni ii ] . An interesting feature located ∼ 1.032 µ mis attributed to a multiplet of [S i ] . Even moderate microscopic mixing of 56 Ni would highly excite the levels of [S i ] producing a very prominent feature. Even higher excitation, due to more mixing would produce a strong [S ii ] feature at 1 02 . -1 04 . µ m. 6
It is well established that the late-time NIR line profiles of [Fe ii ] features trace the energy input by radioactive decay (Hoeflich et al. 2004; Motohara et al. 2006; Maeda et al. 2010). Because of the structure of the atomic levels, the relatively unblended [Fe ii ] 1.64 µ m line is a particularly good tracer of the inner region. In a handful of the SNe Ia observed several hundred days past maximum, NIR emission lines have been found to exhibit pot-bellied profiles. This observational characteristic is the hallmark of M Ch WD progenitors that explode with high central densities, of the order /lessorsimilar 10 9 g cm -3 (Hoeflich et al. 2004; Motohara et al. 2006; Hoeflich et al. 2006; Fesen et al. 2007; Gerardy et al. 2007; Maeda et al. 2010; Hoeflich et al. 2013). At masses close to the M Ch, central densities can become sufficiently high that electron capture processes shift the NSE distribution to stable isotopes of Fe/Co/Ni, leaving a 56 Ni hole at the center of the supernova ejecta, and thereby generate the pot-bellied profiles. This produces a central hole in 56 Ni of about 3000 km s -1 without mixing, very similar to SN 2003du and SN 2003hv (Hoeflich et al. 2004; Motohara et al. 2006). As pointed out in Section 4.4, the [Fe ii ] λ 1.64 µ m feature of SN2012Zalso exhibits a pot-bellied profile (see Figure 10), providing an additional indication that its progenitor was at or near the M Ch.
For clarity on this point, in Figure 13, we show the effect of the 56 Ni distribution on the late phase ( + 269d) [Fe ii ] λ 1.64 µ m line profile. Here the observed feature (black line) is compared to the model calculations of 5p0z22.16 at + 220d for two cases: (1) a hole in the 56 Ni distribution of ∼ 3,000 km s -1 (blue line) and, (2) with the 56 Ni placed at the center of the ejecta (red line). A magnetic field of 10 6 G has been assumed. The blue line is clearly pot-bellied, as in the case of subluminous SNe Ia at the same epoch, because some ≈ 6% of all gamma rays excite the high density, central region (Hoeflich et al. 1991, 2002).
The case for a M Ch explosion is quite strong, lower initial WDmasses would produce 56 Ni confined to the central regions, due to the lower densities in lower mass WDs. More centrally confined 56 Ni would have longer diffusion times, thus falling closer to the the light-curve brightness decline-rate relation and it would not produce the observed spectra since the longer diffusion times would lead to cooler (redder) light curves and spectra more similar to fast decliners like SN 1991bg. Moreover, the persistent pot-bellied line profiles in the NIR require high densities only found in M Ch explosions.
6 Microscopic means on scales smaller than the mean free path for positrons (Hoeflich et al. 2002; Penney 2014).
## 6. Further Comparison Between SN 2005hk and SN 2012Z
In this section we further explore the commonalities between the observational properties of SNe 2005hk and 2012Z. The former of these objects is a well-observed SN Iax that has been previously documented to be similar to the prototypical Type Iax SN2002cx(Phillips et al. 2007; Sahu et al. 2008). As previously highlighted, SNe 2005hk and 2012Z exhibit comparable declinerate values (Figure 6), color curves (Figure 7), and peak luminosity (Figure 8). We now proceed to compare the shape of their light curves, and their optical and NIR spectra at various epochs.
Plotted in Figure 14 is the comparison of the uvw 1 -and ugriBV -band light curves of both supernova plotted vs. T B ( ) max , where the filtered light curves of SN 2005hk have been shifted to match the peak magnitudes of SN 2012Z. Overall the shape of the optical light curves are exceedingly similar, with perhaps the only subtle difference being that SN 2012Z appears to rise to maximum brightness at a slightly faster rate. This is consistent with the 56 Ni being distributed somewhat closer to the surface, leading to shorter diffusion times.
Figure 15 is a comparison between the -8 d, + 2d and + 21d visual-wavelength spectra of SN 2012Z, and similar epoch spectra of SN 2005hk (Phillips et al. 2007). Overall the two objects are remarkably similar, exhibiting similar continua, and nearly identical spectral features at all epochs. Comparing the expansion velocities of SN 2005hk (see Phillips et al. 2007, their Figure 9) and SN 2012Z (see Figure 9), shows that the latter consistently exhibits v exp values which are about 1500 km s -1 higher for all ions. The similar magnitude v exp values between the two events is indicative that they had comparable explosion energies, while the broader line widths observed in SN 2012Z can be attributed to an optical depth effect caused by its 56 Ni distribution extending to higher velocities. This suggests that the 56 Ni was produced in different regions in SN 2005hk and SN 2012Z, implying that the explosion mechanism varied either due to some variation in a violent merger or more likely, that they both occurred due to a PDD explosion (Khokhlov 1991; Khokhlov, Mueller & Hoeflich 1993; Hoeflich & Khokhlov 1996), where the mass of the pulsational shell varied leading to 56 Ni production further-in in SN 2005hk and further-out in SN 2012Z. Thus, we can think of the variation between SN 2005hk and SN 2012Z as variations in the kinetic energy of the 56 Ni , and not variations in the total kinetic energy of the explosion.
Figure 16 shows a comparison between the NIR spectrum of SN 2012Z obtained at maximum (top panel) and three weeks later (bottom panel) to similar epoch spectra of SN 2005hk (Kromer et al. 2013). At both epochs the spectra of the two objects are very similar, with the main difference again being the moderately higher blueshifts of the absorption features in SN 2012Z. Another noticeable difference observed in the bottom panel is that the strength of many of the absorption features appears stronger in SN 2005hk, particularly the [Co ii ] lines located between ≈ 1.6-1.8 µ m. This is due to the Fe-group elements in SN 2005hk being slightly more confined in velocity space. In fact, all of the line profiles are somewhat more confined in velocity space due to the more central nickel distribution in SN 2005hk.
## 7. Discussion
From the findings presented in the previous sections it is clear that SN 2012Z is among the brightest and most energetic SN Iax yet observed, further extending the parameter space of this peculiar class of transients. It is also apparent that objects populating the bright end of the SN Iax luminosity distribution exhibit exceedingly similar observational parameters. In Section 4.2 we derived explosion parameters based on a number of underlying assumptions that provide model fits to the UVOIR light curves of SNe 2005hk and 2012Z. The model fits suggest 56 Ni values ranging between 0.2-0.3 M /circledot , E K values of ≈ 10 51 erg, and ejecta masses consistent with M Ch.
We have presented a possible explosion scenario that explains the nature of at least some SNe Iax. This involves a PDD explosion of a near M Ch WD . The production of Fe-group elements occurs in the deflagration phase (in distinction to normal SNe Ia where the Fe-group elements are primarily produced in the detonation phase, leading to a more layered or radially stratified structure). Our scenario explains both the photospheric velocities of the IMEs, as well as the high densities required by the late phase NIR spectra. The production of 56 Ni closer to the photosphere also simply explains the blue colors and hot spectra. From our reference model and the series of PDD models of Hoeflich et al. (1995) we would find the total mass of 56 Ni to be 0.15 and 0.2 M /circledot , respectively.
SN 2012Z provides a uniquely large data set which allows for a detailed analysis covering a wide range of observables unparalleled even for typical SNe Ia. The nature of SN 2012Z is a theoretical challenge, because several key physics aspects in thermonuclear explosions are not well understood, such as: the progenitor evolution; the thermonuclear runaway; rotation; and so on, making fully self consistent models a goal for the future. These problems lead to the wide variety of models suggested for SNe Iax.
Using both analytic and numerical models as benchmarks, we were able to develop constraints for SN 2012Z based on multiple observables. The light curves suggest a massive, close to M Ch, WD as the progenitor. From the peak brightness we may estimate the mass to be within ∼ 0 2 . M /circledot of M Ch. This view is supported by the pot-bellied, late time NIR line profiles that can be understood by high-density burning taking place at densities in excess of 10 9 g cm -3 , which puts the progenitor within about 0 02 . M /circledot of the M Ch. The other, less stringent constraint, M > 1 2 . M /circledot , is from the lack of Ca II (and other IMEs) at low velocities, which even at solar abundance, produces very strong features in the optical and NIR. Thus, solar abundance primordial calcium must have been burned at low velocities.
Light curves are mostly determined by diffusion time scales, that is, their appearance is dictated by the nature of the innermost regions. The light curve rise time is determined by the nickel distribution. Even with M Ch progenitors, fast rise times and slow declines can be reproduced (Baron et al. 2012). Within this model elements produced in the deflagration will be mixed, while those produced in the detonation will have a layered structure. Note that the detonation running through the material previously burned in the deflagration will reproduce the layered structure.
PDDs are similar to deflagrations, but have a detonation phase which produces a layered structure without requiring a bound remnant. Because PDD models may vary in the ratio of energy released to binding energy (Quimby et al. 2007) they produce a wide range of 56 Ni masses. Also since the shell mass may vary, a range of unburned C+O in the outer layers can also be in PDDs.
The similarity between SN 2012Z and SN 2005hk suggests very similar properties of the exploding WDs. However, SN 2012Z has undergone relatively enhanced mixing to bring
Fe-group elements to the base of the incomplete oxygen burning (Si/S) region. All multidimensional models predict mixing by RT instabilities which produce an almost homogeneous mixing of the layers of IMEs with those below. In SN 2012Z, we see a layered structure, the signature of detonations, for calcium, silicon, and magnesium. This suggests that these elements are produced in a detonation phase after the mixing has occurred and the majority of Fe-group elements have been synthesized. RT instabilities in the deflagration phase seem to be a stable feature from the hydrodynamical view making deflagration models much less favorable as compared to models that result in the layered structure observed in SNe Ia. Detonation fronts in a WD will produce IMEs in a wide velocity range due to the smooth density structure and thus, slowly varying burning conditions. An exception to this picture can be found in the PDD models, where low densities in an shell prevent burning to the Fe-group by a detonation, and limit the expansion velocities of the burning products. For SN 2012Z, we need a shell mass of about 0 05 0 12 . -. M /circledot to be consistent with the low range of velocities observed in IMEs of SN 2012Z. We also need little or no production of 56 Ni during the detonation phase. PDDs model will not show signatures of narrow (low velocity) oxygen features at late phases because oxygen is consumed during the detonation phase. Moreover, as the expansion velocities of SNe Iax are roughly half of what is observed in normal SNe Ia, densities are kept sufficiently high to suppress spectra dominated by forbidden lines of Fe-group elements out to eight times later than observed in SNe Ia, or roughly two to three years past explosion.
Because most of the 56 Ni production is during the deflagration phase, we expect a wide variation in the ∆ m 15 relation compared to normal SNe Ia in which most of the 56 Ni is produced during the detonation phase. This is due to the fact that the amount of 56 Ni produced and its spatial distribution will depend sensitively on the mixing during the deflagration phase, rather than the smooth layered nickel distribution produced in the detonation phase.
PDD models synthesize abundances of 56 Ni that range from near zero to 0.8 M /circledot , and therefore offer the flexibility to produce both bright and dim events. While our proposed model seems to do a reasonable job of explaining the moderately bright SN 2005hk and SN 2012Z, and perhaps also the low luminosity SN 2010ae, we should stress the limitations of our analysis. It is not yet clear whether all SNe Iax are a homogeneous class, or like SNe Ia, will show wide variations. Our analysis here has only been a qualitative description of our scenario. To understand how widely applicable it is to SNe Iax as a class a detailed parameter study with radiative transfer needs to be compared to large datasets of individual SNe Iax.
A common problem with both the PDD model and models with C + O shells, depending on the mass and the amplitude of the explosion, may be the optical depth of the shell during early times which could lead to high velocity calcium and other background metal lines. Accretion from a helium or low metallicity star may reduce this problem due to the low opacity of both unexcited helium (since gamma-rays are shielded from the surface). If there is a significant unburned C+O shell then CO molecules may form which would provide an important coolant and should be visible in the mid infrared (Hoeflich et al. 1995).
Pure deflagration models have been suggested as a candidate scenario for SNe Iax since they can produce a wide range of kinetic energies, and peak brightnesses (Phillips et al. 2007; Sahu et al. 2008). Depending on the initial conditions, pure deflagrations may or may not leave a bound remnant (Reinecke, Hillebrandt & Niemeyer 2002; Fink et al.
2013). Recently a suite of model calculations has been presented, especially designed for SNe Iax, that consist of C + O WDs that undergo deflagration driven disruption leaving bound remnants (Jordan et al. 2012; Fink et al. 2013; Kromer et al. 2013).
Kromer et al. (2013) presented a pure deflagration model leading to a bound remnant of nearly 1 M /circledot , which produces enough 56 Ni to account for the peak brightness of SN 2005hk. Their model does a reasonable job of fitting the photospheric phase optical spectral around maximum light, and the fast rise and slow late-time decline of the light curve of SN 2005hk. In this model the remnant serves to explain the absence of late-time oxygen lines that are produced by the incomplete burning and mixing inherent in pure deflagration models.
Any model that involves a pure deflagration will give a non-layered structure with mixing of unburned material, IMEs, and Fe-group elements throughout the entire ejecta. In particular, this would produce low-velocity oxygen lines at late times (Kozma et al. 2005). The model of Kromer et al. (2013) avoids the appearance of low velocity unburned material by leaving the central regions behind in the bound remnant, while not producing a radially stratified ejecta. Late-time spectra of SN 2012Z (see Figure 11) show shell like line profiles which may be produced in the bound remnant scenario, but the low mass may lead to the formation of true shell-like profiles, that is the formation of horned emission lines. It remains for future calculations to determine if the bound remnant model is able to produce disruptions that synthesize the small amounts of 56 Ni measured in the 2008ha/2010ae-like SNe Iax. In the same context, the PDD models that are found in the literature have also not been extended to the dimmest extent in order to be compared with the observations of the faintest members of the SNe Iax class.
In many respects, other models invoking detonations, such as dynamical merging scenarios and helium detonations can not be ruled out for SNe Iax as they may produce similar configurations as PDD models (Benz et al. 1990; Livio 1989; Lor´ en-Aguilar, Isern & Garc´ ıa-Berro 2009; Pakmor et al. 2 011; Shen et al. 2010; Ruiter et al. 2014). However, they may also be somewhat disfavored due the lack of polarization observed in SNe Ia and SNe Iax (Chornock et al. 2006; Maund et al. 2010; Patat et al. 2012). Additionally the late phase spectra of SN 2012Z, seem to disfavor merger models due to lower densities in the progenitors which leads to 56 Ni all the way to the center (see Section 4.4).
Characteristics in favor of PDD-like models come from the advantages of combining attributes of deflagration and detonation models. As discussed above, the late phase spectra of SN 2012Z show evidence for burning at high densities.
The observational properties described in this paper for SN2012Zhave also not been probed for the majority of SNe Iax, due to incomplete data sets. We note that combining the observed characteristics from different SNe Iax could lead to wrong conclusions. Therefore we cannot exclude that a wide variety of scenarios may be realized within the SNe Iax class.
Understanding the nature of SNe Iax will also come from determinations about the progenitor system in individual cases. McCully et al. (2014b) have observed the progenitor system of SN 2012Z with HST via serendipitous measurements of the host galaxy. They find a stellar source for the progenitor, S1, which is luminous and blue. Examining the color magnitude diagram for S1, it is found that it is consistent with a massive blue supergiant, which is a very unlikely progenitor for SN 2012Z in any scenario. They conclude that they may be seeing the donor companion star, likely a ∼ 2 M /circledot helium star. While the current
observations are not completely aligned with our expectations of a hot disk as the source of S1, it is possible that the environment is contaminated enough that at least some of the light from S1 is indeed from the hot accretion disk. Nevertheless, the observed progenitor fits in with a single degenerate explosion which would be consistent with our proposed scenario. More will be learned from future observations of the putative progenitor when the light from the supernova has faded.
In conclusion, we stress that future efforts to model the SN Iax class should aim to provide robust radiative transfer calculations in order to compare to the detailed optical and NIR spectroscopic sequences now being routinely obtained at both early and late phases. We have shown that there are constraints that models must obey, but quantitative comparison with the full SNe Iax class requires a wide parameter study of the variety of mixing and pulsation properties. Such a study is underway, and should provide confirmation or not of whether PDD models may indeed provide a viable explosion model for SNe Iax.
Acknowledgements. We thank M. Tanaka and V. Stanishev for providing access to their published spectra of SN 2005hk. A special thanks to L. W. Hsiao for providing assistance in some of the observations presented in this study, as well as to the Las Campanas technical staff for their continued support over the years. M. D. Stritzinger, E. Hsiao, and C. Contreras gratefully acknowledge generous support provided by the Danish Agency for Science and Technology and Innovation realized through a Sapere Aude Level 2 grant. M. D. Stritzinger, F. Taddia and S. Valenti acknowledge funding provided by the Instrument Center for Danish Astrophysics (IDA). We would like to express our thanks to Peter van Hoof for creating the Atomic Line List V2.05B18 at http://www.pa.uky.edu/˜ peter/newpage/ . The CSP is supported by the NSF under grants AST-0306969, AST-0607438 and AST-1008343. This work was also supported by the NSF to P. Hoeflich through grants AST-22111 and AST-23432. E. Baron was supported in part by NSF grant AST-0707704. S. Benetti, is partially supported by the PRIN-INAF 2011 with the project 'Transient Universe: from ESO Large to PESSTO'. G. Pignata acknowledges support provided by the Millennium Institute of Astrophysics (MAS) through grant IC120009 of the Programa Iniciativa Cientifica Milenio del Ministerio de Economia, Fomento y Turismo de Chile. This research has made use of the NASA/IPAC Extragalactic Database (NED), which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
## References
Anderson, L. S. 1989, ApJ, 339, 558
Arnett, W. D. 1982, ApJ, 253, 785 Axelrod, T. S. 1980, Ph.D. Thesis, Baron, E., Bongard, S., Branch, D., & Hauschildt, P. H. 2006, ApJ, 645, 480 Baron, E., Hoeflich, P., Krisciunas, K., et al. 2012,ApJ, 753 Benz, W. Cameron, A. G. W., Press, W. H., Bowers, R. L. 1990, ApJ, 348, 647 Bernstein, R. Shectman, S. A., Gunnels, S. M., et al. 2003, Proc. SPIE, 4841, 1694 Bowers, E. J. C., Meikle, W. P. S., Geballe, T. R., et al. 1997, MNRAS, 290, 663 Branch, D. 1992, ApJ, 392, 35 Branch, D., Baron, E., Thomas, R. C., et al. 2004, PASP, 116, 824 Breeveld, A. A., Landsman, W., Holland, S. T., et al. 2011, in AIP Conf. Proc. 1358, Gamma-Ray Bursts 2010, ed. J. E. McEnergy, J. L. Racusin, & N. Gehrels (Melville, NY: AIP), 373 Brown, P. J., Holland, S. T., Immler, S., et al. 2009, AJ, 137, 4517 Brown, P. J., Roming, P. W. A., Milne, P., et al. 2010, ApJ, 721, 1608 Burns, C. R., Stritzinger, M. Phillips, M. M., et al. 2011, AJ, 141, 19 Burrows, D. N., et al. 2005, Space Sci. Rev., 120, 165 Cartier, R., Hamuy, M., Pignata, G., et al. 2014, ApJ, 789, 89 Cenko, S. B., et al. 2012, CBET 3014 Chornock, R., Filippenko, A. V., Branch, D., et al. 2006, PASP, 118, 722 Colella, P., & Woodward, P. R. 1984, Journal of Computational Physics, 54, 174 Colgate, S.A., & McKee, C. 1969, ApJ, 157 Contreras C., Hamuy, M., Phillips, M. M., et al. 2010, AJ, 139, 519 Cyburt, R. H., Amthor, A. M., Ferguson, R., et al. 2010, ApJS, 189, 240 Diamond, T. 2014, Ph.D. Thesis, The Flordia State University, in preparation De, S., Baron, E., & Hauschildt, P. H. 2010, MNRAS, 407, 658 De, S., Baron, E., & Hauschildt, P. H. 2010, MNRAS, 401, 2081 Dessart, L., Hillier, D. J., Blondin, S., & Khokhlov, A. 2014, MNRAS, 439, 3114
- Dessart, L., Blondin, S., Hillier, J. D., & Khokhlov, A. 2014, MNRAS, 441, 532 Dreizler, S., & Werner, K. 1993, A&A, 278, 199
- Fesen, R. A., Hoeflich, P. A., Hamilton, A. J. S., et al. 2007, ApJ, 658, 396
- Filippenko, A. V., Li, W. D., Treffers, R. R., Modjaz, M. 2001, in ASP Conf. Ser. 246, Small-Telescope Astronomy on Global Scales, ed. W. P. Chen, C. Lemme, & B. Paczynski (San Francisco, CA: ASP), 121
- Filippenko, A. V. 2005, in ASP Conf. Ser. 332, The Fate of the Most Massive Stars, ed. R. Humphreys & K. Stanek (San Francisco, CA: ASP), 33
- Fink, M., Kromer, M., Seitenzahl, I., et al., 2013, MNRAS, 429, 2287 Fitzpatrick, E. L. 1999, PASP, 111, 63
- Foley, R. J., Chornock, R., Filippenko, A. V ., et al. 2009, AJ, 138, 376
- Foley, R. J., Rest, A., Stritzinger, M., et al. 2010, AJ, 140, 1321
- Foley, R. J., Challis, P., Chornock, R., et al. 2013, ApJ, 767, 57
- Fransson, C., Houck, J., Kozma, C. 1994, Stockholm Observatory/Sweden, Vol. 32
- Friesen, B., Baron, E., Wisniewski, J. P., et al. 2014, ApJ, (in press), arXiv:1407.7732
- Gamezo, V. N., Khokhlov, A. M.,Oran, E. S. et al. 2003, Science, 299, 77
- Gamezo, V. N., Khokhlov, A. M., Oran, E. S. 2005, ApJ, 623, 337
- Gerardy, C. L., Hoeflich, P., Fesen, R. A., et al. 2004, ApJ, 607, 391
- Gerardy, C. L., Meikle, W. P. S., Kotak, R., et al. 2007, ApJ, 661, 995
- Goldhaber, G., & Perlmutter, S. 1998, Phys. Rep., 307, 325
- Hamuy, M., Phillips, M. M., Suntzeff, N. B., et al. 1996, AJ, 112, 2438
- Hamuy, M., Folatelli, G., Morrell, N., et al., 2006, PASP, 118, 2
- Hillier D.J. 2003, Stellar Atmosphere Modeling, ed. I. Hubeny, D. Mihalas, & K. Werner Astronomical Society of the Pacific Conference Series, 288 (San Francisco: ASP), 199
- Hoeflich, P. 1990, A&A, 229, 191
- Hoeflich, P., Mueller, E., & Khokhlov, A. 1991, A&A, 248L, 7
- Hoeflich, P., Khokhlov, A., & Mueller, E. 1992, A&A, 259, 549
- Hoeflich, P., Khokhlov, A. M., & Wheeler, J. C. 1995, ApJ, 444, 831
- Hoeflich, P., Mueller, E., & Khokhlov, A. 1993, A&A, 268, 570
- Hoflich, P. 1995, ApJ, 443, 89
- Hoeflich, P., Khokhlov, A., Wheeler, J.C., et al. 1996 ApJ, 472, L81
- Hoeflich, P., & Khokhlov, A. 1996, ApJ, 457, 500
- Hoeflich, P., Gerardy, C. L., Fesen, R. A., Sakai, S. 2002, ApJ, 568, 791
- Hoeflich, P. 2003, in Stellar Atmosphere Modeling, ed. I. Hubeny, D. Mihalas, & K. Werner Astronomical Society of the Pacific Conference Series, 288 (San Francisco: ASP), 185
- Hoeflich, P., Gerardy, C., Nomoto, K., et al 2004, ApJ, 617, 1258
- Hoeflich, P., Gerardy, C. L., Marion, H., Quimby, R. 2006, New A Rev., 50, 470 H¨ oflich, P. 2006, Nucl.Phys.A, 777, 579
- Hoeflich, P. 2009a, in Recent Directions In Astrophysical Quantitative Spectroscopy And Radiation Hydrodynamics, ed. I. Hubeny, J. M. Stone, K. MacGregor, & K. Werner (New York: American Inst. of Physics)
Hoeflich,
P.
2009b, in American Institute of Physics Conference
Series,
Vol.
- 1171, American Institute of Physics Conference Series, ed. I. Hubeny, J. M. Stone, K. MacGregor, & K. Werner, 161-172
- Hoeflich, P., Dragulin, P., Mitchell, J., et al. 2013, Frontiers of Physics, 8, 144 Hsiao, E. Y., Marion, G. H., Phillips, M. M., et al. 2013, ApJ, 766, 72 Hubeny, I., & Lanz, T. 1995, ApJ, 439, 875
- Jeffery, D. J., & Branch, D. 1990, Supernovae, Jerusalem Winter School for Theoretical Physics, 149
Jha, S., Branch, D., Chornock, R., et al. 2006, AJ, 132, 189 Jordan I. V ., G. C., Perets, H. B., Fisher, R. T., et al. 2012, ApJ, 761, L23 Kasen, D., R¨ opke, F.K., Woosley, S. E. 2009, Nature, 460, 86 9 Kelson, D. D. 2003, PASP, 115, 688
- Khokhlov, A. M. 1991, A&A, 245L, 25
- Khokhlov, A. M. 1995, ApJ, 449, 695
- Khokhlov, A. 2001, APS Meeting Abstracts, 3003
Khokhlov, A., Mueller, E., & Hoeflich, P. 1992, A&A, 253, L9
- Khokhlov, A. M., Mueller, E., & Hoeflich, P. 1993, A&A, 270, 223
- Komatsu, E., Dunkley, J., Nolta, M. R., et al. ApJS, 180, 330
- Kozma, C., Fransson, C., Hillebrandt, W., et al. 2005, A&A, 437, 983
- Kromer, M., Fink, M., Stanishev, V., et al. 2013, MNRAS, 429, 2287
Kurucz, R. 1994, Atomic Data for Ca, Sc, Ti, V, and Cr.
Kurucz CD-ROM
- No. 20. Cambridge, Mass.: Smithsonian Astrophysical Observatory, 1994., 20
- Landolt, A. U. 1992, AJ, 104, 340
- Leloudas, G., Stritzinger, M. D., Sollerman, J., et al. 2009, A&A, 505, 265L Lentz, E. J., Baron, E., Branch, D., Hauschildt, P. H., Nugent, P. E. 2000, ApJ, 530, 966
Li, W., Filippenko, A. V ., Chornock, R., et al. 2003, PASP, 115, 453L Li W., Leaman, J., Chornock, R., et al., 2011, MNRAS, 412, 1441 Liu, W., Jeffery, D. J., & Schultz, D. R. 1997, ApJ, 483, L107 Livio, M. 1989, NATO ASIC Proc. 290: Theory of Accretion Disks, 135 Livne, E., & Arnett, D. 1993, ApJ, 415, L107
- Livne, E. 1993, ApJ, 406, L17
- Livne, E. 1999, ApJ, 527, L97
- Livne, E., Asida, S. M., & Hoeflich, P. 2005, ApJ, 632, 443L
Lor´ en-Aguilar, L., Isern, J., & Garc´ ıa-Berro, E. 2009, A& A, 500, 1193
- Maeda, K., Mazzali, P. A., Deng, J., et al. 2003, ApJ, 593, 931
- Maeda, K., Benetti, S., Stritzinger, M., et al. 2010, Nature, 466, 82
- Maeda, K., Leloudas, G., Taubenberger, S., et al. 2011, MNRAS, 413, 3075
- Maund, J., Wheeler, J.-C., Wang, L., et al. 2010, ApJ, 722, 1162
- Mazzali, P., Cappellaro, E., Danziger, I. et al., 1998, MNRAS, 499, L49
- McClelland, C. M., Garnavich, P. M., Galbany, Luis, et al. 2010, ApJ, 720, 704 McCully, C., Jha, S. W., et al. 2014, ApJ, 786, 134
- McCully, C., Jha, S. W., Foley, R. J., et al. 2014, Nature, in press
- Mihalas, D., & Mihalas, B. W. 1984, Foundations of Radiation Hydrodynamics (Oxford: Oxford University)
- Milne, P. A., Kurfess, J. D., Kinzer, R. L., & Leising, M. D. 2002, New A Rev., 46, 553
- Motohara, K., Maeda, K., Gerardy, C. L., et al. 2006, ApJ, 652, 101
- Mueller, E., Hoeflich, P., & Khokhlov, A. 1991, A&A, 249, L1
- Narayan, G., Foley, R. J., Berger, E., et al. 2011, ApJ, 731, L11
- Nomoto, K. 1982, ApJ, 253, 798
- Nomoto K., Thielemann, F.-K., & Yokoi K. 1984, ApJ, 286, 644
- Nugent, P., Branch, D., Baron, E., et al. 1995, Physical Review Letters, 75, 394
- Nugent, P., Phillips, M., Baron, E., Branch, D., & Hauschildt, P. 1995, ApJ, 455, L147
- Nugent, P., Baron, E., Branch, D., et al. 1997, ApJ, 485, 812
- Owocki, S. P., & Cohen, D. H. 2001, ApJ, 559, 1108
- Pakmor, R., Hachinger, S., R¨pke, F. K., et al. 2011, A&A, 52 8, A117 o
- Patat, F.,Hoeflich, P., Baade, D., et al. 2012, A&A, 545, 7
- Penney, R. B. 2011, Ph.D. Thesis, The Flordia State University
- Penney, R. B. & Hoeflich, P. 2014, ApJ(submitted).
- Persson, S. E., et al. 1998, AJ, 116, 2475
- Phillips, M. M. 1993, ApJ, 413, L105
- Phillips, M. M., Lira, P., Suntzeff, N. B., et al. 1999, AJ, 118, 1766
- Phillips, M. M., Li, W. D., Frieman, J. A., et al. 2007, PASP, 119, 360
- Phillips, M. M., Simon, J. D., Morrell, N., et al. 2013, ApJ, 779, 38
- Piersanti, L., Gagliardi, S., Iben, I., Jr., & Tornamb´, A. 2003, ApJ, 598, 1229 e
Pinto, P. A., & Eastman, R. G. 2001, New A, 6, 307
- Poznanski, D., Prochaska, J. X., & Bloom, J. S. 2012, MNRAS, 426, 1465
- Quimby, R., Hoeflch, P., Wheeler, J.-C. 2007, ApJ, 666, 1083
- Reinecke, M., Hillebrandt, W., & Niemeyer, J. C. 2002, A&A, 391, 1167
- Riess, A. G., Macri, L., Li, W., et al. 2009, ApJS, 183, 109
- Riess, A. G., Macri, L., Casertano, S., et al. 2011, ApJ, 730, 119
- F. K. R¨ opke, F. K., Woosley, S. E., Hillebrandt, W. 2007, ApJ , 660,1344
- Roming, P. W. A., Kennedy, T. E., Mason, K. O., et al. 2005, Space Sci. Rev., 120, 95
- Ruiter, A. J., Belczynski, K., Sim, S. A., Seitenzahl, I. R., & Kwiatkowski, D. 2014, MNRAS, 440, L101
- Sadler, B. Hoeflich, P., Baron, E. et al. 2012, in: eds.: R. Di Stephano and M. Orio, IAU-symposium 281, p. 309
- Sahu, D. K., Tanaka, M., Anupama, G. C. et al. 2008, ApJ, 680, 592 Schlafly, E. F., & Finkbeiner, D. P. 2011, ApJ, 737, 103
- Schweitzer, A., Hauschildt, P. H., & Baron, E. 2000, ApJ, 541, 1004
- Seaton, M. J. 2005, MNRAS, 362, L1
- Seitenzahl, I. R., Meakin, C. A., Lamb, D. Q., & Truran, J. W. 2009, ApJ, 700, 642
- Seitenzahl, I., Ciaraldi-Schoolmann, F., R¨pke, F., et al . MNRAS, 429, 1156 o Shen, K. J., Kasen, D., Weinberg, N. N., Bildsten, L., & Scannapieco, E. 2010, ApJ, 715, 767
- Silverman, J. M., Ganeshalingam, M., & Filippenko, A. V. 2013, MNRAS, 430, 1030
Sim, S. A., Fink, M., Kromer, M., et al. 2012, MNRAS, 420, 3003 Simon, J. D., Frebel, A., McWilliam, A., et al. 2010, ApJ, 715, L446 Smartt, S., Valenti, S., Fraser, M., et al. 2013, The Messenger, 154, 50 Smith, J. A., Tucker, D., Kent, S., et al. 2002, AJ, 123, 2121 Sollerman, J., Lindahl, J., Kozma, C., et al. 2004, A&A, 428, 555 Spyromilio, J., Meikle, W. P. S., Allen, D.A., Graham, J.R. 1992, MNRAS, 258, 53
Stone, J. M., Mihalas, D., & Norman, M. L. 1992, ApJS, 80, 819 Stritzinger, M., & Sollerman, J. 2007, A&A, 470L, 1 Stritzinger, M., Mazzali, P., Phillips, M. M., et al. 2009, ApJ, 696, 713 Stritzinger, M., Phillips, M. M., Boldt, L., et al. 2011, AJ, 142, 156 Stritzinger, M. D., Hsiao, E., Valenti, S., et al. 2014, A&A, 561, 146 Turatto, M., Benetti, S., Cappellaro, E., et al. 1996, MNRAS, 283, 1 Umeda, H.,Nomoto, K., Kobayashi, C., et al. 1999, ApJ, 522, L43 Valenti, S., Benetti, S., Cappellaro, E., et al. 2008, MNRAS, 383, 1485 Valenti, S., Pastorello, A., Cappellaro, E., et al. 2009, Nature, 459, 674 Wang, B., Justham, S., & Han, Z. 2013, A&A, 559, 94 White, C. J., Kasliwal, M., Nugent, P., et al. 2014, submitted to ApJ
- Woosley, S. E., & Weaver, T. A. 1994, ApJ, 423, 371
Woosley, S. E., Weaver, T. A., & Taam, R. E. 1980, Texas Workshop on Type I Supernovae, 96
Yamaoka, H., Nomoto, K., Shigeyama, T., Thielemann, F. 1992, ApJ, 393, L55
Figure 1: Composite image of the SA(s)bc galaxy NGC 1309, with the position of SN 2012Z indicated. North is up and East is left.
max
Figure 2: UV, optical and NIR light curves of SN 2012Z.
Figure 3: Early phase, visual-wavelength spectroscopy of SN 2012Z plotted in the rest frame of the host galaxy. A subset of spectra from Foley et al. (2013) are plotted in black, and our previously unpublished spectra are shown in red. Phase relative to T B ( ) max is listed to the right of each spectrum. Telluric features are indicated with an Earth symbol. Note that spectra obtained on + 2d, + 9d, and + 16d suffer from moderate fringing.
Figure 4: Late phase, visual-wavelength spectra of SN 2012Z obtained on + 194d, + 215d and + 248d. Each spectrum has been slightly smoothed for presentation purposes.
Figure 5: NIR-wavelength spectroscopy of SN 2012Z. Phase relative to T B ( ) max is listed to the right of each spectrum.
Figure 6: M B plotted vs. ∆ m 15 ( B ) for a sample of CSP SNe Ia (black dots), a handful of SNe Iax (blue squares) that span their full range in luminosity, and SN 2012Z (red star). Note the error bars associated with SN 2012Z are smaller than its symbol size. The comparison SNe Iax plotted are SN 2002cx (Li et al. 2003; Phillips et al. 2007), SN 2003gq (see Foley et al. 2013, and references therein), SN 2005hk (Phillips et al. 2007), SN 2007qd (McClelland et al. 2010), SN 2008A (see Foley et al. 2013, and references therein), SN 2008ge (Foley et al. 2013), SN 2008ha (Stritzinger et al. 2014), SN 2010ae (Stritzinger et al. 2014), and SN 2011ay (Brown, private communication).
Figure 7: ( B -V ) , ( V -r ) , and ( V -i ) color evolution of SN 2012Z (red stars), compared to the Type Iax SN 2005hk (blue dots) and the un-reddened, normal Type Ia SN 2006ax (solid line). The photometry of SN 2006ax is taken from Contreras et al. (2010). The color curves of SNe 2005hk and 2012Z have been corrected for Milky Way and host extinction adopting E B ( -V ) tot = 0 112 . mag (Chornock et al. 2006; Phillips et al. 2007) and E B ( -V ) tot = 0 105 . mag, respectively, while the color curves of SN 2006ax have been corrected for Milky Way reddening.
Figure 8: UVOIR light curves of SNe 2005hk and 2012Z, overplotted by model fits based on Arnett's equations. The model fits provide estimates of the M Ni , M ej and E K (see Section 4.2). Also plotted as filled cyan squares are UVOIR flux points of SN 2012Z computed using its optical and NIR photometry. These points are in excellent agreement with the UVOIR flux points of SN 2012Z, as computed using the NIR flux contribution estimated from SN 2005hk. Note that the error bars are appreciably larger for SN 2005hk, due to its larger uncertainty in distance.
Figure 9: Evolution of the Doppler velocity at maximum absorption of several prominent ions in the spectra of SN 2012Z. The C ii feature is not prominent in the observed spectra and the line identification is suspect. Hence, the inferred velocity is somewhat questionable (see the text for more discussion).
Figure 10: Comparison of similar epoch late phase NIR spectrum of the normal Type Ia SN 2012fr and the Type Iax SN 2012Z.
Figure 11: Comparison of late phase visual-wavelength spectra of a number of SNe Iax and other SN types. The first 5 spectra, sorted from faintest to brightest (top to bottom), are the Type Iax SNe 2010ae (Stritzinger et al. 2014), 2008ge (Foley et al. 2010), 2002cx (Jha et al. 2006), 2005hk (Sahu et al. 2008), and 2012Z. Also shown are similar epoch spectra of the normal Type Ia SN 1998bu, (Silverman, Ganeshalingam & Filippenko 2013), the Type Ib SN 2007Y (Stritzinger et al. 2009), and the subluminous Type IIP SN 2008bk (Stritzinger et al. 2014). Note in SN 2008bk Balmer lines are associated with emission from the SN ejecta, rather than with the narrower lines associated with host nebular lines, as is the case for SNe 2002cx and 2010ae.
Figure 12: Late phase optical (right) and NIR (right) spectrum of SN 2012Z compared to the synthetic spectrum computed from the reference model.
Figure 13: Effect of the 56 Ni distribution upon the late phase [Fe ii ] λ 1.64 µ m feature. Comparison of this feature from the + 269d spectrum of SN 2012Z (black) to the expected line profile computed from the 5p0z22.16 model at + 220d for two different 56 Ni distribution. Plotted in blue is the synthetic spectra for the 'original' model which has a 'hole' in the 56 Ni distribution of ≈ 3000 km s -1 , while in red is the synthetic spectrum with 56 Ni concentrated in the center.
Figure 14: Comparison of uvw 1 - and ugriBV -band light curves of SNe 2005hk (black plus symbols) and 2012Z (red circles). The light curves of SN 2005hk have been shifted to match the peak magnitudes of SN 2012Z.
Figure 15: Comparison of visual-wavelength spectra of SN 2012Z at phases of -8d, + 2d, and + 24d to similar epoch spectra of SN 2005hk (Phillips et al. 2007).
Figure 16: Comparison of NIR-wavelength spectra of SN 2012Z at phases of 0d (top) and + 22d (bottom), to similar epoch spectra of SN 2005hk (Kromer et al. 2013). Prevalent features attributed to ions of Fe ii , Si iii , and Co ii are indicated with labels.
Stritzinger et al.: The bright and energetic Type Iax SN 2012Z.
Table 1. UVOT Ultraviolet Photometry of SN 2012Z.
| JD-2450000 | Phase a | uvw 2 | uvm 2 | uvw 1 |
|--------------|-----------|-------------|-------------|-------------|
| 5959.77 | - 8.12 | 17.217(122) | 17.001(091) | 15.853(075) |
| 5961.95 | - 5.94 | 17.257(114) | 17.191(093) | 15.802(070) |
| 5965.02 | - 2.87 | 17.371(116) | 17.435(096) | 15.892(082) |
| 5967 | - 0.89 | 17.567(118) | 17.811(102) | 16.241(091) |
| 5968.66 | 0.77 | 17.683(118) | 17.855(103) | 16.362(087) |
| 5971.11 | 3.22 | · · | · · · | 17.296(104) |
| 5976.96 | 9.07 | · · · · | 19.565(296) | · · · |
| 5977.72 | 9.83 | 19.073(226) | · · · | 17.663(107) |
| 5979.6 | 11.71 | 19.128(230) | · · · | 18.034(137) |
| 5982.36 | 14.47 | 19.600(376) | · · · | 18.093(171) |
| 5989.79 | 21.90 | · · | · · · | 19.094(219) |
| 5992.13 | 24.24 | · · · · | · · · | 19.341(262) |
| 5994.01 | 26.12 | · · · | · · · | 19.483(290) |
| 5998.08 | 30.19 | · · · | · · · | 19.532(302) |
Note. - One sigma uncertainties given in parentheses are in millimag.
a Days past T B ( ) max , i.e. JD-2455967.89.
Table 2. Optical and NIR photometry of the local sequence of SN 2012Z in the natural system.
| STAR | α (2000) | δ (2000) | u (mag) | N | g (mag) | N | r (mag) | N | i (mag) | N | B (mag) | N | V (mag) | N | Y (mag) | N | J (mag) | N | H N (mag) | |
|--------|-------------|--------------|-------------|-------|-------------|-------|-------------|-------|-------------|-------|-------------|-------|-------------|-------|-------------------|-------|-------------|-------|-------------|-------------------|
| 001 | 03:22:15.58 | -15:19:11.93 | 15.026(007) | 17 | · · · | · · · | · · · | · · · | · · · | · · · | 13.339(005) | 7 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · |
| 002 | 03:22:19.50 | -15:19:19.63 | 14.038(006) | 18 | · · · | · · · | · · · | · · · | · · · | · · · | 13.282(005) | 7 | 12.762(011) | 1 | · · · | · · · | · · · | · · · | · · · | · · · |
| 003 | 03:22:00.97 | -15:19:28.24 | 14.358(006) | 18 | · · · | · · · | · · · | · · · | · · · | · · · | 13.399(005) | 8 | 12.759(011) | 1 | · · · | · · · | · · · | · · · | · · · | · · · |
| 004 | 03:21:52.18 | -15:20:52.44 | 17.045(010) | 19 | 16.002(003) | 22 | 15.636(003) | 20 | 15.501(003) | 25 | 16.304(003) | 26 | 15.791(003) | 28 | · · · | · · · | · · · | · · · | · · · | · · · |
| 005 | 03:22:05.62 | -15:22:34.86 | 17.007(010) | 19 | 16.022(003) | 22 | 15.618(003) | 20 | 15.471(003) | 25 | 16.330(003) | 27 | 15.795(003) | 29 | · · · | · · · | · · · | · · · | · · · | · · · |
| 006 | 03:21:54.34 | -15:21:12.06 | 17.220(011) | 19 | 16.101(003) | 22 | 15.648(003) | 20 | 15.477(003) | 25 | 16.435(003) | 27 | 15.849(003) | 29 | · · · | · · · | · · · | · · · | · · · | · · · |
| 007 | 03:22:18.71 | -15:19:27.70 | 19.076(044) | 10 | 16.957(004) | 19 | 16.159(004) | 17 | 15.844(004) | 20 | 17.441(007) | 22 | 16.565(004) | 23 | · · · | · · · | · · · | · · · | · · · | · · · |
| 008 | 03:22:09.62 | -15:22:13.87 | 18.311(020) | 18 | 17.188(004) | 22 | 16.780(005) | 20 | 16.616(005) | 25 | 17.532(006) | 27 | 16.965(004) | 29 | · · · | · · · | · · · | · · · | · · · | · · · |
| 009 | 03:22:18.09 | -15:24:01.51 | · · · | · · · | 18.486(010) | 22 | 17.342(006) | 20 | 16.861(006) | 25 | 18.905(018) | 20 | 17.920(008) | 29 | · · · | · · · | · · · | · · · | · · · | · · · |
| 010 | 03:21:53.30 | -15:27:14.29 | · · · | · · · | 19.155(024) | 11 | 17.840(012) | 12 | 16.165(006) | 15 | 19.778(071) | 5 | 18.510(018) | 17 | · · · | · · · | · · · | · · · | · · · | · · · |
| 011 | 03:22:12.29 | -15:19:41.63 | · · · | · · · | 19.133(019) | 17 | 17.742(009) | 18 | 16.700(005) | 22 | 19.788(047) | 9 | 18.442(013) | 28 | · · · | · · · | · · · | · · · | · · · | · · · |
| 012 | 03:22:16.89 | -15:20:44.88 | 18.917(082) | 1 | 18.880(014) | 21 | 18.060(011) | 20 | 17.628(010) | 25 | 18.962(024) | 14 | 18.463(014) | 29 | · · · | · · · | · · · | · · · | · · · | · · · |
| 013 | 03:21:48.55 | -15:21:14.36 | 19.452(094) | 3 | 18.817(018) | 13 | 18.476(019) | 14 | 18.291(023) | 16 | 18.755(022) | 14 | 18.587(017) | 22 | · · · | · · · | · · · | · · · | · · · | · · · |
| 014 | 03:21:52.45 | -15:27:02.48 | 18.958(048) | 7 | 18.991(017) | 16 | 18.666(019) | 17 | 18.674(031) | 18 | 18.970(025) | 13 | 18.881(023) | 20 | · · · | · · · | · · · | · · · | · · · | Stritzinger · · · |
| 015 | 03:22:22.44 | -15:20:44.09 | · · · | · · · | 19.444(024) | 16 | 18.628(017) | 19 | 18.263(018) | 24 | 19.445(044) | 8 | 18.885(022) | 20 | · · · | · · · | · · · | · · · | · · · | · · · |
| 016 | 03:21:55.41 | -15:20:24.11 | · · · | · · · | 20.111(044) | 12 | 18.848(021) | 19 | 17.568(010) | 25 | 20.147(104) | 3 | 19.480(036) | 15 | · · · | · · · | · · · | · · · | · · · | et · · · |
| 017 | 03:22:14.78 | -15:27:04.36 | · · · | · · · | 20.372(097) | 3 | 19.082(032) | 11 | 18.222(020) | 17 | · · · | · · · | 19.656(059) | 7 | · · · | · · · | · · · | · · · | · · · | al.: · · · |
| 018 | 03:22:05.60 | -15:22:35.06 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 14.808(016) | 4 | 14.572(017) | 5 | 14.246(014) | 5 |
| 019 | 03:22:03.04 | -15:22:27.83 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 16.431(024) | 4 | 16.035(028) | 5 | 15.384(023) | The 5 |
| 020 | 03:22:09.59 | -15:22:13.71 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 15.973(020) | 4 | 15.708(024) | 5 | 15.398(022) | 5 |
| 021 | 03:22:04.73 | -15:23:44.15 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 16.357(033) | 4 | 15.901(032) | 5 | 15.245(029) | bright 5 |
| 022 | 03:22:01.66 | -15:22:30.86 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 18.399(096) | 2 | · · · | · · · | · · · | · · · |
| 023 | 03:22:05.94 | -15:24:38.59 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | | · · · 17.229(100) | 1 | 16.854(104) | 1 | 16.383(119) | 1 |
Note. - Uncertainties given in parentheses in thousandths of a magnitude correspond to an rms of the magnitudes obtained on photometric nights.
Stritzinger et al.: The bright and energetic Type Iax SN 2012Z.
Table 3. Swope optical photometry of SN 2012Z in the natural system.
| JD - 2 , 450 , 000+ | u (mag) | g (mag) | r (mag) | i (mag) | B (mag) | V (mag) |
|-----------------------|-------------|-------------|-------------|-------------|-------------|-------------|
| 5959.57 | 15.631(011) | 15.533(004) | 15.582(005) | 15.753(008) | 15.591(008) | 15.676(007) |
| 5960.57 | 15.448(008) | · · · | 15.322(009) | 15.488(009) | 15.368(009) | 15.416(009) |
| 5961.57 | 15.323(006) | 15.106(004) | 15.124(005) | 15.282(005) | 15.171(008) | 15.214(008) |
| 5962.56 | 15.233(007) | · · · | · · · | 15.115(005) | 15.027(006) | 15.028(005) |
| 5963.58 | 15.190(011) | 14.839(004) | 14.814(004) | 14.985(004) | 14.925(007) | 14.896(004) |
| 5964.6 | · · · | 14.746(006) | 14.704(006) | 14.871(008) | 14.857(006) | 14.785(005) |
| 5965.61 | 15.165(016) | 14.697(006) | 14.608(008) | 14.783(008) | 14.812(009) | 14.698(007) |
| 5966.59 | 15.156(013) | · · · | · · · | 14.698(006) | 14.759(006) | 14.619(005) |
| 5967.59 | 15.199(012) | · · · | 14.472(006) | 14.623(007) | 14.750(007) | 14.562(005) |
| 5968.59 | 15.254(008) | 14.584(007) | 14.417(006) | 14.578(008) | 14.753(004) | 14.512(006) |
| 5969.59 | 15.321(009) | 14.572(004) | 14.369(004) | 14.517(008) | 14.780(004) | 14.479(007) |
| 5970.56 | 15.425(009) | 14.579(004) | 14.336(005) | 14.467(007) | 14.830(008) | 14.458(008) |
| 5971.57 | 15.520(010) | 14.586(007) | 14.288(009) | 14.431(009) | 14.829(006) | 14.412(006) |
| 5972.57 | 15.629(009) | 14.646(004) | 14.287(005) | 14.423(008) | 14.915(004) | 14.422(006) |
| 5973.55 | 15.751(009) | 14.698(004) | 14.279(004) | 14.389(005) | 15.007(005) | 14.428(006) |
| 5974.57 | 15.893(007) | 14.750(004) | 14.268(004) | 14.370(004) | 15.086(004) | 14.437(004) |
| 5979.58 | 16.859(009) | 15.235(005) | 14.322(005) | · · · | 15.699(007) | 14.671(006) |
| 5985.54 | 17.998(015) | 15.934(004) | 14.604(004) | 14.470(004) | 16.498(006) | 15.121(005) |
| 5986.57 | 18.138(017) | 16.063(008) | · · · | 14.517(004) | 16.601(007) | 15.204(005) |
| 5988.57 | 18.350(024) | 16.211(005) | 14.800(004) | 14.608(004) | 16.792(007) | 15.364(005) |
| 5990.58 | 18.496(051) | 16.352(006) | 14.925(005) | 14.716(005) | · · · | · · · |
| 5991.58 | 18.672(076) | 16.422(008) | 14.983(005) | 14.769(004) | 17.018(011) | 15.559(004) |
| 5992.56 | 18.656(041) | 16.511(006) | 15.047(008) | 14.811(006) | 17.074(011) | 15.643(008) |
| 5993.53 | · · · | 16.562(006) | 15.100(004) | 14.875(004) | 17.129(009) | 15.694(005) |
| 5994.54 | 18.764(067) | 16.599(007) | 15.157(004) | 14.922(006) | 17.168(011) | 15.749(006) |
| 5995.54 | 18.880(064) | 16.668(008) | 15.215(006) | 14.969(007) | 17.237(010) | 15.797(007) |
| 5996.53 | 18.945(038) | 16.713(009) | 15.267(008) | 15.023(008) | 17.250(014) | 15.848(006) |
| 5997.52 | 19.017(023) | 16.743(007) | 15.310(007) | 15.072(008) | 17.299(009) | 15.890(006) |
| 5998.52 | 19.063(030) | 16.773(004) | 15.349(004) | 15.101(004) | 17.347(009) | 15.923(006) |
| 5999.52 | 19.024(021) | 16.819(011) | 15.396(009) | 15.168(010) | 17.375(010) | 15.994(009) |
| 6000.56 | · · · | 16.826(008) | 15.418(004) | 15.172(006) | 17.395(008) | 15.995(004) |
| 6001.55 | · · · | 16.847(005) | 15.467(004) | 15.214(005) | 17.419(009) | 16.024(007) |
| 6002.55 | · · · | 16.853(005) | 15.502(004) | 15.252(006) | 17.422(009) | 16.045(006) |
| 6003.55 | · · · | 16.877(005) | 15.530(004) | 15.275(004) | 17.451(009) | 16.085(004) |
| 6004.55 | · · · | 16.912(006) | 15.563(004) | 15.316(004) | 17.474(009) | 16.103(004) |
| 6005.55 | · · · | · · · | · · · | 15.341(004) | 17.478(009) | 16.121(005) |
| 6006.54 | · · · | 16.954(008) | 15.623(006) | 15.385(008) | 17.520(009) | 16.157(006) |
| 6007.52 | · · · | 16.975(009) | 15.676(009) | 15.404(009) | 17.522(009) | 16.188(009) |
| 6009.52 | · · · | 16.978(006) | 15.702(006) | 15.452(006) | 17.545(007) | 16.213(006) |
| 6010.51 | · · · | 17.000(007) | 15.741(006) | 15.498(008) | 17.551(008) | 16.235(007) |
| 6011.52 | · · · | · · · | · · · | · · · | 17.596(010) | 16.253(007) |
| 6012.53 | · · · · · · | · · · | · · · | · · · | 17.581(009) | 16.273(005) |
| 6014.51 | | 17.056(007) | 15.850(005) | 15.606(005) | · · · | · · · |
| 6015.53 | · · · | · · · | · · · | · · · | · · · | 16.335(008) |
| 6017.52 | · · · | · · · | · · · | · · · | 17.625(015) | 16.358(007) |
| 6018.49 | · · · | 17.097(010) | 15.942(004) | 15.714(004) | · · · | · · · |
| 6019.49 | · · · | · · · | · · · | · · · | 17.644(014) | 16.386(008) |
| 6233.82 | · · · | · · · | · · · | 20.381(119) | 21.411(318) | 20.876(206) |
Note. - Values in parentheses are 1 σ measurement uncertainties in millimag.
Table 4. du Pont NIR photometry of SN 2012Z in the natural system.
| JD - 2 , 450 , 000+ | Y (mag) | J (mag) | H (mag) |
|-----------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------|-------------------------------------------------------------------------|
| 5992.51541 5996.52808 5997.49164 6000.49606 6001.49526 6022.46819 6167.85063 6168.82954 | 13.899(013) · · · 14.047(012) 14.139(012) 14.185(013) 14.912(011) 19.818(055) · · · | 14.208(013) · · · 14.402(013) 14.557(013) 14.565(013) · · · 19.269(046) 19.160(065) | 13.963(012) 14.119(011) · · · 14.288(011) 14.277(011) · · · · · · · · · |
Note. - Values in parentheses are 1 σ measurement uncertainties in millimag.
Stritzinger et al.: The bright and energetic Type Iax SN 2012Z.
