id
stringlengths 11
13
| text
stringlengths 1
74.7k
| image
imagewidth (px) 1.7k
3.75k
|
|---|---|---|
230613499/31 | By (72), (92), and (152) the following is true.
\begin{eqnarray*} &&\text{If} \quad \frac{r}{d_1}\le \frac{1}{p}-\frac{1}{q}+\frac{1}{2},\quad \text{then}\quad\theta=0, \notag\
[.4cm] &&\text{if}\quad\frac{1}{p}-\frac{1}{q}+\frac{1}{2}< \frac{r}{d_1}\le 1-\frac{1}{q},\quad \text{then} \
[.2cm] &&\qquad\quad\theta=\frac{-r+\left(\frac{1}{p}-\frac{1}{q}\right)d_1-\frac{d_2}{2}} {d_1+d_2}-\left(-\frac{r}{d_1}+\frac{1}{p}-\frac{1}{q}\right)=\frac{\left(\frac{r}{d_1}-\left(\frac{1}{p}-\frac{1}{q}+\frac{1}{2}\right)\right)d_2} {d_1+d_2},\quad \
[.4cm] &&\text{if} \quad 1-\frac{1}{q}<\frac{r}{d_1},\quad \text{then} \notag\
[.2cm] &&\qquad\quad\theta=\frac{-r+\left(\frac{1}{p}-\frac{1}{q}\right)d_1-\frac{d_2}{2}} {d_1+d_2}-\frac{-r+\big(\frac{1}{p}-\frac{1}{q}\big)d_1-\big(1-\frac{1}{p}\big)d_2}{d_1+d_2}=\frac{\left(\frac{1}{2}-\frac{1}{p}\right)d_2} {d_1+d_2}. \notag \end{eqnarray*}
For $\frac{1}{p}-\frac{1}{q}+\frac{1}{2}<\frac{r}{d_1}< 1-\frac{1}{q}$ we have
$$
\theta=\frac{\left(\frac{r}{d_1}-\left(\frac{1}{p}-\frac{1}{q}+\frac{1}{2}\right)\right)d_2} {d_1+d_2}
<\frac{\left(\frac{1}{2}-\frac{1}{p}\right)d_2} {d_1+d_2},
$$
while for $\frac{r}{d_1}\ge 1-\frac{1}{q}$
$$
\theta= \frac{\left(\frac{1}{2}-\frac{1}{p}\right)d_2} {d_1+d_2}\le\frac{\left(\frac{1}{2}-\frac{1}{p}\right)\left(\frac{1}{p}-\frac{1}{q}\right)}{\frac{1}{2}-\frac{1}{q}}\le \frac{1}{8},
$$
see (154). We conclude that the maximal value of $\theta$ is reached iff $\theta=\frac{1}{8}$ iff holds.
So assume (149). Here we estimate the gap including log factors. By (14), (71), (96), and (113) we have
\begin{equation*} \sigma_1=0,\quad\beta_1=0,\quad\beta_2=\left\{\begin{array}{lll} 1 & \quad\mbox{if}\quad r=d_1, \\
0 & \quad\mbox{if}\quad r>d_1, \end{array} \right. \quad\sigma_2=1. \end{equation*}
From (72) and (92) we conclude
\begin{equation*} \Phi_1(n)=n^{-\frac{r}{2d_1}-\frac{1}{8}},\quad\Phi_2(n)=n^{-\frac{r}{2d_1}-\frac{1}{4}}, \end{equation*}
and consequently
\begin{equation*} \frac{\Phi_1(n)}{\Phi_2\left(\frac{n}{\log(n+1)}\right)(\log(n+1))^{\beta_2\left(2-\frac{1}{p}\right)}} =n^{\frac{1}{8}}(\log(n+1))^{-\frac{7\beta_2}{4}-\frac{r}{2d_1}-\frac{1}{4}} \end{equation*}
\begin{equation*} \frac{\Phi_1(n)(\log (n+1))^{\beta_1\left(2-\frac{1}{p}\right)}}{\Phi_2(n)(\log(n+1))^{\sigma_2/2}} =n^{\frac{1}{8}}(\log(n+1))^{-\frac{1}{2}}. \end{equation*}
This together with (152) yields (150) and (151). | |
230612848/7 | On the Direct Construction of MDS and Near-MDS Matrices
$\mathrm{rank}(S^{\perp}(I))=|I|-\mathrm{rank}(S(I))=r$.