Table 5. Journal of Spectroscopic Observations.
| Date | JD - 2,450,000+ | Phase a | Telescope | Instrument |
|---------------|-------------------|-----------|-------------|--------------|
| Optical | Optical | Optical | Optical | Optical |
| 2012 Feb 01.2 | 5958.7 | - 9.2 | Lick | Kast |
| 2012 Feb 02.2 | 5959.7 | - 8.2 | Lick | Kast |
| 2012 Feb 04.1 | 5961.5 | - 6.4 | Clay | MIKE |
| 2012 Feb 12.2 | 5969.7 | + 1.8 | NOT | Alfosc |
| 2012 Feb 16.1 | 5973.6 | + 5.7 | FLWO | FAST |
| 2012 Feb 18.9 | 5976.4 | + 8.5 | NOT | Alfosc |
| 2012 Feb 19.1 | 5976.6 | + 8.7 | FLWO | FAST |
| 2012 Feb 21.0 | 5978.6 | + 10.7 | FLWO | FAST |
| 2012 Feb 21.1 | 5979.6 | + 11.7 | FLWO | FAST |
| 2012 Feb 22.1 | 5980.5 | + 12.6 | Clay | LDSS3 |
| 2012 Feb 25.8 | 5983.4 | + 15.5 | NOT | Alfosc |
| 2012 Mar 02.1 | 5988.6 | + 20.7 | Clay | LDSS3 |
| 2012 Mar 15.2 | 6001.7 | + 33.8 | Keck | LRIS |
| 2012 Aug 22.1 | 6161.6 | + 193.7 | SALT | RSS |
| 2012 Sep 12.3 | 6182.8 | + 214.9 | du Pont | WFCCD |
| 2012 Oct 15.5 | 6216.0 | + 248.1 | Keck | DEIMOS |
| NIR | NIR | NIR | NIR | NIR |
| 2012 Feb 03.0 | 5960.6 | - 7.3 | Baade | Fire |
| 2012 Feb 10.0 | 5967.5 | 0.4 | Baade | Fire |
| 2012 Feb 15.0 | 5972.5 | - + 4.6 | NTT | Sofi |
| 2012 Feb 21.0 | 5978.5 | + 10.6 | VLT | ISAAC |
| 2012 Feb 27.0 | 5984.5 | + 16.6 | VLT | ISAAC |
| 2012 Mar 03.0 | 5989.5 | + 21.6 | Baade | Fire |
| 2012 Mar 14.9 | 6001.5 | + 33.6 | VLT | ISAAC |
| 2012 Mar 29.9 | 6016.5 | + 48.6 | VLT | ISAAC |
| 2012 Apr 07.9 | 6025.5 | + 57.6 | Baade | Fire |
| 2012 Nov 05.2 | 6236.7 | + 268.8 | Baade | Fire |
a Days since T B ( ) max , i.e., JD -2455967.8.
Stritzinger et al.: The bright and energetic Type Iax SN 2012Z.
Table 6. Optical Light Curve Parameters of SNe 2005hk and 2012Z.
| Filter | Peak Time (JD - 2,450,000) | Peak Obs. (mag) | ∆ m 15 (mag) | Peak Abs. (mag) |
|----------|------------------------------|-------------------|----------------|-------------------|
| | SN 2005hk | SN 2005hk | SN 2005hk | |
| u | 3682.54 ± 0.36 | 16.27 ± 0.01 | 1.96 ± 0.07 | - 17.73 ± 0.25 |
| g | 3685.92 ± 0.08 | 15.78 ± 0.01 | 1.36 ± 0.01 | - 18.08 ± 0.25 |
| r | 3691.66 ± 0.23 | 15.68 ± 0.01 | 0.70 ± 0.02 | - 18.07 ± 0.25 |
| i | 3695.27 ± 0.16 | 15.80 ± 0.01 | 0.60 ± 0.01 | - 17.88 ± 0.25 |
| B | 3685.16 ± 0.07 | 15.91 ± 0.01 | 1.62 ± 0.01 | - 18.00 ± 0.25 |
| V | 3688.28 ± 0.14 | 15.73 ± 0.01 | 0.92 ± 0.01 | - 18.07 ± 0.25 |
| | SN 2012Z | SN 2012Z | SN 2012Z | |
| u | 5965.48 ± 0.35 | 15.16 ± 0.01 | 1.92 ± 0.08 | - 17.95 ± 0.09 |
| g | 5969.78 ± 0.12 | 14.57 ± 0.01 | 1.30 ± 0.01 | - 18.40 ± 0.09 |
| r | 5975.26 ± 0.33 | 14.27 ± 0.01 | 0.66 ± 0.02 | - 18.60 ± 0.09 |
| i | 5978.46 ± 0.53 | 14.34 ± 0.01 | 0.54 ± 0.04 | - 18.46 ± 0.09 |
| B | 5967.89 ± 0.11 | 14.75 ± 0.01 | 1.43 ± 0.02 | - 18.27 ± 0.09 |
| V | 5972.74 ± 0.14 | 14.42 ± 0.01 | 0.89 ± 0.01 | - 18.50 ± 0.09 |
## Appendix A: Definitive Photometry of SN 2005hk
A portion of the CSP photometry of SN 2005hk originally published by Phillips et al. (2007) was computed from science images which underwent template subtraction using provisional template images. We returned to the field of UGC 272 and obtained updated optical BV -and NIR Y JHK s -band images. These were used to compute proper template subtraction for the respective science images. In addition, the field of local sequence stars was also re-calibrated based on a larger number of nights in which standard star fields were observed. Table A1 contains the re-calibration of our optical and NIR local sequences, while our revised optical and NIR photometry of SN 2005hk is presented in Table A2 and Table A3, respectively. Note that all of our photometry of the local sequence stars and the SN are presented in the natural system (Stritzinger et al. 2011).
Table A1. Optical and NIR photometry of the local sequence of SN 2005hk in the natural system.
| STAR | α (2000) | δ (2000) | u (mag) | N | g (mag) | N | r (mag) | N | i (mag) | N | B (mag) | N | V (mag) | N | Y (mag) | N | J (mag) | N | H (mag) | N | (m | (m | (m | (m |
|--------|-------------|--------------|-------------|-------|-------------|-------|-------------|-------|-------------|-------|-------------|-------|-------------|---------------|-------------|---------------|---------------|---------------|------------------|---------------|--------|-------|-------|-------|
| 001 | 00:27:58.40 | -01:13:12.19 | 15.556(019) | 5 | 14.458(009) | 2 | 14.499(007) | 3 | 14.590(006) | 4 | 14.602(006) | 5 | 14.441(006) | 4 | 12.692(008) | 5 | 12.480(007) 5 | | 12.157(007) | 5 | · | · | · | · |
| 002 | 00:27:59.34 | -01:09:36.38 | 16.625(023) | 5 | 15.020(007) | 3 | 14.420(009) | 2 | 14.201(010) | 1 | 15.422(006) | 5 | 14.673(006) | 4 | 13.467(009) | 4 | 13.226(008) | 5 | 12.848(008) | 5 | · | · | · | · |
| 003 | 00:28:01.15 | -01:11:29.09 | 17.883(037) | 4 | 16.459(006) | 5 | 15.820(006) | 5 | 15.582(006) | 5 | 16.858(009) | 5 | 16.090(008) | 5 | 14.824(010) | 5 | 14.530(008) | 5 | 14.114(009) | 5 | · | · | · | · |
| 004 | 00:28:02.93 | -01:11:23.56 | 19.259(164) | 1 | 17.086(007) | 5 | 15.859(006) | 5 | 15.355(006) | 5 | 17.663(014) | 5 | 16.462(009) | 5 | 14.344(009) | 5 | 13.970(007) | 5 13.394(007) | | 5 | · | · | · | · |
| 005 | 00:27:38.83 | -01:09:16.30 | 17.641(036) | 4 | 16.521(006) | 5 | 15.953(007) | 5 | 15.733(006) | 5 | 16.879(009) | 5 | 16.199(009) | 5 | · · · | · · · | · · · | · · · | · · · | · · · | · | · | · | · |
| 006 | 00:27:38.38 | -01:11:30.82 | 18.537(060) | 3 | 16.723(006) | 5 | 15.911(006) | 5 | 15.575(006) | 5 | 17.180(011) | 5 | 16.283(009) | 5 | · · · | · · · | · · · | · · · | · · · | · · · | · | · | · | · |
| 007 | 00:27:51.58 | -01:14:36.05 | 17.245(027) | 5 | 16.258(006) | 5 | 16.057(007) | 5 | 16.036(007) | 5 | 16.471(008) | 5 | 16.127(008) | 5 | 15.497(016) | 4 | 15.361(013) | 5 | 15.209(019) | 5 | · | · | · | · |
| 008 | 00:27:50.88 | -01:08:02.40 | 17.669(032) | 4 | 16.706(006) | 5 | 16.334(007) | 5 | 16.165(007) | 5 | 17.009(010) | 5 | 16.467(009) | 5 | 15.543(017) | 4 | 15.347(017) | 5 | 15.007(022) | 5 | · | · | · | · |
| 009 | 00:28:00.54 | -01:15:50.50 | · · · | · · · | 17.609(009) | 4 | 16.492(008) | 4 | 16.052(008) | 4 | 18.175(019) | 4 | 17.058(012) | 4 | 15.092(015) | 4 | 14.722(012) | 5 | 14.123(013) | 5 | · | · | · | · |
| 010 | 00:27:47.46 | -01:07:55.83 | 18.483(057) | 3 | 17.288(007) | 5 | 16.685(008) | 5 | 16.452(009) | 5 | 17.696(014) | 5 | 16.949(011) | 5 | 15.651(018) | 4 | 15.417(017) | 5 | 15.060(023) | 5 | · | · | · | · |
| 011 | 00:27:47.82 | -01:13:59.88 | 19.505(198) | 1 | 17.683(009) | 5 | 16.823(008) | 5 | 16.469(009) | 5 | 18.109(018) | 5 | 17.238(012) | 5 | 15.624(013) | 5 | 15.357(010) | 6 | 14.836(014) | 6 | 14.65 | 14.65 | 14.65 | 14.65 |
| 012 | 00:28:02.72 | -01:08:07.84 | 19.271(116) | 2 | 17.845(010) | 5 | 17.009(009) | 5 | 16.686(010) | 5 | 18.332(022) | 5 | 17.448(014) | 5 | 15.821(020) | 4 | 15.553(019) | 5 | 14.982(022) | 5 | · | · | · | · |
| 013 | 00:27:58.56 | -01:12:13.40 | 18.957(082) | 3 | 17.640(009) | 5 | 17.019(009) | 5 | 16.737(010) | 5 | 18.031(017) | 5 | 17.298(013) | 5 | 15.910(015) | 5 | 15.656(014) | 5 | 15.212(019) | 5 | · | · | · | · |
| 014 | 00:28:05.55 | -01:14:45.17 | 18.346(049) | 3 | 17.413(008) | 5 | 17.031(009) | 5 | 16.895(011) | 5 | 17.727(014) | 5 | 17.194(012) | 5 | 16.244(024) | 4 | 16.005(023) | 5 | 15.787(035) | Stritzinger 5 | · | · | · | · |
| 015 | 00:27:41.73 | -01:10:57.01 | 19.363(187) | 1 | 17.871(010) | 5 | 17.221(010) | 5 | 16.951(012) | 5 | 18.263(021) | 5 | 17.511(014) | 5 | · · · | 0 | · · · | · · · | · · · | · · · | · | · | · | · |
| 016 | 00:28:05.45 | -01:12:24.64 | · · · | · · · | 18.848(022) | 4 | 17.466(012) | 5 | 16.602(010) | 5 | 19.437(055) | 3 | 18.109(022) | 5 | 15.400(012) | 5 | 14.998(010) | 5 | 14.408(011) | et 5 | · | · | · | · |
| 017 | 00:27:47.99 | -01:12:55.62 | · · · | · · · | 18.556(016) | 5 | 17.719(014) | 5 | 17.356(015) | 5 | 19.006(037) | 3 | 18.116(020) | 5 | 16.467(023) | 5 | 16.170(017) | 6 | 15.545(023) | al.: 6 | 16.20 | 16.20 | 16.20 | 16.20 |
| 018 | 00:27:38.35 | -01:09:00.28 | · · · | · · · | 18.886(022) | 4 | 17.588(013) | 5 | 16.130(007) | 5 | 19.637(066) | 3 | 18.254(023) | 5 | · · · | | 0 · · · | · · · | · · · | · · · | · | · | · | · |
| 019 | 00:27:37.32 | -01:10:55.99 | · · · | · · · | 19.093(026) | 4 | 17.845(015) | 5 | 17.237(014) | 5 | 19.672(067) | 3 | 18.313(024) | 5 | · · · | 0 | · · · | · · · | · · · | The · · · | · | · | · | · |
| 020 | 00:27:55.68 | -01:14:38.15 | 18.003(053) | 3 | 18.467(015) | 5 | 18.104(018) | 5 | 17.856(024) | 5 | 18.714(029) | 3 | 18.279(024) | 5 | · · · | · · · | · · · | · · · | · · · | · · · | · | · | · | · |
| 021 | 00:27:50.49 | -01:09:04.46 | 18.805(107) | 1 | 18.844(021) | 4 | 18.504(025) | 5 | 18.314(035) | 5 | 19.093(041) | 4 | 18.658(032) | 5 | · · · | · · · | · · · | · · · | · · · | · · · | · | · | · | · |
| 022 | 00:27:55.61 | -01:07:57.62 | · · · | · · · | 19.220(029) | 4 | 18.886(034) | 5 | 18.649(050) | 4 | 19.546(060) | 3 | | 19.099(045) 3 | · · · | · · · | · · · | · · · | · · · | bright · · | · | · | · | · |
| 023 | 00:27:36.28 | -01:11:58.38 | · · · | · · · | 19.911(054) | 3 | 18.732(029) | 4 | 17.859(025) | 4 | 20.634(198) | 1 | 19.533(072) | 3 | · · · | · · · | · · · | · · · | · · · | · · · | · | · | · | · |
| 024 | 00:27:58.91 | -01:15:00.70 | · · · | · · · | 20.097(064) | 3 | 18.748(031) | 4 | 17.793(022) | 5 | · · · | · · · | 18.482(041) | 4 | 16.614(032) | 4 | 16.169(027) | 5 | 15.667(033) | and 5 | · | · | · | · |
| 025 | 00:27:46.17 | -01:09:55.61 | · · · | · · · | 20.380(093) | 2 | 19.105(042) | 4 | 18.224(032) | 5 | · · · | · · · | 19.740(084) | 3 | · · · | · · · | · · · | · · · | · · · | · · · | · | · | · | · |
| 026 | 00:27:57.20 | -01:13:03.18 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 12.692(008) | 5 | 12.480(007) | 5 | 12.157(007) | energetic 5 | · | · | · | · |
| 027 | 00:28:04.74 | -01:12:35.09 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 13.077(008) | 5 | 12.848(007) | 5 | 12.498(007) | 5 | · | · | · | · |
| 028 | 00:28:02.42 | -01:09:21.58 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 13.261(010) | 4 | 12.975(008) | 5 | 12.510(009) | 5 | · | · | · | · |
| 029 | 00:27:47.38 | -01:13:05.76 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 16.482(023) | 5 | 16.063(016) | 6 | 15.492(022) | 6 | | | | 15.63 |
| 030 | 00:28:06.38 | -01:15:11.20 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 16.239(024) | 15.852(021) | 5 | 15.343(026) | 5 | · | · | · | · | · |
| 031 | 00:28:02.09 | -01:08:55.42 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 16.850(042) | 4 4 | 16.548(043) | | 5 | 16.058(057) | Type 5 | | · | · |
| 032 | 00:28:12.89 | -01:08:33.94 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 15.474(017) | 4 | 15.056(014) | 5 | | 14.456(016) | · | · | · | · |
| 033 | 00:27:45.47 | -01:10:09.93 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 16.803(036) | 4 | 16.356(028) | 4 | 15.893(049) | | | Iax 3 | · | · |
| 034 | 00:28:06.44 | -01:07:07.95 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 15.414(020) | 3 15.050(016) | 4 | | SN 14.467(017) 4 | · | · | · | · | · |
| 035 | 00:28:13.85 | -01:09:28.74 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 15.082(015) | 3 | 14.704(011) | 5 | 14.120(012) 5 | · | | | | |
| 036 | 00:27:51.47 | -01:13:23.41 | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | · · · | 17.227(043) | 5 | 16.855(031) | 16.338(046) | 5 | 6 | | | · | · |
Note. - Uncertainties given in parentheses in thousandths of a magnitude correspond to an rms of the magnitudes obtained on photometric nights.
Stritzinger et al.: The bright and energetic Type Iax SN 2012Z.
Table A2. Revised CSP optical photometry of SN 2005hk in the natural system.
| JD - 2 , 453 , 000+ | u (mag) | g (mag) | r (mag) | i (mag) | B (mag) | V (mag) |
|-----------------------|-------------|-------------|-------------|-------------|-------------|-------------|
| 675.61 | 16.759(016) | 16.759(008) | 16.893(009) | 17.106(012) | 16.830(009) | 16.857(010) |
| 682.58 | 16.265(014) | 15.883(004) | 15.945(005) | 16.170(005) | 15.978(006) | 15.933(006) |
| 683.58 | 16.280(017) | 15.828(006) | 15.878(007) | 16.104(009) | 15.930(006) | 15.871(007) |
| 684.62 | 16.326(018) | 15.801(008) | 15.833(009) | 16.064(010) | 15.920(007) | 15.821(007) |
| 687.6 | 16.541(019) | 15.804(008) | 15.734(008) | 15.943(010) | 15.967(000) | 15.727(009) |
| 689.66 | 16.774(024) | 15.886(009) | 15.701(007) | 15.885(009) | 16.118(009) | 15.761(008) |
| 690.61 | 16.917(023) | 15.958(006) | 15.684(004) | 15.870(005) | 16.185(008) | 15.759(007) |
| 694.65 | 17.619(026) | 16.327(006) | 15.709(006) | 15.779(009) | 16.688(009) | 15.928(008) |
| 695.64 | 17.812(029) | 16.446(007) | 15.726(008) | 15.783(008) | 16.829(008) | 15.976(007) |
| 698.59 | 18.349(041) | 16.827(005) | 15.869(004) | 15.846(005) | 17.286(009) | 16.235(006) |
| 699.62 | 18.688(056) | 16.964(005) | 15.924(004) | 15.875(006) | 17.428(009) | 16.319(007) |
| 702.63 | 19.083(066) | 17.310(007) | 16.108(006) | 16.005(008) | 17.793(014) | 16.580(008) |
| 706.66 | 19.538(067) | 17.650(008) | 16.376(009) | 16.210(009) | 18.166(015) | 16.884(009) |
| 712.57 | 19.896(123) | 17.961(012) | 16.669(009) | 16.495(010) | 18.429(018) | 17.177(011) |
| 720.57 | 20.265(288) | 18.134(015) | 16.966(007) | 16.811(009) | 18.647(030) | 17.468(015) |
| 725.6 | 20.261(117) | 18.272(011) | 17.127(008) | 16.982(010) | 18.766(020) | 17.569(012) |
| 728.61 | · · · | 18.332(009) | 17.202(008) | 17.066(009) | 18.823(022) | 17.663(010) |
| 730.62 | 20.394(121) | 18.309(009) | 17.267(007) | 17.135(008) | 18.797(021) | 17.697(011) |
| 736.57 | 20.508(071) | 18.411(007) | 17.396(007) | 17.289(008) | 18.875(015) | 17.782(009) |
| 739.57 | 20.373(085) | 18.466(008) | 17.475(007) | 17.367(009) | 18.933(015) | 17.839(010) |
| 745.54 | · · · | 18.498(016) | 17.588(011) | 17.500(012) | 18.951(048) | 17.928(019) |
| 748.56 | · · · | 18.528(016) | 17.657(012) | 17.564(013) | 19.034(067) | 17.993(024) |
| 751.55 | · · · | 18.572(011) | 17.733(011) | 17.666(014) | 19.069(028) | 18.031(017) |
| 761.53 | · · · | 18.665(019) | 17.983(010) | 17.841(012) | 19.194(026) | 18.181(013) |
Note. - Values in parentheses are 1 σ measurement uncertainties in millimag.
Stritzinger et al.: The bright and energetic Type Iax SN 2012Z.
Table A3. Revised NIR photometry of SN 2005hk in the natural system.
| JD - 2 , 453 , 000+ | Y (mag) | J (mag) | H (mag) | K s (mag) |
|-----------------------|-------------|-------------|-------------|-------------|
| 676.49 | 16.818(024) | 16.780(031) | 16.929(061) | · · · |
| 677.52 | 16.660(021) | 16.545(028) | 16.759(050) | · · · |
| 678.51 | 16.498(019) | 16.434(027) | 16.593(064) | · · · |
| 680.52 | 16.294(016) | 16.177(024) | 16.390(058) | · · · |
| 681.57 | 16.225(013) | 16.167(012) | 16.280(016) | 16.208(054) |
| 685.6 | 16.024(021) | 15.963(021) | 16.083(038) | · · · |
| 688.63 | · · · | 15.912(020) | 15.907(025) | · · · |
| 691.6 | · · · | 15.908(056) | · · · | · · · |
| 692.58 | · · · | · · · | 15.776(026) | · · · |
| 696.61 | 15.531(016) | 15.888(020) | 15.654(021) | · · · |
| 697.55 | 15.468(013) | 15.956(025) | 15.600(027) | · · · |
| 701.57 | · · · | 15.910(023) | 15.563(021) | · · · |
| 705.53 | 15.525(015) | 15.979(024) | 15.644(021) | · · · |
| 710.61 | 15.628(014) | 16.098(022) | 15.805(026) | · · · |
| 717.59 | 15.855(013) | 16.447(027) | 16.074(042) | · · · |
| 719.6 | 15.906(021) | 16.582(033) | 16.152(047) | · · · |
| 722.57 | 15.999(015) | 16.710(013) | 16.258(016) | 16.545(123) |
| 723.53 | 16.026(020) | 16.814(235) | · · · | · · · |
| 727.56 | 16.182(019) | 16.973(050) | 16.431(059) | · · · |
| 731.54 | 16.322(019) | 17.166(059) | 16.599(063) | · · · |
| 743.57 | 16.746(015) | 17.605(023) | 16.976(036) | · · · |
| 756.55 | 17.184(027) | 17.998(052) | · · · | · · · |
Note. - Values in parentheses are 1 σ measurement uncertainties in millimag.
|
10.1051/0004-6361/201424168
|
[
"M. D. Stritzinger",
"S. Valenti",
"P. Hoeflich",
"E. Baron",
"M. M. Phillips",
"F. Taddia",
"R. J. Foley",
"E. Y. Hsiao",
"S. W. Jha",
"C. McCully",
"V. Pandya",
"J. D. Simon",
"S. Benetti",
"P. J. Brown",
"C. R. Burns",
"A. Campillay",
"C. Contreras",
"F. Forster",
"S. Holmbo",
"G. H. Marion",
"N. Morrell",
"G. Pignata"
] |
2014-08-05T20:00:12+00:00
|
2014-08-05T20:00:12+00:00
|
[
"astro-ph.HE",
"astro-ph.SR"
] |
Comprehensive Observations of the Bright and Energetic Type Iax SN 2012Z: Interpretation as a Chandrasekhar Mass White Dwarf Explosion
|
We present UV through NIR broad-band photometry, and optical and NIR
spectroscopy of Type Iax supernova 2012Z. The data set consists of both early
and late-time observations, including the first late phase NIR spectrum
obtained for a spectroscopically classified SN Iax. Simple model calculations
of its bolometric light curve suggest SN 2012Z produced ~0.3 M_sun of (56)Ni,
ejected about a Chandrasekhar mass of material, and had an explosion energy of
~10^51 erg, making it one of the brightest and most energetic SN Iax yet
observed. The late phase NIR spectrum of SN 2012Z is found to broadly resemble
similar epoch spectra of normal SNe Ia; however, like other SNe Iax,
corresponding visual-wavelength spectra differ substantially compared to all
supernova types. Constraints from the distribution of IMEs, e.g. silicon and
magnesium, indicate that the outer ejecta did not experience significant mixing
during or after burning, and the late phase NIR line profiles suggests most of
the (56)Ni is produced during high density burning. The various observational
properties of SN 2012Z are found to be consistent with the theoretical
expectations of a Chandrasekhar mass white dwarf progenitor that experiences a
pulsational delayed detonation, which produced several tenths of a solar mass
of (56)Ni during the deflagration burning phase and little (or no) (56)Ni
during the detonation phase. Within this scenario only a moderate amount of
Rayleigh-Taylor mixing occurs both during the deflagration and fallback phase
of the pulsation, and the layered structure of the IMEs is a product of the
subsequent denotation phase. The fact that the SNe Iax population does not
follow a tight brightness-decline relation similar to SNe Ia can then be
understood in the framework of variable amounts of mixing during pulsational
rebound and variable amounts of (56)Ni production during the early subsonic
phase of expansion.
|
1408.1094v1
|
## Boundary-law scaling of entanglement entropy in diffusive metals
Andrew C. Potter
Department of Physics, University of California, Berkeley,
CA 94720, USA
Entanglement structure serves as a powerful way to characterize quantum many-body phases. This is particularly so for gapless quantum liquids, where entanglement-based tools provide one of the only means to systematically characterize these complicated phases. For example, the Fermisurface structure of Fermi-liquids is revealed in entanglement entropy by a log-correction to the typical boundary-law scaling of simpler quantum ground-states. In this paper, I analyze the entanglement structure of a disordered, but delocalized diffusive metal. Using a combination of analytic arguments and numerical calculations, I show that, despite having the same number of extended gapless excitations as a clean Fermi-liquid, the diffusive metal exhibits only boundary-law entanglement scaling. This result pinpoints the sharp Fermi-surface structure, rather than the finite density of gapless excitations, as the origin of the log-correction in the Fermi-liquid entanglement scaling.
Introduction -Understanding the structure of entanglement provides a powerful basis for characterizing quantum many-body systems. In ground-states of local Hamiltonians in d -dimensions, the entanglement entropy of a sub region of size L with the rest of the system typically scales like the boundary of the region S L ( ) ∼ L d -1 where d . By contrast excited states of thermal systems have extensive 'volume-law' entropy scaling, S L ( ) ∼ L d . The limited entanglement structure of ground-states enables efficient simulations techniques like DMRG and higher-dimensional tensor-network analogs[1, 2]. Moreover, the distinction between boundary and volume law scaling allows a sharp definition of non-thermal manybody localized phases - for which even excited states exhibit boundary-law behavior[5, 6].
Entanglement scaling can also be used to identify special properties of quantum ground-states. For example gapped topologically ordered phases exhibit a universal sub-leading constant, γ , correction to boundary-law scaling: S L ( ) ∼ aL -γ + . . . , whose value contains information about the topological excitations of the phase[3, 4].
Whereas such gapped topological ground-states are quite well understood, the understanding of gapless quantum phases remains more rudimentary. Here again, arguments based on entanglement have provided a powerful toolset for characterizing gapless phases. For example, entanglement monotonicity properties such as the c-, F-, and atheorems[8-12] place strong constraints on possible phase diagrams[13], allowing one to non-perturbatively rule out various scenarios in contexts where there are limited or no applicable field-theoretic techniques.
Entanglement scaling can also be used to directly distinguish different gapless phases in dimensions higher than one. For example, Fermi-liquids with an extended momentum-space surface of gapless excitations are distinguished from scale invariant quantum critical points with gapless excitations only at a single point by a logarithmically diverging correction to the boundary-law scaling[14-16]:
$$S _ { \text{FL} } ( L ) \approx ( k _ { F } L ) ^ { d - 1 } \log k _ { F } L \text{ \quad \ \ } ( 1 )$$
where k F is the Fermi wave vector. There is also numerical evidence that this log-correction to boundarylaw scaling survives when the quasi-particle nature of the electrons are destroyed resulting in a fractionalized nonFermi liquid with a Fermi-surface worth of non-quasiparticle excitations[17].
Motivated by the fundamental role of entanglement based concepts in our understanding of higherdimensional gapless phases, I analyze the entanglement entropy scaling of a diffusive metal. Here, weak disorder 'erases' the Fermi-surface structure from (disorderaveraged) correlation functions on length-scales longer than the elastic mean-free path glyph[lscript] . However, despite the smearing of momentum space Fermi-surface structure, the system still contains a 'Fermi-surface's worth' of gapless and spatially extended excitations. Hence, it is not a priori obvious whether entanglement entropy in a diffusive metal displays the log-violation to boundary-law scaling characteristic of a Fermi-surface.
In this work, I demonstrate that the log-correction to boundary-law scaling is absent in a diffusive metal, showing that the log-correction requires a sharp Fermi-surface structure[18], and is not simply a property of having a finite density of extended gapless excitations.
To establish this result, I employ a few complementary lines of attack. First, based on the scaling of number fluctuations, I show that all Renyi entropies with Renyi index α ≥ 2 follow a boundary-law. This strongly suggests boundary law behavior for the von-Neumann entropy, which is obtained from the Renyi entropies by analytically continuing α → 1. This prediction is then verified by direct numerical computation for a non-interacting disordered Fermion system. Next, I suggest an intuitive explanation for these results, based on a hydrodynamic description of a Fermi-liquid, that provides insight into the important physical distinction between the entanglement properties of diffusive and ballistic Fermi-liquids.
Lastly, I use a different but related example of a quadratically dispersing semimetal to sharpen the distinction between having a finite density of extended gapless excitations and having a Fermi surface.
Number fluctuation based proof of boundary law Renyi entropies -A non-local quantity, entanglement entropy cannot be completely reconstructed from a finite number of local observables. Nevertheless, in free fermion systems, fluctuations of electron number, N A = ∫ r ∈ A c † r c r , within a sub region A :
$$\sigma _ { N _ { A } } ^ { 2 } & = \langle N _ { A } ^ { 2 } \rangle - \langle N _ { A } \rangle ^ { 2 } = \int _ { r, r ^ { \prime } \in A } \langle c _ { r } ^ { \dagger } c _ { r ^ { \prime } } \rangle \langle c _ { r } c _ { r ^ { \prime } } ^ { \dagger } \rangle \quad \text{
& = \text{tr} M (1 - M ) \quad \text{
& = \tt { t r } M ( 1 - M ) \quad \text{
& \quad \text{(2)} \quad \text{ \text{num} \quad \text{tron} \quad \text{$$
tend to track the scaling of the entanglement between A and its complement. Here M r,r ′ = 〈 c † r c ′ r 〉 with r, r ′ ∈ A is the matrix of two-body correlation functions.
For example, the ground-state of an ordinary insulator is a product state of exponentially well localized orbitals. Then, uncertainty in the number of electrons in a region A comes only from orbitals that straddle the boundary of A , and hence like entanglement exhibits boundary-law scaling. As a second example, in a clean (ballistic) metal, like entanglement entropy, number fluctuations also exhibit a log-correction to boundary law scaling[20] (see also Appendix A):
$$\sigma _ { N _ { A } } ^ { 2 } | _ { \text{ballistic} } \approx \alpha ( k _ { F } L ) ^ { d - 1 } \log ( k _ { F } L ) \quad \ \ ( 3 )$$
where k F is the Fermi wave-vector and α is a numerical constant, and d is the number of spatial dimensions.
Generally, number fluctuations provide only a lower bound for the entanglement entropy:
$$S = - \text{tr} \left [ M \log M + ( 1 - M ) \log ( 1 - M ) \right ] \text{ \quad \ \ } ( 4 ) \text{ \quad,}$$
$$\infty$$
$$= \sigma _ { N } ^ { 2 } + \sum _ { n = 2 } ^ { \infty } \frac { 1 } { n } \text{tr} \left [ M ( 1 - M ) ^ { n } + ( 1 - M ) M ^ { n } \right ] \geq \sigma _ { N } ^ { 2 } \quad \text{so} \\ \text{rad} \\ \text{soft}$$
where tr M = ∑ r ∈ A M r,r . However, the reason the scaling of σ 2 N closely tracks that of S is that for the ground-state of a local quantum Hamiltonian, the majority of eigenvalues { λ j } of M are essentially 0 or 1, and make negligible contribution to both entropy and number fluctuation. The remaining sub-extensive fraction of non-negligible eigenvalues give roughly comparable O (1) contributions to both quantities. Equivalently, in the ground-state of a local Hamiltonian, the entanglement Hamiltonian: H E = log [ ρ -1 A -1 ] has far fewer degrees of freedom than the sub region A itself, where
ρ A = tr A c ρ is the reduced density matrix of A . Hence, both entanglement entropy and number fluctuations are essentially just 'counting' the number of non-negligible eigenvalues λ j , i.e. the number of degrees of freedom in H E , and tend to scale the same way.
In Appendix A, I compute the disorder averaged number fluctuations, σ 2 N A , for a spherical sub region of radius L in a three-dimensional diffusive metal and find that it follows the boundary-law scaling:
$$\overline { \sigma _ { N _ { A } } ^ { 2 } } | _ { \text{diffusive} } \approx \beta \left ( k _ { F } L \right ) ^ { d - 1 } \log ( \gamma k _ { F } \ell ) \quad \ \ ( 5 )$$
where β, γ are numerical constants and glyph[lscript] is the elastic mean-free path. The absence of a log-correction in number fluctuations suggests that the entanglement entropy for a diffusive metal follows boundary law scaling, but does not directly rule out the possibility of a logcorrection for the entanglement entropy.
Stronger evidence for boundary-law scaling of entanglement comes from considering the Renyi entropies:
$$\mathcal { R } _ { \alpha } = \frac { 1 } { 1 - \alpha } \log \text{tr} \rho _ { A } ^ { \alpha }$$
These are closely related to the entanglement entropy, which is recovered in the limit of α → 1. For a free fermion system, the Renyi entropies can be expressed in terms of the two-particle correlation matrix as[7, 19]:
$$\mathcal { R } _ { \alpha } = \frac { 1 } { 1 - \alpha } \text{tr} \log \left [ M ^ { \alpha } + ( 1 - M ) ^ { \alpha } \right ] \quad \ \ ( 7 )$$
I will now show that the number fluctuations bound the (disorder averaged) Renyi entropies for all real values of Renyi index α ≥ 2, i.e. that:
$$\overline { \mathcal { R } _ { \alpha } } = f ( \alpha ) L ^ { d - 1 } + \dots$$
where ( . . . ) indicate subleading in L terms, and f ( α ) is some bounded function of Renyi index, α . It is natural to expect that f ( α ) has a unique analytic continuation to the full interval α ∈ [1 , ∞ ]. In this case, the von-Neumann entanglement entropy is given by, S = f (1) L d -1 + . . . , and also satisfies boundary-law scaling. These expectations will subsequently be checked numerically in the following section to ensure that no unexpected pathology arises in taking the limit of α → 1.
To establish Eq. 8, it is first useful to note the eigenvalues of M with unrestricted indices are just the occupation numbers of the energy eigenmodes of the system, which take values either 0 or 1. Hence, by the Cauchy interlacing theorem, the eigenvalues of M restricted to the sub region A are also bounded between 0 and 1.
With this in mind, let us expand the 2 nd Renyi entropy as a power series in M and (1 -M ):
$$\mathcal { R } _ { 2 } = \sum _ { n = 1 } ^ { \infty } \frac { 2 ^ { n } } { n } \text{tr} M ^ { n } ( 1 - M ) ^ { n } = 2 \text{tr} M ( 1 - M ) + \sum _ { n = 1 } ^ { \infty } \frac { 1 } { n + 1 } \frac { 2 ^ { n + 1 } \text{tr} \left [ M ^ { n + 1 } ( 1 - M ) ^ { n + 1 } \right ] } { \text{tr} M ( 1 - M ) } \leq 2 \sigma _ { N } ^ { 2 } + 2 \log ( 4 / e ) \quad ( 9 )$$
where in the last step I have used 0 ≤ λ j ≤ 1 to bound ∑ j λ n +1 j (1 -λ j ) n +1 ≤ 1 4 n ∑ j λ j (1 -λ j ). Moreover, R 2 bounds all other Renyi entropies with α ≥ 2, since by Jensen's inequality 1 1 -α ∑ j log [ λ α j +(1 -λ j ) α ] is a decreasing function of α . Importantly, since these bounds hold for each given disorder realization, they are also true for disorder-averaged quantities.
Since even gapped systems display boundary law scaling, we must also have R α ≥ f < ( α L ) d -1 . Together with Eq. 9, this establishes Eq. 8 with f < ( α ) ≤ f ( α ) ≤ 2 σ 2 N L d -1 . Having established that all Renyi entropies with α ≥ 2 follow boundary-law scaling, I now turn to a numerical computation of the entanglement entropy.
avoid the exponential buildup of numerical inaccuracies, I employ a repeated orthogonalization procedure[21, 22] (see also Appendix B).
Numerical Results for the Spatial Scaling of Entanglement To confirm the above analytic arguments, I numerically compute the entanglement entropy for noninteracting disordered metal from Eq. 4. Specifically, I consider a square lattice tight-binding model with dimensions L x × L y (in units of the lattice constant). The uniform part of the Hamiltonian consists of nearest neighbor hopping with strength t and chemical potential µ ≈ 0 (measured with respect to the band-center, i.e. near halffilling). Disorder is modeled by a normally distributed random on-site potential with variance W , that is uncorrelated between different lattice sites. Periodic boundary conditions are used to eliminate boundary effects. The entanglement sub-region A is chosen as a cylinder extending from x = 1 to x = L , and which wraps around full system in the y -direction.
To completely evade localization and obtain a asymptotically diffusive metal on all length scales one would need to simulate either a fully three-dimensional system or a spin-orbit coupled two-dimensional system. However, to avoid the computational difficulty of simulating such higher-dimensional systems, I instead simulate quasi-one dimensional systems of length L x much larger than width L y . In a strictly one dimensional system, there is no separation of scales between localization length ξ and mean-free path glyph[lscript] . However for fixed glyph[lscript] , ξ grows with increasing strip width, L y , and by choosing moderate values of L y one can obtain an order of magnitude or more separation between ξ and glyph[lscript] .
The mean-free path glyph[lscript] is estimated perturbatively in disorder strength as glyph[lscript] = 2 πN (0) W 2 where N (0) = 1 √ L y ∑ j, occ k ( j ) F 2 πv ( j ) F is the density of states at the Fermilevel, and k ( j ) F and v ( j ) F are respectively the Fermi wave vector and velocity for the j th sub-band. The localization length is obtained by numerically extracting the smallest Lyapunov exponent of the 1D transfer matrices , ˆ T x of 1 × L y slices of the 2D system at fixed x [21, 22]. To
Representative results are shown in Fig. 1a for glyph[lscript] = 10, for L y = 10 ( ξ ≈ 50) and L y = 24 ( ξ ≈ 100). One clearly sees that the initial log-growth of entanglement expected for a clean system is cutoff by the mean-free path glyph[lscript] such that S L ( ) saturates to a constant for L glyph[greatermuch] glyph[lscript] (see also Fig. 1b for plots of S L ( ) for various glyph[lscript] with L y = 10). The analytic and numerical results show that the entanglement scales like L y log L x for L x glyph[lessmuch] glyph[lscript] , and saturates to a constant for L x glyph[greatermuch] glyph[lscript] . The following function naturally interpolates between these two limits:
$$S ( L ) = \frac { c _ { \text{eff} } } { 3 } \log \frac { 1 } { \sqrt { L ^ { - 2 } + \ell _ { * } ^ { - 2 } } } + \text{const} \quad ( 1 0 )$$
As shown in the inset, the entanglement curves are well fit by the functional form in Eq. 10 with c eff being essentially equal to the number of occupied sub-bands (i.e.
FIG. 1. Numerical results for the scaling of entanglement entropy S E for length L segment of a disordered quasi-1D strip of dimension L x × L y near half-filling ( µ = -0 1 ). . t Each curve is averaged over sufficiently many disorder realizations to reduce the error to glyph[lessorsimilar] 1%. Perturbatively computed mean free path, glyph[lscript] , and numerically determined localization length ξ are indicated by ◦ , and glyph[squaresolid] respectively. Dashed curves show clean glyph[lscript] = ∞ results. Panel a) shows entanglement scaling for L x = 200, L y = 10 24. , The inset shows a fit to cut-off log-scaling of the form (inset) c eff 3 log 1 √ L -2 + glyph[lscript] -2 ∗ + const for L x = 150, L y = 24, yielding glyph[lscript] ∗ ≈ 15 ± 0 6 and . c eff ≈ 11 3 . ± 1 1, consistent . with the perturbatively calculated glyph[lscript] ≈ 10 and the number of occupied sub-bands (12). Panel (b) shows S E ( L ) for various disorder strengths corresponding to glyph[lscript] = 10 20 50 100 and , , , ∞ for L x = 200, L y = 10. The curves saturate at a scale consistent with the perturbatively computed glyph[lscript] .
is the central charge in the clean limit), and glyph[lscript] ∗ in reasonable agreement with the perturbatively estimated glyph[lscript] . Note that little change is seen as the system as L passes through the localization length ξ , as S L ( ) has already saturated at this point (note that in the localized regime L glyph[greatermuch] ξ S L ( ) may not decrease below its asymptotic value in the diffusive regime due to the sub-additivity property of entanglement entropy[25]).
In agreement with the analytic considerations, the direct numerical simulations clearly show an absence of log-corrections to area-law scaling of the ground-state entanglement entropy of a diffusive metal. The subsequent sections aim to provide additional physical intuition into these results.
Intuition from a hydrodynamic description of the Fermi-liquid Further insight into the above results can obtained from a 'hydrodynamic' description of the diffusive metal in terms of bosonic fields describing shape fluctuations of the Fermi-surface[23]. by a 'hydrodynamic' description in terms of bosonic fields[23] describing shape-fluctuations of the Fermi-surface. Consider an isotropic 3D Fermi surface. The low-energy longwavelength excitations of a clean Fermi-liquid can be approximately described by fluctuations of the Fermi-wave vector δk F ( ˆ )( Ω r , t ) that depend on both momentumspace angle ˆ on the Fermi-surface and which vary slowly Ω in real space with characteristic scales L glyph[greatermuch] k -1 F . The Fermi-surface disturbances can be equivalently parameterized by the extra density with momentum along solid angle Ω on the Fermi-surface: ˆ
$$\delta \rho ( \hat { \Omega } ) = \int _ { | k | } \text{sgn} ( k - k _ { F } ) c _ { \mathbf k } ^ { \dagger } c _ { \mathbf k } = \frac { k _ { F } ^ { 2 } } { ( 2 \pi ) ^ { 3 } } \delta k _ { F } ( \hat { \Omega } ) \quad ( 1 1 ) \quad \text{Co}.$$
The non-interacting Hamiltonian then reads: H 0 = 16 π v 4 F k 4 F ∫ d ˆ Ω ρ 2 ˆ Ω . Assuming rotational invariance for simplicity, and introducing angular Harmonics, ρ l,m = ∫ d ˆ Ω Y l,m ( ˆ ) Ω ρ ( ˆ ) Ω gives:
$$H _ { 0 } = \frac { 1 6 \pi ^ { 4 } v _ { F } } { k _ { F } ^ { 4 } } \sum _ { l, m } \rho _ { l, m } ( \hat { \Omega } ) ^ { 2 } \quad \quad ( 1 2 ) \quad \text{ary} \\ \text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text$$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text$$\text$$\text$$\text$$\text$\text$$\text$$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text $\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$
$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$\text$
} . . + - - < </ : "$$
In the presence of random impurities, the overall density mode ρ 0 0 , propagates diffusively:
$$\overline { \langle \rho _ { 0, 0 } ( t, \mathbf r ) \rho _ { 0, 0 } ( 0, 0 ) \rangle } \sim e ^ { - r ^ { 2 } / D | t | } \quad \ \ ( 1 3 ) \quad \text{com}.$$
whereas the remaining modes (for example ρ 1 1 , , which corresponds to net electrical current along the z -direction) decay exponentially in both space and time
$$\overline { \langle \rho _ { l, m } ( t, { \mathbf r } ) \rho _ { l, m } ( 0, 0 ) \rangle } \sim e ^ { - r / \ell _ { l, m } } e ^ { - t / \tau _ { l, m } } \quad ( 1 4 ) \quad 1 D \, \epsilon$$
Here ( . . . ) indicates average over disorder, D is the diffusion constant, and τ l,m ≈ glyph[lscript] l,m /v F are disorder induced relaxation times, which are equal for point-like disorder but in general can depend on ( l, m ).
Therefore, in a loose sense, that the disorder averaged system contains only a single propagating hydrodynamic mode with dynamical exponent z = 2, which in dimensions d > 1 contributes only area law contribution to entanglement (plus possible subleading corrections). By contrast, the clean system contains an infinite number of ballistically propagating hydrodynamic modes each contributing a logarithmically diverging entanglement entropy that add up to give L d -1 log L entanglement scaling. The hydrodynamic picture provides an intuitive, albeit slightly hand-waving, understanding for the fact that a diffusive metal has only area-law contributions to entanglement entropy. In particular, whereas the above methods apply only for free fermion systems, the hydrodynamic picture explained here is unchanged by electronelectron interactions, suggesting that interacting diffusive metals will also satisfy boundary-law entanglement scaling.
## Distinction Between Finite Density of Extended
States and a Fermi-Surface In a diffusive metal, I have shown that disorder removes log-corrections to boundary law entanglement scaling on the same length scale, glyph[lscript] , as it smears the Fermi-surface structure. This strongly suggests that such log-corrections are a unique entanglement signature of a sharp Fermi-surface, and are not produced simply by having a large number of extended gapless excitations. In this section, I provide further support for this picture by noting a second example of a metallic system with finite density of extended gapless excitations but no Fermi-surface, and which exhibits boundary law scaling with no log-correction.
Consider a (clean) two-dimensional semimetal, such as bilayer graphene, with a quadratic band touching (QBT) described by the one-body Hamiltonian:
$$H _ { \text{QBT} } ( p _ { x }, p _ { y } ) = \frac { 1 } { 2 m ^ { * } } \begin{pmatrix} 0 & p _ { - } ^ { 2 } \\ p _ { + } ^ { 2 } & 0 \end{pmatrix} \quad \ \ ( 1 5 )$$
where p ± = p x ± ip y , and m ∗ is the effective band mass. For concreteness, consider a system with periodic boundary conditions, infinite extent in the x-direction, and large, but finite size L y in the y-direction. Let us compute the entanglement between the cylindrical sub region: A = { 0 < x < L, | | ∀ y } and its complement. Due to the y -direction translation symmetry of A H , QBT decomposes into O ( L y ) transverse sub-channels with fixed
$$\text{which} \quad p _ { y } \in \{ \frac { 2 \pi n } { L _ { y } } \} \text{ and dispersion: } E _ { n } ( p _ { x } ) = \sqrt { \left ( \frac { p _ { x } ^ { 2 } } { 2 m ^ { * } } \right ) ^ { 2 } + \Delta _ { n } ^ { 2 } },$$
where ∆ n = π n 2 2 2 m L ∗ 2 y . The entanglement of A with the rest of the system is just the sum of the corresponding 1D entanglement entropies for each sub-band:
$$\stackrel { \sim } { S } ( L _ { x }, L _ { y } ) & = \sum _ { n } ^ { \cdot } S _ { 1 \text{D} } ( L _ { x }, \Delta _ { n } ) \\ & \approx \frac { \sqrt { 2 m ^ { * } } } { \pi } L _ { y } \int \frac { d \Delta } { \sqrt { \Delta } } S _ { 1 \text{D} } ( L _ { x }, \Delta ) \quad \ ( 1 6 )$$
While the p y = 0 channel is gapless, it actually just decomposes into completely filled and empty bands that happen to touch at p x = 0, and hence gives only boundary-law contributions to S . Moreover, all of the sub-channels with non-zero p y are gapped onedimensional systems and also should give only boundarylaw contributions. To shore up this argument, one needs to ensure that there are no anomalous contributions from vicinity of small but non-zero p y . For this purpose, it is sufficient to focus on the regime of small ∆, where all length scales, L x and 1 √ m ∗ ∆ , are much longer than the lattice scale, a , and scaling arguments (see also Ref. [20]) may be applied.
Specifically, consider the contribution to Eq. 16 from ∆ ∈ [0 , ∆ ] where ∆ 0 0 is any suitable cutoff satisfying 1 m a ∗ 2 glyph[greatermuch] ∆ 0 glyph[greatermuch] 1 m L ∗ 2 x . For simplicity let us choose units such that 2 m /π ∗ 2 = 1 to drop the length-independent prefactor in Eq. 16. Then, at low energies, scale invariance of the theory with dynamical exponent z = 2 requires that: S 1D ( L , x ∆) = S (reg) 1D + S (∆ L 2 ) where S (reg) 1D is a non-universal non-singular piece from short-distance physics which may contain at most area law contributions, and S ( x ) is a universal scaling function containing long-wavelength contributions. Then the derivative of the long-wavelength contributions to S from ∆ glyph[lessmuch] ∆ is: 0
$$\frac { \partial S _ { \text{LW} } ( L _ { x }, L _ { y } ) } { \partial L _ { x } } \approx L _ { y } \int d \Delta \ \mathcal { S } ^ { \prime } ( x ) \approx \frac { L _ { y } } { L _ { x } ^ { 2 } } \int d x \ x \mathcal { S } ^ { \prime } ( x ) \quad \ \ [ 4 ] { \ A } \\ ( 1 7 ) \quad \ \ \ \ t h \\ [ 5 ] { \ B }$$
Since S ( x ) is the entanglement entropy of a region of size x the gapped 1D system with Hamiltonian H n (with length measured in units ξ ∆ ) it must follow an boundary law, i.e. S ( x ) ≈ const + O ( 1 x ) . Hence the integrand x S ′ ( x ) ≤ O ( 1 x ) , and ∂S ∂L x ≤ √ L y L 2 x ∫ dx x ≈ L y √ ∆ log 0 L x L 2 x where I have used x max = ∆ 0 L x as a short distance cutoff. Since the 1 L y ∂S LW ∂L y vanishes in the limit of L x → ∞ , S LW contains at most area-law contributions ∼ const × L y .
This demonstrates that, like the diffusive metal, the quadratically dispersing semimetals follow boundary-law entanglement scaling, lending further support to idea that an extended Fermi-surface rather than just finite density of states is required to produce a log-violation of boundary-law scaling.
Discussion In this work, I have established that diffusive metals obey boundary-law scaling of entanglement entropy, in contrast to clean metals with a Fermi-surface, which exhibit a characteristic log-violation of boundary law scaling. While the calculations undergone here are valid for non-interacting fermions, the hydrodynamic picture explained above suggests that these results to hold also for interacting diffusive metals.
In conjunction with general monotonicity requirements for entanglement entropy, these results could possibly be used to constrain the entanglement scaling and possible phase diagrams for phases that are continuously connected to a diffusive metal. To this end, a potentially interesting subject for future work would be to understand where there are any universal structure to the sub-leading corrections, e.g. analogous to the subleading constant piece, F , of 2D conformal systems (see however [26]), and if so whether they satisfy analogous monotonicity properties under renormalization group flows.
Acknowledgements I would like to thank Ashvin Vishwanath, Romain Vasseur, Siddharth Parameswaran, and especially Brian Swingle for helpful conversations. This work was supported by the Gordon and Betty Moore Foundation.
- [1] see e.g. U. Sch¨llwock o The density-matrix renormalization group in the age of matrix product states. U. Schollwck, Ann. Phys. (N.Y.) 326 , 96 (2011) and references therein.
- [2] see e.g. J.I. Cirac, F. Verstraete J. Phys. A: Math. Theor. 42 , 504004 (2009) and references therein
- [3] M. Levin and X.-G. Wen Detecting Topological Order in a Ground State Wave Function Phys. Rev. Lett. 96 , 110405 (2006).
- [4] A. Kitaev and J. Preskill Topological Entanglement Entropy Phys. Rev. Lett. 96 , 110404 (2006).
- [5] B. Bauer, C. Nayak, Area laws in a many-body localized state and its implications for topological order. J. Stat. Mech. 2013 , P09005 (2013)
- [6] B. Swingle, A simple model of many-body localization. arXiv:1307.0507 (2013).
- [7] H.F. Song, S. Rachel, C. Flindt, I. Klich, N. Laflorencie, and K. Le Hur, Bipartite fluctuations as a probe of manybody entanglement. Phys. Rev. B. 85 035409 (2012).
- [8] A. B. Zamolodchikov, JETP Lett. 43, 730 (1986) [Pisma Zh. Eksp. Teor. Fiz. 43, 565 (1986)].
- [9] H. Casini and M Huerta, A c-theorem for entanglement entropy. J. High Energy Phys. 40 , 7031 (2007).
- [10] H. Casini and M Huerta, Renormalization group running of the entanglement entropy of a circle. Phys. Rev. D 85 , 125016 (2012)
- [11] J.L. Cardy, Is There a c Theorem in Four-Dimensions? Phys. Lett. B 215 749 (1988).
- [12] Z. Komargodski and A. Schwimmer, On Renormalization Group Flows in Four Dimensions arXiv:1107.3987
- [13] T. Grover Chiral Symmetry Breaking, Deconfinement, and Entanglement Monotonicity Phys. Rev. Lett. 112 , 151601 (2014).
- [14] D. Gioev and I. Klich, Entanglement Entropy of Fermions in Any Dimension and the Widom Conjecture. Phys. Rev. Lett. 96 , 100503 (2006)
- [15] W. Li, L. Ding, R. Yu, T. Roscilde, and S. Haas, Scaling behavior of entanglement in two- and three-dimensional free-fermion systems. Phys. Rev. B 74 , 073103 (2006).
- [16] B. Swingle, Entanglement Entropy and the Fermi Surface. Phys. Rev. Lett. 105 , 050502 (2010)
- [17] Y. Zhang, T. Grover, A. Vishwanath Entanglement en-
tropy of critical spin liquids Phys. Rev. Lett. 107 , 067202 (2011).
- [18] Here I define 'sharp' Fermi surface as having a continuous momentum space surface, for example the 2 k F components of density associated with Friedel oscillations, at which certain correlation functions exhibit singularities. This definition includes non-Fermi liquids that lack coherent quasi-particles but which may nevertheless have a 'sharp' Fermi surface structure in the sense defined here, and which also exhibit log-corrections to boundary law scaling[17].
- [19] I. Peschel and V. Eisle, Reduced density matrices and entanglement entropy in free lattice models. J. Phys. A: Math. Theor. 42 , 504003 (2009)
- 87 , 045123 (2013).
- [21] A. MacKinnon and B. Kramer One-Parameter Scaling of Localization Length and Conductance in Disordered Systems. Phys. Rev. Lett. 47 , 1546 (1981).
- [22] J L Pichard and G Sarma Finite size scaling approach to Anderson localisation. J. Phys C: Solid State Phys. 14 , L127 (1981).
- [23] A.H. Castro Neto and E. Fradkin Bosonization of the low energy excitations of Fermi liquids. Phys. Rev. Lett. 72 , 1393 (1994).
- [24] E. H. Lieb, M. B. Ruskai, J. Math. Phys. 14, 1938 (1973).
- [25] T. Grover Certain General Constraints on the ManyBody Localization Transition. arXiv:1405.1471
- [20] B. Swingle and T. Senthil, Universal crossovers between entanglement entropy and thermal entropy. Phys. Rev. B
- [26] B. Swingle Entanglement does not generally decrease under renormalization. arXiv:1307.8117 (2013).
## Appendix A - Computation of number fluctuations in ballistic and diffusive metals
In this appendix I compute number fluctuations in a sub region of size L for a ballistic metal and a diffusive metal. Consider a spherical sub region, A , of radius L . The variance of electron number in A is given by:
$$\sigma _ { N _ { A } } ^ { 2 } & = \langle N _ { A } ^ { 2 } \rangle - \langle N _ { A } \rangle ^ { 2 } = \int _ { r, r ^ { \prime } \in A } \langle c _ { r } ^ { \dagger } c _ { r ^ { \prime } } \rangle \langle c _ { r } c _ { r ^ { \prime } } ^ { \dagger } \rangle \\ & = \int _ { r, r ^ { \prime } \in A } \int _ { k, \omega, k ^ { \prime }, \omega ^ { \prime } } e ^ { i k \cdot ( r - r ^ { \prime } ) } e ^ { i k ^ { \prime } \cdot ( r ^ { \prime } - r ) } A _ { k, \omega } A _ { k ^ { \prime }, \omega ^ { \prime } n _ { F } } ( \omega ) ( 1 - n _ { F } ( \omega ^ { \prime } ) )$$
where A k,ω = g R ( k,ω ) -g A ( k,ω ) 2 πi , and g R/A are the retarded/advanced one-electron Green functions.
$$\sigma _ { N _ { A } } ^ { 2 } = \int _ { q, \omega, \omega ^ { \prime } } & n _ { F } ( \omega ) ( 1 - n _ { F } ( \omega ^ { \prime } ) ) \left | \int _ { r < L } e ^ { i q \cdot r } \right | ^ { 2 } \times \\ & \times \left ( \frac { 1 } { 2 \pi ^ { 2 } } \right ) \int _ { k } \left ( g ^ { R } ( \omega, k + q ) g ^ { A } ( \omega ^ { \prime }, k ) + g ^ { A } ( \omega, k + q ) g ^ { R } ( \omega ^ { \prime }, k ) \right )$$
where I have used that ∫ k g R R g = 0 = ∫ k g A A g .
FIG. 2. Diagrammatic representation number fluctuations for a ballistic clean metal (a) and the corresponding disorder averaged quantity in a diffusive metal (b). Solid lines indicate bare Green functions (a) or disorder averaged Green functions (b). Encircled A vertices indicate integration over points within the region A : { r : | r | < L } . The dashed line with an × in the middle indicates scattering from an impurity.
## A.1 - Number fluctuations in a clean ballistic metal
In a clean metal: g R/A = ( ω -ε k ± i 0 + ) -1 . Linearizing ε k near the Fermi-surface gives:
$$\lambda _ { k } ^ { \quad \lambda } ( \omega, k + q ) g ^ { A } ( \omega ^ { \prime }, k ) & \approx \int d \hat { \Omega } _ { k } \ \nu ( 0 ) \int d \xi \frac { 1 } { \omega - \xi - v _ { F } \cdot q + i 0 ^ { + } } \frac { 1 } { \omega ^ { \prime } - \xi - i 0 ^ { + } } \\ & = \nu ( 0 ) \int d \hat { \Omega } _ { k } \ \frac { 2 \pi i } { \omega - \omega ^ { \prime } + v _ { F } \cdot q + i 0 ^ { + } }$$
where ν (0) ≈ k 2 F 2 πv F is the density of states at the Fermi-surface, and k F , v F are respectively the Fermi wave vector and velocity.
Then I find:
$$& \text{Then I find} \, \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\quad & \int _ { k } \left ( g ^ { R } ( \omega, k + q ) g ^ { A } ( \omega ^ { \prime }, k ) + g ^ { A } ( \omega, k + q ) g ^ { R } ( \omega ^ { \prime }, k ) \right ) = \int _ { k } \left ( g ^ { R } ( \omega, k + q ) g ^ { A } ( \omega ^ { \prime }, k ) + c. c \right ) \\ & = 4 \pi ^ { 2 } \nu ( 0 ) \int d \hat { \Omega } _ { k } \delta \left ( \omega - \omega ^ { \prime } + v _ { F } \cdot q \right ) = 4 \pi ^ { 2 } \nu ( 0 ) \int _ { - 1 } ^ { 1 } \frac { d \mu } { 2 } \delta \left ( \omega - \omega ^ { \prime } + v _ { F } q \mu \right ) \\ & = \frac { 2 \pi ^ { 2 } \nu ( 0 ) } { v _ { F } q } \theta \left ( 1 - \frac { | \omega - \omega ^ { \prime } | } { v _ { F } q } \right ) \\ & \text{which, upon performing the frequency integrals becomes} \, \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\quad & \text{with, upon performing the frequency integrals becomes} \, \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
which, upon performing the frequency integrals becomes:
$$\int _ { \omega, \omega ^ { \prime } } n _ { F } ( \omega ) ( 1 - n _ { F } ( \omega ^ { \prime } ) ) \theta \left ( 1 - \frac { | \omega - \omega ^ { \prime } | } { v _ { F } q } \right ) = \frac { 1 } { 4 \pi ^ { 2 } } \int _ { \omega ^ { \prime } = 0 } ^ { v _ { F } q } \int _ { \omega = - v _ { F } q } ^ { \omega ^ { \prime } - v _ { F } q } = \frac { 1 } { 4 \pi ^ { 2 } } \int _ { \omega ^ { \prime } = 0 } ^ { v _ { F } q } \omega ^ { \prime } = \frac { 1 } { 8 \pi ^ { 2 } } ( v _ { F } q ) ^ { 2 }$$
Inserting these expressions into Eq. 18 gives:
$$\sigma _ { N _ { A } } ^ { 2 } = \int _ { q } \left | \int _ { r < L } e ^ { i q \cdot r } \right | ^ { 2 } \frac { \nu ( 0 ) } { 4 } v _ { F } q$$
The Fourier-transformed projection into the sub region evaluates to:
$$\int _ { r < L } d ^ { 3 } r e ^ { i q \cdot r } = \int _ { r < L } r ^ { 2 } d r \frac { \sin q r } { q r } = \frac { \sin q L - q L \cos q L } { q ^ { 3 } }$$
Assembling these expressions gives:
$$\sigma _ { N _ { A } } ^ { 2 } = \frac { \nu ( 0 ) } { 4 } \int _ { q } \left ( \frac { \sin q L - q L \cos q L } { q ^ { 3 } } \right ) ^ { 2 } v _ { F } q$$
In the limit of large sub region size, L →∞ , the number fluctuations read:
$$\sigma _ { N _ { A } } ^ { 2 } \approx \frac { v _ { F } \nu ( 0 ) } { 8 \pi } \ L ^ { 2 } \int d q \frac { \cos ^ { 2 } q L } { q } \approx \alpha ( k _ { F } L ) ^ { 2 } \log k _ { F } L$$
where α is a numerical constant, and I have used the Fermi-momentum as a short-distance cutoff, remembering that our expressions were obtained by linearizing the electron dispersion near the Fermi-surface. Note that, like the entanglement entropy, the number fluctuations scale like ( k F L ) 2 log( k F L ). The extra factor of log( k F L ) comes from a log-divergence at small wave-vectors q , which is cut-off only by the finite size of the sub region under consideration. In the next section, I perform the analogous disorder-averaged calculation for the diffusive metal and find that there is no such long-wavelength singularity in that case, and correspondingly no log correction to the boundary law scaling of number fluctuations.
## A.2 - Number fluctuations in a disordered diffusive metal
In the presence of impurity scattering, the disorder averaged Green's functions develop an elastic lifetime τ :
$$\bar { g } ^ { R / A } = \frac { 1 } { \omega - \varepsilon _ { k } \pm \frac { i } { 2 \tau } \pm i 0 ^ { + } }$$
To obtain the disorder averaged scaling of number fluctuations, however, one must disorder average the entire Eq. 18 (shown diagrammatically in Fig. 2b).
For simplicity, let us assume a density n imp of impurities with a short-range potential approximated by a delta-function of strength V 0 (related to the elastic scattering rate by τ -1 = 2 πν (0) V 2 0 n imp ). In the limit ε F τ glyph[greatermuch] 1 the disorder averaged product of Green functions Π(Ω , q ) = ∫ k g R ( ω +Ω , k + ) q g A ( ω, k ), can be written as a geometric series: Π = Π + 0 V 2 0 n imp Π = Π 0 1 -V 2 0 n imp Π 0 in the product of disorder averaged Green functions:
$$\Pi _ { 0 } ( \Omega, q ) & = \int _ { k } \overline { g } ^ { R } ( \omega + \Omega, k + q ) \overline { g } ^ { A } ( \omega, k ) = \nu ( 0 ) \int d \xi \int _ { - 1 } ^ { 1 } \frac { d \mu } { 2 } \frac { 1 } { \omega + \Omega - \xi - v _ { F } q \mu + \frac { i } { 2 \tau } } \frac { 1 } { \omega - \xi - \frac { i } { 2 \tau } } \\ & = \nu ( 0 ) \int _ { - 1 } ^ { 1 } \frac { d \mu } { 2 } \frac { 2 \pi i } { \Omega + v _ { F } q \mu + \frac { i } { \tau } } \approx \frac { 2 \pi i \nu ( 0 ) } { \Omega + i D q ^ { 2 } + \frac { i } { \tau } } \\ & \quad \quad \quad$$
Where, the last line is valid in the limit of Ω , Dq 2 glyph[lessmuch] τ .
Summing the geometric series gives the familiar diffusive form of Π:
$$\Pi ( \Omega, q ) = \Pi _ { 0 } + V _ { 0 } ^ { 2 } n _ { \text{imp} } \Pi = \frac { \Pi _ { 0 } } { 1 - V _ { 0 } ^ { 2 } n _ { \text{imp} } \Pi _ { 0 } } = \frac { 2 \pi i \nu ( 0 ) } { \Omega + i D q ^ { 2 } }$$
The real part of Π is then:
$$\Pi + c. c. = \frac { 2 \pi \nu ( 0 ) D q ^ { 2 } } { \left ( \omega - \omega ^ { \prime } \right ) ^ { 2 } + \left ( D q ^ { 2 } \right ) ^ { 2 } }$$
Inserting this expression into Eq. 18 gives:
$$\overline { \sigma _ { N _ { A } } ^ { 2 } } \approx \int _ { \omega < 0, \omega ^ { \prime } > 0 } \int _ { q } \left ( \frac { \sin q L - q L \cos q L } { q ^ { 3 } } \right ) ^ { 2 } \Pi ( \omega - \omega ^ { \prime }, q )$$
For the clean metal, the log-correction to area law came from a long-wavelength (i.e. infra-red or IR) divergence from the small q , which was cut off by the finite size of the sub region when q glyph[lessorsimilar] 1 L . Let us examine the IR behavior coming from ω, Dq 2 glyph[lessmuch] τ -1 , in the limit of L →∞ :
$$q \lesssim \frac { 1 } { L }. \text{ Let us examine the IR behavior coming from } \omega, D q ^ { * } \ll \tau ^ { - * }, \text{ in the limit of } L \to \infty \colon \\ \overline { \sigma _ { N _ { A } } ^ { 2 } } | _ { \underline { t } \gg q \gg \underline { t } } \, \text{ Regime} \, \varepsilon \, \nu ( 0 ) \int _ { \omega < 0, \omega ^ { \prime } > 0 } \int q ^ { 2 } d q \frac { L ^ { 2 } } { q ^ { 4 } } \frac { D q ^ { 2 } } { ( \omega ^ { \prime } - \omega ) ^ { 2 } + ( D q ^ { 2 } ) ^ { 2 } } \\ \approx \nu ( 0 ) D L ^ { 2 } \int d q \int _ { 0 } ^ { \infty } \, _ { d \omega } \frac { 1 } { D q ^ { 2 }$$
Here I have introduced the elastic mean-free path glyph[lscript] = v F τ , and glyph[lscript] -1 and τ -1 serve as short wavelength and high frequency cutoffs to the momentum and frequency integrals respectively since the diffusive form of Π applies only for ω, Dq 2 glyph[lessmuch] τ -1 . For the ballistic regime, ω, Dq 2 glyph[greatermuch] τ -1 , disorder is unimportant and the contributions from this regime coincide with those of a clean-metal computed in the previous section (but with IR cutoff glyph[lscript] ).
Piecing together the contributions from the diffusive and ballistic regimes, gives number fluctuations scaling as:
$$\overline { \sigma _ { N _ { A } } ^ { 2 } } \approx L ^ { 2 } \left [ c _ { 1 } \frac { \nu ( 0 ) D } { \ell } + c _ { 2 } k _ { F } ^ { 2 } \log k _ { F } \ell \right ] \approx \beta \left ( k _ { F } L \right ) ^ { 2 } \log ( \gamma k _ { F } \ell )$$
where β, γ, c 1 2 , are numerical constants. Hence, the number fluctuations of the diffusive metal do not have a log-correction to area law scaling.
The Lyapunov exponents are given by the log of the eigenvalues of the limiting matrix:
$$\hat { T } _ { \infty } = \lim _ { N \to \infty } \left ( \prod _ { j = 1 } ^ { N } \hat { T } _ { x } \right ) ^ { 1 / N }$$
The smallest Lyapunov exponent corresponds to the inverse localization length. However, as written, Eq. 34 is unsuitable for direct numerical computation, since repeated products of transfer matrices are dominated by the largest Lyapunov exponent, resulting in rapid (exponential in N ) accumulation of numerical errors for the localization length.
I employ the standard trick[21, 22] of periodic orthogonalization to avoid this numerical instability. This approach begins by choosing a block size of N ∼ 10 sites and computing: ˆ T 1 = ∏ N x =1 ˆ . T x A QR-decomposition of ˆ T (1) = Q (1) R (1) is then performed, where R is an upper right-triangular matrix whose diagonal entries contain the eigenvalues of ˆ T (1) (sorted from highest to lowest) and Q (1) is an orthogonal matrix. The Lyapunov exponents are then computed iteratively as follows: define ˆ T ( j +1) = ∏ ( N +1) j x = Nj +1 Q ( j ) . Here Q ( j ) is the orthogonal matrix from the QR composition: ˆ T ( j ) = Q ( j ) R ( j ) . In this procedure, as j is increased, the k th column of Q ( j ) is increasingly oriented along the eigenvector of ˆ T ∞ with k th largest eigenvalue. This approximate orthogonalization reduces the random contamination of small Lyapunov exponents by larger ones in subsequent steps.
The Lyapunov exponents are then given by λ n = lim j →∞ 1 Nj ∑ j log( R ( j ) ) n,n . The localization length is then the inverse of the smallest Lyapunov exponent ξ = 1 min n λ n . In practice, a large number glyph[greaterorsimilar] 10 5 of iterations are required to determine localization length with accuracy of ≈ 1%.
| null |
[
"Andrew C. Potter"
] |
2014-08-05T20:00:14+00:00
|
2014-08-05T20:00:14+00:00
|
[
"cond-mat.str-el",
"cond-mat.dis-nn",
"quant-ph"
] |
Boundary-law scaling of entanglement entropy in diffusive metals
|
Entanglement structure serves as a powerful way to characterize quantum
many-body phases. This is particularly so for gapless quantum liquids, where
entanglement-based tools provide one of the only means to systematically
characterize these complicated phases. For example, the Fermi-surface structure
of Fermi-liquids is revealed in entanglement entropy by a log-correction to the
typical boundary-law scaling of simpler quantum ground-states. In this paper, I
analyze the entanglement structure of a disordered, but delocalized diffusive
metal. Using a combination of analytic arguments and numerical calculations, I
show that, despite having the same number of extended gapless excitations as a
clean Fermi-liquid, the diffusive metal exhibits only boundary-law entanglement
scaling. This result pinpoints the sharp Fermi-surface structure, rather than
the finite density of gapless excitations, as the origin of the log-correction
in the Fermi-liquid entanglement scaling.
|
1408.1096v1
|
## Model Realization and Numerical Studies of a Three-Dimensional Bosonic Topological Insulator and Symmetry-Enriched Topological Phases
Scott D. Geraedts and Olexei I. Motrunich
Department of Physics, California Institute of Technology, Pasadena, California 91125, USA
(Dated: August 7, 2014)
We study a topological phase of interacting bosons in (3+1) dimensions which is protected by charge conservation and time-reversal symmetry. We present an explicit lattice model which realizes this phase and which can be studied in sign-free Monte Carlo simulations. The idea behind our model is to bind bosons to topological defects called hedgehogs. We determine the phase diagram of the model and identify a phase where such bound states are proliferated. In this phase we observe a Witten effect in the bulk whereby an external monopole binds half of the elementary boson charge, which confirms that it is a bosonic topological insulator. We also study the boundary between the topological insulator and a trivial insulator. We find a surface phase diagram which includes exotic superfluids, a topologically ordered phase, and a phase with a Hall effect quantized to one-half of the value possible in a purely two-dimensional system. We also present models that realize symmetry-enriched topologically-ordered phases by binding multiple hedgehogs to each boson; these phases show charge fractionalization and intrinsic topological order as well as a fractional Witten effect.
## I. INTRODUCTION
The study of topological phases of matter has been a major component of condensed matter research in recent decades. Amongthe many phases studied, the topological insulator (TI) is one of the most prominent. 1,2 The TI is a three-dimensional phase of free fermions. Though it is insulating in the bulk, its topological behavior can be deduced from its unusual surface properties, in particular the odd number of Dirac cones it has on its surface. The topological insulator is an example of a symmetry-protected topological phase (SPT). Like all SPT's, it has short-ranged entanglement, which implies that it has only conventional excitations in the bulk and a unique ground state on any closed manifold. This is in contrast to intrinsically topologically ordered states like the fractional quantum Hall states. The relevant symmetries for the topological insulator are charge conservation and time-reversal, and if either of these symmetries is violated, the phase loses its topological properties.
One obvious extension of research into topological insulators is to consider the effects of interactions on their properties. This is however a difficult task. Many of the methods used to study TI's involve the properties of their band structure, and these methods obviously do not apply if interactions are strong. As an introduction to this difficult problem, one can try to study an analog of the topological insulator, constructed of interacting bosons instead of fermions. In bosonic systems we know that the non-interacting case would be a condensate, so we can be sure that the topological behavior is due to the interactions. In addition, certain theoretical techniques, like the Monte Carlo studies employed in this paper, work mainly for bosonic systems.
The study of topological phases of interacting bosons is relatively recent, but much progress has been made. 3-13 Chen, Liu, Gu, and Wen 3 have used group cohomology theory to determine which symmetries and dimensions can lead to nontrivial topological phases. However, this approach tells us little about the properties of these phases, which must be determined through other methods. One well-studied case is that of SPT phases with U (1) symmetry in two dimensions. Lu and Vishwanath 4 proposed a phenomenological Chern-Simons field theory to describe both the bulk and edges of these states and showed that such states have a Hall effect quantized to an even integer (in units of e /h 2 ) and possess gapless counter-propagating edge modes. Therefore such states are called 'bosonic integer quantum Hall phases'. Senthil and Levin 5 proposed how to realize such phases in a quantum Hall-like setting by starting with two species of bosons in a magnetic field and using mutual flux attachment. One drawback of the flux attachment technique is that it is difficult to relate it precisely to a microscopic model. To address this, in our work 7 we provided an alternative exactly solvable model that replaces flux attachment by a (precisely formulated) dynamical binding of bosons to vortices. We studied this model in Monte Carlo and confirmed that it realizes the integer Quantum Hall phase of bosons, including observation of the gapless edge modes. We also showed that by binding multiple vortices to bosons we can realize SymmetryEnriched Topological phases (SET) 14,15 with long-ranged entanglement.
In this paper we focus on understanding interacting bosonic topological phases which are analogs of the electronic topological insulator in three dimensions. Motivated by the formal cohomology results, 3 Vishwanath and Senthil in an inspirational paper 6 found effective field theories which can describe both the bulk and the surface of a three-dimensional 'bosonic topological insulator' with charge conservation and time-reversal symmetry. They found exotic behavior on the surface which can be used to assert the topological behavior in the bulk.
In particular, Vishwanath and Senthil found three kinds of exotic phases on the surface of the bosonic TI. These surface phases cannot exist in a purely two-dimensional system, and their existence on the surface of a three-dimensional system shows that the system is topological. The first kind of phase is a superfluid, which spontaneously breaks charge conservation symmetry. There are actually several different types of superfluids which can exist on the surface. The gapped vortex excitations in these superfluids have properties which cannot
exist in a purely two-dimensional system. Phase transitions between the different superfluids are predicted to be deconfined critical points. Another kind of surface phase appears when we break time-reversal symmetry on the surface. This phase has a Hall conductivity quantized to one-half of the elementary value possible in a purely two-dimensional bosonic system. Since the Hall conductivity in the bosonic integer quantum Hall effect is quantized to even integers, this surface phase is expected to have an odd integer Hall conductivity. Finally, it is possible to have a surface phase which breaks no symmetries but has a symmetry-enriched intrinsic topological order of a kind impossible in a purely two-dimensional system with these symmetries.
In both the two- and three-dimensional cases, the topological behavior can be thought of as coming from the binding of bosons to point topological defects, and the condensation of such objects. In our construction in two dimensions the topological defects were vortices, and the binding is essentially an exact dynamical realization of the flux attachment. In the three-dimensional systems that we will study here, the topological defects are hedgehogs which we introduce by adding an additional SO (3) 'spin' symmetry, and we construct a bosonic TI by binding hedgehogs of this symmetry to the bosons. 6 In an independent work, Metlitski and Fisher also produced a construction which explicitly binds hedgehog-like 'monopole' objects to bosons without enlarging the continuous symmetry and showed that it gives the three-dimensional bosonic TI, 16 while the present setting is a bit simpler to analyze and is amenable to Monte Carlo studies.
In this work, we construct explicit models which realize the interacting bosonic analog of the topological insulator. The models have both charge conservation U (1) symmetry and a time reversal Z T 2 symmetry to be discussed in the main text. We present two different models which realize this physics. In the first model, described in Section II, the spins are represented by SO (3) degrees of freedom with Heisenberg interactions. We introduce a term in our action which energetically binds hedgehogs to bosons, and we show that this term can lead to a phase (which we call the 'binding phase') where these bound states are proliferated.
In order to show that the binding phase is indeed the bosonic topological insulator, we will attempt to find the phases predicted by Vishwanath and Senthil on its surface. Initially we find only superfluids on the surface. In our Monte Carlo simulations, we do not have simple access to properties of the gapped vortex excitations in these surface phases. Therefore we cannot determine whether the superfluids we observe are the exotic superfluids predicted by Vishwanath and Senthil, or more conventional superfluids. We find that the surface superfluids in our model are connected by a direct transition. If this transition were second-order, then it could be the predicted deconfined critical point. However, we cannot access large enough system sizes to determine the order of the observed phase transition.
We can try to find other exotic surface phases by explicitly breaking the Z T 2 symmetry of our model on the surface. If the bulk of the system is in a topological phase, this will lead to a Hall conductivity quantized to an odd integer. In order to measure this Hall conductivity in our system, we need to introduce external gauge fields. In the model in Section II, this is complicated by our definition of the hedgehog number, which is a discontinuous function of the Heisenberg spins. In this case, we do not know how to properly couple to the external gauge fields, which prevents us from measuring the surface Hall conductivity.
To remedy this, in Section III we introduce a second model which binds hedgehogs to bosons. In this model, we represent the spins with an easy-plane CP 1 model. We can identify hedgehogs with the monopoles of the internal compact gauge field of this CP 1 model. This formulation will allow us to make several measurements which demonstrate that our phase is a bosonic topological insulator.
One measurement that we can make is called the Witten effect. 16-19 This is the binding of one-half of a boson charge to external monopoles in the spin sector. We introduce such external monopoles into the bulk of our system and find that each binds precisely half of a boson charge.
We then determine the surface phases of our model. We again find superfluids connected by a direct transition. We can also break the Z T 2 symmetry on the surface. In the CP 1 model, we can measure the surface Hall conductivity and we find it to be quantized to odd integers as predicted. We reiterate that this Hall conductivity cannot be observed in a purely two-dimensional model, and so we must be measuring the Hall conductivity on the surface of a topological phase. Finally, we also find a surface phase which breaks none of the symmetries of the model. We suspect that this phase has symmetry-enriched intrinsic topological order of the kind predicted by Vishwanath and Senthil, 6 but we do not know how to test this using Monte Carlo.
In our studies of the (2+1)-dimensional 'bosonic quantum Hall' phases, 7 we found that we can obtain symmetryenriched phases with intrinsic topological order by binding multiple topological defects (vortices) to bosons. In Section IV we consider such a binding in (3+1) dimensions. We find that binding of multiple hedgehogs to bosons leads to a bulk phase with intrinsic topological order, thus opening the study of bosonic SET phases in (3+1) dimensions. To help understand the properties of this SET phase, in the Appendix we consider a more abstract model where monopoles of a compact quantum electrodynamics (CQED) are bound to bosons; we analyze the resulting SPT-like and SET-like phases in such a CQED × boson theory, and also discuss how things change upon including matter fields coupled to the CQED, which is the situation in the CP 1 representation of the spins.
## II. REALIZING THE TOPOLOGICAL INSULATOR BY BINDING BOSONS TO HEDGEHOGS OF SO (3) SPINS
We first study the binding between bosons and topological defects by using hedgehogs of a Heisenberg model as our topological defects. We demonstrate the existence of a phase which can be loosely viewed as a condensate of bound states of bosons and hedgehogs, and explore its properties. This model provides an intuitive introduction to the physics of such
binding. Additional properties of the resulting SPT phase will be considered in Section III.
## A. Model and its Bulk Phase Diagram
We study the following action, in (3+1)D Euclidean spacetime:
$$S = S _ { \text{spin} } + \frac { \lambda } { 2 } \sum _ { r, \mu } [ J _ { \mu } ( r ) - Q _ { \mu } ( r ) ] ^ { 2 }. \quad \text{ (1)} \quad \text{$\text{$\text{$\text{$\text{$exan}$}$}$} \\ \quad \dots \dots$$
S spin is an action which controls fluctuations in the SO (3) spins. The second term provides the binding interaction between bosons and hedgehogs. The bosons are represented by integer-valued conserved currents, J µ ( r ) , defined on the links of a four-dimensional cubic lattice, where r is a site label on the lattice and µ ∈ ( x, y, z, τ ) is a direction. These currents represent the world-lines of the bosons in the (3+1)dimensional space-time. The Q µ ( r ) variables represent the hedgehogs, which are also integer-valued conserved currents and will be defined shortly. When the real number λ is large, this term will bind bosons and hedgehogs together. We work with periodic boundary conditions and require no net boson charge and no net boson spatial currents, so that the J µ spacetime currents have zero total winding number; such conditions are automatically satisfied for the hedgehog currents Q µ defined below. Imposing such conditions on the boson currents is just a convenient choice, which does not affect the bulk physics, but allows a precise change of variables involving both J and Q currents [see Eq. (10) below].
In this section we represent the spin degrees of freedom by SO (3) unit vectors. For S spin , we take the following action, which describes a Heisenberg model:
$$S _ { \text{spin} } = - \beta \sum _ { R, \mu } \vec { n } ( R ) \cdot \vec { n } ( R + \hat { \mu } ). \quad \ \ ( 2 ) \quad \text{Error}$$
Here /vector n R ( ) = ( n , n , n a b c )( R ) are three-component unit vectors which represent the spins. They reside on a different lattice from the J µ ( r ) variables above. This lattice has its sites labelled by R and located at the centers of the (hyper)cubes of the r -lattice in Eq. (1), i.e., the R and r lattices are dual to each other.
From these /vector n R ( ) , we can define the hedgehog currents Q µ ( r ) using the prescription in the literature 20-23 generalized to four dimensions. We summarize this prescription here. We first define variables α µ ( R ) , which reside on links connecting the spins /vector n i ≡ /vector n R ( ) and /vector n j ≡ /vector n R ( + ˆ) µ :
$$e ^ { i \alpha _ { \mu } ( R ) } = \frac { 1 + \vec { n } _ { i } \cdot \vec { n } _ { j } + \vec { n } _ { i } \cdot \vec { N } _ { 0 } + \vec { n } _ { j } \cdot \vec { N } _ { 0 } + i \vec { N } _ { 0 } \cdot ( \vec { n } _ { i } \times \vec { n } _ { j } ) } { \sqrt { 2 ( 1 + \vec { n } _ { i } \cdot \vec { n } _ { j } ) ( 1 + \vec { n } _ { i } \cdot \vec { N } _ { 0 } ) ( 1 + \vec { n } _ { j } \cdot \vec { N } _ { 0 } ) } }. \quad \text{current} \\ \text{Here } \vec { N } _ { 0 } \text{ is a reference vector which we can choose arbitrarily.} \quad \text{where}.$$
Here /vector N 0 is a reference vector which we can choose arbitrarily. We then define placket 'fluxes' ω µν ∈ ( -π, π ] as follows:
$$e ^ { i \omega _ { \mu \nu } ( R ) } = e ^ { i [ \nabla _ { \mu } \alpha _ { \nu } ( R ) - \nabla _ { \nu } \alpha _ { \mu } ( R ) ] }. \quad \quad ( 4 )$$
.
Here ∇ µ α ν ( R ) ≡ α ν ( R + ˆ) µ -α ν ( R ) . One can show that changing the reference vector /vector N 0 corresponds to a gauge transformation of the α µ ( R ) variables so that ω µν ( R ) are independent of the reference vector. Finally, we can define the hedgehog current:
$$Q _ { \mu } ( r ) = \frac { 1 } { 4 \pi } \epsilon _ { \mu \nu \rho \sigma } \nabla _ { \nu } \omega _ { \rho \sigma },$$
with implied summation over repeated indices. Consider for example
$$Q _ { \tau } [ r = R + ( \frac { 1 } { 2 }, \frac { 1 } { 2 }, \frac { 1 } { 2 }, - \frac { 1 } { 2 } ) ] = \frac { \nabla _ { x } \omega _ { y z } + \nabla _ { y } \omega _ { z x } + \nabla _ { z } \omega _ { x y } } { 2 \pi }.$$
The right-hand-side is defined on a cube [ R,R +ˆ x, R +ˆ y, R + ˆ] z , and can also be associated with a point R +(ˆ + ˆ+ ˆ) 2 x y z / in the center of this cube. The ω variables are fluxes passing through the plaquettes of this cube, and the net flux out of the cube is guaranteed to be integer multiple of 2 π . We then define hedgehog number to be the net outgoing flux divided by 2 π . When all four dimensions are considered, the center of this cube can be equivalently associated with a link in the τ direction from a dual lattice site r = R +(ˆ + ˆ+ ˆ x y z -ˆ) τ / 2 , and the hedgehog number becomes the τ -component, Q τ ( r ) , of the hedgehog four-current.
We have studied the above action in Monte Carlo simulations. In addition to simple independent updates of /vector n and J µ , when λ is large one needs to try updates which change both J µ and Q µ in such a way that the second term in Eq. (1) is unchanged. To do this we choose a spin and an amount to update it, check to see if any Q µ variables will change, and include matching changes in the J µ variables as part of the attempted Monte Carlo move.
In our simulations, we monitor the 'internal energy per site,' /epsilon1 = S/ Vol , where Vol ≡ L 4 is the volume of the system, which we take to have linear size L in all four directions. From this, we can determine the specific heat per site:
$$C = ( \langle \epsilon ^ { 2 } \rangle - \langle \epsilon \rangle ^ { 2 } ) \times \text{Vol.}$$
We can locate phase transitions in our model by looking for singularities in the specific heat. We also monitor the magnetization per spin:
$$m = \frac { \langle | \sum _ { R } \vec { n } ( R ) | \rangle } { \text{Vol} }. \quad \text{(7)} \\ \text{inc area discard that communication is wrong}$$
When the spins are disordered the magnetization is proportional to 1 / √ Vol , while in the ordered phase the magnetization remains non-zero in the thermodynamic limit. Therefore we can use measurements of the magnetization at different sizes to determine if the spins are ordered.
To study the behavior of the boson currents, we monitor current-current correlators, defined as:
$$\rho _ { J } ( k ) = \left \langle J _ { \mu } ( k ) J _ { \mu } ( - k ) \right \rangle,$$
where k is a wave vector, µ is a fixed direction, and
$$J _ { \mu } ( k ) \equiv \frac { 1 } { \sqrt { \text{Vol} } } \sum _ { r } J _ { \mu } ( r ) e ^ { - i k \cdot r }.$$
In the space-time isotropic system, ρ J ( k ) is independent of the direction µ , and when we show numerical data we average over all directions to improve statistics. In an ensemble which would allow non-zero total winding number, ρ J (0) would be the familiar superfluid stiffness. In our model J k ( = 0) = 0 , so this measurment is not available in our simulations. Instead, we evaluate the correlators at the smallest non-zero k . For example, if µ is in the x direction, we can take k min = (0 , 2 π L , 0 0) , , (0 , 0 , 2 π L , 0) , and (0 , 0 0 , , 2 π L ) and average over these; we exclude ( 2 π L , 0 0 0) , , because in our ensemble the net winding of the J x current is zero, so the J x ( k ) evaluated at this wavevector is also zero. We also monitor current-current correlators of the hedgehog currents, ρ Q ( k ) , which are defined in the same way as for the boson currents.
In the phase where the J µ are gapped, only small loops contribute to the current-current correlators and ρ J ( k min ) ∼ k 2 min ∼ 1 /L 2 , while when the J µ proliferate ρ J is independent of the system size. Therefore we can use finite-size scaling of this quantity to determine the locations of phase transitions. For the hedgehog currents, ρ Q ( k min ) ∼ k 2 min in all phases, so we cannot use finite-size scaling of this quantity to find phase transitions; this is similar to the properties of vortex currents in a (2+1)-dimensional XY model and originates from effective long-range interactions of these topological defects.
The bulk phase diagram of this model is shown in the inset in Fig. 1 and can be understood essentially analytically. Indeed, in the case where the entire system is in the same phase (i.e., the boson-hedgehog binding is applied everywhere), we can change to new variables in the partition sum
$$\tilde { J } _ { \mu } ( r ) \equiv J _ { \mu } ( r ) - Q _ { \mu } ( r ) \, \quad \ \ ( 1 0 ) \quad \ \ b o u$$
which satisfy the same conditions as the original boson currents. Expressed in these variables, the first and second terms in Eq. (1) decouple, and we can study them independently. At small β , the /vector n spins are disordered, which also implies that the hedgehog currents are proliferated. As β is increased, the spins order. This means that there is a large energy cost for hedgehog currents to exist. We say that the hedgehog currents are gapped, and only small loops of them are present. We can determine the location of the spin-ordering phase transition by finding singularities in the heat capacity, or by performing finite-size scaling on the magnetization as described above. The value found by our numerics agrees with the literature on the 4D classical Heisenberg model. 24
At small λ , the ˜ J µ variables are proliferated. In our original variables, this means that the boson currents are effectively independent of the hedgehog currents and are condensed. At large λ , the ˜ J µ variables are gapped. The boson currents are bound to the hedgehog currents. The result of this can be seen in Fig. 1, which shows the current-current correlators for the boson and hedgehog currents. For small λ , there is no relation between the bosons (red solid line) and the hedgehogs (blue dashed line). As λ is increased, the current-current correlators become essentially identical as the bosons and hedgehogs are bound together. We can determine the location of the phase transition in λ by studying singularities in the heat capacity, or by peforming finite-size scaling on ρ ˜ J as described above.
ρ
FIG. 1. Inset: Bulk phase diagram for the model in Eqs. (1) and (2) with bosons and Heisenberg spins. The phase diagram is mathematically equivalent to a system of decoupled currents and spins, hence the straight line boundaries. At λ = 0 , the system has a paramagnetferromagnet transition as β is increased, while the bosons are superfluid throughout. As λ is increased, the boson currents bind to the hedgehog currents. The loop pictures in the phases show a 'snapshot' of the phase. Red solid loops mean that boson currents are proliferated in the phase, while blue dashed loops indicate proliferated hedgehog currents. The phase of interest is the 'binding' phase in the upper left corner where bosons are bound to hedgehogs. The main figure shows the current-current correlations of the bosons and hedgehogs as λ is increased while β = 0 , for a system of linear dimension L = 6 . Wesee that the correlators become essentially equal as the system enters the upper left phase, indicating that bosons have bound to hedgehogs.
We can now summarize the phase diagram shown in the inset in Fig. 1. At small λ and β both the boson and hedgehog currents are proliferated and are not bound, while at large λ and β all currents are gapped. At small λ and large β the boson currents are condensed but the hedgehog currents are gapped. Finally, when λ is large and β is small the system is in the 'binding' phase with 'proliferated' bound states, which we will argue is a (3+1)D topological phase protected by the appropriate symmetries.
It is also helpful to think about the 'easy-plane' regime for the spin variables, in which the spins, /vector n = ( n , n , n a b c ) , are roughly in the ab -plane, with only small c components; we will denote the corresponding global symmetry of spin rotations in the ab -plane as U (1) spin . In the easy-plane case we can define 'vortices' of the XY spins ( n , n a b ) (i.e., phase windings of the complex order parameter ∼ n a + in b ). The vortices are defined on the plackets of the cubic lattice. Therefore in the (3+1)D space-time they are represented as twodimensional 'world-sheets.' When dealing with vortices, one can gain intuition by thinking in terms of only the three spatial dimensions of our (3+1) dimensional space-time. In this picture the bosons and hedgehogs are represented by point particles, while the vortices are represented by lines. The next two paragraphs can be most easily understood by thinking in this
FIG. 2. As discussed in the text, the easy-plane spin system has two species of vortex lines, ↑ and ↓ , depending on n c at the core. A hedgehog is a transition point between the two types of lines, with the sign of the hedgehog determined by the orientation of the type and vorticity of the vortex lines, as shown with examples. This figure shows only the spatial dimensions of the system, therefore the hedgehogs are point particles and the vortices are lines. Applying a Zeeman field to the surface means allowing only one type of vortex line through the surface. This leads to a correlation between the hedgehog number (and therefore the boson charge) and the vorticity at the surface, which is the origin of the Hall conductivity.
picture.
Though ordinary XY spins are not defined at the core of a vortex, our spins have a c -component which can point either up or down at the vortex core. We can define two species of vortices, which we call the ↑ and ↓ species, depending on whether n c is positive or negative at the core. This description is useful since an ↑ vortex ending and continuing as a ↓ vortex is a hedgehog. Therefore our system can be thought of as a system of vortex lines having two different species. These vortex lines can change species, and the locations where this happens are hedgehogs. 25 This is illustrated in Fig. 2.
Hedgehog number can be either positive or negative, and this is determined by the properties of the vortex line the hedgehog is attached to. If, when looking along the line from the ↑ core to the ↓ core the vorticity is clockwise (counterclockwise), we define a positive (negative) hedgehog. This is pictured in Fig. 2.
## 1. Importance of Discrete Symmetry
The action in Eq. (1) has a U (1) symmetry which comes from the conservation of the J µ currents; this is boson charge conservation symmetry. It also has an SO (3) symmetry from the spins. Both of these symmetries participate in the protection of the topological phase, in the sense that if they are broken it is possible to continuously connect the topological phase to a trivial phase. In addition, the action has a Z 2 symmetry, which is obtained by reflecting the /vector n spins in a plane in the spin space. To see how this affects the hedgehog current, we can examine Eq. (3), taking the reference vector /vector N 0 to be in the plane of reflection. We see that reflecting the spins changes the sign of the imaginary part of e iα µ , and therefore the hedgehog current changes sign under such a reflection. For our entire action to be invariant, we therefore need to combine such reflections with an operation which changes the sign of the boson currents. For concreteness we will consider the Z 2 symmetry corresponding to reflections of the /vector n variables in the ab plane of the spin space. In this case the Z 2 symmetry can be summarized as:
$$\begin{array} { c c c c c } n _ { a }, n _ { b } & \rightarrow & n _ { a }, n _ { b } \\ n _ { c } & \rightarrow & - n _ { c } \\ Q _ { \mu } & \rightarrow & - Q _ { \mu } \\ J _ { \mu } & \rightarrow & - J _ { \mu } \end{array}. \quad \begin{array} { c } ( 1 1 ) \\ J _ { \mu } & \rightarrow & - J _ { \mu } \end{array} \\ \colon \text{also possible to reflect the $\vec{n}$ engine around} \, a$$
Note that it is also possible to reflect the /vector n spins around a different plane, but this is not a distinct symmetry since it is simply the product of the above Z 2 symmetry and an element of SO (3) .
By analogy with the electronic topological insulator, we would like the Z 2 symmetry described above to be a 'timereversal' symmetry, i.e. it should be anti-unitary. The symmetry in Eq. (11) can be either a unitary or anti-unitary symmetry. Note that Eq. (1) is a real, (3+1)-dimensional action which is assumed to arise from the Trotter decomposition of the imaginary time propagator (i.e., Euclidean path integral) of a three-dimensional quantum Hamiltonian. The symmetry operations in Eq. (11) can be derived from the action of a symmetry operation on the quantum Hamiltonian. Therefore asking whether the symmetry in Eq. (11) is anti-unitary is the same as asking whether the symmetry of the quantum Hamiltonian which generates Eq. (11) is anti-unitary. This is a difficult question for us to answer as we do not know the quantum Hamiltonian which has Eq. (1) as its Euclidean path integral. Nevertheless, we take the perspective where we can check whether the original Hamiltonian has time reversal by complex-conjugating the action combined with the appropriate variable transformations, and in this way we can view the above symmetry of the action also as time reversal.
In the easy plane case, our system has U (1) boson × U (1) spin in addition to this discrete symmetry. We can think of the U (1) spin as also coming from a boson, and n a + in b gives the phase degree of freedom of this boson. We can imagine allowing tunnelling between the two U (1) symmetries. If we do this, then the discrete symmetry should act the same way on the two species, and we can see that the above symmetry changes the number of the bosons but not their phase. In a system consisting of one species of bosons, the symmetry which acts in this way is anti-unitary. From now on the symmetry described above will also be treated as anti-unitary and denoted by Z T 2 .
The action in Eq. (1) is invariant under several Z 2 symmetries, but for our purposes we will consider only the Z T 2 symmetry described above as important, as it protects the topological behavior. To see this we can break this symmetry and
argue that the topological phase is destroyed. We break the Z T 2 symmetry by introducing a Zeeman field into our action:
$$S _ { Z \text{eeman} } = - h \sum _ { R } n _ { c } ( R ). \quad \quad ( 1 2 ) \quad S = \\. \quad. \ \dots \ - \quad \dots \ \dots \. \ \dots \ \dots \ \rho ^ { \lambda }$$
Here h is the strength of the Zeeman field, which points in the c -direction. Note that the Zeeman field does not break U (1) spin or U (1) boson . In our picture of two species of vortices (Fig. 2), the Zeeman field forbids one of the species. Since the vortex lines cannot change species, hedgehogs are forbidden and the binding phase is destroyed. This can be made more precise if we replace the binding term in Eq. (1) with the following term:
$$\frac { \lambda } { 2 } \sum _ { r, \mu } [ J _ { \mu } ( r ) - \eta Q _ { \mu } ( r ) ] ^ { 2 } \,, \quad \quad ( 1 3 ) \quad \text{Wh} _ { \quad \text{ of } t } \\. \quad. \quad. \quad. \quad. \quad. \quad. \quad.$$
where η is a real number. If we choose the parameters β and λ so that the system is initially in the binding phase, the introduction of η allows us to tune the system between the binding phase ( η = 1 ) and the trivial insulator ( η = 0 ). Without a Zeeman field, the system undergoes a phase transition as η is changed between 1 and 0 . When a Zeeman field is applied, the hedgehogs are effectively forbidden, and so we can tune parameters h and η without going through a phase transition. Indeed, when η = 1 , the change of variables in Eq. (10) leads to decoupled ˜ J currents and spins, so there is no phase transition when making h arbitrarily large to align all spins. We can then tune η to zero and finally reduce h back to zero, all without undergoing a phase transition. Thus, in the presence of the Zeeman field, the phase with η = 1 is not distinct from the trivial insulator.
## 2. Binding of Multiple Bosons to a Hedgehog
The above methods also allow us to answer the question of what happens to the system if η in Eq. (13) is an integer larger than 1. In a U (1) × U (1) system in two dimensions, we studied the binding of multiple bosons to vortices (realized by taking integer η ) and found that each number of bound bosons led to a different symmetry-protected topological phase. 7 There were therefore as many SPTs as there are integers, in agreement with the cohomology classification. 3,4 In the present three-dimensional case the classification of Chen et al. 3 for U (1) and Z T 2 symmetry predicts the existence of only a finite number of such symmetry protected topological phases, implying that not every value of η would lead to a distinct phase. Indeed, we find that all systems with η an even integer are topologically trivial, while when η is odd we have the same topological phase as η = 1 .
We can justify this claim by showing that η = 2 can be continuously connected to η = 0 which is a trivial insulator. This argument can then be extended to show that any two systems where η differs by 2 are in the same phase. Our argument is inspired by Ref. 26, except that here we are working with a microscopic model rather than a topological field theory. We start by considering two copies of our action, each with its own bosons and Heisenberg spins and with η = 1 :
$$S & = - \beta \sum _ { R, \mu } \left [ \vec { n } ^ { ( 1 ) } ( R ) \cdot \vec { n } ^ { ( 1 ) } ( R + \hat { \mu } ) + \vec { n } ^ { ( 2 ) } ( R ) \cdot \vec { n } ^ { ( 2 ) } ( R + \hat { \mu } ) \right ] \\ & \lambda \swarrow \left ( \lambda \sum _ { R, \mu } \right ) \right ) \right ) \left ( \lambda \sum _ { R, \mu } \right ) \right ) \left ( \lambda \sum _ { R, \mu } \right ) \right ) \right ) \right ) \right ) \right ) \right..$$
$$\text{ints in} \quad + \frac { \lambda } { 2 } \sum _ { r, \mu } \Big { ( } [ J _ { \mu } ^ { ( 1 ) } ( r ) - Q _ { \mu } ^ { ( 1 ) } ( r ) ] ^ { 2 } + [ J _ { \mu } ^ { ( 2 ) } ( r ) - Q _ { \mu } ^ { ( 2 ) } ( r ) ] ^ { 2 } \Big { ) } \,, \quad ( 1 4 ) \\ \text{$two$-} \\ \text{of $vor$-} \quad \text{where the sumerscripts indicate which conv a variable is from}$$
where the superscripts indicate which copy a variable is from. We now couple the two copies by adding the following terms:
$$\dots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \\ \delta S & = - A \sum _ { R } \vec { n } ^ { ( 1 ) } ( R ) \cdot \vec { n } ^ { ( 2 ) } ( R ) \\ & - B \sum _ { r } \cos [ \Phi ^ { ( 1 ) } ( r ) - \Phi ^ { ( 2 ) } ( r ) ]. \quad \quad ( 1 5 ) \\ \text{en} \, A \, \text{is large and positive, the first term above locks spins}$$
The Φ variables are 2 π -periodic phases that can be thought of as conjugates to the J µ variables. More precisely, in our path integral we sum over only the configurations of J µ in which the currents are divergenceless. We can instead sum over all configurations of J µ , and include the following term in our path integral:
When A is large and positive, the first term above locks spins of the different copies together, /vector n (1) ≈ /vector n (2) . The hedgehog variables therefore take on the same values, and either can be viewed as the hedgehog number of the whole spin system in this case: Q ≈ Q (1) ≈ Q (2) . On the other hand, when A is large and negative the spins of different types are locked in opposite directions, /vector n (1) ≈ -/vector n (2) , and Q (1) ≈ -Q (2) .
$$\int _ { 0 } ^ { 2 \pi } D \Phi ( r ) e ^ { - i \sum _ { r } \Phi ( r ) ( \sum _ { \mu } \nabla _ { \mu } J _ { \mu } ) ( r ) } \,, \quad ( 1 6 ) \\ \mathrm i o h \, d u n o m i c a l l v \, \text{for each the content} \, \text{that the chosen our}.$$
which dynamically enforces the constraint that the boson currents be conserved. We introduce Φ variables for each copy of the boson currents, and the B term is tunnelling between the two copies. When B is large (of either sign) only the sum of J (1) and J (2) is conserved and can be identified as the current of the whole boson system (i.e., combining both copies).
When A and B are large and positive, we can expand the terms on the second line of Eq. (14), and take J = J (1) + J (2) , Q = Q (1) = Q (2) . In this case we obtain coupling between J and Q which is effectively the same as in Eq. (13) with η = 2 . On the other hand, when A is large and negative, but B is still large and positive, J is unchanged but its coupling to the hedgehogs vanishes as contributions from Q (1) and Q (2) cancel; this gives Eq. (13) with η = 0 . We can continuously deform A = 0 , B = 0 to A = ∞ , B = ∞ without undergoing a phase transition. In addition, we can deform A = 0 , B = 0 to A = -∞ , B = ∞ without undergoing a phase transition. This implies that we can tune from η = 2 to η = 0 without undergoing a phase transition, and so both of these cases are in the trivial insulating phase.
## B. Phase Diagram on the Boundary Between the Binding Phase and a Trivial Insulator
By analogy to the fermionic topological insulator, we expect that one way to investigate the topological nature of our
phase is to study the physics of its surface. In particular, a number of interesting phases have been predicted on the surface of the bosonic TI, 6 and if we can identify these phases on the surface of our binding phase, it will be a powerful argument that the binding phase is a bosonic TI.
We introduce a surface between the binding phase and a trivial insulator by allowing η in Eq. (13) to vary spatially. We vary η in the z direction, so that:
$$\eta [ r = ( x, y, z, \tau ) ] = \begin{cases} \ 1, \ z _ { L } \leq z < z _ { R } \\ 0, \text{ otherwise} \end{cases} \. \ \ ( 1 7 ) \quad \ \Big | \ \ i ^ { \prime } _ { \ell } \\ \dots \. \. \. \. \. \. \end{cases}$$
This leads to the binding phase in the region z L < z < z R , while the trivial phase occupies the rest of the space. Note that there are two surfaces of the binding phase: one at z L and one at z R . The above geometry in our Monte Carlo setup with periodic boundary conditions corresponds to a four-dimensional torus which is divided into two parts along the z -direction.
On the surface, we can measure all of the quantities which we measured in the bulk, but we now only sum over the sites near the surface (and when averaging over directions we only use the ˆ x y , ˆ and ˆ τ directions).
The binding between hedgehogs and bosons in the bulk of the topological phase leads to exotic physics on the surface. In both the binding and trivial phases, hedgehog currents are proliferated, while boson currents are bound to the hedgehog currents in the binding phase and are absent in the trivial phase, as shown in Fig. 3. Consider what happens when a hedgehog loop tries to cross the boundary between the phases. Since the boson currents must form closed loops, the above conditions cannot both be satisfied without something interesting happening on the surface. For example, we could have unbound boson currents on the surface, or hedgehogs could be effectively forbidden from crossing the surface.
In order to study the surface physics and search for the exotic surface behavior predicted in Ref. 6, we first determine the surface phase diagram. To do this, we fix the values of β and λ in the bulk and tune them only on one of the surfaces, by setting:
$$\dot { \beta } _ { \mu } ( X, Y, Z, T ) & = \begin{cases} \begin{array} { c c } \beta _ { \text{surf} }, & Z = z _ { R } - 1 / 2, \, \mu = \hat { x }, \hat { y }, \hat { \tau } \\ \beta _ { \text{bulk} }, & \text{otherwise} \end{array} \end{cases} & \text{should} \\ \lambda _ { \mu } ( x, y, z, \tau ) & = \begin{cases} \begin{array} { c c } \lambda _ { \text{surf} }, & z = z _ { R }, \, \mu = \hat { x }, \hat { y }, \hat { \tau } \end{array} \end{cases} & \text{spins } \begin{array} { c } \text{$\text{$\text{$\text{$\text{$\text{$\text{$bulk}$} } }$} \\ \lambda _ { \text{bulk} }, & \text{otherwise} \end{array} \end{cases}$$
$$\begin{smallmatrix} \\ \\ \\$$
Here we focused on the surface at z R , and µ denotes link orientation for either β or λ terms. We show the surface phase diagram in the inset of Fig. 4. All data was taken with β bulk = 0 , λ bulk = 5 2 . , parameters which put the bulk deep into the binding phase. The surface phase diagram contains three distinct phases. When λ surf is small the bosons are in a superfluid phase, breaking their U (1) symmetry. This is the scenario pictured in Fig. 3, where hedgehogs can cross the surface, and these crossings are connected by boson currents. If β surf is also small the SO (3) symmetry is unbroken as the spins are disordered. As β surf increases at small λ surf , the SO (3) symmetry breaks, and so both symmetries are broken. Finally, at large λ surf bosons see a large energy cost, and so the U (1) symmetry is unbroken. This forbids hedgehogs from crossing
FIG. 3. A snapshot of the system, when it is spatially divided into a region which is a trivial insulator and a region which is in the binding phase. In the trivial phase, boson currents are gapped and hedgehog currents are proliferated. In the binding phase both currents are gapped individually, but their bound states are proliferated. Large hedgehog current loops can exist in either region; however, when such a loop tries to cross the boundary the conservation of boson current leads to a conflict between the two sides, and this can lead to exotic surface physics.
the surface, leading to a system of SO (3) spins with hedgehogs forbidden, which on the three-dimensional cubic latice is known to have magnetic order thus breaking the SO (3) symmetry. 21,23
The locations of the phases and phase transitions in Fig. 4 were determined by studying singularities in the specific heat, as well as by studying the surface magnetization and currentcurrent correlators. As an example of such data, in the main plot of Fig. 4 we show the magnetization, multiplied by the square-root of the volume of the surface ( L 3 / 2 ). This quantity should be constant when the spins are disordered and should grow with system size when they are ordered. The data was taken with β surf = 0 and increasing λ surf . We can see that the spins order at λ surf ≈ 4 .
In order to use the current-current correlators to detect the breaking of the U (1) boson symmetry on the surface, we must think through such measurements carefully. In a D -dimensional system of boson worldlines (space-time currents), the argument that ρ J ( k min ) ∼ k 2 min ∼ 1 /L 2 in the gapped phase relies on the conservation of the currents J . However, this is no longer true if we consider only the surface, as currents can enter and exit from the rest of the system. Therefore we instead measure the current-current correlators in the entire system (though only in the ˆ x y , ˆ , and ˆ τ directions, which are parallel to the surface). When currents are gapped everywhere, this quantity will still be proportional to 1 /L 2 . If we measured the current-current correlator only on the surface in the phase where the surface is a superfluid, we would expect the result to be independent of system size. Since we are measuring the correlator over the whole system, the result should
FIG. 4. The inset shows the phase diagram of the surface of the binding phase, without a Zeeman field. It was obtained by tuning the bulk parameters deep into the binding phase and then varying β and λ only on the surface . We find that in this model our surface always spontaneously breaks a symmetry. At small β surf and λ surf the boson U (1) symmetry is broken and the bosons condense into a superfluid, while at large λ surf the spin SO (3) symmetry is broken and the spins align into a ferromagnet. At small λ surf and large β surf both symmetries are broken. The main plot shows surface magnetization on a sweep in λ surf for β surf = 0 . We show mL 3 / 2 , which is independent of system size in the disordered phase, and grows with system size in the ordered phase. We can clearly see that the SO (3) symmetry is broken as λ surf is increased past a value of approximately 4 . All data in this section was taken with β bulk = 0 , λ bulk = 5 2 . .
be proportional to the fraction of the system which makes up the surface, which is 1 /L . In Fig. 5 we plot ρ J · L 2 . We see that at large λ surf this quantity is independent of system size, which tells us that the currents are gapped everywhere and the surface is an insulator in the boson degrees of freedom. At small λ surf , ρ J · L 2 ∼ L , which tells us that there is a region whose volume fraction is proportional to 1 /L where the bosons are in a superfluid phase. We interpret this as evidence that there is a superfluid at the surface.
The surface phase where the U (1) boson symmetry is broken is clearly a superfluid of bosons. In the easy-plane picture the spins can also be thought of as having a U (1) symmetry which is broken when the spins are ordered, and so the phase where the spins order can also be thought of as a superfluid of new particles representing the spins. We can now ask whether these superfluids are trivial superfluids, or the exotic superfluids thought to exist on the surface of a bosonic TI. 6 The exotic properties have to do with the charge and statistics of gapped vortex excitations, and unfortunately we do not have access to these properties in our Monte Carlo. Therefore our surface superfluids cannot tell us whether the binding phase is a bosonic TI.