Since $|I|-\mathrm{rank}(S(I))=r$, by Theorem, we have $d_r(\mathcal{C})\leqslant \delta$.
If possible, let $d_r(\mathcal{C})=\delta-t$ for some $t\geqslant 1$. Now, by Theorem 3, there exist some $I'\subset \mathopen{}\mathclose\bgroup\originalleft\{1,2,\dotsc,n\aftergroup\egroup\originalright\}$ with $|I'|=\delta-t$ such that
\begin{align*} &|I'|-\mathrm{rank}(S(I'))\geqslant r\\
\;\Rightarrow\; & \mathrm{rank}(S(I'))\leqslant |I'|-r\\
\;\Rightarrow\; & \mathrm{rank}(S(I'))\leqslant \delta-t-r. \end{align*}
Therefore, there exist $|I'|=\delta-t$ many columns, say $H_{i_1},H_{i_2},\dotsc,H_{i_{\delta-t}}$, of $H$ of rank $\leqslant \delta-t-r$. Now, by adding any other $t-1$ columns of $H$ to that $\delta-t$ columns we have $\delta-1$ columns, say $H_{i_1},H_{i_2},\dotsc,H_{i_{\delta-t}},H_{i_{\delta-t+1}},\dotsc,H_{i_{\delta-1}}$, of $H$ of rank $\leqslant (\delta-t-r)+(t-1)=\delta-r-1<\delta-r$. This leads to a contradiction to condition $(i)$. Hence, we must have $d_r(\mathcal{C})=\delta$. $\blacksquare$
Definition 4. (NMDS code) [5] An $[n, k]$ code $\mathcal{C}$ is said to be Near-MDS or NMDS if
\begin{align*} &d_1(\mathcal{C})=n-k~~\text{and}~~d_i(\mathcal{C})=n-k+i,~~\text{for}~i=2,3,\dotsc,k. \end{align*}
Remark 3. From the monotonicity of generalized Hamming weights, we can say that an $[n, k]$ code is NMDS if and only if $d_1(\mathcal{C})=n-k~\text{and}~d_2(\mathcal{C})=n-k+2$.
Remark 4. For an $[n,k,d]$ code $\mathcal{C}$, if $d=n-k$, then $\mathcal{C}$ is called an Almost-MDS or AMDS code. However, it is worth noting that not all AMDS codes are necessarily NMDS codes. For example, consider the linear code $\mathcal{C}$ with a generator matrix
\begin{align*} G&=\begin{bmatrix} 1 & 0 & 0 & \alpha^2 & \alpha & 0 \\
0 & 1 & 0 & \alpha & \alpha & 0 \\
0 & 0 & 1 & \alpha & 0 & \alpha \end{bmatrix} \end{align*}
over the finite field $\mathbb{F}_{2^2}$ constructed by the polynomial $x^2 + x + 1$ and $\alpha$ is a root of $x^2 + x + 1$. Then it can be checked that $\mathcal{C}$ is a $[6,3,3]$ code. Also, by determining the minimal support of all two-dimensional subspaces $\mathcal{D}\subset \mathcal{C}$, we get $d_2(\mathcal{C})=4<5$. This value is achieved by the subspace spanned by the first two rows of the generator matrix $G$. Hence, $\mathcal{C}$ is not an NMDS code.
However, when both $\mathcal{C}$ and its dual $\mathcal{C}^{\perp}$ meet the criteria of being AMDS codes, then $\mathcal{C}$ is classified as an NMDS code [4].
Theorem 4 provides the following useful result on NMDS code.
Lemma 2. [5] Let $H$ be a parity check matrix of an $[n,k]$ code $\mathcal{C}$. Then the code $\mathcal{C}$ is NMDS if and only if $H$ satisfies the conditions
- (i) every $n-k-1$ columns of $H$ are linearly independent, (ii) there exist some $n-k$ columns that are linearly dependent, | |
180800448/10 | Figure7. The stratified abundance profiles from the TARDIS fitting process of our SNe Iax sample (solid lines), compared to those originating from the corresponding (see in Sec. 4.4 - 4.8) pure deflagration model (dashed lines). Note that we use oxygen (dotted line) as a “ filler ” element, while carbon mass fraction (dash-dot line) shows the limit for the prediction of the deflagration models. The Fe/Co/Ni mass fractions are computed at 100 seconds after the explosion. | |
210609120/9 | the reduced building of $G$ (not of $M$!) and by $M(F)_{[x]}$ the stabilizer of $[x]$ in $M(F)$.
Definition4.1.1. Let $x,y \in \mathcal{B}(M,F)$. Let $r \in \mathbb{R}$ and set $s=r/2$.
- 1. For $t=s$ or $t=s+$, we write $(M,G)(F)_{x,(r,t)}$ for the group
$$
G(F) \cap \langleT(E)_{r},\ U_\alpha(E)_{x,r},\ U_\beta(E)_{x,t} \, | \, \alpha \in R(T,M),\ \beta \in R(T,G/M) \, \rangle,
$$
- where $T$ is a maximal torus of $M$ that splits over a tamely ramified exten- sion $E/F$ with $x \in \mathcal{B}(T,F) \subseteq \mathcal{B}(G,F)$, and $U_\alpha(E)_{x,r}$ denotes the Moy – Prasad filtration subgroup of depth $r$ at $x$ of the root group $U_\alpha(E) \subseteq G(E)$ corresponding to the root $\alpha$, and similarly for $U_\beta(E)_{x, t}$. 2. Let $U^x_y$ be the quotient of
$$
((M, G)(F)_{x, (r, s)} \cap (M, G)(F)_{y, (r, s+)})(M, G)(F)_{x, (r, s+)}
$$
- by
$$
(M, G)(F)_{x, (r, s+)}.
$$
- 3. For $m \in M(F)_{[x]} \cap M(F)_{[y]}$ let $\delta^x_y(m) = \text{\rm sgn}_{\mathbb{F}_p}(\det\nolimits_{\mathbb{F}_p}(c_m\mathrm{res}_{U^x_y}))$, where $c_m$ is the action of $m$ on $U^x_y$ by conjugation.
We have used that $U^x_y$ has a natural structure of an $\mathbb{F}_p$ -vector space, being an abelian $p$ -group, and that the conjugation action of $G(F)$ on itself induces an action of $M(F)_{[x]} \cap M(F)_{[y]}$ on $U^x_y$.
We now formulate the key result, Lemma 4.1.2, that connects Theorem 3.4 to Yu’s construction. The usefulness of Lemma 4.1.2 stems from Proposition 4.1.6 below. Recall that we write $M_\tx{sc}$ for the pre-image of $M$ in $G_\tx{sc}$, and $M_\tx{ad}$ for the image of $M$ in $G_\tx{ad}$.
Lemma4.1.2. Assume the existence of an element $X$ of $\tx{Lie}^*(M_\tx{sc,ab})(F)$ that satisfies $\operatorname{ord}(\pair X{H_\alpha})=-r$ for all $\alpha \in R(T,G/M)$. For every $y \in \mathcal{B}(M,F)$ denote by $\varepsilon^{G/M}_y$ the pull back to $M(F)_{[y]}$ of the inflation of $\varepsilon^{G_\tx{ad}/M_\tx{ad}}_{[y]}$ of Theorem 3.4. Then
- 1. For every $y \in \mathcal{B}(M,F)$ and $m \in M(F)$, we have that $\^{m^{-1}}{\smash{\varepsilon_{m\cdot y}^{G/M}}}$ equals $\varepsilon_{y}^{G/M}$. 2. For every $x,y \in \mathcal{B}(M, F),$
$$
(\varepsilon_x^{G/M}\mathrm{res}_{M(F)_{[x]} \cap M(F)_{[y]}})\delta^x_y
= (\varepsilon_y^{G/M}\mathrm{res}_{M(F)_{[x]} \cap M(F)_{[y]}})\delta^y_x.