One predicted feature of the superfluids on the surface of the bosonic TI is that they are connected by a direct transition which is a deconfined critical point. 6 As shown in Figs. 4 and 5, our SO (3) and U (1) symmetries appear to break on the opposite sides of the same point, and so it looks that we also have a direct transition between these phases; however, we would
FIG. 5. Current-current correlators of the entire system, multiplied by L 2 , used to detect symmetry breaking of the bosons on the surface. This quantity is constant when the boson U (1) symmetry is preserved and increases linearly with system size when it is broken on the surface. Parameters used are the same as in Fig. 4. We see that there is a phase transition at λ surf ≈ 4 .
need larger system sizes to see whether there is indeed a direct transition and whether it is continuous. Though this does not definitively show that the binding phase is topological, but it does provide evidence that we have an unusual field theory on the surface.
## C. Surface with Zeeman Field
Another exotic phase predicted to exist on the surface of the bosonic TI is a phase which breaks Z T 2 symmetry and has a quantized Hall reponse. We can try to realize this phase by applying a Zeeman field on the surface of our model to explicitly break the Z T 2 symmetry. We add a term similar to Eq. (12) to our action, but only on the surface of the model. We then expect that there is a surface phase which does not break any U (1) symmetry-i.e., an insulator-but which has the surface Hall conductivity quantized to an odd integer, which is different from the even values expected in the (2+1)D bosonic integer quantum Hall effect. For easy-plane spins, we expect that a finite Zeeman field is needed to destroy the surface superfluids to reach this phase, while for SO(3) spins used here and starting in the spin-ordered phase, we expect to immediately transition to the quantum Hall insulator. 23
In our Monte Carlo study, we actually apply Zeeman field h to both surfaces but take the fields on the two surfaces to have opposite signs. We take λ = 5 2 . and β = 0 everywhere (including at the surfaces), which puts the bulk regions deep into the binding or trivial phases respectively, while the surfaces start in the spin-ordered phase at h = 0 .
Since the Hall conductivity in the surface system of spins and bosons will be due to correlations between the vorticity of the spins and the boson charges, 7 we will measure these correlations directly before moving on to the more complicated Hall conductivity measurement. In the absence of the Zeeman
field, we can use the Z T 2 symmetry, which is reflection of the /vector n variables in the ab -plane, to change the sign of hedgehog and hence the boson charge without changing the spin vorticity, and so such correlations vanish. This can be seen in Fig. 2, where, for example, on the top surface there is one clockwise vortex attached to a positive hedgehog (which is in turn bound to a positively charged boson), and another clockwise vortex is attached to a negative hedgehog. Applying a Zeeman field corresponds to only allowing one type of vortex ( ↑ or ↓ ) to pass through the surface. In Fig. 2, this means that only solid lines are allowed to pass through the top surface. We can see that this leads to a correlation between the vorticity of the vortex on the surface and the charge of the boson it is binding nearby.
We can further think about Fig. 2 as depicting a slab of the binding phase. Opposite Zeeman fields on the two surfaces give us a quasi-two-dimensional slab on which vortices are bound to charges with definite relation between the vorticity and charge, and these bound states are proliferated. We have studied such a system in a previous work, 7 and found that its Hall conductivity is quantized to be equal to two. It is reasonable to assume that this conductivity is evenly distributed between the two surfaces, leading to the surface Hall conductivity equal to one on each surface. This intuitive argument reproduces the prediction of Ref. 6.
To measure correlations between vortices and bosons we first define the spin vorticity V µν ( R ) as
$$V _ { \mu \nu } ( R )$$
where s µ ( R ) ∈ ( -π, π ] measures the difference between the spin angles at R + ˆ µ and R and brings it to ( -π, π ] :
$$s _ { \mu } ( R ) \equ$$
We can then Fourier transform the vorticity as follows:
$$V _ { x y } ( k ) & =$$
The prime on the sum indicates we are summing over all sites at a fixed z = z R . We measure |〈 V xy ( k min ) J τ ( -k min ) 〉| , where k min = (2 π/L, 0 0 0) , , , and the results are shown in Fig. 6.
We see that as soon as the Zeeman field is applied, the vortices and charges become correlated. Unlike the Hall conductivity, we do not expect these correlations to approach any universal value. We do note that the correlations on the two surfaces of the system are approximately equal, which is encouraging as we would expect the Hall conductivity on these surfaces to be equal as well. The differences between the correlations on the two surfaces is likely due to the fact that the surfaces are realized differently on a lattice, but we would expect that universal values such as the Hall conductivity would not have these differences.
In order to measure Hall conductivity, we need to couple both the bosons and spins to external probing gauge fields,
FIG. 6. Spin vortex - boson charge correlators near the surface of the binding phase in the model Eqs. (1) and (2) using Heisenberg spins. The horizontal axis is the strength of the Zeeman field applied only near the surfaces (and of opposite sign on the two surfaces). We see that as soon as the Zeeman field is turned on, the correlator takes a non-zero value. The value is non-universal, but is approximately the same on the top and bottom surfaces. Data was taken with β = 0 , λ = 5 2 . everywhere and the binding phase occupying half of the system, cf. Eq. (17). Since we do not know how to properly couple external gauge fields to this system, we are unable to compute the Hall conductivity.
and then use linear response theory to determine the conductivity. When we do this we run into a problem, which has to do with the way the hedgehog currents Q µ were defined. We can see from Eq. (4) that ω µν is a discontinuous function of spins since we required it to be brought to ( -π, π ] . When including the probing gauge fields, this causes the path integral to be a discontinuous function of the probing fields, which prevents us from taking the derivatives needed for linear response theory, and so we do not know how to calculate conductivity in this system. In the next section we will find a way around this problem, while in the present case we can only appeal to the intuitive argument presented above.
## III. REALIZING THE TOPOLOGICAL INSULATOR BY BINDING BOSONS TO HEDGEHOGS OF AN EASY-PLANE CP 1 MODEL
The Heisenberg model discussed in the previous section allowed us to realize a 'binding' phase of bosons and hedgehogs. However, in this model we were unable to find definitive evidence that the binding phase is in fact a symmetryprotected bosonic topological insulator, although our indirect arguments are compelling. In this section we introduce a new model, which includes the spin degrees of freedom in a CP 1 representation. This theory has spinor matter fields ('spinons') coupled to a compact gauge field and is a faithful representation of the microscopic spin system with shortrange interactions (i.e., such a lattice 'field theory' is 'emer-
gable' from a local microscopic Hamiltonian). We will see that this formulation allows us to make measurements, such as a Witten effect in the bulk and a quantized Hall conductivity on the surface, which indicate that the binding phase is a bosonic TI.
## A. Bulk Phase Diagram
The following action represents the spins in the CP 1 representation:
$$S _ { \text{spin} } = -$$
Here the spins are represented by two complex bosonic fields z ↑ , z ↓ ('spinons'), which satisfy | z ↑ | 2 + | z ↓ | 2 = 1 . We can write the z fields as a spinor, z ≡ ( z , z ↑ ↓ ) T , and extract the spin /vector n = z † /vector σ z , where /vector σ ≡ ( σ , σ 1 2 , σ 3 ) is a vector of Pauli matrices. The spinon fields are minimally coupled to a compact gauge field a µ ( R ) . The last term is a Maxwell-like term for the compact gauge field, which appears after partially integrating out the spinon fields. The variables in the above action live on a cubic lattice, where R gives the position on the lattice and µ ν , are directions.
The CP 1 model defined above actually has global SO (3) symmetry, similar to the previous section. In this section we find it convenient to break the SO (3) symmetry down to U (1) explicitly by taking the 'easy-plane' limit of the CP 1 model. We align all the spins /vector n in the ab -plane, which corresponds to fixing the magnitude of z ↑ and z ↓ , and allowing only phase fluctuations, i.e., z s ≡ 1 √ 2 e iφ s . The CP 1 model in the easyplane limit becomes:
$$\cdot \\ S _ { \text{spin} } = - \beta \sum _ { s = \uparrow, \downarrow } \sum _ { R, \mu } \cos [ \nabla _ { \mu } \phi _ { s } ( R ) - a _ { \mu } ( R ) ] & \quad \text{freed} \quad \text{ous p} \\ + \frac { K } { 2 } \sum _ { R, \mu < \nu } [ \nabla _ { \mu } a _ { \nu } ( R ) - \nabla _ { \nu } a _ { \mu } ( R ) - 2 \pi B _ { \mu \nu } ( R ) ] ^ { 2 } \,. \, ( 2 3 ) & \quad \text{phase} \\ \quad \text{in the Let} \\ \text{Here } \, \phi _ { * } \, \text{and } \phi _ { + } \, \text{ are } 2 \pi \text{-periodic variables which represent the} & \quad \text{will} \, b$$
Here φ ↑ and φ ↓ are 2 π -periodic variables which represent the phases of the spinon fields. We have also replaced the cosine in the Maxwell term by a quadratic 'Villain' form, with B µν which are unconstrained, integer-valued dynamical variables residing on the plackets of the lattice. Upon summing over B µν in the partition sum, the third term generates a 2 π periodic function of ∇ µ a ν - ∇ ν a µ , and therefore this does not change the universality class of the problem.
Using the Villain form of the Maxwell term is advantageous as it allows us to define the hedgehog current:
$$Q _ { \mu } ( r ) = \frac { 1 } { 2 } \epsilon _ { \mu \nu \rho \sigma } \nabla _ { \nu } B _ { \rho \sigma }. \quad \quad ( 2 4 ) \quad \text{ord} \\ \quad \text{deg}$$
Note that Q µ ( r ) resides on the links of the lattice whose sites are labelled by r which, as in the previous section, is interpenetrating with the lattice labelled by R . The above definition is analogous to Eq. (5) with B ρσ ↔ ω ρσ / (2 π ) . The B µν variables have the meaning of 'Dirac strings' of the hedgehogs. Thus defined, the hedgehogs in the CP 1 model are actually monopoles of the compact gauge field a µ . We will continue to call them hedgehogs in this work to avoid confusion with a different type of monopole introduced later.
We can again study the model described by Eqs. (1) and (23) in Monte Carlo. Equilibration becomes difficult in the regime where K and λ are large, and it is necessary to include composite updates which simultaneously change multiple variables. One such update is to change B µν while also changing J µ so that there is no change in the second term in Eq. (1). Another update is to change both a µ and B µν in such a way as to keep the K term in Eq. (23) small.
We can find phase transitions in this model by looking for singularities in the specific heat, which is defined in the same way as in the previous section. We can identify order in the bosonic degrees of freedom by studying the superfluid stiffness and order in the spins by studying the magnetization. In this easy-plane version of the model, the spin degree of freedom is an XY vector with components ( n , n a b ) . Since ( n , n a b ) = (cos φ spin , sin φ spin ) with φ spin = φ ↑ -φ ↓ , the magnetization is given by:
$$m = \frac { \left \langle \left | \sum _ { R } e ^ { i [ \phi _ { \uparrow } ( R ) - \phi _ { \downarrow } ( R ) ] } \right | \right \rangle } { \text{Vol} }. \quad \text{(25)} \\ \text{phase diagram in this model is parameterized by $\beta$, $K$},$$
The phase diagram in this model is parameterized by β K , , and λ . As in the previous section, we can make the change of variables in Eq. (10) and find that the ˜ J part of the problem decouples from the spin part. The behaviour of the bosons is the same as in the previous section. When λ is small, the physical bosons are essentially independent of the hedgehogs, and are in the superfluid phase. As λ is increased, they become bound to hedgehogs. The transition happens at λ ≈ 4 . The locations of the phase transitions in the spin degrees of freedom are independent of λ , though the nature of the various phases are not. In Fig. 7 we show the phase diagram in the β and K variables, for two cases: λ small and λ large. The phase diagram is consistent with the easy-plane CP 1 model in the literature. 27
Let us first consider the case when λ is small. The bosons will be in a superfluid phase for any β and K . The spin system has the following three phases: i) When β and K are small, the spin degrees of freedom are disordered and the hedgehogs are proliferated. The phase is therefore a conventional paramagnet in the spin degrees of freedom. ii) As K is increased, hedgehogs acquire a large energy cost, and become gapped. This phase was studied in Refs. 23, 25, and 27. It is known as the Coulomb phase 28 because it has an emergent gapless photon and gapped excitations that carry charge 1 / 2 and interact via a Coulomb interaction. iii) Finally, the phase at large β has a large energy penalty for spin fluctuations, and so the spins order. This phase is a conventional ferromagnet in the spin degrees of freedom.
In the case when λ is large, the spin parts of the Coulomb and ferromagnetic phases do not change, since both of these phases suppress hedgehogs. These phases are now trivial insulators in the boson degrees of freedom. On the other hand,
FIG. 7. Bulk phase diagram in the β and K variables for the model defined in Eqs. (1) and (23) where the spins are described by an easy-plane CP 1 model. The top panel shows the phases at small λ , while the bottom panel shows large λ . The locations of the phase boundaries are independent of λ , but the nature of the phases are not. Symbols indicate points where the phase transitions were identified numerically from singularities in the heat capacity. The candidate for the topological phase is the binding phase, which occurs at large λ , small β , and small K .
in the paramagnetic phase the hedgehogs are proliferated and the bosons become bound to them. This is the binding phase, which we will argue is a topological phase protected by the U (1) spin × U (1) boson and Z T 2 symmetries.
## 1. Symmetries When The Spins Are Represented By An Easy-plane CP 1 Model
When representing our spins with an easy-plane CP 1 model we have explicitly broken the SO (3) symmetry from the previous section down to a U (1) symmetry which corresponds to spin rotations in the easy plane. This complicates our discussion of the discrete symmetries of the model. In the previous section all reflections of /vector n were related to each other by an operation in SO (3) , but in the easy-plane case, e.g., reflections in the ab plane ( n , n , n a b c ) → ( n , n , a b -n c ) are now distinct from reflections in the ac plane ( n , n , n a b c ) → ( n , a -n , n b c ) . The result of this is that there are now two Z T 2 symmetries which protect the topological phase, in the sense that as long as any one of these symmetries is preserved, the topological phase cannot be continuously connected to a trivial phase. The first Z T 2 symmetry is the same as that in the previous section; in the variables of Eq. (23) it reads:
$$\begin{array} { c c c c c c c }. & & & \cdot & \cdot & \cdot \\ & \phi _ { \uparrow } & \to & - \phi _ { \downarrow } \\ & \phi _ { \downarrow } & \to & - \phi _ { \uparrow } \\ & a _ { \mu } & \to & - a _ { \mu } \\ & B _ { \mu \nu } & \to & - B _ { \mu \nu } \\ & J _ { \mu } & \to & - J _ { \mu } \\ & i & \to & - i \end{array} \,. \quad \quad \, ( 2 6 )$$
This symmetry corresponds to a reflection in the ab plane.
The second Z T 2 symmetry is a combination of reflections in the ab and ac planes and acts on spins as ( n , n , n a b c ) → ( n , a -n , b -n c ) . In the CP 1 representation, this symmetry is given by:
$$\begin{smallmatrix} \phi _ { \uparrow } & \rightarrow & \phi _ { \downarrow } \\ \phi _ { \downarrow } & \rightarrow & \phi _ { \uparrow } \\ a _ { \mu } & \rightarrow & a _ { \mu } \\ B _ { \mu \nu } & \rightarrow & B _ { \mu \nu } \\ J _ { \mu } & \rightarrow & J _ { \mu } \\ i & \rightarrow & - i \end{smallmatrix}.$$
Note that a symmetry which consists only of a reflection in the ac plane does not protect the topological phase, as under this symmetry we can still apply the Zeeman field in the c direction and connect to a trivial phase as in Sec. II A 1. Note also that we cannot apply fields in the ab plane as this would violate the U (1) spin symmetry.
If either of the above Z T 2 symmetries are preserved, the topological phase cannot be connected to a trivial phase. A system with only the first symmetry has a total symmetry of U (1) spin × U (1) boson × Z T 2 , while if we consider the second symmetry the direct product in front of Z T 2 is replaced with a semidirect product. (see Sec. V for further discussion of these symmetries) The Zeeman field introduced in the previous section (and used again below) breaks both of these symmetries.
Note that in the presence of separate U (1) spin and U (1) boson symmetries, we could in principle consider unitary symmetries given by Eqs. (26), (27), by simply omitting the complex conjugation. However, if we imagine introducing a tunnelling between the two U (1) symmetries, which would reduce the total symmetry to U (1) × Z T 2 [or U (1) /multicloseright Z T 2 ], the discrete symmetry should then treat both spin and boson variables in the same way. We find that this condition is satisfied as long as the above symmetries are understood as antiunitary; hence we will refer to these symmetries as Z T 2 .
## B. Observation of a Witten effect
The CP 1 representation allows us to make a bulk measurement which can detect whether our system is a bosonic topological insulator. This measurement, predicted in both
the fermionic 29 and bosonic 16-19 TIs, is called a Witten effect, and is the tendency of an external magnetic monopole in a TI to bind a charge of one-half. In order to justify our claim that the binding phase is a bosonic TI, we now demonstrate that our system exhibits a Witten effect.
The first step in measuring a Witten effect in our Monte Carlo is to add external U (1) gauge fields to the system. These external fields correspond to the U (1) spin and U (1) boson symmetries of the system. We will fix configurations for these fields before performing the simulations, which corresponds to putting our system in some external electromagnetic field. These external gauge fields are distinct from the internal, dynamical gauge field a µ . Similarly, the external monopoles introduced in this section are different from the hedgehogs (which are monopoles of the field a µ ) discussed previously. We will introduce magnetic monopoles in the U (1) spin gauge field and will measure U (1) boson charge. The Witten effect is the statement that the external monopoles of the U (1) spin gauge field will bind half of a charge of the U (1) boson symmetry.
Let us first consider the gauge field coupled to the spin degrees of freedom. Here it is convenient to think in terms of a parton description. This is a description of a spin model by using the easy-plane CP 1 model in which the phases φ ↑ and φ ↓ represent the phases of different types of bosonic 'partons'. These partons each represent one-half of a physical boson. Each parton carries a unit charge under the internal gauge field a . The physical boson carries unit charge under the external gauge field, which we denote A 1 . The partons carry half a charge under this gauge field, with one parton carrying positive charge and the other negative. To modify Eq. (23) to reflect this, we add ± A 1 µ / 2 inside the cosines on the first line. Partially integrating out the parton fields then gives compact Maxwell terms in the field combinations a µ + A 1 µ / 2 and a µ -A 1 µ / 2 , with equal couplings due to the Z T 2 symmetry. We can write each in the Villain form, which introduces two integer-valued placket variables B + µν and B -µν . Wecan expand and recombine these quadratic terms to get separate terms for the a and A 1 fields, leading to the following action:
$$\begin{smallmatrix} 1 \\ - \end{smallmatrix}$$
$$+ \, \frac { \lambda } { 2 } \sum _ { r, \mu } [ J _ { \mu } ( r ) - Q _ { \mu } ( r ) ] ^ { 2 } + i \sum _ { r, \mu } J _ { \mu } ( r ) A _ { 2 \mu } ( r ). \quad ( 2 8 ) \quad \text{where} \quad \text{the 1} \quad \text{the 1} \quad \text{th}$$
$$\text{the a and } A _ { 1 } \text{ fields, leading to the following action:} \\ S & = - \beta \sum _ { R, \mu } \cos [ \nabla _ { \mu } \phi _ { \uparrow } ( R ) - a _ { \mu } ( R ) - \frac { 1 } { 2 } A _ { 1 \mu } ( R ) ] \\ & - \beta \sum _ { R, \mu } \cos [ \nabla _ { \mu } \phi _ { \downarrow } ( R ) - a _ { \mu } ( R ) + \frac { 1 } { 2 } A _ { 1 \mu } ( R ) ] \\ & + \frac { K } { 2 } \sum _ { R, \mu < \nu } [ \nabla _ { \mu } a _ { \nu } ( R ) - \nabla _ { \nu } a _ { \mu } ( R ) - 2 \pi B _ { \mu \nu } ( R ) ] ^ { 2 } \\ & + \frac { K } { 8 } \sum _ { R, \mu < \nu } [ \nabla _ { \mu } A _ { 1 \nu } ( R ) - \nabla _ { \nu } A _ { 1 \mu } ( R ) - 2 \pi M _ { \mu \nu } ( R ) ] ^ { 2 } \\ & + \frac { \lambda } { 2 } \sum _ { r, \mu } [ J _ { \mu } ( r ) - Q _ { \mu } ( r ) ] ^ { 2 } + i \sum _ { r, \mu } J _ { \mu } ( r ) A _ { 2 \mu } ( r ). \quad ( 2 8 ) \quad \text{while} \quad \text{the to c}$$
Here B µν = ( B + µν + B -µν ) / 2 , and M µν = B + µν -B -µν . Note that B µν and M µν are not completely independent variables but satisfy the conditions that B µν is 1 / 2 × integer, M µν is
$$2 B _ { \mu \nu } ( R ) = M _ { \mu \nu } ( R ) \mod 2. \quad \quad ( 2 9 )$$
The variables M µν can be interpreted as the Dirac strings of the external monopoles. We can see that when we introduce external monopoles of odd integer strength, the internal hedgehog variables become half-integer valued-this is a crucial observation for the discussion of the Witten effect. 16 Note that in the case of no external field A 1 and no external monopoles, the B µν variables are integer-valued and Eq. (28) reduces to Eq. (23). Note also that the coupling we wrote for ( ∇ µ A 1 ν - ∇ ν A 1 µ -2 πM µν ) 2 [the 'Maxwell term' for the external gauge field on the fourth line of Eq. (28)] is special to the preceding spinon-generated argument, while it is expected to be renormalized up by the rest of the universe; in fact, we will assume that the A 1 field is essentially externally controlled and is static.
We introduce a monopole into our system by making a specific choice for the external variables A 1 µ and M µν . First, we choose a configuration of M µν which will lead to a pair of external monopoles. In our system with periodic boundary conditions, it is not possible to have only a single monopole. We will place external monopoles at coordinates ( x, y, z ) = (0 , 0 0) , and (0 , 0 , L/ 2) , on the lattice labelled by r . The external monopoles will have opposite charges, with the positively-charged one at the origin. All configurations of external gauge fields will be constant in the τ direction. In order to place external monopoles at these locations, we set M xy ( R ) = 1 whenever X = -1 / 2 , Y = -1 / 2 , and 1 / 2 ≤ Z ≤ L/ 2 -1 / 2 (the 1 / 2 's come from the R lattice being displaced from the r lattice by half a lattice spacing). All other M µν are set to zero. By Eq. (29), we must also constrain B xy to be odd half-integers on this Dirac string. Now that we have specified the M µν values which introduce external monopoles, we choose values for the A 1 µ variables to minimize the action of the Maxwell term on the fourth line of Eq. (28).
In our simulations we will set A 2 µ = 0 everywhere, so that it does not affect the system. It will be used only when computing linear responses.
There are in fact multiple configurations of the variables M µν which give the same configuration of external monopoles. The physics of the system is independent of which configuration of M µν we choose because the various configurations are related by the following gauge transformation:
$$\begin{array} { c c c } M _ { \mu \nu } ( R ) & \to & M _ { \mu \nu } ( R ) + \nabla _ { \mu } \kappa _ { \nu } ( R ) - \nabla _ { \nu } \kappa _ { \mu } ( R ) \\ B _ { \mu \nu } ( R ) & \to & B _ { \mu \nu } ( R ) + \frac { 1 } { 2 } [ \nabla _ { \mu } \kappa _ { \nu } ( R ) - \nabla _ { \nu } \kappa _ { \mu } ( R ) ] \\ A _ { 1 \mu } ( R ) & \to & A _ { 1 \mu } ( R ) + 2 \pi \kappa _ { \mu } ( R ) \\ a _ { \mu } ( R ) & \to & a _ { \mu } ( R ) + \pi \kappa _ { \mu } ( R ) \\ \end{array}, \quad \dots \dots \, \left ( \pi \vee \colon \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \end{array}$$
where κ µ ( R ) is an integer-valued field living on the links of the lattice labelled by R . One can use Eqs. (24) and (28) to check that this transformation does not change the action, including the configurations and energetics of the external monopoles and gauge fields, so our results are independent of the specific choice of the Dirac string M µν .
Having introduced external monopoles into our system, we can begin to see why they should bind half a charge of the bosons. The argument goes as follows. First, when modifying the Dirac string variables M µν to insert external monopoles, we were also forced to modify the variables B µν in such a way as to introduce one-half of a hedgehog at the same locations as the external monopoles. However, we saw in the previous section that hedgehog-boson 'molecules', which have similar bare long-range interactions as just hedgehogs (i.e., carry hedgehog number), are proliferated in the binding phase. Therefore the 1 / 2 -hedgehog which we introduced will be screened by a 'cloud' of hedgehog-boson molecules drawn from the rest of the system. This screening is analogous to Debye screening in a plasma. The screening cloud will carry a hedgehog number of one-half, but with opposite sign to the first hedghog, leading to a total hedgehog number of zero. Since in the binding phase hedgehogs are bound to charges, the cloud also carries a boson charge of one-half. Therefore we find that half of a boson charge has bound to the external monopole. We have tested this intuition by direct Monte Carlo simulations.
The above discussion is complicated by two degeneracies in Eq. (28). First, there is a degeneracy between Q τ (0 , 0 0 , , τ ) = 1 / 2 and Q τ (0 , 0 0 , , τ ) = -1 / 2 : e.g., when M xy = 1 on the Dirac string, B xy can be either +1 2 / or -1 / 2 with the same energy. This degeneracy is a result of the symmetry in Eq. (27). In what follows it can be helpful to neglect variations in the τ direction and think about Q τ (0 , 0 0 , , τ ) as a stationary hedgehog charge at the origin. Because of this degeneracy the statistical mechanics has each of the two states equally probable, which in an infinitely long simulation would lead to zero net hedgehog charge, and no observation of the Witten effect. Second, if we were able to fix the hedgehog number in one of these two states, for example Q τ (0 , 0 0 , , τ ) = 1 / 2 , we have another degeneracy, between J τ (0 , 0 0 , , τ ) = 0 and J τ (0 , 0 0 , , τ ) = 1 . This leads to an average boson charge of 1 / 2 at the location of the monopole. This charge cancels the boson charge from the screening cloud, leading to no observation of the Witten effect.
Despite these degeneracies, we may still observe a Witten effect if the degenerate states are metastable, and the Monte Carlo is stuck in one of the two states. For example, in order to get from Q τ = +1 2 / to Q τ = -1 / 2 one needs to modify all of the B µν on the Dirac string, and the φ ↑ ↓ , and a µ variables nearby. Such a move would be quite unlikely (and impossible in the limit of an infinite separation between the monopoles). Similarly, to get from J τ = 0 to J τ = 1 one needs to insert a boson loop which passes from one monopole to another, and such a step is highly unlikely with local updates.
The results of our numerics shows that even at the small system sizes that we can access, degeneracy in the J variables is always broken. However, the degeneracy in the Q variables is unbroken, and so we do not observe a Witten effect. Let us consider the Q degeneracy more carefully. One of the two degenerate states has Q = +1 2 / bound to a positive external monopole at r = 0 , and Q = -1 / 2 bound to a negative external monopole at r = L/ 2 . The other state as Q = -1 / 2 bound to a positive external monopole at r = 0 , and Q = +1 2 / bound to a negative external monopole at r = L/ 2 . Note that to change between the two degenerate states we need to modify variables along a string connecting the two external monopoles, and the probability of this happening reduces exponentially with distance, so in the thermodynamic limit this degeneracy would certainly break, even though on our small systems it does not.
We will break this degeneracy in our system by decreasing the probability that the system will flip between degenerate states. To do this we add the following 'biasing' term:
$$\ e s _ { \substack { \mathfrak { u } \\ \mathfrak {$$
where γ is some small real number. We have scanned the system by increasing γ from zero and seen no phase transitions, implying that these small γ do not change the phase we are in. It would be surprising if the above biasing term affected the bulk physical properties of the system, as we are only making a modification to a fraction of the system proportional to L -3 . In addition, though this term breaks the symmetry in Eq. (26), it preserves the symmetry in Eq. (27), and the topological phase is protected as long as either of these symmetries is preserved.
We hope that the above term reduces the probability of switching between degenerate states, so that we can observe a Witten effect. From our numerical results we can see that the above term indeed does this for a large range of γ . Therefore we have broken the problematic degeneracies and removed the obstacle to measuring the Witten effect. We would like to stress that the Witten effect is ordinarily defined for a single monopole, in the thermodynamic limit. The problems we have with degeneracies are artifacts of the fact that we are trying to measure a Witten effect in a finite-size system with two external monopoles. If the action were defined on an infinite system with only one monopole, these problems would not arise as it would take an infinitely long time for the system to change between degenerate configurations.
We observe the Witten effect by measuring the total charge enclosed in a sphere of radius w , centered around the location of the monopole. The precise definition of our measurement is:
$$\begin{array} {$$
Note that the sphere discussed above is only a sphere in the x , y and z directions, and we have averaged over the τ direction. We performed simulations with L = 10 and show the results in Fig. 8. At w = 0 , we are at the location of the monopole. There is nearly no boson charge bound here. At w = 1 , we have already included most of the screening cloud, and therefore measure an enclosed charge close to 1 / 2 , as expected. The fact that w = 1 measures a value close to one half shows that the screening length in the system is quite short. At w = 2 , we have included the entire screening cloud, so the charge is even closer to 1 / 2 . As w is further increased, we start to include the screening cloud from the other monopole, which is located at a distance L/ 2 from the first
FIG. 8. Measurement of the Witten effect. The inset shows the measurement setup. External monopoles are inserted into the system, at a distance L/ 2 apart. They carry hedgehog numbers of ± 1 2 / , and are Debye screened by hedgehogs of equal and opposite number, which also carry half of a boson charge. The main plot shows the boson charge enclosed in a sphere of radius w . For w ≈ 1 -3 , this sphere measures the boson charge in the screening cloud near the origin, and the result is 1 2 / , as expected. For w /greaterorsimilar 5 the sphere includes the charge from the other screening cloud, and the enclosed charge drops to zero. The system size is L = 10 , using bulk parameters β = 0 2 . , K = 0 4 . , λ = 8 (cf. bottom panel in Fig. 7), and local biasing parameter γ = 1 5 . .
one. This cancels some of the charge from the first monopole, and so the total charge starts to decrease. When w > L/ 2 , the sphere encompasses the screening clouds from both external monopoles, and so there is a total charge of nearly zero. The values of charge are negative because we set the biasing parameter γ such that at the origin there is a monopole of positive charge, Q = +1 2 / , and the sign of the charge in the screening cloud is the opposite of the sign of the monopole.
In Fig. 8, we have found that the amount of charge at the site of the monopole ( w = 0 ) is nearly zero. However, this is not universal and in fact depends on the choice of γ in Eq. (31). In our simulations, we find that w = 2 is sufficiently far from the monopole to be unaffected by the change in γ . Figure 9 shows simulations taken with different values of γ . We see that though the amount of charge close to the monopoles (near w = 0 and w = L/ 2 ) can be affected by changing γ , the value at w = 2 is always very close to one-half, regardless of what γ is used.
Various other measurements can be made to support our conclusions. Measuring the total charge on each site shows that the half-charge is distributed around the monopole in an approximately spherically symmetric way. We can also use the Witten effect to detect phase transitions out of the bosonic TI. Figure 10 shows the total charge on all the nearestneighbours, as a function of K . We note that the quantized Witten effect disappears at K = 0 6 . , which is where the phase transition to the Coulomb phase is located, see Fig. 7. (In the Coulomb phase the amount of charge isn't necessarily zero, but it is not quantized and in our simulations we found it to be
FIG. 9. Witten effect for different values of γ , but all other parameters the same as in Fig. 8. We see that near w = 2 , when we are measuring the charge inside a sphere which surrounds exactly one monopole, the amount of charge is approximately one-half, and independent of γ . At smaller w (or at w /greaterorsimilar 4 when the sphere starts overlapping with the screening cloud near the second monopole), the enclosed charge does depend on γ .
FIG. 10. A demonstration of how the Witten effect can be used to detect phase transitions. The plot shows the amount of charge enclosed by a sphere with w = 1 , while changing K but keeping all other parameters the same as in Fig. 8. We can compare to Fig. 7, and see that at K = 0 6 . , when the system transitions from the topological phase to the Coulomb phase, the 1/2-quantization of the enclosed charge abruptly stops.
zero.) We also observe the disappearance of the bound charge when the system undergoes a transition to trivial insulator as η is decreased to zero.
## C. Surface Phase Diagram
The measurement of the Witten effect is evidence that our binding phase is a bosonic topological insulator. We can
now study the exotic physics on the surface of this topological phase. We expect to find the surface phases predicted in Ref. 6.
We begin with some analytical arguments that provide a microscopic derivation of the surface field theory proposed in Ref. 6. We define the surface as in Sec. II B. To uncover the exotic physics, we begin by performing a change of variables from the physical boson currents J µ ( r ) to new integer-valued variables
$$G _ { \mu } ( r ) \equiv J _ { \mu } ( r ) -$$
which satisfy
$$\left ( \sum _ { \mu } \nabla _ { \mu } G _ {$$
The action for the spins and the new G µ ( r ) variables is simply
$$S & = S _ { \text{spin} } + \sum _ { r, \mu }$$
For simplicity, from now on we consider situation where the trivial phase region and the SPT phase region are deep in their respective phases: in particular, λ bulk , defined as in Eq. (18), is very large. At first we further simplify the situation by taking λ surf to be very large, in which case we expect the variables G µ ( R ) to be zero everywhere. We also assume that β is small everywhere, so the spin variables want to be deep in the disordered phase. However, near the two surfaces, the spin configurations must satisfy
$$\begin{array} { c } Q _ { z } ( x, y, z _ { R } - 1, \tau ) = 0, \\ Q _ { z } ( x, y, z _ { L } - 1, \tau ) = 0. \\. \quad. \quad. \end{array}$$
Focusing on the spins near one surface, say at z R , we can view Q z ( x, y, z R -1 , τ ) as simply hedgehog numbers in the corresponding (2+1)D spin system spanned by sites ( X,Y,Z = z R -1 / 2 , T ) , and the above conditions correspond to complete suppression of hedgehogs in this spin system. Such a (2+1)D Heisenberg O (3) spin model with hedgehog suppression was studied in Ref. 23 and argued to be described by a non-compact CP 1 field theory ( NCCP 1 ). On a simple (2+1)D cubic lattice, the Heisenberg model with complete hedgehog suppression actually has spontaneous magnetic order of spins even when the direct spin-spin interactions are zero. 20,21 However, more generic such models can have a spindisordered phase with a propagating 'photon,' 20,23 as well as other phases such as coexistence of the magnetic order and the propagating photon. 30 We will see that our findings in the present simulations on the surface of the bosonic TI region are consistent with these earlier results.
Let us now proceed more systematically and, in particular, show how we obtain a generic NCCP 1 model on the surface of the bosonic TI region. For simplicity, we take λ bulk to be very large. For finite λ surf , we need to keep G ,G ,G x y τ degrees of freedom in the (2+1)D 'layer' at z R , while all other G µ ( r ) are zero. Focusing on the spin variables residing on sites ( X,Y,Z = z R -1 / 2 , T ) , the hedgehogs in this (2+1)D system are given precisely by Q z ( x, y, z R -1 , τ ) , which we will denote simply as Q x, y, τ ( ) . The structure of the surface theory is
$$\dots \quad \dots \quad \dots \quad \dots \quad \dots \\ \intertext { in } S _ { \text{surface} } & = S _ { \text{matter-gauge} } + \frac { K } { 2 } \sum ( \nabla \times a - 2 \pi B ) ^ { 2 } \\ \dots & \quad + \frac { \lambda _ { \text{surf} } } { 2 } \sum G ^ { 2 } \,, \\ \text{subject to constraints}$$
subject to constraints
$$\Phi ^ { \dots }, \quad & \nabla _ { x } G _ { x } + \nabla _ { y } G _ { y } + \nabla _ { \tau } G _ { \tau } \equiv \nabla \cdot G = Q ( x, y, \tau ) \equiv \nabla \cdot B \,. \\ & \quad \Phi ^ { \dots } \sim \dots \sim \dots$$
Here S matter -gauge represents the first term in Eq. (23) restricted to the surface degrees of freedom. The above is a 3D statistical mechanics model, and a µ , ( ∇ × a ) µν ≡ ∇ µ a ν - ∇ ν a µ , and B µν from Eq. (23) can be now defined as 3-vectors and are denoted by bold-face (e.g., µ -th component of B is 1 2 /epsilon1 µνρ B νρ ). We have suppressed position indices to simplify notation.
This surface theory has spins plus integer-valued 'currents' G µ sourced and sinked by the hedgehogs of the spin system. When the 'line tension' λ surf for the lines formed by these 'currents' is large, we intuitively expect that the hedgehogs of the spin system are linearly confined. It is not immediately clear, however, what happens when λ surf is small. Below we argue that the surface is still qualitatively described by the same 'hedgehog-suppressed' field theory, which, however, can be in different regimes and have several different phases.
We can deal with the constraints in the partition sum by changing to new variables
$$B ^ { \prime } = B - G \,,$$
which satisfy
$$\nabla \cdot B ^ { \prime } = 0 \,.$$
The action becomes
$$\omega, z = \quad \text{$\text{$\text{$the account connections}$} } \text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text $\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text
$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text#\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{ $\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\test{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\test$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text}$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text.$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text'$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text`$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text<$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text1$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text $\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text-$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$
$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$\text{$$$
$$\inf _ { 1 } \left ( \L a s _ { \text{surface} } = S _ { \text{matter-gauge} } + S _ { \text{gauge,eff} } [ \nabla \times \L a - 2 \pi B ^ { \prime } ] ( 4 0 ) \right ) \\ \cdot \quad \Phi \right ) \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Psi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi\Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phis \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi$$
There are now no constraints on the G variables, and we can formally sum these out to obtain a local action which is a function of ∇ × a -2 π B ′ ,
However, any such action with the compact variables a and divergenceless, integer-valued B ′ can be formally viewed as describing a non-compact gauge field! In the limit of large λ surf , the effective action will have essentially lattice Maxwell form with stiffness K , while for intermediate to small λ surf the gauge field energy will have more complicated but still local form. Thus, the field theory at the surface has the spinon matter fields coupled to a non-compact gauge field. In particular, we expect that the surface can be in the same phases as the (2+1)D easy-plane NCCP 1 model.
We can also study how the surface action is coupled to the external gauge fields introduced in Eq. (28). The minimal coupling between J µ and A ext 2 µ , combined with the change of variables in Eq. (33), leads to the following term:
$$i \sum _ { r, \mu } G _ { \mu } ( r ) A _ { 2 \mu } ^ { \text{ext} } ( r ) + i \sum _ { r, \mu } \eta ( r ) Q _ { \mu } ( r ) A _ { 2 \mu } ^ { \text{ext} } ( r ) \,. \quad ( 4 1 ) \quad \text{scribe} \\ _ { - } \,. \quad. \quad. \quad. \quad. \quad \text{phase}$$
It is convenient here to represent Q µ ( r ) as
$$Q _ { \mu } ( r ) = \frac { 1 } { 2 } \epsilon _ { \mu \nu \rho \sigma } \nabla _ { \nu } \left ( B _ { \rho \sigma } - \frac { \nabla _ { \rho } a _ { \sigma } - \nabla _ { \sigma } a _ { \rho } } { 2 \pi } \right ) \,, \quad ( 4 2 ) \quad \text{$\text{which} \\ N C C \\ \text{are tal} }. \quad \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \cdots \colon$$
where we have added a formal zero to the defining Eq. (24). Using this in the preceding equation and integrating the second term by parts, we find both an additional bulk term as well as a surface term which results from taking a derivative of η r ( ) . Focusing again on the (2+1)D layer at z R , we can write the surface contributions as
$$i \sum \left ( G + \frac { \nabla \times a } { 2 \pi } - B \right ) \cdot A _ { 2 } = i \sum \frac { \nabla \times \alpha } { 2 \pi } \cdot A _ { 2 } \,, \, ( 4 3 ) \, \begin{array} { c } \text{tons} z \\ N C C \\ \text{ken bu} \\ \beta _ { \text{surf} } \, i \end{array}$$
where we defined
$$\nabla \times \alpha \equiv \nabla \times a - 2 \pi B ^ { \prime } \,, \quad \quad ( 4 4 ) \quad \ z _ { \uparrow } \, \varepsilon \, \\ \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \colon$$
which is precisely the flux of the non-compact gauge field identified in Eq. (40). When this is combined with Eq. (40), we are left with an effective action for the surface with schematic Lagrangian density
$$& \left | \left ( \nabla - i \alpha \mp i \frac { A _ { 1 } ^ { \text{ext} } } { 2 } \right ) z _ { \uparrow / \downarrow } \right | ^ { 2 } + \frac { \kappa } { 2 } ( \nabla \times \alpha ) ^ { 2 } + i \frac { \nabla \times \alpha } { 2 \pi } \cdot A _ { 2 } \,. \quad \text{surfac} \\ & \text{This action, which we derived from our lattice model, has pre-} \\ & \text{$i$} \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \, \dots \end{cases} \, \stackrel { \text{comp.} } { \text{$\text{$\text{$n$}$} } } \, \text{$\text{$\text{$n$} } } \, \text{$\text{$n$} } \, \text{$\text{$n$} } \, \text{$\text{$n$} } \, \text{$\text{$n$} } \, \text{$\text{$n$} } \, \text{$\text{$n$} } \, \text{$\text{$n$} } \, \text{$\text{$n$} } \, \text{$\text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \, \text{$n$} } \,$$
∣ ∣ This action, which we derived from our lattice model, has precisely the easy-plane NCCP 1 form proposed in Ref. 6. In claiming that this is the correct effective action of the surface, we have neglected bulk terms which in general may also contribute to the surface response properties. We do not have an analytical justification for this choice, though from the Monte Carlo study presented in Sec. III E we find that essentially only the surface terms given above contribute to the measured response properties, and it seems plausible that our argument applies in the limit of a sharp boundary between the topological and trivial phases deep in their respective regimes.
Note that the above arguments were based on the assumption that the U (1) and Z T 2 symmetries were preserved in the bulk. If the U (1) symmetry is broken in the bulk, the entire derivation of Eq. (36) based on conserved currents is invalid. On the other hand if only the time reversal is broken in the bulk, the derivation naively holds, but in the matter-gauge sector there is no reason for z ↑ and z ↓ to enter symmetrically-in particular, no reason for them to carry precise +1 2 / and -1 / 2 charges, and the field theory written above is not valid. (In fact, the system will have non-quantized σ xy proportional to the length of the system in the z-direction). Since when the symmetry is broken the bulk ceases to be a topological phase, we of course should not expect exotic physics on the surface in this case.
We can confirm the above arguments, which were made in some simplifying limits, by studying the system in Monte Carlo. We can determine the phase diagram of the surface by looking at singularities in the heat capacity. We can also study the magnetization and current-current correlators as described in the previous section. In this phase diagram we set the bulk parameters so that the system is in the topological phase; specifically, we take β bulk = 0 2 . , K bulk = 0 2 . , and λ = 8 (cf. bottom panel of Fig. 7). We then vary the surface parameters and obtain the phase diagram shown in Fig. 11, which is in good agreement with the phase diagram of the NCCP 1 model in the literature. Labels on the phase diagram are taken from Ref. 30, and their relation to labels in Fig. 4 is described below. There are three phases in the diagram. At small β surf the φ ↑ ↓ , variables are disordered, conserving the U (1) spin symmetry; the U (1) boson symmetry is broken and the corresponding Goldstone mode is precisely the propagating photon in the NCCP 1 theory on the surface, hence the label 'Photon Phase' in Fig. 11. At large β surf , K surf the partons z ↑ , z ↓ are condensed, leading to a 'Higgs Phase' (in the NCCP 1 language) in which the U (1) spin symmetry is broken but the U (1) boson symmetry is preserved. Finally at large β surf and small K surf both the U (1) spin and U (1) boson symmetries are broken. In the NCCP 1 language, the individual z ↑ and z ↓ are gapped so the gauge field a µ is free to fluctuate, but the 'molecular field' Ψ mol ∼ z z † ↑ ↓ , which is precisely the easy-plane spin field, Ψ mol ∼ n a + in b , becomes ordered, hence the label 'Molecular Phase' in Fig. 11. We emphasize that the microscopic model we are simulating in (3+1)D has a compact gauge field, and we are detecting the presence or absence of U (1) spin and U (1) boson symmetry breaking on the surface by direct measurements. It is remarkable that the surface phase diagram is captured by the NCCP 1 field theory with non-compact gauge field!
All of the phases in Fig. 11 break either a U (1) spin or a U (1) boson symmetry and are therefore superfluids. As in the previous section, without access to the properties of their gapped excitations we cannot directly confirm that they are the predicted surface phases. As in the previous section, our phase diagram contains a direct transition between the superfluid phases, which can be viewed as providing some evidence for the proposed surface physics and is also predicted to be a deconfined critical point. 6
## D. Symmetric Surface Phase with Topological Order
Vishwanath and Senthil proposed that it is also possible to have a surface phase which is gapped and breaks no symmetries but has intrinsic topological order. 6 Since this phase is not featured in Fig. 11, we need to add another term to our surface action in order to push the system into this phase. The term we need to add is the following parton 'pair hopping term:' 6,16
$$S _ { \text{pair} } = - t _ { \text{pair} } \sum _ { R, \mu } \cos [ \nabla _ { \mu } ( \phi _ { \uparrow } + \phi _ { \downarrow } ) - 2 a _ { \mu } ], \quad ( 4 5 )$$
1
FIG. 11. Surface phase diagram on the boundary of the SPT phase in the model with a CP 1 version of the spins. The bulk parameters are β bulk = 0 2 . , K bulk = 0 2 . , λ = 8 , and the surface parameters are varied. The phase diagram has the same structure as the one found in the NCCP 1 model in Ref. 30. The phases are the same as those in Fig. 4, though in this figure we have replaced the SO (3) symmetry with U (1) spin , and the β surf axis is oriented horizontally in Fig. 4 and vertically in the present figure.
where we have included proper coupling to the gauge fields. Note particularly that the pair field Ψ pair ∼ z z ↑ ↓ does not carry U (1) spin charge. We can now see what happens to the surface phase diagram (Fig. 11) when we increase t pair trying to induce condensation of Ψ pair . For sufficiently large t pair = 2 , we get the phase diagram in Fig. 12. We see that a new phase has opened up at small β surf and large K surf , where, as we will argue, Ψ pair is condensed while the individual z ↑ and z ↓ are gapped.
When all the couplings β surf , K surf , and t pair are small, there is nothing which can order the spins or gap the bosons. Therefore we are in the photon phase, which conserves U (1) spin and breaks U (1) boson . To get a pairing phase with no broken symmetries, we need to restore the U (1) boson symmetry without breaking the U (1) spin symmetry, which can be acheived by condensing Ψ pair . Let us first recall how the various terms in the action change the system. The β surf term allows hopping of the partons z ↑ ↓ , , but even when this hopping is strong the fluctations in the gauge field a µ when K surf is small prevent the partons from condensing. The combination Ψ mol ∼ z z † ↑ ↓ can condense, and this breaks U (1) spin and takes us to the molecular phase. We can see from Fig. 11 that the K surf term on its own does not change the phase of the system if β surf is kept small. However, when it is combined with the β surf term it can prevent fluctations in the a µ field. This gaps the photon, and allows z ↑ and z ↓ to condense. This gives us the Higgs phase, where the U (1) boson symmetry has been restored, but the U (1) spin symmetry has been broken.
With this in mind, we can see why the t pair term brings us into the topological phase. The t pair term is similar to the β surf in that it is also a hopping term, though it hops pairs of partons. Therefore when both t pair and β surf are large the two terms cooperate, which is why the U (1) spin symmetry breaks at lower β surf in Fig. 12 than in Fig. 11. However, when β surf is absent, the t pair only hops pairs of spinons, and so it can condense Ψ pair without condensing the individual z ↑ ↓ , or Ψ mol . When the t pair term is combined with the K surf term the fluctuations of a µ are gapped. Therefore the phase at large t pair , large K surf , and small β surf can restore U (1) boson without breaking U (1) spin , and this is the phase we are looking for.
We have confirmed that the pairing phase breaks neither U (1) symmetry by direct measurements in the spin and boson sectors. Also, Ψ pair is invariant under the Z T 2 symmetry in Eq. (26), so this symmetry is not broken either, and we indeed do not observe any Hall response on the surface. Therefore we believe that this phase is the fully symmetric gapped phase envisioned by Vishwanath and Senthil, which they argued has intrinsic topological order. Specifically, we expect to have gapped spinon excitations carrying 1/2 of the U (1) spin charge; at the same time, we also have vortices in the Ψ pair field which carry 1/2 of the unit flux of the a µ gauge field and hence 1/2 of the U (1) boson charge; finally, the spinons and the vortices in Ψ pair clearly have mutual statistics of π . Unfortunately, we do not have simple direct measurements to confirm the topological order on the surface. However, the indirect evidence for this scenario is very strong.
Thus, it is suggestive to compare Fig. 12 with the phase diagram obtained in Fig. 1 of Ref. 31. In that work, we studied a (2+1)D model with U (1) × U (1) symmetry and mutual statistics between two different species of bosons. When the mutual statistics is π , we get a phase diagram with the same topology as Fig. 12. The phase diagram contains a topological phase, two phases where one of the U (1) symmetries is broken, and a phase where both symmetries are broken. There is a direct transition between the phases with one broken symmetry, which, if it were continuous, is a candidate for a deconfined critical transition. 32 The surface of our bosonic topological phase is thought to have a similar field theory to that in our previous work, 31,32 and so we expect that the interpretations of the phases and phase transitions are the same in both models. We also remark that in Ref. 7 we presented a microscopic local Hamiltonian which has a topological phase with the same content of excitations. However, that phase also breaks Z T 2 symmetry and in particular has σ xy = 1 2 / , while the present surface phase respects Z T 2 and has no Hall response.
## E. Time-Reversal Breaking and Hall Effect on the Surface
/negationslash
Vishwanath and Senthil 6 predict another exotic phase on the surface of the bosonic topological insulator-a phase which breaks the Z T 2 symmetry and has a Hall conductivity quantized to an odd integer (in units of e 2 h ). We can test this prediction in our Monte Carlo simulations. The first step is to break the Z T 2 symmetry on the surface. By examining Eq. (26), we see that one way to break the symmetry is to replace the parameter β in Eq. (23) by the parameters β ↑ and β ↓ , which appear in the terms containing φ ↑ and φ ↓ respectively. When β ↑ = β ↓ , the Z T 2 symmetry is broken; this is roughly like applying the Zeeman field in Sec. II A 1.
We start with K = 0 4 . and λ = 8 everywhere, β bulk = 0 2 . ,
FIG. 12. Surface phase diagram for the easy-plane CP 1 version of the model, with the additional pairing term on the surface given by Eq. (45). This diagram was obtained for β bulk = 0 2 . , K bulk = 0 2 . , λ = 8 , and on the surface t pair = 2 , β surf and K surf varied. Compared to Fig. 11, we see that there is a new phase at small β surf and large K surf . Weexpect that this surface phase is fully symmetric and has intrinsic topological order.
and β ↑ = β ↓ = 0 2 . on the surface. This system will have its bulk in the topological phase, and its surface in the photon phase. We break the Z T 2 symmetry on one of the surfaces by increasing β ↑ . We expect that a small increase will not change the properties of the system very much, since z ↑ ∼ e iφ ↑ and z ↓ ∼ e iφ ↓ will still be gapped. 23 However, as β ↑ is further increased, z ↑ ∼ e iφ ↑ 'condenses' and vortices in the φ ↑ variables will become gapped. In our simulations we see a singularity in the specific heat measured on the surface, indicating that the system has entered a new phase. We expect that this is the phase that will have Hall conductivity quantized at odd integer. In our simulations we also break time-reversal symmetry in the opposite direction on the other surface by increasing β ↓ . As discussed in Sec. II B, in this setup the top and bottom surface taken together have Hall conductivity adding to two.
We can see that unlike in the Heisenberg model, Eq. (23) is a differentiable function of the probing fields A 1 and A 2 , and so we know how to properly couple the external gauge fields and can use linear response theory to compute the Hall conductivity. If our system has a non-zero Hall conductivity, then we can imagine integrating out the internal degrees of freedom to get the following effective action in terms of the external fields at the surface:
$$S _ { \text{eff} } = i \sum _ { \text{surface} } \frac { \sigma _ { x y } ^ { 1 2 } } { 4 \pi } [ A _ { 1 } \cdot ( \nabla \times A _ { 2 } ) + A _ { 2 } \cdot ( \nabla \times A _ { 1 } ) ] \,, \ ( 4 6 ) \quad \text{$part of} \\ \text{share hold face donation of the reason commonant vector on non-protonia} \quad \text{$z$ =.}$$
where bold face denotes three-component vectors appropriate for the (2+1)D surface, e.g., A 1 = ( A 1 x , A 1 y , A 1 τ ) , and the above form specifies our convention for σ 12 xy (these units are such that e /h 2 = 1 ). By taking, e.g., A 1 = ( A 1 x , 0 0) , and A 2 = (0 , A 2 y , 0) , we have
$$S _ { \text{eff} } = - i \sum _ { \text{surface} } \frac { \sigma _ { x y } ^ { 1 2 } } { 2 \pi } A _ { 1 x } \nabla _ { \tau } A _ { 2 y } \,. \quad \ ( 4 7 ) \quad \text{as tl}$$
Going to momentum space, we can obtain the Hall conductivity by:
$$\Big | \quad \sigma _ { x y } ^ { 1 2 } ( k ) = \lim _ { A _ { 1 }, A _ { 2 } \rightarrow 0 } \frac { 2 \pi } { 2 \sin ( \frac { k _ { \tau } } { 2 } ) } \frac { \partial ^ { 2 } \ln Z } { \partial A _ { 1 x } ( k ) \partial A _ { 2 y } ( - k ) } \,, \quad ( 4 8 ) \\ \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad.$$
where Z is the partition sum, and we also took k = (0 0 , , k τ ) . Note that A 1 and A 2 reside on lattices dual to each other, and when defining the Fourier transforms we take the convention to transform in the absolute coordinates of the origins of the links (namely, lattice sites on the dual lattice have absolute coordinates displaced from the direct lattice by half of lattice spacing). Starting from the microscopic model, we can evaluate this conductivity from the current-current correlation functions:
$$\sigma _ { x y } ^ { 1 2 } ( k ) = \frac { 2 \pi } { 2 \sin ( \frac { k _ { \tau } } { 2 } ) } \left \langle \xi _ { x } ( - k ) J _ { y } ( k ) \right \rangle, \quad ( 4 9 )$$
where the U (1) spin current on a link R,µ is
$$\text{extra} \quad \xi _ { \mu } ( R ) \equiv \frac { i } { 2 } \left [ \beta _ { \uparrow } \sin ( \nabla _ { \mu } \phi _ { \uparrow } - a _ { \mu } ) - \beta _ { \downarrow } \sin ( \nabla _ { \mu } \phi _ { \downarrow } - a _ { \mu } ) \right ] \,.$$
The measurements are performed at the smallest wave-vector k min = (0 0 2 , , π/L ) , as described in Section II A.
Note that the above conductivity measures the response of the U (1) spin currents to applied fields coupled to the bosons (or vice-versa). To study SPTs with single U (1) , we can take the usual approach in the literature 6 and 'glue' the U (1) spin and U (1) charge by identifying A 1 and A 2 in Eq. (46); the conventional definition of σ xy in the case with single U (1) is then related to the above σ 12 xy via σ xy = 2 σ 12 xy . In particular, when the gauge fields are identified the Hall conductivity of a twodimensional system of bosons is quantized to 2 times an integer (in units of e /h 2 ), while the Hall conductivity on the surface of a topological phase is an odd integer. Therefore where we present numerical values we show 2 σ 12 xy so that the results can be easily compared to the literature values.
Figure 13 shows our numerical measurements of the Hall conductivity. The horizontal axis is the strength of the Z T 2 symmetry breaking, which we will loosely call Zeeman field. We see that initially there is no quantized Hall conductivity, until the Zeeman field is strong enough to forbid one species of vortex (realized here by 'condensing' the corresponding spinon species). At this point the Hall conductivity increases and reaches the value of approximately 1 . Though the value observed is actually slightly less than 1 , we believe that a large part of this is a finite-size effect, and indeed as the system size is increased the Hall conductivity gets closer to the expected value. Note that we performed this measurement at precisely z = z R . It is a priori possible that the Hall conductivity could be spread among several values of z near the surface, but this spreading is apparently very small, so that including additional layers does not change the result. In addition to the plot shown, we have performed the same measurement for several different values of K β , and λ , and found that the quantized result is independent of these parameters as long as the bulk stays in the topological phase. This odd integer cannot be observed in a purely two-dimensional system with
|
↑
-
↓
|
FIG. 13. Hall conductivity on the surface of the binding phase realized in the CP 1 version of the model, measured in units of e /h 2 (see text for details). On each surface we find that the conductivity is quantized to 1 , and this odd-integer value shows that we are on the surface of a bosonic TI. Data was taken for K = 0 4 . , λ = 8 , β ↓ = β bulk = 0 2 . . We measure the same conductivity on both the top and bottom surfaces of the topological phase.
only short-ranged entanglement, and therefore this observation shows that we are measuring the Hall conductivity on the surface of a bosonic TI.
## IV. REALIZING SYMMETRY-ENRICHED TOPOLOGICAL PHASES BY BINDING MULTIPLE HEDGEHOGS TO A BOSON
In all of the above sections, we have studied a system where a single boson is bound to a single hedgehog. In this section we will describe the new physics which arises when our system contains bound states of a boson and multiple hedgehogs. We induce such a binding by making the following modification to Eq. (1):
$$S = S _ { \text{spin} } + \frac { \lambda } { 2 } \sum _ { r, \mu } [ d J _ { \mu } ( r ) - Q _ { \mu } ( r ) ] ^ { 2 }. \quad ( 5 0 ) \quad \text{cont} \quad \\ \dots \quad. \quad. \quad. \quad. \quad. \dots \.$$
Here d is an integer, and for large λ the action will bind a boson to d hedgehogs, since the λ term is minimized by ( Q,J ) = ( d, 1) × integer .
/negationslash
/negationslash
When d = 1 the change of variables in Eq. (10) can no longer be applied. Therefore the phase diagram in this case will be different from the d = 1 case. We can determine the phase diagram by performing Monte Carlo simulations. As an example, the phase diagram for d = 3 is shown in Fig. 14. Phase boundaries were determined from singularities in the specific heat. Note that in the Heisenberg model we can define a maximum of one hedgehog per lattice site, and so this model cannot easily be used to describe the binding of multiple hedgehogs. Therefore all results for d = 1 come from the easy-plane CP 1 model for the spins.
FIG. 14. Bulk phase diagram for the model described in Eqs. (23) and (50), with d = 3 and small β = 0 1 . . We can compare this to the case of d = 1 , where the middle phase is absent and the phase boundaries are straight vertical and horizontal lines; and to d = 2 , where the middle phase is absent and there is a line of phase transitions between the paramagnet/superfluid phase in the lower left corner and the Coulomb/insulator phase in the upper right corner.
Figure 14 presents the phase diagram in the variables λ and K , with fixed β = 0 1 . . At small λ and K , there is no energy cost for either hedgehogs or bosons, and they are independent. This leads to a paramagnet of spins and a superfluid of bosons. The U (1) spin symmetry is preserved, and the U (1) boson symmetry is broken. In contrast, at large λ and K both hedgehog and boson currents are forbidden, leading to a Coulomb phase of spins and an insulator of bosons. Here both the U (1) spin and U (1) boson symmetries are preserved, but the spin system has intrinsic topological order. At large K but small λ , the Coulomb phase of the spin system survives and the hedghogs are gapped, but the bosons are condensed breaking the U (1) boson symmetry. On the other hand, at very small K and large λ the hedgehogs are proliferated and bosons are bound to them, and we are in the binding phase; here, neither symmetry is broken, and we will argue shortly that this is a Symmetry Enriched Topological (SET) phase.
/negationslash
Note that similar phase diagram (at fixed small β ) for d = 1 contains only such four phases, where the binding phase is the SPT phase discussed in the previous sections. For d = 1 these are the only phases, and due to the change of variables in Eq. (10), the phase boundaries are all straight lines. With d = 1 , we have a new phase in the middle of the diagram, and no direct transition from the phase in the lower right to the binding phase. The middle phase can be understood as one in which hedgehog currents are proliferated, but their interactions are still too costly for objects with three hedgehogs and a boson to exist. Therefore such bound states are not proliferated, and individual bosons are also gapped. We expect that this phase preserves the U (1) symmetries from both the spins and bosons and is conventional paramagnet/insulator. The topological phase only arises when K is lowered to the point that it does not penalize significantly objects with a hedgehog current of three and λ is increased to strongly penalize any objects other than the ( Q,J ) = (3 1) , bound states; at this
point these bound states can form and proliferate, and the system enters the topological phase. For other values of d , the phase diagram is expected to have a similar form, with the exception of d = 2 , where in our studies of small system sizes the middle phase is not observed and there is a line of phase transitions between the lower left and upper right phases.
Let us now focus on the properties of the binding phase at large λ and small K , which binds d hedgehogs to a boson. This phase is distinct from the topological phase discussed earlier in this work. In particular, it has intrinsic topological order. Condensing bound states of d hedgehogs causes the electric field lines in the phase to fractionalize, i.e., it is possible to have electric field lines of strength 1 /d . These fractionalized electric field lines are one of the gapped excitations of the system. The other elementary gapped excitation is a single hedgehog, which binds a boson charge of 1 /d . The hedgehog has well-defined statistics as it encircles the electric field lines, and we expect it to acquire a phase of 2 π/d when this happens around the elementary fractionalized line. The matter fields z ↑ and z ↓ are confined, but still act as sources and sinks for the electric field lines of integer strength, therefore the line topological excitations in the system are only defined up to an integer, and we can say that the system has Z d topological order. 33,34
In the Appendix we formally demonstrate the above properties by first removing the spinon matter fields and considering a CQED × U (1) boson system in which the monopoles of the compact electrodynamics are bound to bosons. Such CQED × U (1) boson models allow changes of variables similar to those possible in U (1) × U (1) models demonstrating SPT and SET phases of bosons in two dimensions 7 , which allow their properties to be readily determined. After the change of variables has been performed, we can couple the additional spinon matter fields to the CQED sector, and this gives us precisely the CP 1 × U (1) boson model studied in this paper.
The Witten effect and Hall effect measurements can be extended to the cases with multiple hedgehogs. For the Witten effect, the amount of bound charge will be modified, since now for each hedgehog there is a boson charge of 1 /d . Therefore the screening cloud will have a charge of 1 / (2 d ) . We have studied the cases of d = 2 3 , in Monte Carlo and our results, shown in Fig. 15, confirm this expectation. Recall that when measuring the Witten effect we used a 'biasing' term to break degeneracy between positive and negative internal monopoles. This biasing term may also introduces some excess internal monopole or boson charge at the location of the external monopole. This excess is screened by the surrounding system. In the d = 1 case, we found that for most values of γ , such as those shown in Fig. 9, the screening length is quite short and so the Witten effect can still be clearly observed. In the case of d > 1 , the screening length seems to be much larger, which can lead to fluctuations of charge larger than the Witten effect we are trying to observe. We have chosen the biasing parameter in Fig. 15 in such a way as to minimize these fluctuations. At other values of γ the Witten effect can still be observed but the observation is less clear due to these fluctuations.
We have also measured the surface Hall effect upon break-
FIG. 15. Witten effect measurement, similar to Fig. (8), but for the case where d hedgehogs are bound to each boson. The amount of bound charge is given by 1 (2 / d ) , as expected (indicated by dashed lines). Parameters K β , and λ were chosen to put the bulk into the binding phase, while γ was chosen to minimize the amount of charge away from w = 2 , though the value at w = 2 is independent of this choice.
ing the Z T 2 symmetry on the surface by applying the Zeeman field as in Sec. III E. Our results for the Hall conductivity are shown in Fig. 16. We find that the surface Hall conductivity is given by 1 /d , which is one-half of the value found for a two-dimensional bosonic fractional quantum Hall effect. 7 We can again rationalize this observation by considering a slab of the binding phase as in Fig. 2 with the opposite Zeeman fields on the two surfaces, cf. discussion in the paragraph preceding Eq. (20).
## V. DISCUSSION AND CONCLUSIONS
In this work we have realized a three-dimensional bosonic topological insulator in a lattice model which can be studied in Monte Carlo. The model works by binding bosons to point topological defects (hedgehogs). We determined the phase diagram in both the bulk and on the surface of the model. We were able to numerically extract signatures of the topological behavior: In the bulk of the model we observed a Witten effect, while on the surface with broken time-reversal symmetry we found a quantized Hall conductance with values distinct from those possible in a purely (2+1)D system. We also found other surface properties consistent with the bosonic TI, including a direct transition between surface phases with different broken symmetries, and a surface phase which breaks no symmetries and may possess topological order of a kind that would break Z T 2 in a purely (2+1)D system. Finally, we can also realize phases with intrinsic topological order in the bulk by binding multiple topological defects to each boson.
Our model can in principle be used to extract other properties of the bosonic topological insulator. One possible future direction is to determine the properties of the surface phase
|
↑
-
↓
|
FIG. 16. Surface Hall conductivity at a boundary of an SET phase for diffent values of d , in units of e /h 2 . Wesee that the Hall conductivity is given by 1 /d . Both surfaces have been averaged over to improve statistics. Dashed lines are drawn at 1 /d to guide the eye. Data was taken with K = 0 2 . , λ = 8 . Different values of β bulk were used for different values of d , since the phase diagram changes as d changes (see for example Fig. 14) and β bulk needs to be chosen so that the system is in the topological phase.
transitions, especially the transition between the two different surface superfluids. Another direction would be to find direct evidence of the exotic properties of the surface superfluids and surface topologically ordered phase. Such surface phases have generated much recent interest, as their excitations are expected to have properties not possible in a purely two-dimensional system. 6,35,36 It would also be interesting to investigate the surface physics of the SET phases discussed in Sec. IV.
In an earlier work on the two-dimensional bosonic topological phases, 7 we were also able to reconstruct (starting from Euclidean space-time actions) explicit microscopic Hamiltonians which realized the bosonic integer and fractional quantum Hall effects. We have been unable to do the same in the present three-dimensional case, but this would be a very interesting result. More broadly, the idea of binding bosons to topological defects may continue to yield precise models of bosonic topological phases.
The model in the main text can be thought of as having either U (1) /multicloseright Z T 2 or U (1) × Z T 2 symmetry and in both cases we are realizing the same phase. Based on the cohomology classification 3 , the symmetry U (1) /multicloseright Z T 2 has a Z 2 2 classification in three dimensions; i.e. there are two base phases with non-trivial topology, but one of the base phases comes from Z T 2 only. There is only one phase that involves U (1) in a non-trivial manner, and it is this phase that we are realizing in our construction and its signature is the Witten effect, which we discuss and observe in Sec. III B. Our construction cannot access the phase that comes from the Z T 2 symmetry only because we require the U (1) symmetry. In the U (1) × Z T 2 case, the cohomology classification gives Z 3 2 , where one of the base phases is again from the Z T 2 alone, while the other two base phases involve U (1) in a non-trivial manner. Of the latter two, again we are realizing the one which has the statistical Witten effect as its signature. The other base phase does not have the Witten effect, its signature is that the monopole is the Kramers doublet under time reversal. 6,10 Our model in principle has more symmetries and could be also deformed to produce this other phase. We do not consider this here but it is a possible direction for future work. We are also not considering beyond cohomology phases, which bring yet another phase due to Z T 2 only in either case. 6,12 .
The Appendix contains a more abstract compact quantum electrodynamics (CQED) model where multiple hedgehogs are bound to a boson, which realizes generalizations of SPT and SET phases in more abstract settings where lattice gauge theories are included as microscopic degrees of freedom. Such generalizations have been discussed formally recently, 33,37-39 and similar ideas can be useful for constructing explicit models and analysing physical properties of such phases. As a simple demonstration, in a forthcoming publication we will consider a lattice CQED model where multiple monopoles condense and lead to a novel topological phase of CQED with fractionalized Faraday lines. 34
## ACKNOWLEDGMENTS
We would like to thank M. Metlitski, M. P. A. Fisher, F. Burnell, A. Kapustin, C. von Keyserlingk, J. Preskill, T. Senthil, and A. Vishwanath for many useful discussions. We are particularly indebted to Max Metlitski and Matthew Fisher for sharing their unpublished results and crucial insights guiding our work, and also for carefully reading the manuscript and suggesting many improvements. This research is supported by the National Science Foundation through grant DMR-1206096, and by the Caltech Institute of Quantum Information and Matter, an NSF Physics Frontiers Center with support of the Gordon and Betty Moore Foundation.
## Appendix A: SPT- and SET-like phases in a CQED × boson model in (3+1)D
Here we consider a model whose degrees of freedom are compact quantum electrodynamics (CQED) residing on a (3+1)-dimensional cubic space-time lattice labelled by R , and bosons residing on a lattice labeled by r . Topological excitations in the CQED are monopoles, which are quantum particles in three spatial dimensions, and our model realizes condensation of bound states of d monopoles and c bosons. We will argue that when d = 1 , these phases are analogs of Bosonic Symmetry-Protected Topological phases for such CQED × boson systems and have quantized cross-transverse response given by the integer c . On the other hand, when d > 1 these phases are analogs of Bosonic Symmetry-Enriched Topological phases and have fractional cross-transverse response given by a rational number c/d ; these phases also feature fractionalized Faraday line excitations of the CQED and
fractionalized boson particle excitations, as well as non-trivial mutual statistics between the particle and line excitations. 33,34
While the CQED × boson setting may appear artificial, this problem is relevant to the problem of SPT and SET phases of bosons in (3+1)D. Specifically, in the main text we had a model with U (1) × U (1) symmetry, and we represented the first U (1) system as an easy-plane CP 1 model, which has two matter fields ('spinons') coupled to a compact gauge field. We then considered binding of the monopoles in the compact gauge field and physical bosons of the second U (1) symmetry. The crucial difference with the CQED × boson model is the additional matter fields present in the CP 1 × boson model. We will compare the systems without and with such matter fields and will argue that the matter fields destroy the distinctions of the former model, except when protected by additional discrete symmetries.
Our CQED × boson model is written in terms of compact gauge fields for the CQED part and integer-valued conserved currents for the boson part:
$$Z [ h _ { \mu \nu } ^ { \text{ext} } ( R ), A _ { \rho } ^ { \text{ext} } ( r ) ] & = \sum _ { J _ { \rho } ( r ) } ^ { \prime } \int _ { 0 } ^ { 2 \pi } D a _ { \mu } ( R ) \int _ { 0 } ^ { 2 \pi } \prod _ { \mu < \nu } d _ { \gamma _ { \mu \nu } } \sum _ { B _ { \mu \nu } ( R ) = - \infty } ^ { \infty } e ^ { - S [ a _ { \mu } ( R ), \gamma _ { \mu \nu }, B _ { \mu \nu } ( R ), J _ { \rho } ( r ) ; h _ { \mu \nu } ^ { \text{ext} } ( R ), A _ { \rho } ^ { \text{ext} } ( r ) ] } \,, \quad ( A _ { 1 } ) \\ z & = \sum _ { f } ^ { \text{ext} } \omega \quad \frac { 1 } { 2 }$$
$$\omega _ { \mu \nu } ( R ) \equiv ( \nabla _ { \mu } a _ { \nu } - \nabla _ { \nu } a _ { \mu } ) ( R ) - \delta _ { R _ { \mu } = 0 } \delta _ { R _ { \nu } = 0 } \gamma _ { \mu \nu } \,.$$
$$& \quad \, \rho ( \nu ) \,, \quad \, \ r \rangle \, - \, \infty \\ & S = & \frac { K } { 2 } \sum _ { R, \mu < \nu } \left [ \omega _ { \mu \nu } ( R ) - h _ { \mu \nu } ^ { \text{ext} } ( R ) - 2 \pi B _ { \mu \nu } ( R ) \right ] ^ { 2 } + \frac { \lambda } { 2 } \sum _ { r, \rho } \left [ c Q _ { \rho } ( r ) + c \frac { \mathfrak { g } _ { \rho } ^ { \text{ext} } ( r ) } { 2 \pi } - d J _ { \rho } ( r ) \right ] ^ { 2 } + i \sum _ { r, \rho } J _ { \rho } ( r ) A _ { \rho } ^ { \text{ext} } ( r ), \quad ( A 2 ) \\ & \omega _ { \mu \nu } ( R ) \equiv ( \nabla _ { \mu } a _ { \nu } - \nabla _ { \nu } a _ { \mu } ) ( R ) - \delta _ { R _ { \mu } = 0 } \delta _ { R _ { \nu } = 0 } \gamma _ { \mu \nu } \,.$$
The above action can be compared to the action of Eqs. (23) and (50) in the main text. Compared to the main text, the above action is missing the β term which couples the gauge field to 'spinons'. We also allow the option of binding multiple bosons-here binding c bosons and d monopoles. The boson sector has conserved particles and is coupled to a probing field A ext ρ in the standard way, as in the main text. Unlike the main text, the CQED sector has conserved electric field lines and is coupled to a probing rank-2 field h ext µν ( R ) . In the absence of the binding term, the coupling of the CQED sector to h ext µν is standard; 40 the additional piece in the binding term, while not important for much of the discussed longdistance physics, is the correct form keeping together the 'Dirac string' B µν and the probing field h ext µν . Note that in the main text the additional matter fields destroy the conservation of the electric field lines, and therefore the CQED sector cannot be probed by such an external rank-2 field. Variables γ µν realize specific 'fluctuating boundary conditions' in the compact gauge field variables; in a representation in terms of an integer-valued electromagnetic field tensor, this corresponds to requiring zero total field for each component. Such details of the boundary conditions are not important for the bulk properties but are nice for a precise mathematical treatment in a system with periodic connectedness assumed here.
From variables B µν ( R ) , we define the monopole four-current Q ρ ( r ) as in Eq. (24) in the main text. We similarly define the four-vector g ext ρ ( r ) from h ext µν ( R ) :
$$\mathfrak { g } _ { \rho } ^ { e x t } ( r ) \equiv \frac { 1 } { 2 } \epsilon _ { \rho \sigma \mu \nu } \nabla _ { \sigma } h _ { \mu \nu } ^ { e x t } \,. \quad \quad ( A 4 )$$
Note that thus defined Q ρ ( r ) are conserved currents satisfying ∑ ρ ∇ ρ Q ρ ( r ) = 0 , and they also satisfy the condition of zero total current in all directions: Q tot ,ρ ≡ r Q ρ ( r ) = 0 .
We first separate out the monopoles in the CQED sector as follows:
∑ For the boson sector, we use a representation in terms of integer-valued conserved currents J ρ ( r ) , which satisfy ∑ ρ ∇ ρ J ρ ( r ) = 0 . We also require that the total boson current is equal to zero for all directions, J tot ,ρ = 0 . The primed sum over J ρ ( r ) in the partition sum signifies all such constraints. The condition of zero total current is again convenient for precise treatment on finite systems [namely, for performing change of variables involving J and Q currents, Eq. (A6) below], while for bulk properties one can ignore these details. The 'binding' term parametrized by λ is the key interaction in the action Eq. (A2) and wants to have bound states of d monopoles and c bosons when λ is large [i.e., it is minimized when ( Q,J ) = ( d, c ) × integer ].
$$\sum _ { B _ { \mu \nu } ( R ) = - \infty } ^ { \infty } [ \dots ] = \sum _ { Q _ { \nu } ( r ) = \frac { 1 } { 2 } \epsilon _ { \mu \nu } \nabla _ { \varepsilon } B _ { \mu \nu } ^ { ( 0 ) } } ^ { \prime } \sum _ { V _ { \mu } ( R ) = - \infty } ^ { \infty } \sum _ { N _ { \mu \nu } = - \infty } ^ { \infty } [ B _ { \mu \nu } ( R ) = B _ { \mu \nu } ^ { ( 0 ) } ( R ) + ( \nabla _ { \mu } V _ { \nu } - \nabla _ { \nu } V _ { \mu } ) ( R ) + \delta _ { R _ { \mu } = 0 } \delta _ { R _ { \nu } = 0 } N _ { \mu \nu } ] \,. ( A 5 )$$
Here B (0) µν ( R ) is an integer-valued field whose monopolicity gives Q ρ ( r ) and is treated as a fixed function of Q ρ ( r ) . V µ ( R )
and N µν are independent integer-valued fields, where the latter appear from careful treatment of the boundary conditions.
Schematically, the above arises by dividing all configurations of B µν ( R ) into classes defined by Q ρ ( r ) and establishing how to recover all members of a given class from one representative. Furthermore, any Q ρ ( r ) satisfying ∑ ρ ∇ ρ Q ρ ( r ) = 0 and Q tot ,ρ = 0 can be represented as above using some integer-valued B µν ( R ) , so the primed sum over Q ρ ( r ) can be viewed as signifying these constraints.
A subtle point in the above is the redundancy in V µ ( R ) . One way to address it precisely is as follows. We can argue that from all links of the (3+1)D hyper-cubic lattice, we can select a subset of links such that we can take V µ ( R ) as independent integer-valued variables on these links, while V µ ( R ) = 0 on all the other links. We can also argue that the original CQED sector in Eq. (A1) can be equivalently formulated using compact gauge fields a µ ( R ) that are non-zero only on exactly the same links as the independent V µ ( R ) . We will assume this implicitly in all manipulations below, both for these variables and for the related variables V ′ µ ( R ) , ˜ V µ ( R ) , v µ ( R ) , k µ ( R ) , and ˜ a µ ( R ) that will appear below.
We now operate with the constrained sums over the integervalued currents Q ρ ( r ) and J ρ ( r ) , taken to be 'outside-most' sums in the partition sum. We change to new independent summation variables on each link: 32
$$\begin{array} { c c c } P = a Q - b J, & Q = d P - b G, \\ G = c Q - d J, & \leftrightarrow & J = c P - a G \,. & \quad & ( \text{A6} ) & \quad \omega _ { \mu \nu } \\ & & d \, \Big | _ { i } \end{array}$$
Here c and d are the same integers as in the binding term in Eq. (A2), while a and b are new integers such that ad -bc = 1 , which makes the above transformation invertible in Z . If c and d are mutually prime, which we will assume throughout, we can always find such a and b , while the arbitrariness a → a + kc, b → b + kd does not affect any physical properties discussed below.
The new variables P ρ ( r ) and G ρ ( r ) are also conserved integer-valued currents with zero total currents. For each P ρ ( r ) and G ρ ( r ) , we can uniquely determine Q ρ ( r ) and hence the corresponding fixed B (0) µν ( R ) . Let us also find for each P ρ ( r ) and G ρ ( r ) some fixed integer-valued B (0) P,µν ( R ) and B (0) G,µν ( R ) such that
$$P _ { \rho } ( r ) = \frac { 1 } { 2 } \epsilon _ { \rho \sigma \mu \nu } \nabla _ { \sigma } B _ { P, \mu \nu } ^ { ( 0 ) } \,, \quad G _ { \rho } ( r ) = \frac { 1 } { 2 } \epsilon _ { \rho \sigma \mu \nu } \nabla _ { \sigma } B _ { G, \mu \nu } ^ { ( 0 ) } \,. \quad \text{appear} \\ \mathfrak { w } _ { - \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \sigma \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \ dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots\dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \delta \dots \delta \dots \delta \dots \delta \dots \delta \dots \delta \dots \delta \dots \delta \delta \dots \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \Delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta$$
We can guarantee that given the constraints satisfied by P ρ ( r ) and G ρ ( r ) , such two-forms B (0) P,µν ( R ) and B (0) G,µν ( R ) always exist; while there are many possible choices, it does not matter which one we use as long as it stays fixed. We then have
$$\epsilon _ { \rho \sigma \mu \nu } \nabla _ { \sigma } \left [ B ^ { ( 0 ) } _ { \mu \nu } - d B ^ { ( 0 ) } _ { P, \mu \nu } + b B ^ { ( 0 ) } _ { G, \mu \nu } \right ] = 0 \,, \quad ( A 8 ) \\ \text{which imnlies that there exist into over-valued $V^{\prime}(R)\N^{\prime}$ such}$$
which implies that there exist integer-valued V ′ µ ( R ) , N ′ µν such that
$$\mathfrak { z } ). \quad B ^ { ( 0 ) } _ { \mu \nu } ( R ) & = d B ^ { ( 0 ) } _ { P, \mu \nu } ( R ) - b B ^ { ( 0 ) } _ { G, \mu \nu } ( R ) \quad ( A 9 ) \\ \intertext { a r - \intertext { s e } } \quad + ( \nabla _ { \mu } V ^ { \prime } _ { \nu } - \nabla _ { \nu } V ^ { \prime } _ { \mu } ) ( R ) + \delta _ { R _ { \mu } = 0 } \delta _ { R _ { \nu } = 0 } N ^ { \prime } _ { \mu \nu } \,. \\ \intertext { a r - \intertext { s e } } \quad \text{s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \| s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - t e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \intertext { s e - \eta - \intertext { s e - \intertext { s e - \intertext { s e - \eta - \intertext { s e - \eta - \intertext { s e - \eta - \intertext { s e - \eta - \intertext { s e - \eta - \intertext { s e - \eta - \intertext { s e - \eta - \intertext { s e - \eta - \intertext { s e - \eta - \intertext { s e - \eta - \intertext { s e - \eta - \intertext { s e - \eta - \intertext { s e - \eta - \eta - \intertext { s e - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - s e - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \beta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - \eta - 1 } } }$$
We can again take V ′ µ ( R ) , N ′ µν as fixed functions of P ρ ( r ) and G ρ ( r ) . Now we can express B µν ( R ) in Eq. (A5) as
$$\text{at} \quad B _ { \mu \nu } ( R ) & = d B _ { P, \mu \nu } ^ { ( 0 ) } ( R ) - b B _ { G, \mu \nu } ^ { ( 0 ) } ( R ) \quad \text{(A10)} \\ \text{to} \quad & + ( \nabla _ { \mu } \tilde { V } _ { \nu } - \nabla _ { \nu } \tilde { V } _ { \mu } ) ( R ) + \delta _ { R _ { \mu } = 0 } \delta _ { R _ { \nu } = 0 } \tilde { N } _ { \mu \nu } \, \\ \text{to} \quad & \text{where a two chosen and that communication on available from $V$ / $D$} \.$$
where we have changed the summation variables from V µ ( R ) , N µν to ˜ V µ ( R ) = V µ ( R ) + V ′ µ ( R ) , ˜ N µν = N µν + N ′ µν . Let us write
$$\tilde { V } _ { \mu } ( R ) = d v _ { \mu } ( R ) + k _ { \mu } ( R ), \ \tilde { N } _ { \mu \nu } = d n _ { \mu \nu } + \ell _ { \mu \nu }, ( \text{A11} )$$
where v µ ( R ) and n µν are arbitrary integers, while k µ ( R ) and /lscript µν are integers 0 1 , , . . . , d -1 . Upon simple grouping of terms, we now have
$$( \mathbf A 6 ) & \quad \omega _ { \mu \nu } ( R ) - h _ { \mu \nu } ^ { \text{ext} } ( R ) - 2 \pi B _ { \mu \nu } ( R ) = & ( \mathbf A 1 2 ) \\ & \quad d \left [ \tilde { \omega } _ { \mu \nu } ( R ) - \frac { 1 } { d } h _ { \mu \nu } ^ { \text{ext} } ( R ) + \frac { 2 \pi b } { d } B _ { G, \mu \nu } ^ { ( 0 ) } ( R ) - 2 \pi \tilde { B } _ { \mu \nu } ( R ) \right ] \,, \\ \text{rm in} \\ \cdot = \, 1 & \quad \text{where we defined}$$
where we defined
$$1, \quad & \text{where we eennea} \\ \i d _ { \vee } \tilde { \omega } _ { \mu \nu } ( R ) & \equiv ( \nabla _ { \mu } \tilde { a } _ { \nu } - \nabla _ { \nu } \tilde { a } _ { \mu } ) ( R ) - \delta _ { R _ { \mu } = 0 } \delta _ { R _ { \nu } = 0 } \tilde { \gamma } _ { \mu \nu } \,, \\ \rightarrow & \text{$\tilde{a}_{\mu}(R) \equiv \frac{a_{\mu}(R)}d-\frac{2\pik_{\mu}(R)}d$} \,, \\ \tilde { \gamma } _ { \mu \nu } & \equiv \frac { \gamma _ { \mu \nu } } d + \frac { 2 \pi \ell _ { \mu \nu } } d \,, \\ \i d _ { \tilde { B } _ { \mu \nu } ( R ) } & \equiv B _ { P, \mu \nu } ^ { ( 0 ) } ( R ) + ( \nabla _ { \mu } v _ { \nu } - \nabla _ { \nu } v _ { \mu } ) ( R ) \\ & \quad + \delta _ { R _ { \mu } = 0 } \delta _ { R _ { \nu } = 0 } n _ { \mu \nu } \,. \\ \ast & \text{$\text{$\text{$\text{$\text{for}_{\mu}(R)$ from $0$ to $2\pi$ and sum-}$} } $}.$$
Note that the integration over a µ ( R ) from 0 to 2 π and summation over k µ ( R ) from 0 to d -1 effectively corresponds to integration over ˜ a µ ( R ) from 0 to 2 π . The same holds for ˜ γ µν . We can also express P ρ ( r ) from Eq. (A7) as P ρ ( r ) = 1 2 /epsilon1 ρσµν ∇ σ ˜ B µν and see that B (0) P,µν ( R ) , v µ ( R ) , and n µν only appear in the above combination that gives ˜ B µν ( R ) . Furthermore, similarly to Eq. (A5), the summation over constrained P ρ ( r ) and unconstrained v µ ( R ) and n µν is equivalent to a summation over unconstrained ˜ B µν . The partition sum becomes
$$Z [ h _ { \mu \nu } ^ { \text{ext} } ( R ), A _ { \rho } ^ { \text{ext} } ( r ) ] & = \sum _ { G _ { \rho } ( r ) } ^ { \prime } \int _ { 0 } ^ { 2 \pi } D \tilde { a } _ { \mu } ( R ) \int _ { 0 } ^ { 2 \pi } \prod _ { \mu < \nu } \sum _ { \tilde { B } _ { \mu \nu } ( R ) = - \infty } ^ { \infty } e ^ { - S [ \tilde { a } _ { \mu \nu } ( R ), \tilde { \gamma } _ { \mu \nu }, \tilde { B } _ { \mu \nu } ( R ), G _ { \rho } ( r ) ; h _ { \mu \nu } ^ { \text{ext} } ( R ), A _ { \rho } ^ { \text{ext} } ( r ) ] }, ( A 1 4 ) \\ \intertext { r n o } \frac { r } { r } \left ( \Delta \right ) \frac { 1 } { 2 } \right ) \frac { 1 } { 2 } \right ) \frac { 1 } { 2 } \right ) \frac { 1 } { 2 } \right ) \frac { 1 } { 2 } \right )$$
$$S & = \frac { K d ^ { 2 } } { 2 } \sum _ { R, \mu < \nu } \left [ \tilde { \omega } _ { \mu \nu } ( R ) - \frac { 1 } { d } h _ { \mu \nu } ^ { e x t } ( R ) + \frac { 2 \pi b } { d } B _ { G, \mu \nu } ^ { ( 0 ) } ( R ) - 2 \pi \tilde { B } _ { \mu \nu } ( R ) \right ] ^ { 2 } + \frac { \lambda } { 2 } \sum _ { r, \rho } \left [ G _ { \rho } ( r ) + c \frac { \mathfrak { g } _ { \rho } ^ { e x t } ( r ) } { 2 \pi } \right ] ^ { 2 } \\ & + i \sum _ { r. o } \left [ c \left ( \frac { 1 } { 2 } \epsilon _ { \rho \sigma \mu \nu } \nabla _ { \sigma } \tilde { B } _ { \mu \nu } \right ) ( r ) - a G _ { \rho } ( r ) \right ] A _ { \rho } ^ { e x t } ( r ) \.$$
$$S = \frac { K d ^ { 2 } } { 2 } \sum _ { R, \mu < \nu } \left [ \tilde { \omega } _ { \mu \nu } ( R ) - \frac { 1 } { d } h _ { \mu \nu } ^ { \text{ext} } ( R ) + \frac { 2 \pi b } { d } B _ { G, \mu \nu } ^ { ( 0 ) } ( R ) - 2 \pi \tilde { B } _ { \mu \nu } ( R ) \right ] ^ { 2 } + \frac { \lambda } { 2 } \sum _ { r, \rho } \left [ G _ { \rho } ( r ) + c \frac { \theta _ { \rho } ^ { \text{ext} } ( r ) } { 2 \pi } \right ] ^ { 2 } \quad \quad ( A 1 5 )$$
$$+ \, i \sum _ { r, \rho } \left [ c \left ( \frac { 1 } { 2 } \epsilon _ { \rho \sigma \mu \nu } \nabla _ { \sigma } \tilde { B } _ { \mu \nu } \right ) ( r ) - a G _ { \rho } ( r ) \right ] A _ { \rho } ^ { \text{ext} } ( r ) \.$$
In this reformulation, we have a new CQED system represented by the compact gauge fields ˜ a µ ( R ) ∈ [0 , 2 π ) [and fluctuating boundary conditions realized with ˜ γ µν ∈ [0 , 2 π ) ], and a new boson system represented by the conserved current variables G ρ ( r ) . We will now argue that these new variables give a direct representation of gapped excitations in the topological phase obtained for small K and large λ .
field, we can also show that the G particles carry charge of 1 /d with respect to A ext (i.e., we also have boson charge fractionalization in this phase if d > 1 ), and that the system has a quantized cross-transverse response characterized by a rational number c/d . First, we note that the current which couples to A ext ρ ( r ) can be written as
For a quick illustration, let us set A ext ρ ( r ) = 0 everywhere. In this case, we can perform summation over ˜ B µν ( R ) independently on each placket and obtain the Villain cosine for the combination
$$\Theta _ { \mu \nu } ( R ) \equiv \tilde { \omega } _ { \mu \nu } ( R ) - \frac { 1 } { d } h _ { \mu \nu } ^ { \text{ext} } ( R ) + \frac { 2 \pi b } { d } B _ { G, \mu \nu } ^ { ( 0 ) } ( R ) \,. ( A 1 6 ) \quad \text{where}$$
Such cosine terms represent dynamics (motions) of quantum lines whose segments are created by e ia ˜ µ ( R ) . From the coupling to h ext µν ( R ) , we conclude that the new lines carry electric field strength 1 /d of the original electric field unit. On the other hand, from the appearance of B (0) G,µν ( R ) in Θ µν ( R ) we conclude that the new lines 'see' each G particle as if it is carrying b/d of monopole charge of the new CQED system. We can state this equivalently as follows: When a G particle is taken around such a new line, there is a phase of 2 πb/d , i.e., there is non-trivial mutual statistics between the new lines and new particles. For sufficiently small K and large λ , the quantum lines created by e ia ˜ µ ( R ) and the particles represented by G ρ ( r ) are clearly gapped and are the only excitations in this phase.
$$\mathfrak { z }. & c P _ { \rho } ( r ) - a G _ { \rho } ( r ) = c \left [ P _ { \rho } ( r ) - \frac { b } { d } G _ { \rho } ( r ) \right ] - \frac { 1 } { d } G _ { \rho } ( r ) \\ \mathfrak { r } & = c \left ( \frac { 1 } { 2 } \epsilon _ { \rho \sigma \mu \nu } \nabla _ { \sigma } \left [ \tilde { B } _ { \mu \nu } - \frac { b } { d } B _ { G, \mu \nu } ^ { ( 0 ) } \right ] \right ) ( r ) - \frac { 1 } { d } G _ { \rho } ( r ) \, \\ \mathfrak { z } ) \quad _ { - \Re \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \$$
where we used ad -bc = 1 and manipulated so as to separate out the same combination of ˜ B µν and B (0) G,µν as in the K -term in Eq. (A15). Second, on each placket R,µ < ν we go from summation over integers ˜ B µν ( R ) to integration over real values as follows:
$$\underset { \substack { \text{$f$ it is} \\ \text{system.} \\ \text{article} \\ \overset { \tilde { B } _ { \mu \nu } ( R ) = - \infty } { \text{$f$} } } } \sum _ { \tilde { B } _ { \mu \nu } ( R ) = - \infty } ^ { \infty } e ^ { - i 2 \pi \tilde { F } _ { \mu \nu } ( R ) \tilde { B } _ { \mu \nu } ( R ) } [ \dots ],$$
By working a bit harder and keeping track of the A ext gauge which can be viewed as an intermediate step in a Poisson resummation from ˜ B µν ( R ) to ˜ F µν ( R ) . Finally, we perform Gaussian integration over the ˜ B µν ( R ) , which can be simplified, e.g., by changing to new integration variable x µν ( R ) ≡ 2 πB ˜ µν ( R ) -2 πb d B (0) G,µν ( R ) + 1 d h ext µν ( R ) -˜ ω µν ( R ) . The resulting action is:
$$S & = \frac { 1 } { 2 K d ^ { 2 } } \sum _ { R, \mu < \nu } \left [ \tilde { F } _ { \mu \nu } ( R ) + c \frac { ( \epsilon _ { \mu \nu \sigma \rho } \nabla _ { \sigma } A _ { \rho } ^ { \text{ext} } ) ( R ) } { 2 \pi } \right ] ^ { 2 } + i \sum _ { R, \mu < \nu } \tilde { F } _ { \mu \nu } ( R ) \tilde { \omega } _ { \mu \nu } ( R ) + \frac { \lambda } { 2 } \sum _ { r, \rho } \left [ G _ { \rho } ( r ) + c \frac { \mathfrak { g } _ { \rho } ^ { \text{ext} } ( r ) } { 2 \pi } \right ] ^ { 2 } \quad ( A 1 7 ) \\ & + i \sum _ { R, \mu < \nu } \frac { 2 \pi b } { d } \tilde { F } _ { \mu \nu } ( R ) B _ { G, \mu \nu } ^ { ( 0 ) } ( R ) - i \sum _ { R, \mu < \nu } \frac { 1 } { d } \tilde { F } _ { \mu \nu } ( R ) h _ { \mu \nu } ^ { \text{ext} } ( R ) - i \sum _ { r, \rho } \frac { 1 } { d } G _ { \rho } ( r ) A _ { \rho } ^ { \text{ext} } ( r ) - i \sum _ { r, \rho } \frac { c } { 2 \pi d } A _ { \rho } ^ { \text{ext} } ( r ) \mathfrak { g } _ { \rho } ^ { \text{ext} } ( r ) \,.$$
We see that the formally introduced ˜ F µν ( R ) can be interpreted as integer-valued electromagnetic field tensor variables conjugate to the compact ˜ a µ ( R ) variables. Integrating out ˜ a µ ( R ) gives 'Maxwell equations in the absence of charges', ∑ ν ∇ ν ˜ F µν ( R ) = 0 , while integrating out ˜ γ µν gives the condition ∑ R δ R µ =0 δ R ν =0 ˜ F µν ( R ) = 0 , which is equivalent to zero total field for each component. For small K , the quantum lines whose worldsheets are represented by ˜ F µν ( R ) are gapped, and for large λ the quantum particles whose worldlines are represented by G ρ ( r ) are gapped, so we have managed to express the partition sum completely in terms of gapped excitations in this phase. From the coupling of ˜ F µν
to h ext µν we explicitly see that the new lines carry fraction 1 /d of the unit electric field strength, while from the coupling of G ρ to A ext ρ we see that the new particles carry charge 1 /d of the original bosons. The coupling between ˜ F µν and B (0) G,µν encodes 2 πb/d statistical interaction between the new elementary line and particle excitations. Finally, the last term i c 4 πd ∑ /epsilon1 ρσµν A ext ρ ∇ σ h ext µν is a property of the 'vacuum' in these variables and represents a kind of cross-transverse response in this phase, which we see is quantized to the rational number c/d in appropriate units.
When d = 1 we have a quantized response, but the quantum numbers of the gapped excitations are not fractionalized, and therefore we have an SPT-like phase characterized by the integer c . When d > 1 the quantum numbers of gapped excitations are fractionalized, and we have intrinsic topological order and an SET-like phase.
The cross-transverse response is due to unusual surface states when the above phase borders a trivial phase ( d = 1 , c = 0 ). Analysis of the surface theory in Sec. III C can be carried over to this case in the absence of the spinon matter, and we see that the surface of the SPT-like phase ( d = 1 , c = 1) in the CQED × boson system has an emergent non-compact electrodynamics whose flux couples to A ext as in Eq. (43). This tells us that the gapless surface mode can be viewed as a (2+1)D photon. Equivalently, the surface can be viewed as a boson system without vortices, which is always in the spinwave phase, and therefore the (2+1)D gapless mode can also be viewed as a (2+1)D phonon.
We can now make the connection to the CP 1 × boson model in the main text, which we will refer to as U (1) spin × U (1) boson model. Here we have additional 'spinon' matter fields coupled to the compact gauge field, schematically
$$\delta S & = S _ { \text{spinons} } [ J _ { \uparrow }, J _ { \downarrow } ] + i \sum _ { R, \mu } [ J _ { \uparrow, \mu } ( R ) + J _ { \downarrow, \mu } ( R ) ] a _ { \mu } ( R ), \quad \text{$item$} \\ \cdot \quad \tau \quad \cdot \tau \quad \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \text{$host$}$$
where J ↑ and J ↓ are integer-valued spinon currents conjugate to the φ ↑ and φ ↓ variables in Eq. (23). Using Eq. (A13) and dropping all contributions of the form i 2 π × integer, we can write
$$\delta S = S _ { \text{spinons} } [ J _ { \uparrow }, J _ { \downarrow } ] + i d \sum _ { R, \mu } & [ J _ { \uparrow, \mu } ( R ) + J _ { \downarrow, \mu } ( R ) ] \tilde { a } _ { \mu } ( R ) \,. \quad \text{in use} \quad \text{track}$$
- 1 M. Z. Hasan and C. L. Kane, Rev. Mod. Phys. 82 , 3045 (Nov 2010)
- 2 X.-L. Qi and S.-C. Zhang, Rev. Mod. Phys. 83 , 1057 (Oct 2011)
- 3 X. Chen, Z.-C. Gu, Z.-X. Liu, and X.-G. Wen, Science 338 , 1604 (2012)Phys. Rev. B 87 , 155114 (Apr 2013)
- 4 Y.-M. Lu and A. Vishwanath, Phys. Rev. B 86 , 125119 (Sep 2012)
- 5 T. Senthil and M. Levin, Phys. Rev. Lett. 110 , 046801 (Jan 2013)
- 6 A. Vishwanath and T. Senthil, Phys. Rev. X 3 , 011016 (Feb 2013)
- 7 S. D. Geraedts and O. I. Motrunich, Annals of Physics 334 , 288 (2013), ISSN 0003-4916
- 8 A. Turner and A. Vishwanath, arXiv:cond-mat, 1301.0330(2013)
- 9 T. Senthil, arXiv:cond-mat, 1405.4015(2014)
- 10 Z. Bi, A. Rasmussen, and C. Xu, arXiv:cond-mat, 1309.0515(2013)
All manipulations leading to Eq. (A17) do not touch the gauge field ˜ a µ ( R ) and remain valid also in the presence of the spinon matter. Now, since the original electric field lines have sources and sinks, the probing field h ext µν ceases to be meaningful and will be dropped; in particular, the characterization using rational c/d for the cross-transverse response collapses. On the other hand, the coupling of the new G particles to A ext remains unaffected, and we see that they carry 1 /d fractional charge with respect to U (1) boson . Furthermore, upon integrating out the ˜ a µ ( R ) fields, we now have
$$\sum _ { \nu } \nabla _ { \nu } \tilde { F } _ { \mu \nu } ( R ) = - d [ J _ { \uparrow, \mu } ( R ) + J _ { \downarrow, \mu } ( R ) ] \,, \quad ( \mathbf A 1 8 )$$
i.e., the spinons act as sources and sinks of the new field lines encoded by the integer-valued ˜ F µν ( R ) , but these sources and sinks are in multiples of d . We assume that the spinons are gapped, and they will clearly be confined in the regime of small K ; however, they destroy the structure that we had of the integer-labeled closed lines. Nevertheless, this destruction happens only for multiples of d , while it is still meaningful to talk about closed line excitations of strengths modulo d , i.e., we have line excitations labeled by Z d . The final result is that there is no fractionalization of the U (1) spin , but we still have Z d quantum line excitations and fractionalized particle excitations carrying charge 1 /d with respect to the U (1) boson , with mutual statistics between these lines and particles.
Note that in either the d = 1 or d > 1 case, when the system contains the U (1) spin and Z T 2 symmetries and c is odd, the resulting phase is distinct from a phase which is a direct product of a trivial spin state times a trivial (or 1 /d fractionalized) boson state. This distinction can be observed by measuring a quantized Witten effect as discussed in the main text. To see this one also needs to carefully include the external gauge field coupled to U (1) spin as we did in Eq. (28), particularly in the presence of external monopoles, which we did not keep track of in this Appendix.
- 11 C. Wang and T. Senthil, Phys. Rev. B 87 , 235122 (Jun 2013)
- 12 A. Kapustin, arXiv:cond-mat, 1404.6659(2014)
- 13 J. Wang, Z.-C. Gu, and X. G. Wen, arXiv:cond-mat, 1405.7689(2014)
- 14 A. M. Essin and M. Hermele, Phys. Rev. B 87 , 104406 (Mar 2013)
- 15 A. Mesaros and Y. Ran, Phys. Rev. B 87 , 155115 (Apr 2013)
- 16 M. Metlitski and M. P. A. Fisher, unpublished notes(2013)
- 17 M. A. Metlitski, C. L. Kane, and M. P. A. Fisher, Phys. Rev. B 88 , 035131 (Jul 2013)
- 18 P. Ye and X.-G. Wen, Phys. Rev. B 89 , 045127 (2014)
- 19 P. Ye and J. Wang, Phys. Rev. B 88 , 235109 (2013)
- 20 M. Kamal and G. Murthy, Phys. Rev. Lett. 71 , 1911 (1993)
- 21 M.-h. Lau and C. Dasgupta, Phys. Rev. B 39 , 7212 (Apr 1989)
- 22 S. Sachdev and K. Park, Annals of Physics N. Y. 298 , 58 (2002)
- 23 O. I. Motrunich and A. Vishwanath, Phys. Rev. B 70 , 075104 (Aug 2004)
- 24 S. McKenzie, C. Domb, and D. L. Hunter, J. Phys. A. Math. Gen. 15 , 3909 (1982)
- 25 O. I. Motrunich and T. Senthil, Phys. Rev. Lett. 89 , 277004 (Dec 2002)
- 26 Z. Bi, A. Rasmussen, and C. Xu, arXiv, 1309.0515(2013)
- 27 O. I. Motrunich and T. Senthil, Phys. Rev. B 71 , 125102 (2005)
- 28 M. Hermele, M. P. A. Fisher, and L. Balents, Phys. Rev. B 69 , 064404 (2004)
- 29 G. Rosenberg and M. Franz, Phys. Rev. B 82 , 035105 (Jul 2010)
- 30 O. I. Motrunich and A. Vishwanath, arXiv:0805:1494(2008)
- 31 S. D. Geraedts and O. I. Motrunich, Phys. Rev. B 85 , 045114 (Jan 2012)
- 32 S. D. Geraedts and O. I. Motrunich, Phys. Rev. B 86 , 245121 (Dec
- 2012)
- 33 S. Gukov and A. Kapustin, arXiv/hep-th, 1307.4793(2013)
- 34 S. Geraedts and O. Motrunich, to appear
- 35 X. Chen, F. Burnell, A. Vishwanath, and L. Fidkowski, arXiv, 1403.6491(2014)
- 36 G. Cho, J. Teo, and S. Ryu, arXiv, 1403.2018(2014)
- 37 A. Kapustin and R. Thorngren, arXiV/hep-th, 1309.4721(2013)
- 38 S. M. Kravec and J. McGreevy, Phys. Rev. Lett. 111 , 161603 (2013)
- 39 C. W. von Keyserlingk and F. J. Burnell, ArXiv e-prints(2014), arXiv:1405.2988 [cond-mat.str-el]
- 40 A. M. Polyakov, Gauge Fields and Strings (Hardwood Academic Publishers, 1987)
|
10.1103/PhysRevX.4.041049
|
[
"Scott Geraedts",
"Olexei Motrunich"
] |
2014-08-05T20:00:51+00:00
|
2014-08-05T20:00:51+00:00
|
[
"cond-mat.stat-mech",
"cond-mat.str-el",
"hep-lat"
] |
Model Realization and Numerical Studies of a Three-Dimensional Bosonic Topological Insulator and Symmetry-Enriched Topological Phases
|
We study a topological phase of interacting bosons in (3+1) dimensions which
is protected by charge conservation and time-reversal symmetry. We present an
explicit lattice model which realizes this phase and which can be studied in
sign-free Monte Carlo simulations. The idea behind our model is to bind bosons
to topological defects called hedgehogs. We determine the phase diagram of the
model and identify a phase where such bound states are proliferated. In this
phase we observe a Witten effect in the bulk whereby an external monopole binds
half of the elementary boson charge, which confirms that it is a bosonic
topological insulator. We also study the boundary between the topological
insulator and a trivial insulator. We find a surface phase diagram which
includes exotic superfluids, a topologically ordered phase, and a phase with a
Hall effect quantized to one-half of the value possible in a purely
two-dimensional system. We also present models that realize symmetry-enriched
topologically-ordered phases by binding multiple hedgehogs to each boson; these
phases show charge fractionalization and intrinsic topological order as well as
a fractional Witten effect.
|
1408.1097v2
|
## IMFIT: A FAST, FLEXIBLE NEW PROGRAM FOR ASTRONOMICAL IMAGE FITTING
PETER ERWIN 1,2
Draft version December 4, 2014
## ABSTRACT
I describe a new, open-source astronomical image-fitting program called IMFIT, specialized for galaxies but potentially useful for other sources, which is fast, flexible, and highly extensible. A key characteristic of the program is an object-oriented design which allows new types of image components (2D surface-brightness functions) to be easily written and added to the program. Image functions provided with IMFIT include the usual suspects for galaxy decompositions (Sérsic, exponential, Gaussian), along with Core-Sérsic and brokenexponential profiles, elliptical rings, and three components which perform line-of-sight integration through 3D luminosity-density models of disks and rings seen at arbitrary inclinations.
Available minimization algorithms include Levenberg-Marquardt, Nelder-Mead simplex, and Differential Evolution, allowing trade-offs between speed and decreased sensitivity to local minima in the fit landscape. Minimization can be done using the standard χ 2 statistic (using either data or model values to estimate perpixel Gaussian errors, or else user-supplied error images) or Poisson-based maximum-likelihood statistics; the latter approach is particularly appropriate for cases of Poisson data in the low-count regime. I show that fitting low-S/N galaxy images using χ 2 minimization and individual-pixel Gaussian uncertainties can lead to significant biases in fitted parameter values, which are avoided if a Poisson-based statistic is used; this is true even when Gaussian read noise is present.
Subject headings: methods: data analysis - techniques: image processing - techniques: photometric galaxies: structure - galaxies: bulges - galaxies: photometry
## 1. INTRODUCTION
Galaxies are morphologically complex entities. Even seemingly simple systems like elliptical galaxies can have outer envelopes and distinct cores or nuclei, while so-called 'bulgeless' spiral galaxies can still have nuclear star clusters and disks with complex radial or vertical profiles. In order to accurately describe the structure of galaxies, it is often necessary to decompose galaxies into component substructures. Even single-component systems are often modeled with analytic functions in order to derive quantitative measurements such as scale lengths or half-light radii, Sérsic indices, etc.
1981; Saglia et al. 1993; Trujillo et al. 2001; Rusli et al. 2013). Furthermore, there are often intrinsic degeneracies involved: images of galaxies with non-axisymmetric components such as bars can yield 1D profiles resembling those from galaxies with axisymmetric bulges, which makes for considerable ambiguity in interpretation. Finally, if one is interested in the properties of non-axisymmetric components (bars, elliptical rings, spiral arms) themselves, it is generally impossible to extract these from 1D profiles.
The traditional method for dealing with this complexity has been to model 1D surface-brightness profiles of galaxies derived from 2D images - as the sum of separate, additive components (e.g., bulge + disk); pioneering examples of this include work by Kormendy (1977), Burstein (1979), Tsikoudi (1979, 1980), Boroson (1981), Send (1982), and Hickson et al. (1982). While this 1D approach can be conceptually and computationally simple, it has a number of limitations, above and beyond the fact that it involves discarding most of the data contained in an image. To begin with, there are uncertainties about what type of 1D profile to use should one use major-axis cuts or profiles from ellipse fits to isophotes, should the independent variable be semi-major axis or mean radius, etc. It is also difficult to correctly account for the effects of image resolution when fitting 1D profiles; attempts to do so generally require simple analytic models of the point-spread function (PSF), extensive numerical integrations, and the assumption of circular symmetry for the PSF, the surface-brightness function, or both (e.g., Pritchet & Kline
1 Max-Planck-Insitut für extraterrestrische Physik, Giessenbachstrasse, 85748 Garching, Germany
2 Universitäts-Sternwarte München, Scheinerstrasse 1, 81679 München, Germany
Abetter approach in many cases is to directly fit the images with 2D surface-brightness models. Early approaches along this line include those of Capaccioli et al. (1987), Shaw & Gilmore (1989), and Scorza & Bender (1990). The first general, self-consistent 2D bulge+disk modeling of galaxy images - that is, constructing a full 2D model image, comparing its intensity values with the observed image pixel-by-pixel, and iteratively updating the parameters until the χ 2 is minimized - was that of Byun & Freeman (1995), with de Jong (1996) being the first to include extra, non-axisymmetric components (bars) in fitting galaxy images. An interesting alternate approach developed at roughly the same time was the Multi-Gaussian Expansion method (Monnet et al. 1992; Emsellem et al. 1994; Cappellari 2002), which involves modeling both PSF and image as the sum of an arbitrary number of elliptical Gaussians; the drawback is the difficulty that lies in trying to associate sets of Gaussians with particular structural components and parameters.