$$
Proof. We claim that the quotient map $G \rightarrow G_\tx{ad}$ maps $U^x_y$ isomorphically onto the analogous space defined in terms of $G_\tx{ad}$. To see this, write $\bar U^x_y$ for the latter space. Since the quotient map restricts to a map $(M,G)(F)_{x,(r,s)} \rightarrow (M_\tx{ad},G_\tx{ad})(F)_{x,(r,s)}$ and analogously for $(r,s+)$, it induces a map $U^x_y \rightarrow \bar U^x_y$. According to [Yu01,Corollary2.3], $U^x_y = (\prescript{}EU^x_y)^{\operatorname{Gal}(E/F)}$ and $\bar U^x_y = (\prescript{}E{\smash{\bar U}}^x_y)^{\operatorname{Gal}(E/F)}$, where $E/F$ is a finite tamely ramified Galois extension splitting a maximal torus $T$ of $M$ with $x \in \mathcal{B}(T,F)$, and $\prescript{}EU^x_y$ and $\prescript{}E{\smash{\bar U}}^x_y$ are constructed in the same way as $U^x_y$ and $\bar U^x_y$ but relative to $E$. Therefore it is enough to show that $\prescript{}EU^x_y \rightarrow \prescript{}E{\smash{\bar U}}^x_y$ is an isomorphism. But the quotient map $G \rightarrow G_\tx{ad}$ is an isomorphism when restricted to each root group with respect to $T$ and this iso- morphism respects the filtrations induced by $x$ and $y$, respectively. While the | |
220604413/16 | When $u,v\in C([0,T];\mathbb H^1)$, applying the inequality, we have
$$
\begin{aligned}
\||u(t)|^p - |v(t)|^p\|_{\mathbb H^{-\theta}} \leqslant C (\|u(t)\|^{p-1}_{\mathbb H^1}+\|v(t)\|^{p-1}_{\mathbb H^1})\|u(t)-v(t)\|_{\mathbb H^1}.
\end{aligned}
$$
For $\mathcal Hu (t) = \ell*u$, we have
$$
\|(\chi\cdot\nabla)[\mathcal H u(t) - \mathcal H v(t)]\| \leqslant \|\chi\|_\infty \cdot\|\mathcal H u(t) - \mathcal H v(t)\|_{\mathbb H^1}.
$$
Because $L^2(\Omega)\subset \mathbb H^{-\theta}$, we get
$$
\begin{aligned}
\|(\chi\cdot\nabla)[\mathcal H u(t) - \mathcal H v(t)]\|_{\mathbb H^{-\theta}} \leqslant C_\theta\|\chi\|_\infty \cdot\|\mathcal H u(t) - \mathcal H v(t)\|_{\mathbb H^1},
\end{aligned}
$$
with $C_\theta$ is a positive constant. Combining (3.6) and (3.7), it leads to
\begin{align*} \|f(u(t),\mathcal Hu(t))-f(v(t),\mathcal Hv(t)\|_{\mathbb H^{-\theta}} & \leqslant C (\|u(t)\|^{p-1}_{\mathbb H^1}+\|v(t)\|^{p-1}_{\mathbb H^1})\|u(t)-v(t)\|_{\mathbb H^1} \\
& \quad + C_\theta\|\chi\|_\infty \cdot\|\mathcal H u(t) - \mathcal H v(t)\|_{\mathbb H^1}. \end{align*}
Thus, the function $f$ satisfies condition (F) with $\mu=1$, $\theta=\delta$ and
\begin{align*} L_f(\rho) = 2C\rho^{p-1}, \; K_f(\rho) = C_\theta\|\chi\|_\infty . \end{align*}
Then, we get the conclusion of Theorem 1 v`a 3 when $\|\chi\|_\infty$ is sufficiently small.
## 4. Application
Consider the following problem of identifying parameter
$$
\begin{aligned}
\partial_t u - (1+D_t^{\{m\}})\Delta u & = g(x)p(t) + f_1(u) \text{in} \Omega, t\in (0,T),\\
u & = 0 \text{on} \partial\Omega,\; t\geqslant 0,\\
u(0) & = \xi \text{in} \Omega, \\
\int_\Omega \kappa(x)u(t,x)dx & = \psi(t),\; t\in [0,T],
\end{aligned}
$$
where $p(t)$, $t\in [0,T]$, is an unknown parameter, $g\in L^2(\Omega)$ is given. In this model, (4.4) is the complementary measurement with $\kappa\in H_0^1(\Omega)$ such that $(g,\kappa)\ne 0$, $\psi\in W^{1,1}(0,T)$.
In this problem, we require further that the kernel $m$ fulfils:
- (M Assumption (M) holds and $m'\in L^1(0,T)$.
Then, problem (4.1) is rewritten as follows
\begin{align*} \partial_t u - (1+m_0) \Delta u - m_1*\Delta u & = g(x)p(t) + f_1(u), \;m_0=m(0), m_1=m'. \end{align*}
Combining with (4.4), we obtain
$$
\psi' + (1+m_0)(\nabla u, \nabla\kappa) + m_1*(\nabla u, \nabla\kappa) = (g,\kappa)p(t) + (f_1(u),\kappa).
$$
Therefore,
$$
p(t) = (g,\kappa)^{-1}[\psi' + (1+m_0)(\nabla u, \nabla\kappa) + m_1*(\nabla u, \nabla\kappa)-(f_1(u),\kappa)].
$$
Set
$$
f_2(u,\mathcal H u):=(1+m_0)(\nabla u, \nabla\kappa) + (\nabla (m_1*u), \nabla\kappa)-(f_1(u),\kappa),
$$
where $\mathcal H$ is the convolution operator with the kernel $m_1$.
Assume that $f_1:\mathbb H^1\rightarrow L^2(\Omega)$ verifies the condition
$$
\|f_1(u)-f_1(v)\| \leqslant L_1 \|u-v\|_{\mathbb H^1}, \text{for all} u,v\in\mathbb H^1.
$$
| |
210410118/1 | # Developing a New Tool to Implement Computer-Supported Active Learning Strategies in the Engineering Classroom
Juan M. Tizón ∗a, Pablo Sierra †a, Luis Sánchez de León ‡a, Emilio Navarro §a, Javier Vilá ¶b, and José F. Moral ‖b
Successful implementation of active learning strategies in the engineering classroom — and in particular in certain subjects which are highly technological in nature such as, for instance, rocket engines and space propulsion — means overcoming certain challenges that arise from the fact that these are extremely complex systems to analyze. In this pa- per, we address the specific means to overcome one of such challenges: the lack of readily available software tools that are suitable for implementing this sort of teaching strategies within the engineering training. In particular, we develop a new tool for the modeling and simulation of liquid-propellant rocket engines specially tailored for the classroom, taking a systematic approach to the development of such tool based on the needs of mod- ern teaching practices. After a thorough review of the available literature on the topic, the few most critical features that our tool should have in order to serve its purported goal are identified. Subsequently, a pilot experience to assess the impact of the usage of said tool on the learners’ performance was carried out, showcasing excellent results, both in terms of the students’ perceived quality of their training as well as in terms of their grade of retention and understanding of the matter. The conclusions of this study, especially the guidelines for the development of software tools aimed at the classroom, nevertheless, should be applicable to any other highly technological discipline, extending the scope of this paper beyond merely the subject of rocket science in engineering.
∗ E-mail: [email protected] — ORCID iD: https://orcid.org/0000-0002-8687-6657 † E-mail: [email protected] — ORCID iD: https://orcid.org/0000-0001-7876-2724 ‡ Address all correspondence to this author: Luis Sánchez de León Peque, Departamento de Mecánica de Flu- idos y Propulsión Aeroespacial, Escuela Técnica Superior de Ingeniería Aeronáutica y del Espacio, Universidad Politécnica de Madrid, Pza. del Cardenal Cisneros 3, 28040 – Madrid, Spain. E-mail: [email protected] — ORCID iD: https://orcid.org/0000-0001-6815-3263 § E-mail: [email protected] ¶ E-mail: [email protected] ‖ E-mail: [email protected] | |
210409923/28 | axion and $\Lambda_D\sim 3$ meV for axion quintessence. In either case, the associated dark-gluons will be relativistic during most of the evolution of the Universe and will act as radiation during BBN, contributing to $\Delta N_{eff}$. In addition, the new Dirac fermion and modulus might also contribute depending on their mass. Despite in most of the parameter space their mass is larger than $T_{BBN}$, we will take a conservative approach and consider that, in addition to the dark gluons, they also contribute to the effective number of neutrinos.