The most commonly used galaxy-fitting codes at the present time are probably GIM2D (Simard 1998; Simard et al. 2002), 3 GALFIT (Peng et al. 2002, 2010), 4 BUDDA (de Souza et al. 2004; Gadotti 2008), 5 and MGE (Emsellem et al. 1994;
3 https://www.astrosci.ca/users/GIM2D/
4 http://users.obs.carnegiescience.edu/ peng/work/galfit/galfit.html
5 http://www.sc.eso.org/˜dgadotti/budda.html
Cappellari 2002). 6 GIM2D is specialized for bulge-disk decompositions and is implemented as an IRAF package, using the Metropolis algorithm to minimize the total χ 2 for models containing an exponential disk and a Sérsic bulge. BUDDA is written in FORTRAN and is also specialized for bulge-disk decompositions, though it includes a wider variety of possible components: exponential disk (with optional doubleexponential profile), Sérsic bulge, Sérsic bar, analytic edgeon disk, and nuclear point source. It uses a version of the Nelder-Mead simplex method (Nelder & Mead 1965), also known as the 'downhill simplex', for χ 2 minimization. GALFIT, which is written in C, is the most general of these codes, since it allows for arbitrary combinations of components (including components with different centers, which allows the simultaneous fitting of overlapping galaxies) and includes the largest set of possible components; the latest version (Peng et al. 2010) includes options for spiral and other parametric modulation of the basic components. GALFIT uses a version of the fast Levenberg-Marquardt gradient-search method (Levenberg 1944; Marquardt 1963) for its χ 2 minimization. MGE, available in IDL and Python versions, is rather different from the other codes in that it uses what is effectively a non-parametric approach, fitting images using the sum of an arbitrary number of elliptical Gaussians (it is similar to GALFIT in using the Levenberg-Marquardt method for χ 2 minimization during the fitting process.)
For most astronomical image-fitting programs the source code is not generally available, or else is encumbered by nonopen-source licenses. Even when the code is available, it is not easy to extend the built-in sets of predefined image components. The simplest codes provide only elliptical components with exponential and Sérsic surface brightness profiles; more sophisticated codes such as BUDDA and (especially) GALFIT provide a larger set of components, including some sophisticated ways of perturbing the components in the case of GALFIT. But if one wants to add completely new functions, this is not easy. (The case of MGE is somewhat different, since it does not allow parametric functions at all.)
As an example of why one might want to do this, consider the case of edge-on (or nearly edge-on) disk galaxies. Both BUDDA and GALFIT include versions of the analytical solution for a perfectly edge-on, axisymmetric, radial-exponential disk of van der Kruit & Searle (1981), with a sech 2 function for the vertical light distribution. But real galaxy disks are not always perfectly edge-on, do not all have single-exponential radial structures, and their vertical structure may in some cases be better described by a sech or exponential profile, or something in between (e.g., van der Kruit 1988; de Grijs et al. 1997; Pohlen et al. 2004; Yoachim & Dalcanton 2006). Various authors studying edge-on disks have suggested that models using radial profiles other than a pure exponential would be best fit via line-of-sight integration through 3D luminositydensity models (e.g., van der Kruit & Searle 1981; Pohlen et al. 2000, 2004). More sophisticated approaches could even involve line-of-sight integrations that account for scattering and absorption by dust (e.g., Xilouris et al. 1997, 1998, 1999).
Another potential disadvantage of existing codes is that they rely on the Gaussian approximation of Poisson statistics for the fitting process. While this is eminently sensible for dealing with many CCD and near-IR images, it can in some cases produce biases when applied to images with low count rates (see Humphrey et al. 2009 and Section 9 of this paper). This is why packages for fitting X-ray data, such as SHERPA (Freeman et al. 2001), often include alternate statistics for fits.
In this paper, I present IMFIT, a new, open-source imagefitting code designed to overcome some of the limitations mentioned above. In particular, IMFIT uses an object-oriented design which makes it relatively easy to add new, userdesigned image components; it also provides multiple fitting algorithms and statistical approaches. It can also be extremely fast, since it is able to take advantage of multiple CPU cores on the same machine to execute calculations in parallel.
The outline of this paper is as follows. Section 2 provides a quick sketch of how the program works, while Section 3 details the process of generating model images and the configuration files which describe the models. The different underlying statistical models and minimization algorithms used in the fitting process are covered in Section 4; methods for estimating confidence intervals for fitted parameters are discussed in Section 5. The default 2D image functions which can be used in models are presented in Section 6; this includes functions which perform line-of-sight integration through 3D luminosity-density models (Section 6.2). After a brief discussion of coding details (Section 7), two examples of using IMFIT to model galaxy images are presented in Section 8: the first involves fitting a moderately-inclined spiral galaxy with disk, bar, and ring components, while the second fits an edgeon spiral galaxy with thin and thick edge-on disk components. Finally, Section 9 discusses possible biases to fitted parameters when the standard χ 2 statistic is used in the presence of low-count images, using both model images and real images of elliptical galaxies. An Appendix discusses the relative sizes and accuracies of parameter error estimates using the two methods available in IMFIT.
To avoid any confusion, I note that the program described in this paper is unrelated to tasks with the same name and somewhat similar (if limited) functionality in pre-existing astronomical software, such as the 'imfit' tasks in the radioastronomy packages AIPS and MIRIAD and the 'images.imfit' package in IRAF.
## 2. GENERAL OUTLINE OF THE PROGRAM
IMFIT begins by processing command-line options and then reads in the data image, along with any optional, userspecified PSF, noise, and mask images (all in FITS format). The configuration file is also read; this specifies the model which will be fit to the data image, including initial parameter values and parameter limits, if any (see Section 3.1).
The program then creates an instance of the ModelObject class, which holds the relevant data structures, instances of the image functions specified by the configuration file, and the general code necessary for computing a model image. If χ 2 minimization (the default) is being done, a noise image is constructed, either from a user-specified FITS file already read in or by internally generating one, assuming the Gaussian approximation for Poisson noise. The noise image is then converted to 1 /σ 2 form and combined with the mask image, if any, to form a final weight image used for calculating the χ 2 value. (If model-based χ 2 minimization has been specified, then the noise image, which is based on the model image, is recalculated and combined with the mask image every time a new model image is computed; if a Poisson maximumlikelihood statistic ( C or PMLR; see Section 4.1.3) is being used for minimization, then no noise image is read or created
and the weight image is constructed directly from the mask image. See Section 4.1 for more on the different statistical approaches.)
The actual fitting process is overseen by one of three possible nonlinear minimization algorithms, as specified by the user. These algorithms proceed by generating or modifying a set of parameter values and feeding these values to the aforementioned model object, which in turn calculates the corresponding model image, convolves it with the PSF (if PSF convolution is part of the model), and then calculates the fit statistic (e.g., χ 2 ) by comparing the model image with the stored data image. The resulting fit statistic is returned to the minimization algorithm, which then updates the parameter values and repeats the process according to the details of the particular method, until the necessary stop criterion is reached e.g., no further significant reduction in the fit statistic, or a maximum number of iterations. Finally, a summary of the fit results is printed to the screen and saved to a file, along with any additional user-requested outputs (final model image, final residual image, etc.).
## 3. CONSTRUCTING THE MODEL IMAGE
## 3.1. Configuration File
The model which will be fit to the data image is specified by a configuration file, which is a text file with a relatively simple and easy-to-read format; see Figure 1 for an example. The basic format for this file is a set of one or more 'function blocks', each of which contains a shared center (pixel coordinates) and one or more image functions. A function block can, for example, represent a single galaxy or other astronomical object, which itself has several individual components (e.g., bulge, disk, bar, ring, nucleus, etc.) specified by the individual image functions. Thus, for a basic bulge/disk decomposition the user could create a function block consisting of a single Sérsic function and a single Exponential function. There is, however, no a priori association of any particular image function or functions with any particular galaxy component, nor is there any requirement that a single object must consist of only one function block. The final model is the sum of the contributions from all the individual functions in the configuration file. The number of image functions per function block is unlimited, and the number of function blocks per model is also unlimited.
Each image function is listed by name (e.g., ' FUNCTION Sersic '), followed by the list of its parameters. For each parameter, the user supplies an initial guess for the value, and (optionally) either a comma-separated, two-element list of lower and upper bounds for that parameter or the keyword 'fixed' (indicating that the parameter will remain constant during the fit). 7
The total set of all individual image-function parameters, along with the central coordinates for each function block, constitutes the parameter vector for the minimization process.
## 3.2. Image Functions
An image function can be thought of as a black box which accepts a set of parameter values for its general setup, and then accepts individual pixel coordinates ( x y , ) and returns a corresponding computed intensity (i.e., surface brightness)
7 Parameter bounds can be used with any of the minimization algorithms; with the Differential Evolution algorithm, they are actually required , though in that case the initial values are ignored; see Section 4.3.
value for that pixel. The total intensity for a given pixel in the model image (prior to any PSF convolution) is the sum of the individual values from each image function.
This design means that the main program needs to know nothing about the individual image functions except the number of parameters they take, and which subset of the total parameter vector corresponds to a given image function. The actual calculations carried out by an image function can be as simple or as complex as the user requires, ranging from returning a constant value for each pixel (e.g., the FlatSky function) to performing line-of-sight integration through a 3D luminosity density model (e.g., the ExponentialDisk3D function); user-written image functions could even perform modest simulations in the setup stage. 8
The list of currently available image functions, along with descriptions for each, is given in Section 6.
## 3.3. PSF Convolution
To simulate the effects of atmospheric seeing and telescope optics, model images can be convolved with a PSF image. The latter can be any FITS file which contains the point spread function. PSF images should ideally be square with sides measuring an odd number of pixels, with the peak of the PSF centered in the central pixel of the image. (Off-center PSFs can be used, but the resulting convolved model images will of course be shifted.) IMFIT automatically normalizes the PSF image when it is read in.
The actual convolution follows the standard approach of using Fast Fourier Transforms of the internally-generated model image and the PSF image, multiplied together, with the output convolved model image being the inverse transform of the product image. The transforms are done with the FFTW library ('Fastest Fourier Transform in the West', Frigo & Johnson 2005), 9 which has the advantage of being able to perform transforms on images of arbitrary size (i.e., not just images with power-of-two sizes); in addition, it is well-tested and fast, and can use multiple threads to take advantage of multiple processor cores.
To avoid possible edge effects in the convolution, the internal model-image array is expanded on all four sides by the width and height of the PSF image, and all calculations prior to the convolution phase use this full (expanded) image. (For example, given a 1000 × 1000-pixel data image and a 15 × 15-pixel PSF image, the internal model image would be 1030 × 1030 pixels in size.) This ensures that model pixels corresponding to the edge of the data image are the result of convolution with an extension of the model, rather than with zero-valued pixels or the opposite side of the model image. This is in addition to the zero-padding applied to the top and right-hand sides of the model image during the convolution phase. (I.e., the example 1030 × 1030-pixel expanded model image would be zero-padded to 1045 × 1045 pixels before computing its Fourier transform, to match with the zeropadded PSF image of the same size.)
## 3.4. Makeimage: Generating Model Images Without Fitting
A companion program called MAKEIMAGE is included in the IMFIT package, built from the same codebase as IMFIT itself. This program implements the complete model-image
8 One should bear in mind that even relatively simple fits will typically require dozens to hundreds of function evaluations during the minimization process, so complex simulations will mean a lengthy fitting process.
9 http://www.fftw.org
- # Comment lines: all lines beginning with "#" are ignored.
- # Anything following a "#" on a line is ignored, too.
- # This is the first function block, which has 2 image functions
- # (1 Sersic + 1 exponential)
- # It begins with the starting guess for pixel coordinates of center (x,y)
- # Each parameter has initial value, then (optional) lower,upper bounds
# Optional keywords describing the image
GAIN 2.7 # A/D
gain for image in e/ADU
READNOISE 4.5
# image read-noise in electrons
X0 150.1
148,152
Y0 149.5
148,152
FUNCTION Sersic
# A 2D elliptical Sersic function
PA
95.0 0,180
# major-axis position angle [deg]
ell 0.05
0,1
# ellipticity
n 2.5
0.5,4.0
# Sersic index
I e 20.0
# intensity [counts/pixel] at r e
r e 5.0
# half-light radius in pixels
FUNCTION Exponential
# A 2D elliptical exponential function
PA 95.0 0,180
ell 0.45
0,1
I 0 90.0 fixed
# this parameter will be held fixed
h 15.0
# (no bounds for this parameter)
## # This is the second function block: just a single exponential
X0 225.0 224,226
Y0 181.7
180,183
FUNCTION Exponential
PA 22.0
0,180
ell 0.25
0,1
I 0 10.0
h 20.0
Figure 1. Example of a configuration file for IMFIT. Comments are colored red.
construction process, including PSF convolution, and then simply saves the resulting model image as a FITS file. It can optionally save separate images, one for each of the individual image functions that make up the model. Since it uses the same configuration-file format as IMFIT, it can use the output best-fit parameter file that IMFIT produces (or even an input IMFIT configuration file).
It also has an optional mode which estimates the fractional flux contributions of the individual components in the model, by summing up the total flux of the individual components on a pixel-by-pixel basis using a very large internal image (by default, 5000 × 5000 pixels). Although analytic expressions for total flux exist for some common components, this is not true for all components - and one of the goals of IMFIT is to allow users to create and use new image functions without worrying about whether they have simple analytic expressions for the total flux. This mode can be used to help determine such things as bulge/total and other ratios after a fit is found, although it is up to the user to decide which of the components is the 'bulge', which is the 'disk', and so forth.
## 4. THE FITTING PROCESS
## 4.1. The Statistical Background and Options
Given a vector of parameter values glyph[vector] θ , a model image is generated with per-pixel predicted data values mi , which are then compared with the observed per-pixel data values di . The goal is to find the glyph[vector] θ which produces the best match between mi and di , subject to the constraints of the underlying statistical model.
The usual approach is based on the maximum-likelihood principle (which can be derived from a Bayesian perspective if, e.g., one assumes constant priors for the parameter values), and is conventionally known as maximum-likelihood estimation (MLE; e.g., Pawitan 2001). To start, one considers the per-pixel likelihood pi ( di | mi ), which is the probability of observing di given the model prediction mi and the underlying statistical model for how the data are generated.
The goal then becomes finding the set of model parameters which maximizes the total likelihood L , which is simply the product over all N pixels of the individual per-pixel likeli-
hoods:
$$\mathcal { L } = \prod _ { i = 1 } ^ { N } p _ { i }. \quad \quad \quad ( 1 )$$
It is often easier to work with the logarithm of the total likelihood, since this converts a product over pixels into a sum over pixels, and can also simplify the individual per-pixel terms. As most nonlinear optimization algorithms are designed to minimize their objective function, one can use the negative of the log-likelihood. Thus, the goal of the fitting process becomes minimization of the following:
$$- \ln \mathcal { L } = - \sum _ { i = 1 } ^ { N } \ln p _ { i }. \quad \quad ( 2 ) \quad \begin{pmatrix} \tilde { \ t h } \\ \tilde { \ t i } \cdot \\ \tilde { P } \cdot \end{pmatrix}$$
During the actual minimization process, this can often be further simplified by dropping any additive terms in ln pi which do not depend on the model, since these are unaffected by changes in the model parameters and are thus irrelevant to the minimization.
In some circumstances, multiplying the negative loglikelihood by 2 produces a value which has the property of being distributed like the χ 2 distribution (e.g., Cash 1979, and references therein); thus, it is conventional to treat -2ln L as the statistic to be minimized.
## 4.1.1. The (Impractical) General Case: Poisson + Gaussian Statistics
The data in astronomical images typically consist of detections of individual photons from the sky + telescope system (including photons from the source, the sky background, and possibly thermal backgrounds in the telescope) in individual pixels, combined with possible sources of noise due to readout electronics, digitization, etc.
Photon-counting statistics obey the Poisson distribution, where the probability of detecting x photons per integration, given a true rate of m , is
$$P ( x ) \, = \, \frac { m ^ { x } e ^ { - m } } { x! }. \quad \quad \ \ ( 3 ) \, \quad \ \ c _ { 0 } \\ \quad \ \ a$$
Additional sources of (additive) noise such as read noise tend to follow Gaussian statistics with a mean of 0 and a dispersion of σ , so that the probability of measuring d counts after the readout process, given an input of x counts from the Poisson process, is
$$P ( d ) \, = \, \frac { 1 } { \sqrt { 2 \pi } \sigma } \exp \left [ \frac { - ( d - x ) ^ { 2 } } { 2 \sigma ^ { 2 } } \right ] \,. \quad \quad ( 4 ) \quad \underset { \gamma ^ { 2 \gamma } } { \text{ba} }.$$
The general case for most astronomical images thus involves both Poisson statistics (for photon counts) and Gaussian statistics (for read noise and other sources of additive noise). Unfortunately, even though the individual elements are quite simple, the combination of a Gaussian process acting on the output of a Poisson process leads to the following rather frightening per-pixel likelihood (e.g., Llacer & Nuñéz 1991; Nuñéz & Llacer 1993):
$$p _ { i } ( d _ { i } | m _ { i } ) \, = \, \sum _ { x _ { i } = 0 } ^ { \infty } \frac { m _ { i } ^ { x _ { i } } e ^ { - m _ { i } } } { x _ { i }! } \frac { 1 } { \sqrt { 2 \pi } \sigma } \exp \left ( \frac { - ( d _ { i } - x _ { i } ) ^ { 2 } } { 2 \sigma ^ { 2 } } \right ). \quad ( 5 ) \quad \stackrel { \infty } { \text{is do} } \\ \text{calcul}$$
The resulting negative log-likelihood for the total image
(dropping terms which do not depend on the model) is
$$( 1 ) \quad - \ln \mathcal { L } \, = \, \sum _ { i = 1 } ^ { N } \left ( m _ { i } - \ln \left [ \sum _ { x _ { i } = 0 } ^ { \infty } \frac { m _ { i } ^ { x _ { i } } } { x _ { i }! } \exp \left ( \frac { - ( d _ { i } - x _ { i } ) ^ { 2 } } { 2 \sigma ^ { 2 } } \right ) \right ] \right ) \,. \ \ ( 6 )$$
Since this still contains an infinite series of exponential and factorial terms, it is clearly rather impractical for fitting images rapidly.
## 4.1.2. The Simple Default: Pure Gaussian Statistics
Fortunately, there is a way out which is often (though not always) appropriate astronomical images. This is to use the fact that the Poisson distribution approaches a Gaussian distribution when the counts become large. In this approximation the Poisson distribution is replaced by a Gaussian with σ = √ m . It is customary to assume this is valid when the counts are glyph[greaterorsimilar] 20 per pixel (e.g., Cash 1979), though Humphrey et al. (2009) point out that biases in the fitted parameters can be present even when counts are higher than this; see Section 9 for examples in the case of 2D fits.
Since the contribution from read noise is also nominally Gaussian, the two can be added in quadrature, so that the perpixel likelihood function is just
$$p _ { i } ( d _ { i } | m _ { i } ) \, = \, \frac { 1 } { \sqrt { 2 \pi } \sigma _ { i } } \exp \left [ \frac { - ( d _ { i } - m _ { i } ) ^ { 2 } } { 2 \sigma _ { i } ^ { 2 } } \right ], \quad \ \ ( 7 )$$
where σ 2 i = σ 2 mi + σ 2 RN = mi + σ 2 RN , with σ RN being the dispersion of the read-noise term. Twice the negative log-likelihood of the total problem then becomes (dropping terms which do not depend on the model) the familiar χ 2 sum:
$$- 2 \ln \mathcal { L } \, = \, \chi ^ { 2 } \, = \, \sum _ { i = 1 } ^ { N } \frac { ( d _ { i } - m _ { i } ) ^ { 2 } } { \sigma _ { i } ^ { 2 } }. \, \quad \quad ( 8 )$$
This is the default approach used by IMFIT: minimizing the χ 2 as defined in Eqn. 8.
The approximation of the Poisson contribution to σ i is based on the model intensity mi . Traditionally, it is quite common to estimate this from the data instead, so that σ 2 i = σ 2 di + σ 2 RN = di + σ 2 RN . This has the nominal advantage of only needing to be calculated once, at the start of the minimization process, rather than having to be recalculated every time the model is updated. 10 However, the bias resulting from using data-based errors in the low-count regime can be worse than the bias introduced by using model-based σ i values (see Section 9). Both approaches are available in IMFIT, with databased σ i estimation being the default. The data-based and model-based approaches are often referred to as 'Neyman's χ 2 ' and 'Pearson's χ 2 ', respectively; in this paper I use the symbols χ 2 d and χ 2 m to distinguish between them.
In the case of 'error' images generated by a data-processing pipeline, the corresponding σ i or σ 2 i (variance) values can easily be used in Equation 8 directly, under the assumption that the final per-pixel error distributions are still Gaussian.
## 4.1.3. The Simple Alternative: Pure Poisson Statistics
So why not always use the Gaussian χ 2 approximation, as is done in most image-fitting packages?
10 In practice, the time spent by IMFIT is dominated by the per-pixel model calculations, so any extra time spent re-estimating the per-pixel σ i values is often negligible.
In the absence of any noise terms except Poisson statistics something often true of high-energy detectors, such as X-ray imagers - the individual-pixel likelihoods are just the probabilities of a Poisson process with mean mi , where the probability of recording di counts is
$$p _ { i } ( d _ { i } | m _ { i } ) \, = \, \frac { m _ { i } ^ { d _ { i } } e ^ { - m _ { i } } } { d _ { i }! } \, & \quad ( 9 ) \, & \quad ( 1 ) \\.$$
This leads to a very simple version of the negative loglikelihood, often referred to as the 'Cash statistic' C , after its derivation in Cash (1979):
$$- 2 \ln \mathcal { L } = C = 2 \sum _ { i = 1 } ^ { N } ( m _ { i } - d _ { i } \ln m _ { i } ), \quad \ ( 1 0 ) \quad \frac { - } { \ m a }$$
where the factorial term has been dropped because it does not depend on the model.
A useful alternative is to construct a statistic from the likelihood ratio test - that is, a maximum likelihood ratio (MLR) statistic - which is the ratio of the likelihood to the maximum possible likelihood for a given dataset. In the case of Poisson likelihood, the latter is the likelihood when the model values are exactly equal to the data values mi = di (e.g., Hauschild & Jentschel 2001), and so the likelihood ratio λ is
$$\lambda \, = \, \mathcal { L } / \mathcal { L } _ { \max } \, = \, \prod _ { i = 1 } ^ { N } \frac { m _ { i } ^ { d _ { i } } e ^ { - m _ { i } } } { d _ { i } ^ { d _ { i } } e ^ { - d _ { i } } } \, \quad \quad ( 1 1 ) \, \quad \, \ m _ { \substack { \mathbf s t \epsilon \\ \mathbf s t \epsilon } }$$
and the negative log-likelihood version (henceforth PMLR) is
$$P M L R \, = \, - 2 \ln \lambda \, = \, 2 \sum _ { i = 1 } ^ { N } ( m _ { i } - d _ { i } \ln m _ { i } + d _ { i } \ln d _ { i } - d _ { i } ) \,. \quad ( 1 2 ) \quad \text{and} \, \underline { \text{for.} } \\ \text{by} \quad \text{age} \, \mathfrak { k }$$
(This is the same as the 'CSTAT' statistic available in the SHERPA X-ray analysis package and the 'Poisson likelihood ratio' described by Dolphin 2002.) Comparison with Equation 10 shows that PMLR is identical to C apart from terms which depend on the data only and thus do not affect the minimization. In the remainder of this paper, I will refer to C and PMLR collectively as Poisson MLE statistics.
Since minimizing PMLR will produce the same best-fitting parameters as minimizing C , one might very well wonder what is the point in introducing PMLR. There are two practical advantages in using it. The first is that in the limit of large N , -2ln λ statistics such as PMLR approach a χ 2 distribution and can thus be used as goodness-of-fit indicators (Wilks 1938; Wald 1943). The second is that they are always ≥ 0 (since λ itself is by construction always ≤ 1); this means they can be used with fast least-squares minimization algorithms. This is the practical drawback to minimizing C : unlike PMLR, it can often have negative values, and thus requires one of the slower minimization algorithms.
Humphrey et al. (2009) point out that using a Poisson MLE statistic (e.g., C ) is preferable to using χ 2 d or χ 2 m even when the counts are above the nominal limit of ∼ 20 per pixel, since fitting pure-Poisson data using the χ 2 d or χ 2 m Gaussian approximations can lead to biases in the derived model parameters. Section 9 presents some examples of this effect using both artificial and real galaxy images, and shows that the effect persists even when moderate (Gaussian) read noise is also present.
Using a Poisson MLE statistic such as C or PMLR is also appropriate when fitting simulated images, such as those
Table 1 Terminology for Fits and Minimization
| Term | Explanation |
|-------------|--------------------------------------------------------------------------------------|
| Poisson MLE | Maximum-likelihood estimation based on Poisson statistics (includes both C and PMLR) |
| C | Poisson MLE statistic from Cash (1979) |
| PMLR | Poisson MLE statistic from maximum likelihood ratio |
| χ 2 d | Gaussian MLE statistic using data pixel values for σ ('Neyman's χ 2 ') |
| χ 2 m | Gaussian MLE statistic using model pixel values for σ ('Pearson's χ 2 ') |
made from projections of N -body models, as long as the units are particles per pixel or something similar.
For convenience, Table 1 summarizes the main symbols and terms from this section which are used elsewhere in the paper.
## 4.2. Implementation: Specifying Per-Pixel Errors and Masking
IMFIT's default behavior, as mentioned above, is to use χ 2 as the statistic for minimization. To do so, the individual, per-pixel Gaussian errors σ i must be available. If a separate error or noise map is not supplied by the user (see below), IMFIT estimates σ i values from either the data values or the model values, using the Gaussian approximation to Poisson statistics. To ensure this estimate is as accurate as possible, the data or model values I i must at some point be converted from counts to actual detected photons (e.g., photoelectrons), and any previously subtracted background must be accounted for.
By default, IMFIT estimates the σ i values from the data image by including the effects of A/D gain, prior subtraction of a (constant) background, and read noise. Rather than converting the image to electrons pixel -1 and then estimating the σ values, IMFIT generates σ values in the same units as the input image:
$$\sigma _ { I, i } ^ { 2 } \, = \, ( I _ { d, i } + I _ { s k y } ) / g _ { e f f } + N _ { c } \, \sigma _ { \mathrm R N } ^ { 2 } / g _ { e f f } ^ { 2 }, \quad \ \ ( 1 3 )$$
where I d i , is the data intensity in counts pixel -1 , I sky is any pre-subtracted sky background in the same units, σ RN is the read noise in electrons, N c is the number of separate images combined (averaged or median) to form the data image, and g eff is the 'effective gain' (the product of the A D / gain, N c , and optionally the exposure time if the image pixel values are actually in units of counts s -1 pixel -1 rather than integrated counts pixel -1 ). If model-based χ 2 minimization is used, then model intensity values I m i , are used in place of I d i , in Equation 4.2. In this case, the σ , I i values must be recomputed each time a new model image is generated, though in practice this adds very little time to the overall fitting process.
If a mask image has been supplied, it is converted internally so that its pixels have values zi = 1 for valid pixels and zi = 0 for bad pixels. Then the mask values are divided by the variances to form a weight-map image, where individual pixels have values of wi = zi /σ , 2 I i . These weights are then used for the actual χ 2 calculation:
$$\chi ^ { 2 } \, = \, \sum _ { i = 1 } ^ { N } w _ { i } ( I _ { d, i } - I _ { m, i } ) ^ { 2 }. \, \text{ \quad \ \ } ( 1 4 )$$
Instead of data-based or model-based errors, the user can
also supply an error or noise map in the form of a FITS image, such as might be produced by a reduction pipeline. The individual pixel values in this image can be Gaussian errors, variances ( σ 2 ), or even pre-computed weight values wi .
In the case of Cash-statistic minimization, the sum C is computed directly based on Equation 10; for PMLR minimization, Equation 12 is used. The 'weight map' in either case is then based directly on the mask image, if any (so all pixels in the resulting weight map have values of zi = 0 or 1). The actual minimized quantities are thus
$$C = 2 \sum _ { i = 1 } ^ { N } z _ { i } ( m _ { i } - d _ { i } \ln m _ { i } ) \quad \quad ( 1 5 ) \quad \stackrel { \text{al} _ { i } } { \& } \\ \text{sc}.$$
and
$$\Gamma _ { \text{MLR} } \, = \, 2 \sum _ { i = 1 } ^ { N } z _ { i } ( m _ { i } - d _ { i } \ln m _ { i } + d _ { i } \ln d _ { i } - d _ { i } ) \quad ( 1 6 ) \quad \underset { \text{rec} } { \text{rec} } \, \text{do} \, \underset { \text{not} } { \text{not} }$$
with mi = g eff ( I m i , + I sky ) and di = g eff ( I d i , + I sky ).
## 4.3. Minimization Algorithms
## 4.3.1. Levenberg-Marquardt
The default minimization algorithm used by IMFIT is a robust implementation of the Levenberg-Marquardt (L-M) gradient search method (Marquardt 1963), based on the MINPACK-1 version of Moré (1978) and modified by Craig Markwardt (Markwardt 2009), 11 which includes optional lower and upper bounds on parameter values. This version of the basic L-M algorithm also includes auxiliary code for doing numerical differentiation of the objective function, and thus the various image functions do not need to provide their own derivatives, which considerably simplifies things when it comes to writing new functions.
The L-M algorithm has the key advantage of being very fast, which is a useful quality when one is fitting large images with a complex set of functions and PSF convolution. It has the minor disadvantage of requiring an initial starting guess for the parameter values, and it has two more significant disadvantages. The first is that like gradient-search methods in general it is prone to becoming trapped in local minima in the objective-function landscape. The second is that it is designed to work with least-squares objective functions, where the objective function values are assumed to be always ≥ 0. In fact, the L-M algorithm makes use of a vector of the individual contributions from each pixel to the total χ 2 , and these values as well (not just the sum) must be nonnegative. For the χ 2 case, this is always true; but this is not guaranteed to be true for the Cash statistic C . Thus, it would be quite possible for the L-M minimizer to fail to find the best-fitting solution for a particular image, simply because the solution has a C value < 0. (Fortunately, minimizing PMLR leads to the same solution as minimizing C , and the individual terms of PMLR are always nonnegative.)
## 4.3.2. Nelder-Mead Simplex
A second, more general algorithm available in IMFIT is the Nelder-Mead simplex method (Nelder & Mead 1965), with constraints as suggested by Box (1965), implemented in the NLopt library. 12 Like the L-M algorithm, this method re-
11 http:// purl.com/net/mpfit
12 Steven G. Johnson, The NLopt nonlinear-optimization package, http://ab-initio.mit.edu/nlopt.
quires an initial guess for the parameter set; it also includes optional parameter limits. Unlike the L-M algorithm, it works only with the final objective function value and does not assume that this value must be nonnegative; thus, it is suitable for minimizing all the fit statistics used by IMFIT. It is also as a rule less likely to be caught in local minima than the L-M algorithm. The disadvantage is that it is considerably slower than the L-M method - roughly an order of magnitude so.
## 4.3.3. Differential Evolution
A third alternative provided by IMFIT is a geneticalgorithms approach called Differential Evolution (DE; Storn & Price 1997). This searches the objective-function landscape using a population of parameter-value vectors; with each 'generation', the population is updated by mutating and recombining some of the vectors, with new vectors replacing older vectors if they are better-performing. DE is designed to be - in the context of genetic algorithms - fast and robust while keeping the number of adjustable algorithm parameters (e.g., mutation and crossover rates) to a minimum. It is the least likely of the algorithms used by IMFIT to become trapped in a local minimum in the objective-function landscape: rather than starting from a single initial guess for the parameter vector, it begins with a set of randomly generated initial-parameter values, sampled from the full range of allowed parameter values; in addition, the crossover-withmutation used to generate new parameter vectors for successive generations helps the algorithm avoid local minima traps. Thus, in contrast to the other algorithms, it does not require any initial guesses for the parameter values, but does require lower and upper limits for all parameters. It is definitely the slowest of the minimization choices: about an order of magnitude slower than the N-M simplex, and thus roughly two orders of magnitude slower than the L-M algorithm.
The current implementation of DE in IMFIT uses the 'DE/rand-to-best/1/bin' internal strategy, which controls how mutation and crossover are done (Storn & Price 1997), along with a population size of 10 parameter vectors per free parameter. Since the basic DE algorithm has no default stop conditions, IMFIT halts the minimization when the best-fitting value of the fit statistic has ceased to change by more than a specified tolerance after 30 generations, or when a maximum of 600 generations is reached.
## 4.3.4. Comparison and Recommendations
For most purposes, the default L-M method is probably the best algorithm to use, since it is fast enough to make exploratory fitting (varying the set of functions used, applying different parameter limits, etc.) feasible, and also fast enough to make fitting large numbers of individual objects in a reasonable time possible. If the problem is relatively small (modest image size, few image functions) and the user is concerned about possible local minima, then the N-M simplex or even the DE algorithm can be used.
Table 2 provides a general comparison of the different minimization algorithms, including the time taken for each to find the best fit for a very simple case: a 256 × 256-pixel cutout of an SDSS r -band image of the galaxy IC 3478, fit with a single Sérsic function and convolved with a 51 × 51-pixel PSF image. For this simple case, the N-M approach takes ∼ 4 times as long as the L-M method, and the DE algorithm takes ∼ 60 times as long. (All three algorithms converged to the same solution, so there was no disadvantage to using the L-M method in this case.)
Table 2 Comparison of Minimization Algorithms
| Algorithm (1) | Initial guess required (2) | Bounds required (3) | Local-minimum vulnerability (4) | Minimize C ? (5) | Speed (6) | Timing example (7) |
|-----------------------------|------------------------------|-----------------------|-----------------------------------|--------------------|-------------|----------------------|
| Levenberg-Marquardt (L-M) | Yes | No No | High Low | No | Fast | 2.2s 9.1s |
| Nelder-Mead Simplex (N-M) | Yes | | Medium | Yes | Slow | |
| Differential Evolution (DE) | No | Yes | | Yes | Very Slow | 2m15s |
Note . -Acomparison of the three nonlinear minimization algorithms available in IMFIT. Column 1: Algorithm name. Column 2: Notes whether an initial guess of parameter values required. Column 3: Notes whether lower and upper bounds on all parameter values are required. Column 4: Vulnerability of the algorithm to becoming trapped in local minima in the χ 2 (or other objective function) landscape. Column 5: Notes whether algorithm can minimize the Cash statistic C in addition to χ 2 and PMLR. Column 6: General speed. Column 7: Approximate time taken for fitting a 256 × 256 pixel SDSS galaxy image (single Sérsic function + PSF convolution), using a MacBook Pro with a quad-core Intel Core i7 2.3 GHz CPU (2011 model).
## 4.4. Outputs and 'Goodness of Fit' Measures
When IMFIT finishes, it outputs the parameters of the best fit (along with possible confidence intervals; see Section 5) to the screen and to a text file; it also prints the final value of the fit-statistic. The best-fitting model image and the residual (data -model) image can optionally be saved to FITS files as well.
automatically produce a set of approximate, 1σ confidence intervals for the fitted parameters as a side product of the minimization process; this comes from inverting the Hessian matrix computed during the minimization process (see, e.g., Section 15.5 of Press et al. 1992).
For fits which minimize χ 2 , IMFIT also prints the reduced χ 2 value, which can be used (with caution) as an indication of the goodness of the fit. (The best-fit value of PMLR can also be converted to a reducedχ 2 equivalent with the same properties.) For fits which minimize the Cash statistic, there is no direct equivalent to the reduced χ 2 ; the actual value of the Cash statistic does not have any directly useful meaning by itself.
All the fit statistics (including C ) can also be used to derive comparative measures of how well different models fit the same data. To this end, IMFIT computes two likelihood-based quantities which can be used to compare different models. The first is the Akaike Information Criterion (AIC, Akaike 1974), which is based on an information-theoretic approach. IMFIT uses the recommended, bias-corrected version of this statistic:
$$\intertext { a c. } \text{aic} _ { c } = - 2 \ln \mathcal { L } + 2 k + \frac { 2 k ( k + 1 ) } { n - k - 1 }, \quad \ \ ( 1 7 ) \quad \ \underset { \text{em} } { \text{aic} }$$
where L is the likelihood value, k is the number of (free) parameters in the model and n is the number of data points. The second quantity is the Bayesian Information Criterion (BIC, Schwarz 1978), which is
$$\text{BIC} \, = \, - 2 \ln \mathcal { L } + k \ln n. \quad \quad ( 1 8 ) \quad \underset { \mathfrak { n } } { \nu }$$
When two or more models fit to the same data are compared, the model with the lowest AIC (or BIC) is preferred, though a difference ∆ AIC or ∆ BIC of at least ∼ 6 is usually required before one model can be deemed clearly superior (or inferior); see, e.g., Takeuchi (2000) and Liddle (2007) for discussions of AIC and BIC in astronomical contexts, and Burnham & Anderson (2002) for more general background. Needless to say, all models being compared in this manner should be fit by minimizing the same fit statistic.
## 5. CONFIDENCE INTERVALS FOR FITTED PARAMETERS
In addition to its speed, the Levenberg-Marquardt minimization algorithm has the convenient advantage that it can
The other minimization algorithms available in IMFIT do not compute confidence intervals. Although one can, as a workaround, re-run IMFIT using the L-M algorithm on a solution that was found using one of the other algorithms, this will not work if C (rather than χ 2 or PMLR) was being minimized (see Section 4.3).
An alternate method of estimating confidence intervals is provided by bootstrap resampling (Efron 1979). Each iteration of the resampling process generates a new data image by sampling pixel values, with replacement, from the original data image. (What is actually generated inside IMFIT is a resampled vector of pixel indices into the image, excluding those indices corresponding to masked pixels; the corresponding x y , coordinates and intensities then form the resampled data.) The fit is then re-run with the best-fit parameters from the original fit as starting values, using the L-M algorithm for χ 2 and PMLR minimization cases and the N-M simplex algorithm when C minimization is being done. After n iterations, the combined set of bootstrapped parameter values is used as the distribution of parameter values, from which properly asymmetric 68% confidence intervals are directly determined, along with the standard deviation. (The 68% confidence interval corresponds to ± 1σ if the distribution is close to Gaussian.)
In addition, the full set of best-fit parameters from all the bootstrap iterations can optionally be saved to a file, which potentially allows for more sophisticated analyses. Figure 2 shows a scatter-plot matrix comparing parameter values for five parameters of a simple Sérsic fit to an image of a model Sérsic galaxy with noise (see Section 9 for details of the model images). One can see that the distributions are approximately Gaussian, have dispersions similar to those from the L-M estimates (plotted as Gaussians using red curves), and also that certain parameter distributions are correlated (e.g., n and re or, more weakly, ellipticity glyph[epsilon1] and re ). Of course, this simple case ignores the complexities and sources of error involved in fitting real images of galaxies; see the Appendix for sample comparisons of L-M and bootstrap error estimates for a small set of real-galaxy images.
The only drawback of the bootstrap-resampling approach is the cost in time. Since bootstrap resampling should ideally
Figure 2. Scatter-plot matrix showing bootstrap resampling analysis of a Sérsic fit to a simple model image with noise (500 rounds of bootstrap resampling, using PMLR for minimization). Each panel plots the best-fit Séric parameter values from individual bootstrap iterations as gray points (panels in the upper right of the plot are rotated duplicates of those in the lower left), except for the panels along the diagonal, which show histograms of individual parameter values (thick blue lines). Plotted on top of the latter are Gaussians with estimated dispersions σ from the Levenberg-Marquardt output of the original fit (thin red curves). Vertical solid gray lines show the parameter values from the fit to the original image; vertical dashed gray lines show the true parameters of the original model.
use a minimum of several hundred to one thousand or more iterations, one ends up, in effect, re-running the fit that many times. (Some time is saved by starting each fit with the original best-fit parameter values, since those will almost always be close to the best-fit solution for the resampled data.)
## 6. IMAGE FUNCTIONS
Image functions are implemented in IMFIT as subclasses of an abstract base class called FunctionObject. The rest of the program does not need to know the details of the individual functions, only that they adhere to the FunctionObject interface. This makes it relatively simple to add new image functions to the program: write a header file and an implementation file for the new function, add a reference to it in another source file, and recompile the program. Further notes on how to do this are included in the documentation.
This section describes the various default image functions that come with IMFIT. Specifications for the actual parameters (e.g., the order that IMFIT expects to find them in) are included in the documentation, and a summary of all available function names and their corresponding parameter lists can be printed using the --list-parameters command-line flag.
## 6.1. 2D Components
Most image functions, unless otherwise noted, have two 'geometric' parameters: the position angle PA in degrees counter-clockwise from the vertical axis of the image 13 and
13 Reproducing the usual convention for images with standard astronomical orientation, where north is up and east is to the left.
the ellipticity glyph[epsilon1] = 1 -b / a , where a and b are the semi-major and semi-minor axes, respectively.
In most cases, the image function internally converts the ellipticity to an axis ratio q = b / a (= 1 -glyph[epsilon1] ) and the position angle to an angle relative to the image x -axis θ (= PA + 90 ), ◦ in radians. Then for each input pixel (or subpixel if pixel subsampling is being done) with image coordinates ( x y , ) a scaled radius is computed as
$$r \, = \, \left ( x _ { p } ^ { 2 } + \frac { y _ { p } ^ { 2 } } { q ^ { 2 } } \right ) ^ { 1 / 2 },$$
where xp and yp are coordinates in the reference frame centered on the image-function center ( x 0 , y 0) and rotated to its position angle:
$$x _ { p } & = ( x - x _ { 0 } ) \cos \theta + ( y - y _ { 0 } ) \sin \theta \quad \quad ( 2 0 ) \\ y _ { p } & = - ( x - x _ { 0 } ) \sin \theta + ( y - y _ { 0 } ) \cos \theta$$
This scaled radius is then used to compute the actual intensity, using the appropriate 1-D intensity function (see descriptions of individual image functions, below).
Pure circular versions of any of these functions can be had by specifying that the ellipticity parameter is fixed, with a value of 0. Some functions (e.g., EdgeOnDisk) have only the position angle as a geometric parameter, and instead of computing a scaled radius, convert the pixel coordinates to corresponding r and z values in the rotated 2D coordinate system of the model function.
## 6.1.1. FlatSky
This is a very basic function which produces a uniform background: I ( x y , ) = I sky for all pixels. Unlike most image functions, it has no geometric parameters.
## 6.1.2. Gaussian
This is an elliptical 2D Gaussian function, with central surface brightness I 0 and dispersion σ . The intensity profile is given by
$$I ( r ) = I _ { 0 } \exp \left ( - \frac { r ^ { 2 } } { 2 \sigma ^ { 2 } } \right ). \text{ \quad \ \ } \text{(2)}$$
$$6. 1. 3. \ M o f a t$$
This is an elliptical 2D function with a Moffat (1969) function for the surface brightness profile, with parameters for the central surface brightness I 0, full-width half-maximum (FWHM), and the shape parameter β . The intensity profile is given by
$$I ( r ) \, = \, \frac { I _ { 0 } } { ( 1 + ( r / \alpha ) ^ { 2 } ) ^ { \beta } },$$
where α is defined as
$$\alpha = \frac { F \text{WHM} } { 2 \sqrt { 2 ^ { 1 / \beta } - 1 } }.$$
In practice, FWHM describes the overall width of the profile, while β describes the strength of the wings: lower values of β mean more intensity in the wings than is the case for a Gaussian (as β →∞ , the Moffat profile converges to a Gaussian).
The Moffat function is often a good approximation to typical telescope PSFs (see, e.g., Trujillo et al. 2001), and MAKEIMAGE can easily be used to generate Moffat PSF images.
## 6.1.4. Exponential and Exponential\_GenEllipse
The Exponential function is an elliptical 2D exponential function, with parameters for the central surface brightness I 0 and the exponential scale length h . The intensity profile is given by
$$I ( r ) \, = \, I _ { 0 } \exp ( - r / h ) ; \quad \quad ( 2 4 ) \quad \ \ i a$$
together with the position angle and ellipticity, there are a total of four parameters. This is a good default for galaxy disks seen at inclinations glyph[lessorsimilar] 80 , though the majority of disk galax-◦ ies have profiles which are more complicated than a simple exponential (e.g., Gutiérrez et al. 2011).
The Exponential\_GenEllipse function is identical to the Exponential function except for allowing the use of generalized ellipses (with shapes ranging from 'disky' to pure elliptical to 'boxy') for the isophote shapes. Following Athanassoula et al. (1990) and Peng et al. (2002), the shape of the elliptical isophotes is controlled by the c 0 parameter, such that a generalized ellipse with ellipticity = 1 -b / a is described by
$$\left ( \frac { | x | } { a } \right ) ^ { c _ { 0 } + 2 } + \left ( \frac { | y | } { b } \right ) ^ { c _ { 0 } + 2 } = 1, \text{ \quad \ \ } ( 2 5 )$$
where | x | and | y | are distances from the ellipse center in the coordinate system aligned with the ellipse major axis ( c 0 corresponds to c -2 in the original formulation of Athanassoula et al.). Thus, values of c 0 < 0 correspond to disky isophotes, while values > 0 describe boxy isophotes; c 0 = 0 for a perfect ellipse.
## 6.1.5. Sersic and Sersic\_GenEllipse
This pair of related functions is analogous to the Exponential and Exponential\_GenEllipse pair above, except that the intensity profile is given by the Sérsic (1968) function:
$$I ( r ) \, = \, I _ { e } \, \exp \left \{ - b _ { n } \left [ \left ( \frac { r } { r _ { e } } \right ) ^ { 1 / n } - 1 \right ] \right \}, \quad \ \ ( 2 6 ) \quad \ \ s e e.$$
where I e is the surface brightness at the effective (half-light) radius re and n is the index controlling the shape of the intensity profile. The value of bn is formally given by the solution to the transcendental equation
$$\Gamma ( 2 n ) \, = \, 2 \gamma ( 2 n, b _ { n } ), \quad \quad ( 2 7 ) \quad \text{$\omega$}$$
where Γ ( a ) is the gamma function and γ ( a x , ) is the incomplete gamma function. However, in the current implementation bn is calculated via the polynomial approximation of Ciotti & Bertin (1999) when n > . 0 36 and the approximation of MacArthur et al. (2003) when n ≤ 0 36. .
The Sérsic profile is equivalent to the de Vaucouleurs ( r 1 / 4 ) profile when n = 4, to an exponential when n = 1, and to a Gaussian when n = 0 5; it has become the de facto standard . for fitting the surface-brightness profiles of elliptical galaxies and bulges. Though the empirical justification for doing so is rather limited, the combination a Sérsic profile with n < 1 and isophotes with a boxy shape is often used to represent bars when fitting images of disk galaxies. In addition, the combination of boxy isophotes and high n values may be appropriate for modeling luminous boxy elliptical galaxies.
## 6.1.6. Core-Sersic
This function generates an elliptical 2D function where the major-axis intensity profile is given by the Core-Sérsic model
(Graham et al. 2003; Trujillo et al. 2004), which was designed to fit the profiles of so-called 'core' galaxies (e.g., Ferrarese et al. 2006; Richings et al. 2011; Dullo & Graham 2012, 2013; Rusli et al. 2013). It consists of a Sérsic profile (parameterized by n and re ) for radii > the break radius rb and a single power law with index -γ for radii < rb . The transition between the two regimes is mediated by the dimensionless parameter α : for low values of α , the transition is very gradual and smooth, while for high values of α the transition becomes very abrupt (a perfectly sharp transition can be approximated by setting α equal to some large number, such as 100). The intensity profile is given by
$$\underset { \text{cal} } { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \t \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \tilde { \t \tilde { \tilde { \tilde { \tilde { \t \tilde { \t \tilde { \tilde { \tilde { \t \tilde { \tilde { \t \tilde { \tilde { \tilde { \t \tilde { \t \tilde { \t \tilde { \t \tilde { \tilde { \t \tilde \t \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \ t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \tt | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t | \t |
} } } \end{cases} \quad I ( r ) \, = \, I _ { b } \left [ 1 + \left ( \frac { r _ { b } } { r } \right ) ^ { \alpha } \right ) ^ { \gamma / \alpha } } \exp \left [ - b \left ( \frac { r _ { ^ { \alpha } } { r } + r _ { b } ^ { \alpha } } { r _ { e } ^ { \alpha } } \right ) ^ { 1 / ( n \alpha ) } \right ], \ \ ( 2 8 )$$
where b is the same as bn for the Sérsic function.
The overall intensity scaling is set by I b , the intensity at the break radius rb :
$$I _ { b } \, = \, I _ { b } \, 2 ^ { - \gamma / \alpha } \exp [ b 2 ^ { 1 / \alpha n } ( r _ { b } / r _ { e } ) ^ { 1 / n } ]. \quad \ ( 2 9 )$$
## 6.1.7. BrokenExponential
This is similar to the Exponential function, but it has two exponential radial zones (with different scalelengths) joined by a transition region at Rb of variable sharpness:
$$I ( r ) \, = \, S I _ { 0 } e ^ { - \frac { r } { h _ { 1 } } } [ 1 + e ^ { \alpha ( r - R _ { b } ) } ] ^ { \frac { 1 } { \alpha } ( \frac { 1 } { h _ { 1 } } - \frac { 1 } { h _ { 2 } } ) }, \quad ( 3 0 )$$
where I 0 is the central intensity of the inner exponential, h 1 and h 2 are the inner and outer exponential scale lengths, Rb is the break radius, and α parameterizes the sharpness of the break. Low values of α mean very smooth, gradual breaks, while high values correspond to abrupt transitions. S is a scaling factor, 14 given by
$$S = ( 1 + e ^ { - \alpha R _ { b } } ) ^ { - \frac { 1 } { \alpha } ( \frac { 1 } { h _ { 1 } } - \frac { 1 } { h _ { 2 } } ) } ; \quad \quad ( 3 1 )$$
see Figure 3 for examples. Note that the parameter α has units of length -1 (pixels -1 for the specific case of IMFIT).
The 1D form of this profile (Erwin et al. 2008) was designed to fit the surface-brightness profiles of disks which are not single-exponential: e.g., disks with truncations or antitruncations (Erwin et al. 2005; Erwin et al. 2008; Muñoz-Mateos et al. 2013).
## 6.1.8. GaussianRing
This function creates an elliptical ring with a Gaussian radial profile, centered at r = R ring along the major axis.
$$I ( r ) = I _ { 0 } \exp \left ( - \frac { ( r - R _ { \text{ring} } ) ^ { 2 } } { 2 \sigma ^ { 2 } } \right ). \quad \quad ( 3 2 )$$
See Figure 4 for an example.
$$6. 1. 9. \ G a u s s i a n R i n g 2 S i d e$$
This function is similar to GaussianRing, except that it uses an asymmetric Gaussian, with different values of σ for r < R ring and r > R ring ). That is, the profile behaves as
$$I ( r ) = I _ { 0 } \exp \left ( - \frac { ( r - R _ { \text{ring} } ) ^ { 2 } } { 2 \sigma _ { \text{in} } ^ { 2 } } \right ) \quad \ \ ( 3 3 )$$
14 As pointed out by Muñoz-Mateos et al. (2013), the original definition of this factor in Eqn. 6 of Erwin et al. (2008) contained a typo.
Figure 3. Examples of the 'broken-exponential' surface-brightness profile used by the BrokenExp and BrokenExp3D image functions. The upper (solid) curves show a profile with inner and outer scale lengths h 1 = 15 and h 2 = 2 pixels, respectively, break radius = 6 pixels, and varying values of α (black = 100, medium gray = 3, light gray = 1). The lower (dashed) curves show the effects of varying the outer scale length only ( h 2 = 4, 3, 2 pixels).
Figure 4. Logarithmically scaled isophotes for examples of the Gaussian ring image functions. The inner, more elliptical ring was generated by the GaussianRing function, with an ellipticity of 0.4, semi-major axis of 80 pixels, and σ = 10 pixels. The larger, rounder ring is an example of the GaussianRing2Side function, with an ellipticity of 0.1, a semi-major axis of 160 pixels, σ in = 10 pixels, and σ out = 20 pixels.
for r < R ring , and
$$I ( r ) = I _ { 0 } \exp \left ( - \frac { ( r - R _ { \text{ring} } ) ^ { 2 } } { 2 \sigma _ { \text{out} } ^ { 2 } } \right ) \quad \ \ ( 3 4 ) \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
for a > R ring ; see Figure 4 for an example.
$$6. 1. 1 0. \ E d g e O n D i s k$$
This function provides the analytic form for a perfectly edge-on disk with a radial exponential profile, using the Bessel-function solution of van der Kruit & Searle (1981)
for the radial profile. Although it is common to assume that the vertical profile for galactic disks follows a sech 2 function, based on the self-gravitating isothermal sheet model of Spitzer (1942), van der Kruit (1988) suggested a more generalized form for this, one which enables the profile to range from sech 2 at one extreme to exponential at the other:
$$L ( z ) \, \in \, \text{sech} ^ { 2 / n } ( n z / ( 2 z _ { 0 } ) ), \quad \quad ( 3 5 )$$
with z the vertical coordinate and z 0 the vertical scale height. The parameter n produces a sech 2 profile when n = 1, sech when n = 2, and converges to an exponential as n →∞ . See de Grijs et al. (1997) for examples of fitting the vertical profiles of edge-on galaxy disks using this formula, and Yoachim & Dalcanton (2006) for examples of 2D fitting of edge-on galaxy images.
In a coordinate system aligned with the edge-on disk, r is the distance from the minor axis (parallel to the major axis) and z is the perpendicular direction, with z = 0 on the major axis. (The latter corresponds to height z from the galaxy midplane.) The intensity at ( r z , ) is given by
$$I ( r, z ) \, = \, \mu ( 0, 0 ) \, ( r / h ) \, K _ { 1 } ( r / h ) \, \text{sech} ^ { 2 / n } ( n z / ( 2 z _ { 0 } ) ) \quad ( 3 6 )$$
where h is the exponential scale length in the disk plane, z 0 is the vertical scale height, and K 1 is the modified Bessel function of the second kind. The central surface brightness µ (0 0) , is given by
$$\mu ( 0, 0 ) \, = \, 2 \, h \, L _ { 0 }, \, \text{ \quad \ \ } ( 3 7 )$$
where L 0 is the central luminosity density (see van der Kruit & Searle 1981). Note that L 0 is the actual input parameter required by the function; µ (0 0) is calculated internally. ,
The result is a function with five parameters: L 0, h , z 0, n , and the position angle; Figure 5 shows three examples with differing vertical profiles parameterized by n = 1, 2, and 100.
## 6.1.11. EdgeOnRing
This is a simplistic model for an edge-on ring, using two offset subcomponents located at distance R ring from the center of the function block. Each subcomponent (i.e., each side of the ring) is a 2D Gaussian with central surface brightness I 0 and dispersions of σ r in the radial direction and σ z in the vertical direction. It has five parameters: I 0, R ring , σ r , σ z , and the position angle. See Figure 6 for examples of this function.
A potentially more correct (though computationally more expensive) model for a ring seen edge-on ring - or at other inclinations - is provided by the GaussianRing3D function, below.
## 6.1.12. EdgeOnRing2Side
This is a slightly more sophisticated variant of EdgeOnRing, where the radial profile for the two components is an asymmetric Gaussian, as in the case of the GaussianRing2Side function, above: the inner ( | r | < R ring ) side of each component is a Gaussian with radial dispersion σ , r in , while the outer side has radial dispersion σ , r out . It thus has six parameters: I 0, R ring , σ , r in , σ , r out , σ z , and the position angle. See the right-hand panel of Figure 6 for an example.
## 6.2. 3D Components
All image functions in IMFIT produce 2D surfacebrightness output. However, there is nothing to prevent one from creating a function which does something quite complicated in order to produce this output. As an example, IMFIT
Figure 5. Examples of the EdgeOnDisk image function, which uses the analytic Bessel-function solution of van der Kruit & Searle (1981) for a perfectly edge-on exponential disk, combined with the generalized sech 2 / n vertical profile of van der Kruit (1988). All panels show models with radial and vertical scale lengths h = 50 and z 0 = 10 pixels, respectively. From left to right, the panels show images with vertical sech 2 / n profiles having n = 1 (sech 2 profile), 2 (sech profile), and 100 ( ≈ exponential profile).
Figure 6. Examples of the EdgeOnRing (left and middle panels) and EdgeOnRing2Side (right panel) image functions, which provide simple approximations for rings seen edge-on. All rings have a radius of 120 pixels. The left-hand panel shows a ring with radial and vertical Gaussian widths of 20 and 10 pixels, respectively; the middle panel shows a model with the radial and vertical widths exchanged. The right-hand panel shows an example of the EdgeOnRing2Side function, where the radial scales are σ = 40 pixels on the inside and 20 pixels on the outside; the vertical scale is 10 pixels.
includes three image functions which perform line-of-sight integration through 3D luminosity-density models, in order to produce a 2D projection.
$$y _ { p } = - ( x - x _ { 0 } ) \sin \theta + ( y - y _ { 0 } ) \cos \theta$$
These functions assume a symmetry plane (e.g., the disk plane for a disk galaxy) which is inclined with respect to the line of sight; the inclination is defined as the angle between the line of sight and the normal to the symmetry plane, so that a face-on system has i = 0 ◦ and an edge-on system has i = 90 ◦ . For inclinations > 0 , the orientation of the line of ◦ nodes (the intersection between the symmetry plane and the sky plane) is specified by a position-angle parameter θ . Instead of a 2D surface-brightness specification (or 1D radial surface-brightness profile), these functions specify a 3D luminosity density j , which is numerically integrated along the line of sight s for each pixel of the model image:
$$I ( x, y ) \, = \, \int _ { - S } ^ { S } j ( s ) \, d s. \quad \quad ( 3 8 ) \quad \stackrel { \dots } { c }$$
To carry out the integration for a pixel located at ( x y , ) in the image plane, the coordinates are first transformed to a rotated image plane system ( xp , yp ) centered on the coordinates of the component center ( x 0 , y 0), where the line of nodes lies along the xp axis (cf. Eqn. 20 in Section 6.1):
$$x _ { p } = ( x - x _ { 0 } ) \cos \theta + ( y - y _ { 0 } ) \sin \theta$$
with θ being the angle between the line of nodes and the image + x axis (as in the case of the 2D functions, the actual userspecified parameter is PA = θ -90, which is the angle between the line of nodes and the + y axis).
The line-of-sight coordinate s is then defined so that s = 0 in the sky plane (an instance of the image plane located in 3D space so that it passes through the center of the component), corresponding to
$$x _ { d, 0 } & = x _ { p } \\ y _ { d, 0 } & = y _ { p } \cos i \\ z _ { d, 0 } & = y _ { p } \sin i$$
in the component's native ( xd , yd , zd ) Cartesian coordinate system. A location at s along the line of sight then maps into the component coordinate system as
$$y _ { d } & = y _ { d, 0 } + s \sin i \\ z _ { d } & = z _ { d, 0 } - s \cos i,$$
with xd = xd , 0 = xp by construction. The luminosity-density value is then j ( s ) = j ( xd , xd , zd ). See Figure 7 for a side-on view of this arrangement.
Although a fully correct integration would run from s = -∞
Figure 7. Asimplified illustration of how line-of-sight integration is handled for 3D image functions. Here, an axisymmetric ExponentialDisk3D component is inclined at angle i with respect to the line of sight, with the line of nodes rotated to lie along the sky-plane xp axis, perpendicular to the page; the disk center is by construction at the intersection of the disk plane and the sky plane. For a pixel with sky-plane coordinates ( xp , yp ), the luminosity-density is integrated along the line of sight (variable s , with s = 0 at the sky plane). For each value of s used by the integration routine, the luminosity density is computed based on the corresponding values of radius r = ( x 2 d + y 2 d ) 1 / 2 and height zd in the disk's native coordinate system.
to ∞ , in practice the limit S is some large multiple of the component's normal largest scale size (e.g., 20 times the horizontal disk scale length h ), to limit the possibility of numerical integration mishaps.
## 6.2.1. ExponentialDisk3D
This function implements a 3D luminosity density model for an axisymmetric disk where the radial profile of the luminosity density is an exponential and the vertical profile follows the sech 2 / n function of van der Kruit (1988) (see the discussion of the EdgeOnDisk function in Section 6.1.10). The line-of-sight integration is done numerically, using functions from the GNU Scientific Library.
In a cylindrical coordinate system ( r z , ) aligned with the disk (where the disk midplane has z = 0), the luminosity density j ( r z , ) at radius r from the central axis and at height z from the midplane is given by
$$j ( r, z ) \, = \, J _ { 0 } \, \exp ( - r / h ) \, \text{sech} ^ { 2 / n } ( n z / ( 2 z _ { 0 } ) ) \quad ( 4 0 ) \quad \tilde { \ p u t } _ { \bullet \, \cdot \, \cdot }$$
where h is the exponential scale length in the disk plane, z 0 is the vertical scale height, n controls the shape of the vertical distribution, and J 0 is the central luminosity density. Note that in the context of the introductory discussion above, z = zd and r = ( x 2 d + y 2 d ) 1 / 2 .
Figure 8 shows three views of the same model, at inclinations of 75 ◦ , 85 , and 89 ◦ ◦ ; the latter is almost identical to the image produced by the analytic EdgeOnDisk with the same radial and vertical parameters (right-hand panel of Figure 5).
## 6.2.2. BrokenExponentialDisk3D
This function is identical to the ExponentialDisk3D function, except that the radial part of the luminosity density function is given by the broken-exponential profile used by the (2D) BrokenExponential function, above (Section 6.1.7). Thus, the luminosity density j ( r z , ) at radius r from the central axis and at height z from the midplane is given by
$$j ( r, z ) \, = \, I _ { \text{rad} } ( r ) \, \text{seh} ^ { 2 / n } ( n z / ( 2 z _ { 0 } ) ) \, \quad \ \ ( 4 1 ) \, \quad \ \$$
where z 0 is the vertical scale height, and the radial part is given by
$$I _ { \text{rad} } ( r ) \, = \, S J _ { 0 } \, e ^ { - \frac { r } { h _ { 1 } } } [ 1 + e ^ { \alpha ( r - R _ { b } ) } ] ^ { \frac { 1 } { \alpha } ( \frac { 1 } { h _ { 1 } } - \frac { 1 } { h _ { 2 } } ) }, \quad ( 4 2 )$$
with J 0 being the central luminosity density and the rest of the parameters as defined for BrokenExponential function (Section 6.1.7).
## 6.2.3. GaussianRing3D
This function creates the projection of a 3D elliptical ring, seen at an arbitrary inclination. The ring has a luminosity density with a radial Gaussian profile (centered at a ring along ring's major axis, with in-plane width σ ) and a vertical exponential profile (with scale height hz ). The ring can be imagined as residing in a plane which has its line of nodes at angle θ and inclination i (as for the ExponentialDisk3D function, above); within this plane, the ring's major axis is at position angle φ relative to the perpendicular to the line of nodes. To derive the correct luminosity densities for the line-of-sight integration, the component coordinate values xd , yd , zd from Equation 39 are transformed to a system rotated about the normal to the ring plane, where the ring's major axis is along the x ring axis:
$$x _ { \text{ring} } & = x _ { d } \cos ( \phi ) + y _ { d } \sin ( \phi ) \\ y _ { \text{ring} } & = - x _ { d } \sin ( \phi ) + y _ { d } \cos ( \phi ) \\ z _ { \text{ring} } & = | z _ { d } |.$$
Figure 9 shows the same GaussianRing3D component (with ellipticity = 0.5) seen at three different inclinations.
## 7. PROGRAMMING NOTES
IMFIT is written in standard C++, and should be compilable with any modern compiler; it has been tested with GCC versions 4.2 and 4.8 on Mac OS X and GCC version 4.6 on Ubuntu Linux systems. It makes use of several open-source libraries, two which are required (CFITSIO and FFTW) and two which are optional but recommended (NLopt and the GNU Scientific Library). IMFIT also uses the Python-based SCons 15 build system and CxxTest 16 for unit tests.
Since the slowest part of the fitting process is almost always computing the model image, IMFIT is written to take advantage of OpenMP compiler extensions; this allows the computation of the model image to be parceled out into multiple threads, which are then allocated among available processor cores on machines with multiple shared-memory CPUs (including single CPUs with multiple cores). As an example of how effective this can be, tests on a MacBook Pro with a quad-core i7 processor, which has a total of eight virtual threads available, show that basic computation of large images (without PSF convolution) is sped up by a factor of ∼ 6 when OpenMP is used. Even when the overhead of an actual fit is included, the total time to fit a four-component model with 21 free parameters (without PSF convolution) to a 500 × 500pixel image is reduced by a factor of ∼ 3 8. .
Additional computational overhead is imposed when one convolves a model image with a PSF. To mitigate this, IMFIT uses the FFTW library to compute the necessary Fourier transforms. This is one of the fastest FFT libraries available, and it can be compiled with support for multiple threads. When the same 500 × 500-pixel image fit mentioned above is done
15 http://www.scons.org
16 http://cxxtest.com
Figure 8. Examples of the ExponentialDisk3D image function, which uses line-of-sight integration through a 3D luminosity-density model of a disk with radial exponential profile and vertical sech 2 / n profile. All panels show the same model, with radial and vertical scale lengths h = 50 and z 0 = 10 pixels, respectively, and a vertical exponential profile ( n = 100). From left to right, the panels show projections with inclinations of 75 ◦ , 85 ◦ , and 89 ◦ ; compare the last panel with the right-hand panel in Figure 5.
Figure 9. Examples of the GaussianRing3D image function, which uses line-of-sight integration through a 3D luminosity-density model of an elliptical ring with Gaussian radial and exponential vertical profiles. This particular ring has an intrinsic (in-plane) ellipticity = 0 5, semi-major axis = 100 pixels, Gaussian . radial width σ = 10 pixels, and exponential scale height hz = 5 pixels. From left to right, panels show face-on, i = 45 ◦ , and edge-on views.
including convolution with a 35 × 35-pixel PSF image, the total time drops from ∼ 280s without any multi-threading to ∼ 120s when just the FFT computation is multi-threaded, and down to ∼ 50s when OpenMP threading is enabled as well.
on disk component (EdgeOnDisk; Section 6.1.10) and the 3D luminosity-density model of an exponential disk (ExponentialDisk3D; Section 6.2.1).
Multithreading can always be reduced or turned off using a command-line option if one does not wish to use all available CPU cores for a given fit.
## 8. SAMPLE APPLICATIONS
IMFIT has been used for several different astronomical applications, including preliminary work on the EUCLID photometric pipeline (Kümmel et al. 2013), testing 1D convolution code used in the analysis of core galaxies (Rusli et al. 2013), fitting kinematically decomposed components of the galaxy NGC 7217 (Fabricius et al. 2014), determining the PSF for Data Release 2 of the CALIFA survey (García-Benito et al. 2014), and separation of bulge and disk components for dynamical modeling of black hole masses in nearby S0 and spiral galaxies (Erwin et al., in prep).
In this section I present two relatively simple examples of using IMFIT to model images of real galaxies. The first case considers a moderately inclined spiral galaxy with a prominent ring surrounding a bar, where use of a separate ring component considerably improves the fit. The second case is an edge-on disk galaxy with both thin and thick disks; I show how this can be fit using both the analytic pure-edge-
## 8.1. PGC 35772: Disk, Bar, and Ring
PGC 35772 is a z = 0 0287, early/intermediate-type spiral . galaxy (classified as SA0/a by de Vaucouleurs et al. 1993 and as Sb by Fukugita et al. 2007) which was observed as part of the H α Galaxy Groups Imaging Survey (HAGGIS; Kulkarni et al., in prep.) using narrow-band filters on the Wide Field Imager of the ESO 2.2m telescope. The upper-left panel of Figure 10 shows the stellar-continuum-filter image (central wavelength ≈ 659 nm, slightly blueward of the redshifted H α line). Particularly notable is a bright stellar ring, which makes this an interesting test case for including rings in 2D fits of galaxies. Ellipse fits to the image show strong twisting of the isophotal position angle interior to the ring, suggesting a bar is also present.
The rest of Figure 10 shows the results three different fits to the image, each successive fit adding an extra component. These fits use a 291 × 281-pixel cutout of the full WFI image, and were convolved with a Moffat-function image with FWHM = 0.98 , representing the mean PSF (based on Mof-′′ fat fits to stars in the same image). The best-fit parameters for each model, determined by minimizing PMLR, are listed in Table 3, along with the uncertainties estimated from the L-
Table 3 Results of Fitting PGC 35772
| Component (1) | Parameter (2) | Value (3) | σ (4) | units (5) |
|---------------------|----------------------------------------|----------------------------------------|----------------------------------------|----------------------------------------|
| | Sersic only (AIC = 18419) | Sersic only (AIC = 18419) | Sersic only (AIC = 18419) | Sersic only (AIC = 18419) |
| Sersic | PA | 138.14 | 0.48 (0.34) | deg |
| Sersic | glyph[epsilon1] | 0.254 | 0.0038 (0.0024) | |
| Sersic | n | 1.041 | 0.0085 (0.027) | |
| Sersic | I e | 39.16 | 0.33 (0.52) | cont. flux |
| Sersic | r e | 8.267 | 0.042 (0.034) | arcsec |
| | Exponential + Bar (AIC = 16569) | Exponential + Bar (AIC = 16569) | Exponential + Bar (AIC = 16569) | Exponential + Bar (AIC = 16569) |
| Exponential (disk) | PA | 137.79 | 0.49 (0.29) | deg |
| Exponential (disk) | glyph[epsilon1] | 0.259 | 0.0039 (0.0021) | |
| Exponential (disk) | I 0 | 194.58 | 1.22 (0.97) | cont. flux |
| Exponential (disk) | h | 5.17 | 0.025 (0.014) | arcsec |
| Sersic (bar) | PA | 16.27 | 3.00 (1.13) | deg |
| Sersic (bar) | glyph[epsilon1] | 0.562 | 0.056 (0.025) | |
| Sersic (bar) | n | 0.897 | 0.282 (0.074) | |
| Sersic (bar) | I e | 171.82 | 25.32 (6.77) | cont. flux |
| Sersic (bar) | r e | 0.713 | 0.041 (0.021) | arcsec |
| | Exponential + Bar + Ring (AIC = 14996) | Exponential + Bar + Ring (AIC = 14996) | Exponential + Bar + Ring (AIC = 14996) | Exponential + Bar + Ring (AIC = 14996) |
| Exponential (disk) | PA | 140.70 | 1.25 (0.53) | deg |
| Exponential (disk) | glyph[epsilon1] | 0.277 | 0.0098 (0.0053) | |
| Exponential (disk) | I 0 | 111.19 | 11.55 (4.05) | cont. flux |
| Exponential (disk) | h | 5.74 | 0.18 (0.052) | arcsec |
| Sersic (bar) | PA | 7.27 | 2.28 (1.04) | deg |
| Sersic (bar) | glyph[epsilon1] | 0.364 | 0.028 (0.092) | |
| Sersic (bar) | n | 1.14 | 0.114 (0.046) | |
| Sersic (bar) | I e | 80.78 | 7.25 (2.47) | cont. flux |
| Sersic (bar) | r e | 1.42 | 0.080 (0.028) | arcsec |
| GaussianRing (ring) | PA | 128.40 | 1.69 (0.69) | deg |
| GaussianRing (ring) | glyph[epsilon1] | 0.258 | 0.013 (0.0053) | |
| GaussianRing (ring) | A | 26.90 | 3.22 (1.10) | cont. flux |
| GaussianRing (ring) | R | 5.50 | 0.36 (0.14) | arcsec |
| GaussianRing (ring) | σ | 3.43 | 0.22 (0.13) | arcsec |
Note . -Results of fitting narrow-band continuum image of spiral galaxy PGC 35771 with progressively more complex models (Sersic; Exponential + Sersic; Exponential + Sersic + GaussianRing). 'AIC' = Akaike Information Criterion values for the fits; lower values imply better fits. Column 1: Component used in fit. Column 2: Parameter. Column 3: Best-fit value for parameter. Column 4: Uncertainty on parameter value from L-M covariance matrix; uncertainty from bootstrap resampling is in parentheses. Column 5: Units ('cont. flux' units are 10 -18 erg s -1 cm 2 Å --1 arcsec -2 ).
Mcovariance matrix. Since the fitting times are short, I also include parameter uncertainties from 500 rounds of bootstrap resampling (in parentheses, following the L-M uncertainties).
The first fit is uses a single Sersic component; the residuals of this fit show a clear excess corresponding to the ring, as well as mis-modeling of the region inside the ring. The fit is improved by switching to an exponential + Sérsic model, with the former component representing the main disk and the latter some combination of the bar + bulge (if any). This two-component model (middle row of the figure) produces less extreme residuals; the best-fitting Sérsic component is elongated and misaligned with the exponential component, so it can be seen to be modeling the bar.
The residuals to this 'disk + bar' fit are still significant, however, including the ring itself. To fix this, I include a GaussianRing component (Section 6.1.8) in the third fit (bottom row of Figure 10). The residuals to this fit are better not just in the ring region, but also inside, indicating that this three-component model is doing a much better job of modeling the inner flux of the galaxy (the three-component also has the smallest AIC value of the three models; see Table 3).
## 8.2. IC 5176: Fitting Thin and Thick Disks in an Edge-on Spiral in 2D and 3D
IC 5176 is an edge-on Sbc galaxy, included in a 'control' sample of non-boxy-bulge galaxies by Chung & Bureau (2004) and Bureau et al. (2006). Chung & Bureau (2004) noted that both the gas and stellar kinematics were consistent with an axisymmetric, unbarred disk; Bureau et al. (2006) concluded from their K -band image that it had a very small bulge and a 'completely featureless outer (single) exponential disk.' This suggests an agreeably simple, axisymmetric structure, ideal for an example of modeling an edge-on galaxy. To minimize the effects of the central dust lane (visible in optical images of the galaxy), I use a Spitzer IRAC1 (3.6 µ m) image from S4G (Sheth et al. 2010), retrieved from the Spitzer archive. For PSF convolution, I use an in-flight point response function image for the center of the IRAC1 field, 17 downsampled to the 0.6 ′′ pixel scale of the post-processed archival galaxy image.
Inspection of major-axis and minor-axis profiles from the IRAC1 image (Figure 11) suggests the presence of both thin and thick disk components; the K -band image of Bureau et al. (2006) was probably not deep enough for this to be seen. The major axis profile and the image both suggest a rather round, central excess, consistent with the small bulge identified by Bureau et al.
Consequently, I fit the image using a combination of two exponential-disk models, plus a central Sérsic component for the bulge. The fast way to fit such a galaxy with IMFIT is to assume that the galaxy is perfectly edge-on and use the 2D analytic EdgeOnDisk functions (Section 6.1.10) for the thin and thick disk components. Table 4 shows the results of this fit. The dominant EdgeOnDisk component, which can be thought of as the 'thin disk', has a nearly sech vertical profile and a scale height of 2 0 . ′′ ≈ 260 pc (assuming a distance of 26.4 Mpc; Tully et al. 2009). The second EdgeOnDisk, with a more exponential-like vertical profile and a scale height of 1.4 kpc, is then the 'thick disk' component; it has a radial scale length ∼ 2 9 times that of the thin disk. .
The central Sérsic component of this model contributes 1.8% of the total flux, while the thin and thick disks account for 70.5% and 27.7%, respectively. The thick/thin-disk luminosity ratio of 0.39 is consistent with the recent study of thick and thin disks by Comerón et al. (2011): using their two assumed sets of relative mass-to-light ratios gives a mass ratio M thick / M thin = 0 47 or 0.94, which places IC 5176 in the . middle of the distribution for galaxies with similar rotation velocities (see their Fig. 13).
A slower but more general approach is to use the EdgeOnDisk3D function (Section 6.2.1) for both components, which allows for arbitrary inclinations. The cost is in the time taken for the fit: ∼ 29 minutes, versus a mere 3m20s for the analytic 2D approach. Using the EdgeOnDisk3D functions does give what is formally a better model of the data than the analytic 2D-component fit, with ∆ AIC ≈ 2305, though most of the parameter values - in particular, the radial and vertical scale lengths - are almost identical to previous fit. The only notable changes are the Sérsic component becoming rounder (with a different and probably not very well-defined position angle) and the vertical profiles of both disk components becoming pure exponentials (the values of n in Table 4 are imposed limits). The relative contributions of the three compo-
17 http://irsa.ipac.caltech.edu/data/SPITZER/docs/irac/calibrationfiles/
Figure 10. Fits of progressively more complex models to a narrow-band continuum image of the spiral galaxy PGC 35772. Top row: data image (displayed with log stretch), log-scaled isophote contours of best-fit Sersic model, residual image (data -model) image, displayed with linear stretch (1 pixel = 0.238 ′′ ). Middle row: isophote contours of best-fit Exponential + Sersic components, residual image. Bottom row: isophote contours of best fit Exponential + Sersic + GaussianRing components, residual image.
Figure 11. Minor-axis profile from Spitzer IRAC1 (3.6 µ m) image of edgeon spiral galaxy IC 5176 (black line), along with corresponding profile from best-fitting two-disk model (red dashed line).
nents are essentially unchanged: 1.8% of the flux from the Sérsic component and 71.7% and 26.5% from the thin and thick disks, respectively. The best-fitting model converges to i ≈ 90 ◦ for the outer (thick) disk component, but does find i = 87 2 . ◦ for the thin-disk component.
The reality is that the combination of low spatial resolution of the IRAC1 image and the presence of residual structure in the disk midplane (probably due to a combination of spiral arms, star formation, and dust) means that we cannot constrain the vertical structure of the disk(s) very well. A vertical profile which is best fit with a sech function when the disk is assumed to be perfectly edge-on can also be fit with a vertical exponential function, if the disk is tilted slightly from edge-on. The low spatial resolution also means that the central bulge is not well constrained, either; the half-light radius of the Sérsic component from either fit is ∼ 2 5 pixels and . thus barely larger than the seeing.
## 9. POTENTIAL BIASES IN FITTING GALAXY IMAGES: χ 2 VERSUS POISSON MAXIMUM-LIKELIHOOD FITS
In Section 4.1, I discussed two different practical approaches to fitting images from a statistical point of view: the standard, Gaussian-based χ 2 statistic and Poisson-based MLE statistics ( C and PMLR). The χ 2 approach can be further subdivided into the common method of using data values to estimate the per-pixel errors ( χ 2 d ) and the alternate method of using values from the model ( χ 2 m ). Outside of certain low-S/N contexts (e.g., fitting X-ray and gamma-ray data), χ 2 minimization is pretty much the default. Even in the case of low S/N, when the Gaussian approximation to Poisson statistics which motivates the χ 2 approach - might start to become invalid, one might imagine that the presence of Gaussian read noise in CCD detectors could make this a non-issue. Is there any reason for using Poisson-likelihood approaches outside of
Figure 12. Top row, left: Logarithmically scaled isophotes of the Spitzer IRAC1 (3.6 µ m) image of edge-on spiral galaxy IC 5176, smoothed with a 5-pixel-wide median filter (1 pixel = 0.6 ′′ ). Top row, middle: Best-fitting, PSF-convolved model (see bottom row). Top row, right: residual image (data -model), displayed with linear stretch. Bottom row: log-scaled isophotes showing PSF-convolved components making up the best-fitting model, consisting of two ExponentialDisk3D components and a Sersic component. All isophote plots use the same logarithmic scaling.
very-low-count, zero-read-noise regimes?
Humphrey et al. (2009) used a combination of analytical approximations and fits of models to artificial data to show how χ 2 fits (using data-based or model-based errors) can lead to biased parameter estimation, even for surprisingly high S/N ratios; these biases were essentially absent when Poisson MLE was used. (Humphrey et al. 2009 used C for their analysis, but minimizing PMLR would yield the same fits, as noted in Section 4.1.3.) A fuller discussion of these issues in the context of fitting X-ray data can be found in that paper, and references therein (e.g. Nousek & Shue 1989). In this section, I focus on the typical optical imaging problem of fitting galaxy images with simple 2D functions and use the flexibility of IMFIT to explore how fitting Poisson (or Poisson + Gaussian) data with different assumptions can bias the resulting fitted parameter values.
## 9.1. Fitting Simple Model Galaxy Images
As a characteristic example, I consider a model galaxy described by a 2D Sérsic function with n = 3 0, . re = 20 pixels, and an ellipticity of 0.5. This model is realized in three count-level regimes: a 'low-S/N' case with a sky background level of 20 counts/pixel and model intensity at the half-light radius I e = 50 counts/pixel; a 'medium-S/N' version which is equivalent to an exposure time (or telescope aperture) five times larger (background level = 100, I e = 250); and a 'highS/N' version with total counts equal to 25 times the lowS/N version (background level = 500, I e = 1250). These values are chosen partly to test the question of how rapidly the Gaussian approximation to Poisson statistics becomes appropriate: 20 counts/pixel is often given as a reasonable lower limit for using this approximation (e.g., Cash 1979), while for
500 counts/pixel the Gaussian approximation should be effectively indistinguishable from true Poisson statistics.
The images were created using code written in Python. The first stage was generating a noiseless 150 × 150-pixel reference image (including subpixel integration, but not PSF convolution). This was then used as the source for generating 500 'observed' images of the same size, using Poisson statistics: for each pixel, the value in the reference image was taken as the mean m for a Poisson process (Equation 4.1.1), and actual counts were (pseudo)randomly generating using code in the Numpy package ( numpy.random.poisson ). 18 For simplicity, the gain was set to 1, so 1 count = 1 photoelectron.
The resulting images were then fit with IMFIT three times, always using the Nelder-Mead simplex method as the minimization algorithm. The first two fits used χ 2 statistics, either the data-based χ 2 d or the model-based χ 2 m approach, with read noise set to 0; the third fit minimized C . (Essentially identical fits are obtained when minimizing PMLR instead of C .) The fitted model consisted of a single 2D Sérsic function with very broad parameter limits and the same starting parameter values for all fits (with the initial I e value scaled by 5 for the mediumS/N images and by 25 for the high-S/N images), along with a fixed FlatSky component for the background.
Figure 13 shows the distribution of best-fit parameters for fits to all 500 individual images in each S/N regime, with thick red histograms for the χ 2 d fits, thinner magenta histograms for the χ 2 m fits, and thin blue histograms for the Poisson MLE ( C ) fits, along with the true parameter values of the original model as vertical dashed gray lines.
Erwin
Figure 13. Distribution of best-fit parameters from fits to 500 realizations of an artificial galaxy image with an elliptical Sérsic component + sky background and pure Poisson noise (solid histograms), or Poisson noise + Gaussian read noise with σ = 5 e -(dashed histograms). Fits used data-based χ 2 minimization ( χ 2 d , red histograms), model-based χ 2 minimization ( χ 2 m , magenta), or Poisson MLE minimization (using C , blue); vertical dashed gray lines indicate parameter values of the original model. Upper row: Results for low-S/N images (sky background = 20 e -/pixel, Sérsic model Ie = 50 e -/pixel). Middle row: Results for medium-S/N images (background = 100 e -/pixel, Ie = 250 e -/pixel). Bottom row: Results for high-S/N images (background = 500 e -/pixel, Ie = 1250 e -/pixel); the additional histograms for fits to images with Gaussian read noise are in this case essentially indistinguishable from the pure-Poisson-noise histograms and are not plotted. Using χ 2 d minimization systematically underestimates (Sérsic n and re ) or overestimates ( Ie ) the parameters, while using χ 2 m minimization produces smaller biases in the opposite directions. These biases diminish as the counts/pixel get larger. The presence of Gaussian read noise reduces the χ 2 bias in the low-S/N regime, but does not eliminate it. In all cases, minimization of the Poisson MLE statistic C is bias-free.
A clear bias for the χ 2 approaches can be seen in the fits to the low-S/N images (top panels of Figure 13). For the χ 2 d approach, the fitted values of n and re are too small: the average value of n is 12.3% low, while the average value of re is 15.4% too small. The fitted values of I e , on the other hand, are on average 34% too large. As can be seen from the figure, these biases are significantly larger than the spread of values from the individual fits. The overall effect also biases the total flux for the Sérsic component, which is underestimated by 10.4% when using the mean parameters of the χ 2 d fit; see Figure 14. The (model-based) χ 2 m approach also produces biases, though these are smaller and are in the opposite sense from the χ 2 d biases: n and re are overestimated on average by 7.0% and 10.4%, respectively, while I e is 15.6% too small; the total flux is overestimated by 6.5%. Finally, the fits using C are unbiased : the histograms straddle the input model values, and the mean values from the fits are all < . 0 1% different from the true values. The other parameters of the fits - galaxy center, position angle, ellipticity - do not show any systematic differences, except for a very slight tendency of the ellipticity to be biased high with the χ 2 d fit, but only at the ∼ 0 5% level. For . the parameters which show biases in the χ 2 fits, the trends are exactly as suggested by Humphrey et al. (2009), including the fact that the χ 2 m biases are smaller and have the opposite sign from the χ 2 d biases.
In the medium-S/N case (middle panels of the same figure), the bias in the χ 2 d and χ 2 m fits is clearly reduced: for the χ 2 d fits, n and re are on average only 2.6% and 3.5% too small, while I e is on average 6.4% too high (in the χ 2 m fits, the deviations are 1.3% and 1.9% too large and 3.2% too small, respectively) though the bias is clearly still present, and in the same pattern. The biases in total flux are smaller, too: 2.3% low and 1.2% high for the data-based and model-based χ 2 fits, respectively (Figure 14). These biases are even smaller in the high-S/N
Table 4 Results of Fitting IC 5176
| Component (1) | Parameter (2) | Value (3) | σ (4) | units (5) |
|----------------------------------|-------------------------|-------------------------|-------------------------|-------------------------|
| Fit with 2D Disks (AIC = 182129) | | | | |
| Sersic (bulge) | PA | 149.7 | 0.0049 | deg |
| Sersic (bulge) | glyph[epsilon1] | 0.206 | 0.014 | deg |
| Sersic (bulge) | n | 0.667 | 0.033 | deg |
| Sersic (bulge) | µ e | 12.90 | 0.000 | mag arcsec - 2 |
| Sersic (bulge) | r e | 1.48 | 0.019 | arcsec |
| EdgeOnDisk (thin disk) | PA | 149.7 | 0.0049 | deg |
| EdgeOnDisk (thin disk) | µ 0 | 11.829 | 0.0008 | mag arcsec - 2 |
| EdgeOnDisk (thin disk) | h | 14.17 | 0.012 | arcsec |
| EdgeOnDisk (thin disk) | n | 2.607 | 0.025 | arcsec |
| | z 0 | 2.01 | 0.0044 | arcsec |
| EdgeOnDisk (thick disk) | PA | 151.3 | 0.019 | deg |
| EdgeOnDisk (thick disk) | µ 0 | 15.557 | 0.0057 | mag arcsec - 2 |
| EdgeOnDisk (thick disk) | h | 40.97 | 0.011 | arcsec |
| EdgeOnDisk (thick disk) | n | 9.89 | 0.700 | |
| EdgeOnDisk (thick disk) | z 0 | 10.88 | 0.036 | arcsec |
| Fit with | 3D Disks (AIC = 179824) | 3D Disks (AIC = 179824) | 3D Disks (AIC = 179824) | 3D Disks (AIC = 179824) |
| Sersic (bulge) | PA | 168.71 | 9.73 | deg |
| Sersic (bulge) | glyph[epsilon1] | 0.046 | 0.016 | deg |
| Sersic (bulge) | n | 0.762 | 0.033 | deg |
| Sersic (bulge) | µ e | 13.10 | 0.023 | mag arcsec - 2 |
| Sersic (bulge) | r e | 1.46 | 0.019 | arcsec |
| ExponentialDisk3D (thin disk) | PA | 149.73 | 0.001 | deg |
| ExponentialDisk3D (thin disk) | i | 87.21 | 0.015 | deg |
| ExponentialDisk3D (thin disk) | µ 0 | 11.475 | 0.0010 | mag arcsec - 2 |
| ExponentialDisk3D (thin disk) | h | 14.44 | 0.011 | arcsec |
| ExponentialDisk3D (thin disk) | n | 50 | - | |
| ExponentialDisk3D (thin disk) | z 0 | 2.04 | 0.004 | arcsec |
| ExponentialDisk3D (thick disk) | PA | 151.43 | 0.019 | deg |
| ExponentialDisk3D (thick disk) | i | 89.40 | 0.126 | deg |
| ExponentialDisk3D (thick disk) | µ 0 | 15.604 | 0.0040 | mag arcsec - 2 |
| ExponentialDisk3D (thick disk) | h | 42.07 | 0.115 | arcsec |
| ExponentialDisk3D (thick disk) | n | 50 | - | |
| ExponentialDisk3D (thick disk) | z 0 | 11.74 | 0.038 | arcsec |
Note . - Results of fitting Spitzer IRAC1 (3.6 µ m) image of the edge-on spiral IC 5176. The first fit uses analytic 2D EdgeOnDisk components (exponential disk seen at i = 90 ◦ ); the second fit uses line-of-sight integration through ExponentialDisk3D components (3D luminosity-density models), for which the inclination i is a free parameter. Size parameters have been converted from pixels to arc seconds; 'AIC' = Akaike Information Criterion values for the two fits. Column 1: Component used in fit. Column 2: Parameter. Column 3: Best-fit value for parameter. Column 4: Uncertainty on parameter value from L-M covariance matrix. Column 5: Units (surface-brightness parameters have been converted from counts/pixel to 3.6 µ mABmagarcsec 2 ; -the µ 0 values for the disk components are equivalent integrated face-on central surface brightnesses).
case: e.g., in the χ 2 d case, n and re are 0.54% and 0.74% too small, while I e is 1.3% too high. In both S/N regimes, the Poisson MLE fits remain unbiased.
What is the effect of adding (Gaussian) read noise to the images? To investigate this, additional sets of images were prepared as before, except that the value from the Poisson process was further modulated by adding a Gaussian with mean = 0 and width σ = 5 e -1 . (This value was chosen as a representative read noise for typical modern CCDs; it is also roughly equal to the dispersion of the Gaussian approximation to the Poisson noise of the background in the low-S/N limit - i.e., σ sky ≈ √ 20.)
The fits were done as before, with the read noise properly included in the χ 2 fitting; the histograms of the resulting fits are shown in Figures 13 and 14 with dashed lines. What is clear from the figure is that while the addition of a Gaussian noise term reduces the bias in the χ 2 fits slightly in the lowS/N regime, the bias is still present. Even though the Poisson MLE approach is no longer formally correct when Gaussian noise is present, the C fits remain unbiased in the presence of moderate read noise.
## 9.2. Quantifying the Bias
How large is the bias produced by χ 2 fits? Humphrey et al. (2009) suggested that the absolute or relative size of the bias might not be as important as the size of the bias relative to the nominal statistical errors of the fits. There are, in principle, three different ways of estimating these errors: from the distribution of the fitted values for all 500 images (similar to what was done by Humphrey et al. for their examples); from the mean of individual-fit error estimates produced by using the L-M algorithm; and from the mean of individualfit error estimates produced by bootstrap resampling. For this simple model, all three approaches produce very similar values. For example, fitting the images in χ 2 d mode with the L-M algorithm produces estimated dispersions within ∼ 10% of the dispersion of values from the individual χ 2 d fits; the latter are in turn very similar to the dispersion of the individual C fits (as is evident from the similar histogram widths in Figure 13). The errors estimated from bootstrap resampling also agree to within ∼ 10% of the other estimates; see Figure 2 for a comparison of bootstrap and L-M error estimates for a fit to a single low-S/N image.
Figure 15 shows the biases for the χ 2 d , χ 2 m , and Poisson MLE fits, plotted against the background value for the different S/N regimes: the top panels show the deviations relative to the true parameter values, while the bottom panels shows the deviations in units of the statistical errors (using the standard deviation of the 500 fitted values). The left and right panels show the cases for χ 2 d and χ 2 m fits, respectively, with the Poisson MLE fits shown in each panel for reference. In all cases, there is a clear trend of the χ 2 biases becoming smaller as the overall exposure level (represented by the mean background level) increases, asymptotically approaching the zero-bias case exhibited by the Poisson MLE fits.
Humphrey et al. (2009) derived an estimate for the bias (relative to the statistical error) that would result from fitting purePoisson data using the χ 2 d statistic, based on the total number of counts Nc and the total number of bins N bins (i.e., the total number of fitted pixels):
$$f _ { b } ( \chi _ { d } ^ { 2 } ) = \frac { | x _ { 0 } - \bar { x } | } { \sigma _ { x } } \sim \frac { N _ { \text{bins} } } { \sqrt { N _ { c } } },$$
where x 0 is the true value, ¯ x is the mean fitted value, and σ x is the statistical error on the parameter value. They did the same for the χ 2 m approach and found
$$f _ { b } ( \chi _ { m } ^ { 2 } ) \sim - 0. 5 \, \frac { N _ { \text{bins} } } { \sqrt { N _ { c } } }.$$
The estimates derived from these equations are plotted as dotted lines in Figure 15. Although the actual biases are systematically smaller than the predictions, the overall agreement is rather good.
## 9.3. Biases in Fitting Images of Real Galaxies
Is there any evidence that the χ 2 -bias effect is significant when fitting images of real galaxies? Figure 16 shows the differences seen when fitting single Sérsic functions to images of
Figure 14. As for Figure 13, but now showing the distribution of the estimated (Sérsic) galaxy luminosity (relative to the true luminosity) from the fits to the model images. In the low-S/N case (left panel), the data-based χ 2 d fits (red) underestimate the true luminosity by 10.4% (9.3% when read noise is present, dashed red histogram), while the model-based χ 2 m fits (magenta) overestimate it by 6.5% (5.8% for the read-noise case); the Poisson MLE fits ( C , blue) are unbiased. These biases diminish in the medium-S/N regime (middle panel) and high-S/N regime (right panel).
Figure 15. Bias in fitted Sérsic parameters, for the low-, medium-, and high-S/N model images (see Figure 13). The bias is plotted versus the background level of the corresponding images. Solid symbols and lines are for parameter values from χ 2 fits to pure-Poisson images (green = n , red = Ie , blue = re ), while semi-filled symbols and dot-dashed lines are from χ 2 fits to images with added read noise. Hollow symbols and dashed lines are from Poisson MLE fits, which show essentially no bias. Top panels : Fractional bias ( ¯ x -x 0 ) / x 0 , where ¯ x is the mean measured parameter value from fits to 500 images and x 0 is the original model value; the left and right panels show χ 2 d and χ 2 m fits, respectively. Bottom panels: Same, but now showing bias relative to statistical error ( ¯ x -x 0 ) / σ x , where σ x is the nominal statistical error from the fit. The upper and lower dotted curves in each panel show the predicted bias from Humphrey et al. (2009), which is based purely on total counts and number of pixels in the images (Equations 43 and 44).
Figure 16. As for the top panels of Figure 15, but now showing relative differences in fitted Sérsic parameters between χ 2 fits and Poisson MLE fits to images of three real elliptical galaxies. Shown is ( x -xC ) / xC , where x is the n re , , or Ie value from a χ 2 fit and xC is the value from a Poisson MLE fit to the same image, plotted against mean sky background for the image. Left panel: Data-based χ 2 fits. Right panel: Model-based χ 2 fits. Solid points are from fits to (in order of increasing background level) SDSS u , g , and r images of NGC 5831; small hollow symbols are from fits to 15s and 60s INT-WFC r -band images of NGC 4697, while larger hollow symbols are from fits to 15s and 40s INT-PFCU V -band images of NGC 3379. Green squares, blue triangles, and red circles indicate Sérsic n , re , and Ie , respectively; solid lines connect results for the same galaxy with different exposure levels.
three elliptical galaxies. In the first case, I fit 996 × 1121-pixel cutouts from SDSS u , g , and r images of NGC 5831; these images correspond to successively higher counts per pixel in both background and galaxy. (The cutouts, as well as the mask images, were shifted in x and y to correct for pointing offsets between the different images.) Although color gradients may produce (genuinely) different fits for the different images, these should be small for an early-type galaxy like NGC 5831; more importantly, the bias estimates I calculate (see next paragraph) are between the χ 2 and Poisson MLE fits for each individual band. In the second and third cases I fit same-filter images with different exposure times: 1801 × 1701-pixel cutouts from short (15s) and long (60s) r -band exposures of NGC 4697, obtained with the Isaac Newton Telescope's Wide Field Camera on 2004 March 17, and 15s and 40s V -band exposures of NGC 3379, obtained with the Prime Focus Camera Unit of the INT on 1994 March 14 (image size = 1243 × 1152 pixels). All images were fit with a single Sérsic function, convolved with an appropriate Moffat PSF image (based on measurements of unsaturated stars in each image). All fits were done with χ 2 d , χ 2 m , and Poisson MLE ( C ) minimization; the χ 2 fits included appropriate readnoise contributions (5.8 e -and 4.5 e -for the WFC and PFCU images, respectively).
Unlike the case for the model images in the preceding section, the 'correct' Sérsic model for these galaxies is unknown (as is, for that matter, the true sky background). Thus, Figure 16 shows the differences between the best χ 2 -fit parameters and the parameters from the Poisson MLE fits, relative to the value of the latter, instead of the difference between all three and the (unknown) 'true' solution. The trends are nonetheless very similar to the model-image case (compare Figure 16 with the top panels of Figure 15): values of n and re from the χ 2 d fits are smaller, and values of I e are larger, than the corresponding values from the Poisson MLE fits, and the offsets are reversed when χ 2 m fitting is done. Although there is some scatter, the tendency of χ 2 d offsets to be larger than χ 2 m offsets is present as well: in fact, the average ratio of the former to the latter is 1.99 (median = 1.65), which is strikingly close to the ratio of 2 predicted by Humphrey et al. (2009). Even the fact that the re offsets are always larger than the n offsets replicates the pattern from the fits to artificial-galaxy images. In addition, the offsets between the χ 2 -fit values and the Poisson MLE fits diminish as the count rate increases, as in the model-image case. If we make the plausible assumption that the higher S/N images are more likely to yield accurate estimates of the true galaxy parameters (to the extent that the galaxies can be approximated by a simple elliptical Sérsic function), then the convergence of estimated parameter values in the high-count regime strongly suggests that the Poisson MLE approach is the least biased in all regimes.
Of course, for many typical optical and near-IR imaging situations, the count rates even in the sky background are high enough that differences between χ 2 and Poisson MLE fits can probably be ignored. For example, typical backgrounds in SDSS g , r , i , and z images range between ∼ 60 and 200 ADU/pixel, or ∼ 300-1000 photoelectrons/pixel. 19 Only for u -band images does the background level become low enough ( ∼ 30-150 photoelectrons/pixel) for the χ 2 -fit bias to become a significant issue.
19 Based on measurements of ∼ 25 SDSS DR7 fields.
## 9.4. The Origins of the Bias
A qualitative explanation for the χ 2 d bias is relatively straightforward. (A more precise mathematical derivation can be found in, e.g., Humphrey et al. 2009.) In the lowcount regime, pixels with downward fluctuations from the true model will have significantly lower σ i values than pixels with similar-sized upward fluctuations; since the weighting for each pixel in the fit is proportional to 1 /σ 2 i , the downward fluctuations will have more weight, and so the best-fitting model will be biased downward.
The χ 2 m bias is slightly more complicated. Here, one has to consider the effects of different possible models being fitted, because the σ i values are determined by the model values, not the data values. In the low-count regime, a model with slightly higher flux than the true model will have higher σ i values, which will in turn lower the total χ 2 . Amodel with lower flux will have smaller σ i values, which will increase the total χ 2 . The overall effect will thus be to bias the best-fitting model upward.
Figure 17 provides a simplified example of how the two forms of bias operate, using a Gaussian + constant background model and a small set of Poisson 'data' points. The original (true) model is shown by the solid gray line, while the dashed red and blue lines show potential models offset above and below the correct model, respectively. The calculated χ 2 d (left) and χ 2 m (right) values for the offset models are also indicated, showing that the downward-offset model has the lowestchisquaredata of the three, while the upward-offset model has the lowest χ 2 m value.
In both cases, the bias is strongest when the mean counts are low, and so for Sérsic fits affects the outer, low-surfacebrightness part of galaxy. In order to accommodate the downward bias of χ 2 d fits, Sérsic models with lower n and smaller re (fainter and steeper outer profiles) are preferred; the I e value increases in compensation to ensure a reasonable fit in the high-surface-brightness center of the galaxy, where the bias is minimal. The opposite trends hold for χ 2 m fits.
## 9.5. Biases (or Lack Thereof) in Other Image-Fitting Software
It is important to note that the χ 2 biases illustrated above apply to specific cases of estimating Gaussian errors from the data or model values on a pixel-by-pixel basis. They do not necessarily apply when an external error map is used in the fitting, unless the error map is itself primarily based on the individual pixel values. Error maps generated in other ways may have little or no effective bias.
For example, the default behavior of GALFIT is to compute an error map from the image by estimating a global background RMS value from the image (after rejecting a fraction of high and low pixel values), combined with per-pixel σ values from a background-subtracted, smoothed version of the data. This means that GALIFT's χ 2 calculation is actually
$$\chi ^ { 2 } \, = \, \sum _ { i = 1 } ^ { N } \frac { ( d _ { i } - m _ { i } ) ^ { 2 } } { d _ { i } ^ { \prime } + \sigma _ { \text{RMS} } ^ { 2 } },$$
where σ RMS is a global value for the image and d i ′ is the background-subtracted, smoothed version of the data; the underlying rationale is that a constant background should have, in the Gaussian approximation, a single σ value (Chien Peng, private communication).
Figure 17. A simplified picture of the origin of the χ 2 d and χ 2 m biases. Both panels shows the same set of data points I ( x ) with Poisson noise generated from a model consisting of a Gaussian plus a constant background (solid gray line). Left panel : Error bars show 1σ Gaussian uncertainties according to the χ 2 d approach ( σ = √ I ). The data points at x = 6 and x = 8 (circled) are equally far from the (true) value ( ∆ I = 2), but the lower ( x = 8) point has a 1 / σ 2 weight almost twice that of the higher ( x = 6) point; this contributes to a lower χ 2 value for a model with a modest downward deviation from the true model (blue dashed curve); an upward-deviating model (red dashed curve) will be an even worse fit. Right panel : Error bars now show uncertainties according the the χ 2 m approach ( σ = √ m ) for the upward-deviating model (red error bars) and the downward-deviating model (blue error bars). The x = 6 and 8 points now have almost equal weights, but the slightly larger error bars in the case of the upward-deviating model help produce a lower χ 2 for that model. The χ 2 values for the three models (fit to all the data points) are shown in the upper-right part of each panel.
Figure 18. Distribution of best-fit Sérsic parameters n and re and total luminosity (relative to the true luminosity) from fits to low-S/N model images; vertical dashed lines show original model values. As in Figures 13 and 14, data-based χ 2 d fits (red histograms) underestimate n , re , and luminosity, while model-based χ 2 m fits (magenta) overestimate them; the Poisson MLE fits ( C , blue) are unbiased. Unweighted χ 2 fits (thick, light gray histograms) are unbiased but less accurate. Finally, fits using GALFIT and its default σ estimation (black) are almost identical to the Poisson MLE IMFIT results. Slight differences in χ 2 d , χ 2 m , and C histograms with respect to those in the 'low-S/N' panels of Figures 13 and 14 are due to using uniform histogram bins within each panel.
This approach has two advantages over the simpler χ 2 d method. First, smoothing the (background-subtracted) data before using it as the basis for σ estimation helps suppress the fluctuations which give rise to the χ 2 d bias, as demonstrated by Churazov et al. (1996) for the case of 1D spectroscopic data. Second, the GALFIT version of the χ 2 statistic is similar in some respects to the so-called modified Neyman's χ 2 :
Hauschild & Jentschel 2001). A hint of this effect can even be seen in the top panel of Figure 13, where the addition of a constant read-noise term to the σ estimation reduces both the χ 2 d and χ 2 m biases.
$$\chi ^ { 2 } _ { N, \text{mod} } \, = \, \sum _ { i = 1 } ^ { N } \frac { ( d _ { i } - m _ { i } ) ^ { 2 } } { \max ( d _ { i }, 1 ) }, \quad \quad ( 4 6 ) \quad \begin{pmatrix} \int _ { \text{fit} } \\ \text{ve} \end{pmatrix}$$
where di is the per-pixel data value. In both cases, the effect of a fixed lower bound to the error term ( σ RMS or 1) is to transition from approximately Poisson weighting of pixels when di or d i ′ is large to equal weighting for pixels when the object counts approach zero (this also removes the problem of pixels with zero counts). This tends to weaken, though not eliminate, the bias in the low-count regime (e.g., Mighell 1999;
Figure 18 shows the distribution of n re , , and total luminosity for Sérsic fits to the low-S/N model images of Section 9.1 for the χ 2 d , χ 2 m , and PMLR ( C ) fits, along with fits to the same images using GALFIT (version 3.0.5). The distributions from fits using GALFIT (black histograms) are almost identical to those from the PMLR fits using IMFIT (blue histograms). A very slight bias in the χ 2 d sense - i.e., underestimation of Sérsic n , re , and luminosity - can still be seen in the GALFIT results, but this is marginal and in any case much smaller than the dispersion of the fits. Similarly, some evidence for the same biases can be seen in the GALFIT simulations of Häussler et al. (2007, their Fig. 5), Hoyos et al. (2011, their Fig. 6), and Davari et al. (2014, their Figs. 4 and 5), but these deviations are only visible at the very lowest S/N levels and are tiny
compared to the overall scatter of best-fit values.
The alert reader may have noticed that the discussion of how the GALFIT approach reduces bias in the fitted parameters implies that unweighted least-squares fits should be unbiased. So would it be better to forgo various σ estimation schemes entirely, and treat all pixels equally? Figure 18 shows that χ 2 fits to the same (simple Sérsic) model images are indeed unbiased when all pixels are weighted (thick gray histograms). But the drawback of completely unweighted fitting is clear in the significantly larger dispersion of fitted results: unweighted fits are less accurate than either the Poisson MLE or GALFIT approaches.
## 10. SUMMARY
I have described a new open-source program, IMFIT, intended for modeling images of galaxies or other astronomical objects. Key features include speed, a flexible user interface, multiple options for handling the fitting process, and the ability to easily add new 2D image functions for modeling galaxy components.
Images are modeled as the sum of one or more 2D image functions, which can be grouped into multiple sets of functions, each set sharing a common location within the image. Available image functions include standard 2D functions used to model galaxies and other objects - e.g., Gaussian, Moffat, exponential, and Sérsic profiles with elliptical isophotes - as well as broken exponentials, analytic edge-on disks, Core-Sérsic profiles, and symmetric (and asymmetric) rings with Gaussian radial profiles. In addition, several sample '3D' functions compute line-of-sight integrations through 3D luminosity-density models, such as an axisymmetric disk with a radial exponential profile and a vertical sech 2 / n profile. Optional convolution with a PSF is accomplished via Fast Fourier Transforms, using a user-supplied FITS image for the PSF.
Image fitting can be done by minimization of the standard χ 2 statistic, using either the image data to estimate the perpixel variances ( χ 2 d ) or the computed model values ( χ 2 m ), or by using user-supplied variance or error maps. Fitting can also be done using Poisson-based maximum-likelihood estimators (Poisson MLE), which are especially appropriate for cases of images with low counts per pixel and low or zero read noise. This includes both the traditional Cash statistic C frequently used in X-ray analysis and an equivalent likelihoodratio statistic (PMLR) which can be used with the fastest (Levenberg-Marquardt) minimization algorithm and can also function as a goodness-of-fit estimator. Other minimization algorithms include the Nelder-Mead simplex method and Differential Evolution. Confidence intervals for fitted parameters can be estimated by the Levenberg-Marquardt algorithm from its internal covariance matrix; they can also be estimated (with any of the minimization algorithms) by bootstrap resampling. The full distribution of parameter values from bootstrap resampling can also be saved to a file for later analysis.
A comparison of fits to artificial images of a simple Sérsicfunction galaxy demonstrates how the χ 2 -bias discussed by Humphrey et al. (2009) manifests itself when fitting images: fits which minimize χ 2 d result in values of the Sérsic parameters n and re (as well as the total luminosity) which are bi- ased low and values of I e which are biased high, while fits which minimize χ 2 m produce smaller biases in the opposite directions; as predicted, these biases decrease, but do not vanish, when the background and source intensity levels increase. Fits using Poisson MLE statistics yield essentially unbiased parameter values; this is true even when Gaussian read noise is present. Sérsic fits to images of real elliptical galaxies with varying exposure times or background levels show evidence for the same pattern of biased parameter values when minimizing χ 2 d or χ 2 m . This suggests that the fitting of galaxy images with IMFIT should generally use Poisson MLE minimization instead of χ 2 minimization whenever possible, especially when the background level is less than ∼ 100 photoelectrons/pixel.
Precompiled binaries, documentation, and full source code (released under the GNU Public License) are available at the following web site:
http://www.mpe.mpg.de/˜erwin/code/imfit/.
Various useful comments and suggestions have come from Maximilian Fabricius, Martin Kümmel, and Roberto Saglia, and thanks are also due to Michael Opitsch and Michael Williams for being (partly unwitting) beta testers. Further bug reports, suggestions, requests, and fixes from Giulia Savorgnan, Guillermo Barro, Sergio Pascual, and (especially) André Luiz de Amorim are gratefully acknowledged. I also thank the referee, Chien Peng, for a very careful reading and pertinent questions which considerably improved this paper. This work was partly supported by the Deutsche Forschungsgemeinschaft through Priority Programme 1177, 'Witnesses of Cosmic History: Formation and evolution of galaxies, black holes, and their environment.'
This work is based in part on observations made with the Spitzer Space Telescope, obtained from the NASA/IPAC Infrared Science Archive, both of which are operated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with the National Aeronautics and Space Administration. This paper also makes use of data obtained from the Isaac Newton Group Archive which is maintained as part of the CASU Astronomical Data Centre at the Institute of Astronomy, Cambridge.
Funding for the creation and distribution of the SDSS Archive has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Aeronautics and Space Administration, the National Science Foundation, the U.S. Department of Energy, the Japanese Monbukagakusho, and the Max Planck Society. The SDSS Web site is http://www.sdss.org/ .
The SDSS is managed by the Astrophysical Research Consortium (ARC) for the Participating Institutions. The Participating Institutions are The University of Chicago, Fermilab, the Institute for Advanced Study, the Japan Participation Group, The Johns Hopkins University, the Korean Scientist Group, Los Alamos National Laboratory, the Max-PlanckInstitute for Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico State University, University of Pittsburgh, University of Portsmouth, Princeton University, the United States Naval Observatory, and the University of Washington.
## APPENDIX
## COMPARING LEVENBERG-MARQUARDT AND BOOTSTRAP ESTIMATES OF PARAMETER UNCERTAINTIES FOR IMAGE FITS
Parameter Estimates for Fits to PGC 35772
Table 3 lists the best-fit parameters for three progressively more complex models of the spiral galaxy PGC 35772 (see Section 8.1 and Figure 10), along with both the Levenberg-Marquardt (L-M) uncertainties and the uncertainties derived from 500 rounds of bootstrap resampling (the latter listed in parentheses after the L-M uncertainties). A comparison of the two types of uncertainty estimates suggests they are similar in size for the simplest model (fitting the galaxy with just a Sérsic component), with mean and median values of σ bootstrap /σ LM = 1 38 and 0.81, respectively. . However, the bootstrap uncertainties are typically about half the size of the L-M uncertainties for the more complex models: the mean and median values for the uncertainty ratios are 0.48 and 0.51, respectively, for the Sérsic + Exponential model and 0.61 and 0.41 for the Sérsic + GaussianRing + Exponential model.
## Parameter Estimates for Multiple Exposures of Elliptical Galaxies
In Section 9.3 I compared different χ 2 fits with Poisson MLE fits for several elliptical galaxies. In this section, I compare L-M and bootstrap parameter error estimates for two of the same elliptical galaxies (plus a third observed under similar conditions), always using Poisson MLE fits in order to avoid χ 2 bias effects. Specifically, I compare best-fit parameters from Sérsic fits to multiple images of the same galaxy, in two ways.
First, I compare best-fit parameter values x (e.g., Sérsic index n ) from fits to short (15s) exposures with values from fits to longer (2 × 40s or 1 × 60s) exposures with the same telescope + filter system on the same night. I do this by comparing differences in parameter values ∆ x = x long -x short with the error estimates for the same parameter σ x from the short exposures (left panel of Figure 19). This can be thought of as a crude answer to the question: how well do the error estimates describe the uncertainty of parameters from short-exposure fits relative to more 'correct' parameters obtained from higher S/N data? (I do not compare I e values because these can vary due to changes in the transparency between exposures; similarly, I do not compare values for the pixel coordinates of the galaxy center because these depend on the telescope pointing and are intrinsically variable.)
To first order, if the uncertainty estimates were reasonably accurate we should expect ∼ 68% of the points to be found at σ 15s / | ∆ x | < 1 and ∼ 32% to be at σ 15s / | ∆ x | > 1. As the left panel of the figure shows, neither approach is ideal, but the bootstrapσ estimates are somewhat better: 50% of those are > 1, while this is true for only 1/8 of the deviations if the L-M σ estimates are used.
The second approach, seen in the right-hand panel of Figure 19, is to compare how well the differences in parameter values obtained from similar samples (i.e., multiple images of the same galaxy with the same exposure time) compare with the error estimates. This is, in a limited sense, a test of the nominal frequentist meaning of confidence intervals: how often do repeated measurements fall within the specified error bounds? In this case, I am comparing parameters from fits to the two 40s V -band exposures of NGC 3379 (squares) and also parameters from fits to three 15s R -band exposures of the lower-luminosity elliptical galaxy NGC 3377 (diamonds), also from the INT-WFC. Again, we should expect ∼ 68% of the points to lie within ± 1 if the error estimates are accurate; as the figure shows, essentially all the bootstrap and L-M estimates lie inside this range and so tend to be too small, particularly for re . The bootstrap estimates do a better job in the case of NGC 3379 and a worse job in the case of NGC 3377.
## Summary
The implication of the preceding subsections is that the L-M and bootstrap estimates of parameter errors are very roughly consistent with each other, though there is some evidence that the latter tend to become smaller than the former as the fitted models become more complex (i.e., more components). In general, both the L-M and bootstrap estimates should probably be considered underestimates of the true parameter uncertainties, something already established for L-M estimates from tests of other image fitting programs (e.g., Häussler et al. 2007).
## REFERENCES
Akaike, H. 1974, IEEE T. Automat. Contr., 19, 716
Athanassoula, E., Morin, S., Wozniak, H., Puy, D., Pierce, M. J., Lombard, J., & Bosma, A. 1990, MNRAS, 245, 130
Boroson, T. 1981, ApJS, 46, 177
Box, M. J. 1965, The Computer Journal, 8, 42
Bureau, M., Aronica, G., Athanassoula, E., Dettmar, R.-J., Bosma, A., & Freeman, K. C. 2006, MNRAS, 370, 753
Burnham, K. P. & Anderson, D. E. 2002, Model Selection and Multimodal Inference: A Practical Information-Theoretic Approach, 2nd Ed. (New York:
Springer-Verlag)
Burstein, D. 1979, ApJ, 234, 435
Byun, Y. I. & Freeman, K. C. 1995, ApJ, 448, 563
Capaccioli, M., Held, E. V., & Nieto, J.-L. 1987, AJ, 94, 1519
Cappellari, M. 2002, MNRAS, 333, 400
Cash, W. 1979, ApJ, 228, 939
Chung, A. & Bureau, M. 2004, AJ, 127, 3192
Churazov, E., Gilfanov, M., Forman, W., & Jones, C. 1996, ApJ, 471, 673
Ciotti, L. & Bertin, G. 1999, A&A, 352, 447
Comerón, S., Elmegreen, B. G., Knapen, J. H., Salo, H., Laurikainen, E., Laine, J., Athanassoula, E., Bosma, A., Sheth, K., Regan, M. W., Hinz, J. L., Gil de
Paz, A., Menéndez-Delmestre, K., Mizusawa, T., Muñoz-Mateos, J.-C., Seibert, M., Kim, T., Elmegreen, D. M., Gadotti, D. A., Ho, L. C., Holwerda, B. W., Lappalainen, J., Schinnerer, E., & Skibba, R. 2011, ApJ, 741, 28
Figure 19. Comparison of different estimates of parameter uncertainties for Sérsic fits to low- and medium-S/N images of elliptical galaxies. Left: Estimated parameter uncertainties σ 15s from fits to 15s images of NGC 3379 ( V -band, INT-PFCU; squares) and NGC 4697 ( -band, INT-WFC; circles), divided by the r difference in fitted parameter values between those fits ( x 15s ) and fits to longer exposures ( x long = mean of two 40s exposures for NGC 3379, single 60s exposure for NGC 4697). Black = using σ from Levenberg-Marquardt covariance-matrix; red = using σ from bootstrap resampling analysis. The general trend is for the σ values to underestimate the deviances; the L-M estimates are always somewhat worse in this sense. Right: Estimated uncertainties from fits to two 40s V -band images of NGC 3379 (squares) and three 15s R -band images of NGC 3377 (INT-WFC; diamonds), divided by the difference in parameter values between the individual fits to the same galaxy (e.g., σ n / ( n 2 -n 1 )); colors as for the left panel. For NGC 3379, the estimated errors from fits to the two images are very similar and only average values are used; for NGC 3377, all combinations of errors and differences from pairs of images are used.
Davari, R., Ho, L. C., Peng, C. Y., & Huang, S. 2014, ApJ, 787, 69
de Grijs, R., Peletier, R. F., & van der Kruit, P. C. 1997, A&A, 327, 966
de Jong, R. S. 1996, A&AS, 118, 557
de Souza, R. E., Gadotti, D. A., & dos Anjos, S. 2004, ApJS, 153, 411
de Vaucouleurs, G., de Vaucouleurs, A., Corwin, H. G., Buta, R. J., Paturel, G., & Fouqué, P. 1993, Third Reference Catalog of Bright Galaxies (New York:
Springer-Verlag)
Dolphin, A. E. 2002, MNRAS, 332, 91
Dullo, B. T. & Graham, A. W. 2012, ApJ, 755, 163
-. 2013, ApJ, 768, 36
Efron, B. 1979, Ann.Statist., 7, 1
Emsellem, E., Monnet, G., & Bacon, R. 1994, A&A, 285, 723
Erwin, P., Beckman, J. E., & Pohlen, M. 2005, ApJL, 626, L81
Erwin, P., Pohlen, M., & Beckman, J. E. 2008, AJ, 135, 20
Fabricius, M. H., Coccato, L., Bender, R., Drory, N., Gössl, C., Landriau, M., Saglia, R. P., Thomas, J., & Williams, M. J. 2014, MNRAS, 441, 2212
Ferrarese, L., Côté, P., Jordán, A., Peng, E. W., Blakeslee, J. P., Piatek, S., Mei, S., Merritt, D., Milosavljevi´, M., Tonry, J. L., & West, M. J. 2006, ApJS, 164, c 334
Freeman, P., Doe, S., & Siemiginowska, A. 2001, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 4477, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, ed. J.-L. Starck & F. D. Murtagh, 76-87
Frigo, M. & Johnson, S. G. 2005, Proceedings of the IEEE, 93, 216, special issue on 'Program Generation, Optimization, and Platform Adaptation'
Fukugita, M., Nakamura, O., Okamura, S., Yasuda, N., Barentine, J. C., Brinkmann, J., Gunn, J. E., Harvanek, M., Ichikawa, T., Lupton, R. H., Schneider, D. P.,
Strauss, M. A., & York, D. G. 2007, AJ, 134, 579
Gadotti, D. A. 2008, MNRAS, 384, 420
García-Benito, R., Zibetti, S., Sánchez, S. F., Husemann, B., de Amorim, A. L., Castillo-Morales, A., Cid Fernandes, R., . Ellis, S. C., Falcón-Barroso, J., Galbany, L., & et al. 2014, A&A, submitted
Graham, A. W., Erwin, P., Trujillo, I., & Asensio Ramos, A. 2003, AJ, 125, 2951
Gutiérrez, L., Erwin, P., Aladro, R., & Beckman, J. E. 2011, AJ, 142, 145
Hauschild, T. & Jentschel, M. 2001, Nucl. Instrum. Methods Phys. Res., 457, 384
Häussler, B., McIntosh, D. H., Barden, M., Bell, E. F., Rix, H.-W., Borch, A., Beckwith, S. V. W., Caldwell, J. A. R., Heymans, C., Jahnke, K., Jogee, S.,
Koposov, S. E., Meisenheimer, K., Sánchez, S. F., Somerville, R. S., Wisotzki, L., & Wolf, C. 2007, ApJS, 172, 615
Hickson, P., Auman, J. R., Ninkov, Z., Fahlman, G. G., Walker, G. A. H., & Menon, T. K. 1982, ApJ, 258, 53
Hoyos, C., den Brok, M., Kleijn, G. V., Carter, D., Balcells, M., Guzmán, R., Peletier, R., Ferguson, H. C., Goudfrooij, P., Graham, A. W., Hammer, D., Karick,
A. M., Lucey, J. R., Matkovi´ c, A., Merritt, D., Mouhcine, M., & Valentijn, E. 2011, MNRAS, 411, 2439
Humphrey, P. J., Liu, W., & Buote, D. A. 2009, ApJ, 693, 822
Kormendy, J. 1977, ApJ, 217, 406
Kümmel, M., Koppenhoefer, J., Riffeser, A., Mohr, J., Desai, S., Henderson, R., Paech, K., & Wetzstein, M. 2013, in Astronomical Society of the Pacific
Conference Series, Vol. 475, Astronomical Data Analysis Software and Systems XXII, ed. D. N. Friedel, 357
Levenberg, K. 1944, Quarterly J. Applied Mathematics, 2, 164
Liddle, A. R. 2007, MNRAS, 377, L74
Llacer, J. & Nuñéz, J. 1991, in The Restoration of HST Images and Spectra, ed. R. L. White & R. J. Allen, 62
MacArthur, L. A., Courteau, S., & Holtzman, J. A. 2003, ApJ, 582, 689
Markwardt, C. B. 2009, in Astronomical Society of the Pacific Conference Series, Vol. 411, Astronomical Data Analysis Software and Systems XVIII, ed. D. A. Bohlender, D. Durand, & P. Dowler, 251
Marquardt, D. 1963, J. SIAM, 11, 431
Mighell, K. J. 1999, ApJ, 518, 380
Moffat, A. F. J. 1969, A&A, 3, 455
Monnet, G., Bacon, R., & Emsellem, E. 1992, A%A, 253, 366
Moré, J. J. 1978, in Lecture Notes in Mathematics, Vol. 630, Numerical Analysis, ed. G. Watson (Springer Berlin Heidelberg), 105-116
Muñoz-Mateos, J. C., Sheth, K., Gil de Paz, A., Meidt, S., Athanassoula, E., Bosma, A., Comerón, S., Elmegreen, D. M., & Elmegreen, B. G. 2013, ApJ, 771, 59
Nelder, J. A. & Mead, R. 1965, The Computer Journal, 7, 308
Nousek, J. A. & Shue, D. R. 1989, ApJ, 342, 1207
Nuñéz, J. & Llacer, J. 1993, PASP, 105, 1192
## Imfit: Galaxy Image Fitting
Pawitan, Y. 2001, In All Likelihood: Statistical Modelling and Inference Using Likelihood (OUP Oxford)
Peng, C. Y., Ho, L. C., Impey, C. D., & Rix, H.-W. 2002, AJ, 124, 266
-. 2010, AJ, 139, 2097
Pohlen, M., Balcells, M., Lütticke, R., & Dettmar, R.-J. 2004, A&A, 422, 465
Pohlen, M., Dettmar, R., Lütticke, R., & Schwarzkopf, U. 2000, A&AS, 144, 405
Press, W. H., Teukolsky, S. A., Vetterling, W. T., & Flannery, B. P. 1992, Numerical Recipes in C: The Art of Scientific Computing, 2nd Ed. (Cambridge:
Cambridge U. Press)
Pritchet, C. & Kline, M. I. 1981, AJ, 86, 1859
Richings, A. J., Uttley, P., & Körding, E. 2011, MNRAS, 415, 2158
Rusli, S. P., Erwin, P., Saglia, R. P., Thomas, J., Fabricius, M., Bender, R., & Nowak, N. 2013, AJ, 146, 160
Saglia, R. P., Bertschinger, E., Baggley, G., Burstein, D., Colless, M., Davies, R. L., McMahan, Jr., R. K., & Wegner, G. 1993, MNRAS, 264, 961
Schwarz, G. 1978, Ann. Statist., 6, 461
Scorza, C. & Bender, R. 1990, A&A, 235, 49
Send, U. 1982, A&A, 112, 235
Sérsic, J. L. 1968, Atlas de galaxias australes
Shaw, M. A. & Gilmore, G. 1989, MNRAS, 237, 903
Sheth, K., Regan, M., Hinz, J. L., Gil de Paz, A., Menéndez-Delmestre, K., Muñoz-Mateos, J.-C., Seibert, M., Kim, T., Laurikainen, E., Salo, H., & et al. 2010, PASP, 122, 1397
Simard, L. 1998, in Astronomical Society of the Pacific Conference Series, Vol. 145, Astronomical Data Analysis Software and Systems VII, ed. R. Albrecht, R. N. Hook, & H. A. Bushouse, 108
Simard, L., Willmer, C. N. A., Vogt, N. P., Sarajedini, V. L., Phillips, A. C., Weiner, B. J., Koo, D. C., Im, M., Illingworth, G. D., & Faber, S. M. 2002, ApJS, 142, 1
Spitzer, Jr., L. 1942, ApJ, 95, 329
Storn, R. & Price, K. 1997, Journal of Global Optimization, 11, 314
Takeuchi, T. T. 2000, Ap&SS, 271, 213
Trujillo, I., Aguerri, J. A. L., Cepa, J., & Gutiérrez, C. M. 2001, MNRAS, 328, 977
Trujillo, I., Erwin, P., Asensio Ramos, A., & Graham, A. W. 2004, AJ, 127, 1917
Tsikoudi, V. 1979, ApJ, 234, 842
-. 1980, ApJS, 43, 365
Tully, R. B., Rizzi, L., Shaya, E. J., Courtois, H. M., Makarov, D. I., & Jacobs, B. A. 2009, AJ, 138, 323
van der Kruit, P. C. 1988, A&A, 192, 117
van der Kruit, P. C. & Searle, L. 1981, A&A, 95, 105
Wald, A. 1943, Trans. Amer. Math. Soc., 43, 426
Wilks, S. S. 1938, Ann. Math. Statist., 9, 60
Xilouris, E. M., Alton, P. B., Davies, J. I., Kylafis, N. D., Papamastorakis, J., & Trewhella, M. 1998, A&A, 331, 894
Xilouris, E. M., Byun, Y. I., Kylafis, N. D., Paleologou, E. V., & Papamastorakis, J. 1999, A&A, 344, 868
Xilouris, E. M., Kylafis, N. D., Papamastorakis, J., Paleologou, E. V., & Haerendel, G. 1997, A&A, 325, 135
Yoachim, P. & Dalcanton, J. J. 2006, AJ, 131, 226
|
10.1088/0004-637X/799/2/226
|
[
"Peter Erwin"
] |
2014-08-05T20:01:50+00:00
|
2014-12-03T20:07:40+00:00
|
[
"astro-ph.IM",
"astro-ph.GA"
] |
Imfit: A Fast, Flexible New Program for Astronomical Image Fitting
|
I describe a new, open-source astronomical image-fitting program called
Imfit, specialized for galaxies but potentially useful for other sources, which
is fast, flexible, and highly extensible. A key characteristic of the program
is an object-oriented design which allows new types of image components (2D
surface-brightness functions) to be easily written and added to the program.
Image functions provided with Imfit include the usual suspects for galaxy
decompositions (Sersic, exponential, Gaussian), along with Core-Sersic and
broken-exponential profiles, elliptical rings, and three components which
perform line-of-sight integration through 3D luminosity-density models of disks
and rings seen at arbitrary inclinations.
Available minimization algorithms include Levenberg-Marquardt, Nelder-Mead
simplex, and Differential Evolution, allowing trade-offs between speed and
decreased sensitivity to local minima in the fit landscape. Minimization can be
done using the standard chi^2 statistic (using either data or model values to
estimate per-pixel Gaussian errors, or else user-supplied error images) or
Poisson-based maximum-likelihood statistics; the latter approach is
particularly appropriate for cases of Poisson data in the low-count regime. I
show that fitting low-S/N galaxy images using chi^2 minimization and
individual-pixel Gaussian uncertainties can lead to significant biases in
fitted parameter values, which are avoided if a Poisson-based statistic is
used; this is true even when Gaussian read noise is present.
|
1408.1098v2
|
Nuclear Physics A
Nuclear Physics A 00 (2014) 1-4
## Probing the non-equilibrium dynamics of hot and dense QCD with dileptons
Gojko Vujanovic a , Jean-Franc ¸ois Paquet , Gabriel S. Denicol , Matthew Luzum a a a,b , Bj¨ orn Schenke c , Sangyong Jeon a , Charles Gale a a Department of Physics, McGill University, 3600 rue University, Montr´ eal, Qu´ ebec H3A 2T8, Canada b Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA c Physics Department, Brookhaven National Lab, Building 510A, Upton, NY, 11973, USA
## Abstract
It is argued that, in heavy ion collisions, thermal dileptons are good probes of the transport properties of the medium created in such events, and also of its early-time dynamics, usually inaccessible to hadronic observables. In this work we show that electromagnetic azimuthal momentum anisotropies do not only display a sensitivity to the shear relaxation time and to the initial shear-stress tensor profile, but also to the temperature dependence of the shear viscosity coe GLYPH<14> cient.
## 1. Introduction
One of the main goals of Relativistic Heavy Ion Colliders, either the Relativistic Heavy Ion Collider (RHIC, at Brookhaven National Laboratory) or the Large Hadron Collider (LHC, at CERN), is to investigate the thermodynamic and transport properties of the hot and dense phase of QCD. Much work has been concentrated on the determination of an e GLYPH<11> ective value of the shear viscosity coe GLYPH<14> cient from analyses of relativistic heavy-ion collisions but so far, such investigations haven been performed mostly by comparing to hadrons produced at the final stages of the collision.
Electromagnetic radiation constitutes a class of complementary and penetrating probes that are sensitive to the entire space-time history of nuclear collisions including its very early stages. In this contribution we show that thermal dileptons are a GLYPH<11> ected by the transport properties of the fluid and by the non-equilibrium aspects of the initial state that are usually inaccessible to hadronic probes. We establish that the azimuthal momentum anisotropies of thermal dileptons are particularly sensitive to the temperature dependence of the shear viscosity coe GLYPH<14> cient. We also show the potential of thermal dileptons in di GLYPH<11> erentiating between possible initial shear-stress tensors and shear relaxation times.
## 2. Fluid-dynamical model
We will discuss only Au-Au collisions at p sNN = 200 GeV. The time evolution of the hot and dense medium created at RHIC is modeled using music , a 3 + 1D hydrodynamical evolution [1]. The main equations of motion are the conservation laws of energy and momentum, given by the continuity equation for the energy momentum tensor, T GLYPH<22>GLYPH<23> , i.e., @GLYPH<22> T GLYPH<22>GLYPH<23> = 0. As usual, T GLYPH<22>GLYPH<23> = " u u GLYPH<22> GLYPH<23> GLYPH<0> GLYPH<1> GLYPH<22>GLYPH<23> P + GLYPH<25> GLYPH<22>GLYPH<23> , with " being the energy density, P the thermodynamic pressure, u GLYPH<22> the fluid four-velocity, GLYPH<25> GLYPH<22>GLYPH<23> the shear-stress tensor, and GLYPH<1> GLYPH<22>GLYPH<23> = g GLYPH<22>GLYPH<23> GLYPH<0> u u GLYPH<22> GLYPH<23> the projection operator onto the 3-space orthogonal to the velocity, with a metric tensor g GLYPH<22>GLYPH<23> = diag(1 ; GLYPH<0> 1 ; GLYPH<0> 1 ; GLYPH<0> 1) : The lattice QCD equation of state is used to relate P and " [2].
The conservation laws are complemented by a relaxation equation for the shear-stress tensor, given by a version of Israel-Stewart (I-S) theory [3, 4],
$$\tau _ { \pi } \Delta _ { \alpha \beta } ^ { \mu \nu } u ^ { \lambda } \partial _ { \lambda } \pi ^ { \alpha \beta } + \frac { 4 } { 3 } \tau _ { \pi } \pi ^ { \mu \nu } \partial _ { \lambda } u ^ { \lambda } = \left ( \pi _ { \text{NS} } ^ { \mu \nu } - \pi ^ { \mu \nu } \right ),$$
where GLYPH<25> GLYPH<22>GLYPH<23> NS = 2 GLYPH<17> GLYPH<27> GLYPH<22>GLYPH<23> = 2 GLYPH<17> GLYPH<1> GLYPH<22>GLYPH<23> GLYPH<11>GLYPH<12> @ GLYPH<11> u GLYPH<12> is the Navier-Stokes limit of the shear-stress tensor, with GLYPH<1> GLYPH<22>GLYPH<23> GLYPH<11>GLYPH<12> = GLYPH<1> GLYPH<16> GLYPH<22> GLYPH<11> GLYPH<1> GLYPH<23> GLYPH<12> + GLYPH<1> GLYPH<22> GLYPH<12> GLYPH<1> GLYPH<23> GLYPH<11> GLYPH<17> = 2 GLYPH<0> GLYPH<1> GLYPH<11>GLYPH<12> GLYPH<1> GLYPH<22>GLYPH<23> = 3 being the double, symmetric, traceless projection operator. In its simplest from, I-S theory has two transport coe GLYPH<14> cients: the shear viscosity GLYPH<17> , also present in Navier-Stokes theory, and the shear relaxation time, GLYPH<28> GLYPH<25> , which only exists in I-S theory.
We use a constant value GLYPH<17>= s = 1 4 = GLYPH<25> as the default value for the shear viscosity over entropy density ratio. In the QGP phase, i.e. for temperatures above a transition temperature T tr = 0 18 GeV, we will also consider an : GLYPH<17>= s with linear temperature dependences of the form GLYPH<17>= s T ( ) = a T ( = T tr GLYPH<0> 1) + 1 4 = GLYPH<25> . The e GLYPH<11> ect of the temperature dependence of GLYPH<17>= s on hadronic and eletromagnetic flow observables is tested by modifying the slope parameter a . The values of a employed in this work are a = 0, 0 2427, and 0 5516, with : : a = 0 corresponding to the constant default value.
The shear relaxation time is assumed to be of the form GLYPH<28> GLYPH<25> = b GLYPH<25> GLYPH<17>= ( " + P ). The role of GLYPH<28> GLYPH<25> is to govern the rate at which GLYPH<25> GLYPH<22>GLYPH<23> evolves and relaxes towards its Navier-Stokes limit. The default value used in this study is b GLYPH<25> = 5. Here, we test the e GLYPH<11> ect of larger relaxation times by also considering b GLYPH<25> = 10 and 20.
The initial energy density profile is determined by the Monte-Carlo Glauber model, with all the free parameters being tuned to describe the multiplicity and elliptic flow of hadronic observables at RHIC's highest energy. The initial value of the shear-stress tensor is also varied in this work and is parametrized in the following way GLYPH<25> GLYPH<22>GLYPH<23> 0 = c GLYPH<2> 2 GLYPH<17>GLYPH<27> GLYPH<22>GLYPH<23> . The parameter c controls the deviation of the initial state from local thermodynamic equilibrium. Here, we set c = 0, 1 2, and 1, with the default value being zero. The initial velocity profile is always set to zero in hyperbolic coordinates. /
## 3. Thermal dilepton rates
Thermal dilepton rates can generically be expressed as:
$$\frac { d ^ { 4 } R } { d ^ { 4 } q } = - \frac { \alpha } { 1 2 \pi ^ { 4 } } \frac { 1 } { M ^ { 2 } } \frac { \text{Im} \Pi _ { \gamma ^ { \prime } } ^ { R } } { e ^ { d ^ { \prime } / T } - 1 }$$
where GLYPH<11> is the electromagnetic structure constant, GLYPH<5> R GLYPH<13> GLYPH<3> = GLYPH<5> R ; GLYPH<22> GLYPH<13> GLYPH<3> ; GLYPH<22> is the trace of the retarded virtual photon self-energy, and M 2 = q 2 , where M is the virtual photon invariant mass. This expression is valid at leading order in GLYPH<11> em, but is exact at all orders of GLYPH<11> s [5]. We have used the Born rate in this paper, which corresponds to the quark-antiquark annihilation rate into dileptons. Viscosity is included via a deviation of the thermal distribution functions n entering in evaluating GLYPH<5> R GLYPH<13> GLYPH<3> such that n ! n + GLYPH<14> n , where GLYPH<14> n p ( ) = G p n p ( ) ( )(1 GLYPH<6> n p ( )) p p GLYPH<22> GLYPH<23> GLYPH<25> GLYPH<22>GLYPH<23> = h 2 T 2 ( " + P ) i , and G p ( ) is a function that must be determined through the use of microscopic physics. In order to find G p ( ), we solved the Boltzmann equation assuming a massless gas of particles with constant 2 ! 2 cross section.
The general form of the thermal dilepton rates Eq. (2) can be applied in the hadronic sector (low temperatures) via the introduction of the Vector Dominance Model (VDM). Through VDM, GLYPH<5> R GLYPH<13> GLYPH<3> is expressed in terms of D V R , the vector meson ( V ) retarded propagator. A key ingredient in evaluating D V R is the vector meson self-energy GLYPH<5> V , the latter being presented in detail in Ref. [6].
## 4. Results
In the left panels of Figures 1, 2, and 3, we show the di GLYPH<11> erential elliptic flow of charged hadrons as a function of transverse momentum, v 2 ( pT ). In the right panel of the same figures we show the integrated elliptic flow of thermal dileptons as a function of their invariant mass, v 2 ( M ). In Figure 1 GLYPH<17>= s was varied, while in Figures 2 and 3 GLYPH<25> GLYPH<22>GLYPH<23> 0 and GLYPH<28> GLYPH<25> were changed, respectively. In each case, the parameters that are not varied are kept at their default values. For each parameter configuration, we computed 200 events, all in the 20-40% centrality class. The color bands in the plots indicate the statistical uncertainties of the calculations. We note that our results for charged hadron v 2 ( pT ) are in good agreement with PHENIX data, which corresponds to the points in the left panels of our figures.
Figure 1. The e GLYPH<11> ects of varying GLYPH<17> s ( T ) in the QGP phase on the elliptic flow of charged hadrons (left panel) and of virtual photons (right panel), in collisions of Au + Au at 200 A GeV, at a 20-40% centrality class.
It was already shown in Ref. [7] that charged hadrons have a small sensitivity to the GLYPH<17>= s T ( ) in the QGP phase at the top RHIC energy. The left panel of Figure 1 illustrates that this behavior still holds true for the temperature dependence of GLYPH<17>= s used in this study. In addition, the results plotted in the left panels of Figures 2 and 3 confirm that the elliptic flow of charged hadrons at RHIC's highest energy has a very small sensitivity also to variations of GLYPH<25> GLYPH<22>GLYPH<23> 0 and of GLYPH<28> GLYPH<25> . Even though it is not shown here, we verified that the same is true for the transverse momentum spectra of charged hadrons.
The situation is not the same when it comes to thermal dileptons. The e GLYPH<11> ect of varying GLYPH<25> GLYPH<22>GLYPH<23> 0 and GLYPH<28> GLYPH<25> is visible on v 2 ( M ), but it is still relatively modest, as seen in the right panels of Figures 2 and 3. However, the magnitude of the slope of GLYPH<17>= s as a function of T has a sizeable influence on the elliptic flow of thermal dileptons; this is shown in the right panel of Figure 1. Recall that, unlike the charged hadrons emitted at the freeze-out hyper-surface, thermal dileptons are emitted throughout the collision history and their elliptic flow retains a memory of the GLYPH<17>= s T ( ) in the QGP phase. For small and intermediate values of invariant mass ( M < 1 2 GeV), increasing the QGP's : GLYPH<17>= s T ( ) leads to an increase in v 2 ( M ). Meanwhile, for larger values of invariant mass ( M > 1 2 GeV) the behavior is inverted and a larger : GLYPH<17>= s T ( ) leads to a smaller elliptic flow coe GLYPH<14> cient. We have found that this inversion also occurs for the momentum anisotropy, GLYPH<15> p = h T xx GLYPH<0> T yy i = h T xx + T yy i , when plotted against time.
Figure 2. The e GLYPH<11> ects of varying GLYPH<25> GLYPH<22>GLYPH<23> 0 on charged hadron's (left panel) and virtual photon's (right panel) elliptic flow created in collisions of Au + Au at 200 A GeV, in the 20-40% centrality class.
The change in the hydrodynamical evolution induced by a T-dependent GLYPH<17>= s is occuring far from the freeze-out surface and therefore is only accessible to electromagnetic probes. At freeze-out (for collisions at RHIC energies), most of the memory of di GLYPH<11> erent values of GLYPH<17>= s in the QGP phase has faded: the charged hadrons v 2 is thus mostly una GLYPH<11> ected (see left panel of Figure 1 and Ref. [7]).
ch
2
v
0.25
0.2
0.15
0.1
0.05
0
0
PHENIX (20%-40%)
τπ
τπ
=5
/(
+P)
τπ
η ε
=10
η ε
/(
+P)
=20
η ε
/(
+P)
0.5
1
1.5
2
2.5
pT [GeV]
Figure 3. The e GLYPH<11> ects of varying GLYPH<28> GLYPH<25> on charged hadron's (left panel) and virtual photon's (right panel) elliptic flow created in collisions of Au + Au at 200 A GeV, in the 20-40% centrality class.
## 5. Conclusions
In this contribution, we showed that thermal dileptons are a GLYPH<11> ected by the transport properties of the QGP and by non-equilibrium aspects of the initial evolution that are usually inaccessible to hadronic probes. For the first time, we explicitly demonstrate that the invariant mass distribution of dileptons and their azimuthal momentum anisotropy have a small but non-negligible dependence on the magnitude of the shear relaxation time and on the value of initial shear-stress tensor. Importantly, virtual photons may also reveal the temperature dependence of the shear viscosity coe GLYPH<14> cient. This endeavor rea GLYPH<14> rms the potential that penetrating probes, such as dileptons, have in furthering our understanding of QCD at high temperatures and densities. We expect that, as experimental uncertainties become smaller, such probes will play a more dominant role in the extraction of the initial state and transport properties of the bulk QCD matter created in ultrarelativistic heavy ion collisions at RHIC and at the LHC.
## 6. Acknowledgements
This work was supported in part by the Natural Sciences and Engineering Research Council of Canada, in part by U. S. DOE Contract No. DE-AC02-98CH10886, and in part by the by the Director, O GLYPH<14> ce of Energy Research, O GLYPH<14> ce of High Energy and Nuclear Physics, Division of Nuclear Physics, of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231. G. Vujanovic acknowledges support by the Canadian Institute for Nuclear Physics, and G.S. Denicol acknowledges support through a Banting Fellowship from the Government of Canada. Computations were performed on the Guillimin supercomputer at McGill University under the auspices of Calcul Qu´ ebec and Compute Canada. The operation of Guillimin is funded by the Canada Foundation for Innovation (CFI), the National Science and Engineering Research Council (NSERC), NanoQu´ ebec, and the Fonds Qu´ eb´ ecois de Recherche sur la Nature et les Technologies (FQRNT).
## References
- [1] B. Schenke, S. Jeon, C. Gale, Phys.Rev.Lett. 106 (2011) 042301.
- [2] P. Huovinen, P. Petreczky, Nucl.Phys. A837 (2010) 26-53.
- [3] W. Israel, Annals of Physics 100 (12) (1976) 310 - 331.
- [4] W. Israel, J. Stewart, Annals Phys. 118 (1979) 341-372.
- [5] J. I. Kapusta, C. Gale, Finite-temperature field theory: Principles and applications , Cambridge, UK, University Press.
- [6] G. Vujanovic, C. Young, B. Schenke, R. Rapp, S. Jeon, et al., Phys. Rev. C 89 (2014) 034904.
- [7] H. Niemi, G. S. Denicol, P. Huovinen, E. Molnar, D. H. Rischke, Phys.Rev.Lett. 106 (2011) 212302.
γ *
|
10.1016/j.nuclphysa.2014.08.078
|
[
"Gojko Vujanovic",
"Jean-Francois Paquet",
"Gabriel S. Denicol",
"Matthew Luzum",
"Bjoern Schenke",
"Sangyong Jeon",
"Charles Gale"
] |
2014-08-05T20:02:05+00:00
|
2014-08-26T13:58:25+00:00
|
[
"nucl-th",
"hep-ph",
"nucl-ex"
] |
Probing the non-equilibrium dynamics of hot and dense QCD with dileptons
|
It is argued that, in heavy ion collisions, thermal dileptons are good probes
of the transport properties of the medium created in such events, and also of
its early-time dynamics, usually inaccessible to hadronic observables. In this
work we show that electromagnetic azimuthal momentum anisotropies do not only
display a sensitivity to the shear relaxation time and to the initial
shear-stress tensor profile, but also to the temperature dependence of the
shear viscosity coefficient.
|
1408.1100v1
|
## Non-Local Transport Mediated by Spin-Supercurrents
Hua Chen, 1 Andrew D. Kent, 2 Allan H. MacDonald, 1 and Inti Sodemann 1
1 Department of Physics, The University of Texas at Austin, Austin, Texas 78712, USA 2 Department of Physics, New York University, New York, New York 10003, USA
In thin film ferromagnets with perfect easy-plane anisotropy, the component of total spin perpendicular to the easy plane is a good quantum number and the corresponding spin supercurrent can flow without dissipation. In this Letter we explain how spin supercurrents couple spatially remote spin-mixing vertical transport channels, even when easy-plane anisotropy is imperfect, and discuss the possibility that this effect can be used to fabricate new types of electronic devices.
PACS numbers: 75.76.+j, 75.70.-i, 72.25.Pn, 85.75.-d
Introduction -Recent progress [1] in the growth of magnetic thin films has made it possible to construct circuits in which materials with perpendicular and in-plane magnetic anisotropy are flexibly combined. As we explain below these advances have improved prospects for the experimental realization of a new class of effects in spintronics in which collective magnetic degrees of freedom play a more active role. In this Letter we discuss effects which rely on the ability of ferromagnetic materials with strong easy-plane order to carry dissipationless spin supercurrents [2-6] that are in many ways analogous to the dissipationless charge currents carried by superconductors. In both XY ferromagnets and superconductors the energy of the ordered state is independent of an angle φ , the azimuthal magnetic orientation angle in the XY ferromagnet case and the Cooper-pair condensate phase angle in the superconducting case. States with a definite value of this angle break spin-rotational symmetry around the ˆ axis and gauge symmetry respectively. The z analogy between XY ferromagnetism and superconductivity extends to superflow behavior. Just as superconductors can support dissipationless charge currents, perfect easy-plane ferromagnets can support dissipationless currents [2-6] of the conserved ˆ component of total spin. z
As an illustration of the types of effects that can occur we address specifically the super-spintronic circuit proposed in Ref. 3 and illustrated in Fig. 1. The circuit consists of two thin film metal stacks containing perpendicular anisotropy magnetic layers and linked by a common easy-plane ferromagnetic thin film, assumed here to be permalloy. As we will describe, collective magnetic behavior induced by exchange interactions within the easy-plane material lead to non-local effects that can be much stronger than the familiar effects associated with spin-diffusion in normal metals. The cross-sectional area of the two stacks and the length of the permalloy link between them are assumed to be sufficiently large to avoid significant charge transport cross-talk [7-9] between stacks. The magnetization of the perpendicularly magnetized metal stacks is assumed to be pinned and only the easy-plane thin film ferromagnet's magnetization is allowed to change in response to the flow of spin-
FIG. 1: (color online). Charge transport through two (Left and Right) metallic multilayer stacks containing both perpendicular-anisotropy films and a common easy-plane anisotropy thin film. Transport electrons in both metal stacks exert spin-transfer torques on the easy-plane anisotropy film. Exchange interactions within that film couple the two vertical transport channels. When the orientation of the easy plane ferromagnet is time-independent the left (L) and right (R) conductances are independent. When the total current I L + I R exceeds a critical value, spin-transfer torques drive the easy-plane thin-film nano-magnet into a dynamic precessional limit cycle in which the transport properties of the L and R stacks are interdependent. If the Gilbert-damping coefficient of the easy-plane magnet is small, the dc conductance of both stacks is strongly suppressed in the dynamic configuration. The specific materials indicated in this illustration are discussed in the text.
polarized currents. The spin-transfer torques exerted by vertical transport drive spin super-currents through the easy-plane nanomagnet. Below we first explain some key ideas by considering simpler macrospin limit in which the spin-stiffness in the thin film nano-magnet is strong enough to inhibit spatial variation in magnetization orientation, and then discuss the long magnetic link limit. Macrospin Limit - Because they are separated by nonmagnetic spacer layers, the magnetizations of different magnetic circuit elements behave independently. When the system is able to reach a time independent state in the presence of bias voltages, the ˆ-direction spin of the easyz plane magnetic layer satisfies the torque-balance equation,
$$\intertext { a } \intertext { d } _ { z a } \quad \hbar { \frac { d S _ { z } } { d t } } = - 2 K \sin ( 2 \phi ) + F _ { L } g _ { L } \mu _ { L } + F _ { R } g _ { R } \mu _ { R } = 0, \quad ( 1 )$$
where the first term on the right hand side is contributed
by in-plane magnetic anisotropy, with -K cos(2 φ ) being the macrospin in-plane anisotropy energy. The remaining terms are spin-transfer torques [10, 11] applied at left and right electrodes by number currents
$$I _ { L, R } = g _ { L, R } \mu _ { L, R } / \hbar {, }$$
where g L,R is the stack conductance in e /h 2 units, and µ L,R the circuit bias energy at the L, R electrode. The currents contribute to dS /dt z because charge flows between electrodes with opposite perpendicular magnetization orientations; the factors F L,R ∼ 1 are materialdependent spin-injection efficiency factors, and the conductances g L,R are relatively large because the in-plane nano magnet can efficiently reverse the spins of electrons which propagate through it. Provided that
$$| F _ { L } I _ { L } + F _ { R } I _ { R } | < I _ { c } = 2 K / \hbar {, } \quad \quad ( 3 ) \quad \ e \Omega$$
Eq. (1) has a solution and the vertical conductance is local in the sense that I L,R depends only on µ L,R . Note that the static solution is allowed for any value of the individual currents as long as the total injected spin-current satisfies Eq. (3).
When the total spin current exceeds ∼ 2 K/h ¯, Eq. (1) can no longer be satisfied. The magnetization can no longer reach a steady state value and begins to precess. Assuming that the easy-plane anisotropy is very large so that the magentization remains in the XY plane, Eq. (1) then generalizes to
$$\hbar { \frac { d S _ { z } } { d t } } & = - \, 2 K \sin ( 2 \phi ) - g _ { i } \, \hbar { \dot { \phi } } & \quad \text{critica} \\ & \quad + F _ { L } g _ { L } ( \mu _ { L } - \hbar { \dot { \phi } } ) + F _ { R } g _ { R } ( \mu _ { R } - \hbar { \dot { \phi } } ) = 0. \quad ( 4 ) & \quad \text{tal spi} \\ & \quad \text{preces}$$
Here the second equality keeps the spin direction close to the easy-plane and the three terms proportional to ˙ φ capture magnetization decay due to Gilbert damping and spin-pumping [12, 13] into the two electrodes. The spin-pumping terms can be understood by performing a unitary transformation to a spin-direction frame which precesses with the magnetization. The exchange field of the ferromagnet is then static, allowing standard transport ideas to be applied, but the energy bias across the vertical transport stacks is shifted. In the precessing case
$$\hbar { I } _ { L, R } = g _ { L, R } \, ( \mu _ { L, R } - \hbar { \dot { \phi } } ). \quad \quad ( 5 ) \quad \ \ ^ { \, \, \cdot \, }$$
The coefficient g i in the Gilbert damping term is related to the usual dimensionless damping coefficient α by ¯ hg i = αS where S is the total macrospin of the easyplane magnetic nano particle.
Given µ L and µ R , Eq. (4) can be solved for φ t ( ). (Solutions for some typical parameters are presented in Fig. 2.) When the total current is comfortably in excess of the critical value, ˙ ( φ t ) is approximately constant. In this
FIG. 2: (color online). Typical Magnetization Dynamics results for g i = g L = g R = g , F L,R = 1 and initial value φ t ( = 0) = 0. The left panels depict the evolution of the macro-spin orientation φ and the right panels illustrate the behavior of the currents in units of 2 K/h ¯. In each case the dashed line shows the long-time average current. a): µ L = µ R = 0 8 . K/g - the total spin-current is below the critical value and the magnetization and currents approach time-independent values. b): µ L = µ R = 2 0 . K/g - the total spin current exceeds the critical value. The magnetization precesses non-uniformly and the current has a large oscillating component. The precession becomes more uniform and the oscillating currents weaker when the currents are increased further. c): µ L = -0 4 . K/g , µ R = 2 0 . K/g - the spin-current injected by the right contact is the same as in case b) but the total spin-current is the same as in case a). The time unit is t 0 = ¯( h g i + g L + g R ) / (4 K ).
case averaging over time yields the following non-local relationships between electrode currents and biases:
$$\text{case} _ { ( 5 ) } & & I _ { L } = \frac { g _ { L } ( g _ { i } + F _ { R } g _ { R } ) } { F _ { L } g _ { L } + F _ { R } g _ { R } + g _ { i } } \ \mu _ { L } + \frac { - F _ { R } g _ { R } g _ { L } } { F _ { L } g _ { L } + F _ { R } g _ { R } + g _ { i } } \ \mu _ { R } \\ \text{re} _ { \text{t} \ \alpha } & & I _ { R } = \frac { - F _ { L } g _ { L } g _ { R } } { F _ { L } g _ { L } + F _ { R } g _ { R } + g _ { i } } \ \mu _ { L } + \frac { g _ { R } ( g _ { i } + F _ { L } g _ { L } ) } { F _ { L } g _ { L } + F _ { R } g _ { R } + g _ { i } } \ \mu _ { R } \\ \text{as} y _ { - }$$
The crossover between Eqs. (2) which apply in the static magnetization regime and Eqs. (6) which apply far into the dynamic magnetization regime occurs rather abruptly as can be seen in Fig. 3 where we plot the timeaveraged current in the left electrode I L vs. µ L for dif-
FIG. 3: (color online). Current-voltage relationship of the left electrode for µ R = 0, µ R = -K/g R and µ R = K/g R . Currents in the dynamic regime are time averages over the limit cycle. Panel a) g L = g R = 5 g i . Panel b) g L = 5 g R = 5 g i . Note that the increase in resistance in the dynamic regime is larger when both g R and g i are much smaller than g L .
ferent values of µ R .
Both the non-locality of transport and the contrast between the dynamic and static magnetization regimes are enhanced when spin-pumping is the dominant magnetization dissipation mechanism, i.e. when g L,R > g i . Achieving sufficiently large spin-injection into easy-plane nano magnets with Gilbert damping that is sufficiently small for this inequality to be satisfied is a challenge which can now be met thanks to recent advances in spin-torque oscillator [14] technology. Because it has a relatively small interface resistance the Co/Cu materials combination is favorable for the perpendicular-magnet/magnetic-spacer elements of the structure. Similarly, because of its small Gilbert damping parameter, permalloy is an attractive material for the easy-plane anisotropy thin film. The Gilbert damping conductance g i of the permalloy nano magnet is proportional to its total spin and therefore to its volume. The inequality we seek is favored by designing samples in which the easy-plane ferromagnet crosssectional area is dominated by the portions within the metal stacks and not by the link portion. In this limit g L,R and g i are both proportional to the stack crosssectional area, while g i is in addition proportional to the permalloy thin film thickness. Assumming a Gilbert damping parameter α ∼ 10 -2 , and a spin per atom ∼ 2, we conclude that g L,R can be larger than g i for films thinner than ∼ 4 nm. It is of course always possible to
FIG. 4: (color online). Two vertical spin injectors coupled by an easy-plane ferromagnet. The yellow arrows indicate magnetization orientation φ within the easy plane. The spin-supercurrent is proportional to the spatial gradient of the magnetization orientation ∂ φ x .
reduce g L,R below maximum values, for example by introducing a thin insulator layer in the metal stacks, to reverse the sense of this inequality in one or both metal stacks. For 1 nm thick films and ∼ 10 -14 m stack cross2 sections, we estimate that g L,R ≤ 300 000 and, assum-, ing that only 20% of the permalloy area is in the link segment, that g i ∼ 80 000. , For a given sample shape, the in-plane magnetic anisotropy of thin film permalloy can be controlled by varying growth conditions [15] to values as small as ∼ 10 -8 eV per permalloy spin. It follows that for the same device dimensions, the bias energies necessary to drive the transition between static and dynamic regimes can be as small as ∼ 10 -7 eV. Electronic transport switches which operate at such small bias voltages are potentially attractive for low-powerconsumption electronic devices.
Long Nanomagnet Limit - When the in-plane magnetic anisotropy is tuned to larger values, spatial variation of the magnetization orientation can become important. In Fig. 4 we illustrate the case of a long narrow permalloy thin film which provides a magnetic link between metal electrodes at left and right. Assuming that the link is sufficiently narrow to force constant magnetization across the bar, the Landau-Liftshitz equation in this case is
$$- \, 2 \frac { K } { L } \sin ( 2 \phi ) - \frac { g _ { i } } { L } \, \hbar { \phi } + A A _ { \perp } \partial _ { x } ^ { 2 } \phi = 0, \quad ( 7 )$$
where A ⊥ is the cross-sectional area of the bar, L is its length, and A is the magnetic stiffness coefficient. The two terms that are balanced in the static limit can be interpreted as contributions to the rate of change of the local S z density from local magnetization precession and from the divergence of the spin-supercurrent,
$$I _ { S } ( x ) = \frac { A A _ { \perp } } { \hbar } \, \partial _ { x } \phi ( x ).$$
Because the exchange splitting of typical ferromagnetic metals is much larger than the anisotropy energy per particle, the spin-current injected at left and right is converted [16, 17] nearly locally into a spin-supercurrent. It
follows that
$$I _ { S } ( x = 0 ) \ & = \ F _ { L } I _ { L } = F _ { L } g _ { L } \ [ \mu _ { L } - \hbar { \dot { \phi } } ( x = 0 ) ], \\ I _ { S } ( x = L ) \ & = \ - F _ { R } I _ { R } = - F _ { R } g _ { R } \ [ \mu _ { R } - \hbar { \dot { \phi } } ( x = L ) ]. ( 9 )$$
As in the macrospin limit vertical transport at each electrode is local when the easy-plane magnetization orientation is static, but the switching boundary is dependent on both bias voltages. Some insight into switching properties can be gained by noting that in the static case
$$\frac { \hbar { ^ { 2 } } I _ { s } ^ { 2 } } { 2 A A _ { \perp } } + \frac { K } { L } \cos ( 2 \phi ) = C, \text{ \quad \ \ } ( 1 0 )$$
where C is a constant independent of position. When F I L L = -F R R I the spin-supercurrent has the same value at L and R bar ends, and static magnetization configurations are allowed at large current magnitudes. When I L and I R have the same sign at the bar ends, however, I s must change sign as a function of position along the bar and static solutions are possible only when:
$$\max ( F _ { L } ^ { 2 } I _ { L } ^ { 2 }, F _ { R } ^ { 2 } I _ { R } ^ { 2 } ) < \frac { 4 K A A _ { \perp } } { L \hslash ^ { 2 } }. \quad \ \ ( 1 1 ) \quad \ \ t h e \quad \\ \quad \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
For large K this limit on the injected spin-currents is more stringent than Eq. 3. In both cases the spininjection limit depends on bias voltages in both metal stacks and the collective magnetic degree of freedom provides the coupling between electrodes.
The ratio of the length of the permalloy link, L , to the typical magnetization variation length scale:
$$\lambda \equiv \frac { 1 } { 2 } \sqrt { \frac { A A _ { \perp } L } { K } }, \quad \quad \quad ( 1 2 ) \quad \begin{array} { c } \text{div} \\ A c \\ \text{h.w.} \end{array} \end{array}$$
controls how far the system is from the macrospin limit. When L glyph[lessmuch] λ the critical currents are given by the macrospin result in Eq. (3). In the opposite limit, L glyph[greatermuch] λ , the currents at which steady solutions exist satisfy Eq. (11) except within a narrow strip along the F R R I = -F I L L line. Figure 5 illustrates this behavior.
Discussion - The I-V characteristics discussed in this Letter provide an example of an electron transport phenomenon in which a collective degree-of-freedom, the magnetization orientation, plays an active role. This type of phenomenon is of potential interest for electronic device applications mainly because it can lead to current flow response to bias voltage that is sharp on scales smaller than k B T/e , an impossibility for the single-particle transport processes exploited in most current electronic devices. We have so far ignored the role of thermal fluctuations in the phenomena discussed here. Switching between static and dynamic magnetization configurations [18-20] is stochastic with a thermal energy barrier due to anisotropy that vanishes at the critical currents or bias voltages. It follows that reliable collective switching can be driven by changes in bias voltage
FIG. 5: (color online). Regions in the I L,R parameter space where static magnetization solutions exist. The blue, red, and green regions correspond to different ratios of the permalloy link length, L , to the length λ defined in Eq. (12). The dashed black lines lines enclose the region predicted by Eq. (3) intended to apply in the macrospin model. The box enclosing the entire picture, the red dashed-dotted box and the black dashed-dotted box enclose the regions described by Eq. (11), intended to apply in the long-link limit, for L = λ , L = 2 λ , L = 5 λ respectively. Note that the L = λ case (blue region) is already well described by the macrospin model, whereas the L = 5 λ (green region) is already well described by Eq. (11), except for a narrow strip near F R R I = -F I L L .
δV with eδV/k B T ∼ eV switch /K ∼ 1 /g L,R glyph[lessmuch] 1, where V switch is the bias voltage typically required for switching, exceeding the limit possible with switches based upon individual independent electron behavior.
Acknowledgements. - HC, IS and AHM were supported by the Welch Foundation under Grant No. TBF1473 and by the SWAN nano electronics program. ADK was supported by NSF-DMR-1309202 and the SRC-INDEX spin logic program.
- [1] A. Brataas, A. D. Kent, and H. Ohno, Nature Materials 11 , 372 (2012).
- [2] For a review of theoretical work on this topics, which has a rather long history, see E.B. Sonin, J. Low Temp. Physics 171 , 757 (2013) and work cited therein.
- [3] J¨rgen K¨ onig, u Martin Chr. Bønsager, and A. H. MacDonald, Phys. Rev. Lett. 87 , 187202 (2001).
- [4] F. S. Nogueira and K.-H. Bennemann, Europhys. Lett. 67 , 620 (2004).
- [5] S. Takei and Y. Tserkovnyak, Phys. Rev. Lett. 112 , 227201 (2014).
- [6] S. Takei, B. I. Halperin, A. Yacoby, and Y. Tserkovnyak, arXiv:1404.3987.
- [7] The resistance of 1 nm thick permalloy thin films at room temperature is [8] ∼ 300 Ω per square. The resistance area product of the Co/Cu metal stacks is typically near 10 -15 Ωm 2 [9]. For a permalloy link widths of ∼ 5 ×
- 10 -8 m, and a lateral dimensions of the metal stacks of ∼ 10 -7 m, the link resistance ( ∼ 500 Ω) is substantially larger than the stack resistance ( ∼ 0 1 Ω). Crosstalk has . little influence on device performance unless the stack resistance is comparable to the link resistance.
- [8] A.F. Mayadas, J.F. Janak, and A. Gangulee, J. Appl. Phys. 45 , 2780 (1974).
- [9] J. Bass and W.P. Pratt Jr., J. Mag. Mag. Mater. 200 , 274 (1999).
- [10] J. C. Slonczewski, J. Mag. Mag. Mat. 159 , L1-L7 (1996); M. Tsoi, A. Jansen, J. Bass, W.-C. Chiang, M. Seck, V. Tsoi and P. Wyder, Phys. Rev. Lett. 80 , 4281 (1998); E. Myers, D. Ralph, J. Katine, R. Louie and R. Buhrman, Science 285 , 867-870 (1999); J. Sun, J. Mag. Mag. Mat. 202 , 157-162 (1999).
- [11] D. C. Ralph and M. D. Stiles, J. Mag. Mag. Mat. 320 , 1190 (2008).
- [12] Y. Tserkovnyak, A. Brataas, G.E.W. Bauer, and B.I. Halperin, Rev. Mod. Phys. 77 , 1375 (2005).
- [13] We have for simplicity assumed that the efficiency factors are identical for spin-transfer torques and spin-pumping. This is strictly speaking true only when the source and drain electrodes are fully spin selective.
- [14] S.I. Kiselev, J.C. Sankey, I.N. Krivorotov, N.C. Emley, R.J. Schoelkopf, R.A. Buhrman, and D.C. Ralph, Nature 425 , 380 (2003); W.H. Rippard, M.R. Pufall, S. Kaka, S.E. Russek, and T.J. Silva, Phys. Rev. Lett. 92 , 027201 (2004).
- [15] Alfredo Garcia-Arribas, Eduardo Fernandez, Andrey V. Svalov, Galina V. Kurlyandskaya, Ane Barrainkua, David Navas, and Jose Manuel Barandiaran, Eur. Phys. J. B 86 , 136 (2013); A. Garcia-Arribas, E. Fernandez, I. Orue, and J. M. Barandiaran, Appl. Phys. Lett., 103 , 142411 (2013).
- [16] Hua Chen, Inti Sodemann, and A.H. MacDonald, to be published (2014).
- [17] Jung-Jung Su and A.H. MacDonald, Nature Physics, 4 , 799 (2008).
- [18] R. H. Koch, J. A. Katine, and J. Z. Sun, Phys. Rev. Lett. 92 , 088302 (2004).
- [19] I. N. Krivorotov, N. C. Emley, A.G. F. Garcia, J.C. Sankey, S. I. Kiselev, D. C. Ralph, and R. A. Buhrman, Phys. Rev. Lett. 93 , 166603 (2004).
- [20] V. Ambegaokar and B.I. Halperin, Phys. Rev. Lett. 22 , 1364 (1969).
|
10.1103/PhysRevB.90.220401
|
[
"Hua Chen",
"Andrew D. Kent",
"Allan H. MacDonald",
"Inti Sodemann"
] |
2014-08-05T20:03:50+00:00
|
2014-08-05T20:03:50+00:00
|
[
"cond-mat.mes-hall"
] |
Non-Local Transport Mediated by Spin-Supercurrents
|
In thin film ferromagnets with perfect easy-plane anisotropy, the component
of total spin perpendicular to the easy plane is a good quantum number and the
corresponding spin supercurrent can flow without dissipation. In this Letter we
explain how spin supercurrents couple spatially remote spin-mixing vertical
transport channels, even when easy-plane anisotropy is imperfect, and discuss
the possibility that this effect can be used to fabricate new types of
electronic devices.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.