The energy density of the new non-abelian dark sector is related to the effective number of neutrinos through [86]:
$$
\Delta N_{eff}=\frac{\rho_D}{\rho_\nu}=\frac{8}{7}g_\ensuremath{*}^{D}\left(\frac{T_D^4}{T_\nu^4}\right)\,,
$$
where $g_\ensuremath{*}^{D}$ is the number of relativistic d.o.f. of the dark sector and $T_D$, $T_\nu$ are the temperatures of the dark sector and neutrinos, respectively. If we consider that the dark gauge group is $G_D=SU(N)$, then we have:
$$
\Delta N_{eff}=\frac{8}{7}\left[(N^2-1)+1+\frac{7}{8}\times 2\right]\left(\frac{T_D^4}{T_\nu^4}\right)\leqslant 0.33\,,
$$
where $\Delta N_{eff}\leqslant 0.33$ corresponds to the current limit from primordial $^4$ He abundance at 95% C.L [87 – 89]. Even for the smallest non-abelian group, $G_D=SU(2)$, this clearly indicates the need of a dark sector colder than the visible sector:
$$
\Delta N_{eff}=6.57 \left(\frac{T_D^4}{T_\nu^4}\right)\leqslant 0.3\rightarrow \frac{T_D}{T_\gamma}=(4/11)^{1/3}\left(\frac{T_D}{T_\nu}\right)\leqslant 0.32\,.
$$
This could be explained, for example, if there is a preferential reheating for our visible sector through a more efficient coupling to the inflaton sector.
The previous estimate corresponds to the ratio of temperatures at BBN. We now estimate the ratio of temperatures of both sectors at the time of reheating. Assuming entropy conservation in each sector [90],
$$
\frac{g_{\ensuremath{*}, s}^D(T^D_{BBN}) T^{D\,3}_{BBN}}{g_{\ensuremath{*}, s}^D(T^D_{RH}) T^{D\,3}_{RH}}=\frac{g_{\ensuremath{*}, s}^{vis}(T_{BBN}) T^{3}_{BBN}}{g_{\ensuremath{*}, s}^{vis}(T_{RH}) T^{3}_{RH}}\,,
$$
we relate the temperatures of the visible and hidden sector (denoted with a ” D”) right after reheating:
$$
\frac{T^D_{RH}}{T_{RH}}=\left(\frac{g_{\ensuremath{*}, s}^{vis}(T_{RH})}{10.75}\right)^{1/3}\left(\frac{g_{\ensuremath{*}, s}^{D}(T_{BBN}^D)}{g_{\ensuremath{*}, s}^D(T^D_{RH})}\right)^{1/3}\frac{T^D_{BBN}}{T_{BBN}}\,.
$$
Assuming a relatively simple dark sector with few degrees of freedom, $g_{\ensuremath{*}, s}^D(T^D_{RH})\sim g_{\ensuremath{*}, s}^D(T^D_{BBN})$, and $g_{\ensuremath{*}, s}^{vis}(T_{RH})=106.75$, the previous limit translates into:
$$
\frac{T^D_{RH}}{T_{RH}}\lesssim 0.55\,.
$$
| |
231018381/18 | RIKI SHIMIZU AND
HAU-TIENG WU
The overall performance of ConceFT-S is equivalent to, if not better than, ex- isting results in the literature. Notably, You et al. [29] and Jiang et al. [52] have assessed an extensive range of spindle detectors on the same dataset, utilizing the identical Leave-One-Subject-Out Cross-Validation (LOSO-CV) methodology ap- plied in our research. Their studies reveal a wide spectrum of F1 scores, ranging from 0.175 for DETOKS to 0.739 for Spindle U-Net, with Spindle U-Net sharing structural similarities with SUMO and both being rooted in the U-Net neural net- work framework. While deep learning techniques have consistently displayed supe- rior performance over traditional methods in sleep spindle detection, our findings underscore that straightforward and interpretable thresholding techniques, when combined with an improved TFR, can yield comparable results to deep learning methods. Notably, compared with “ black-box ” deep learning models, an advan- tage of ConceFT-S is its interpretability, which is essential for scientific research. Grounded in the amplitude and power within the sigma frequency band, it offers transparent insights into the detection process, fostering understanding and trust, and bridging the gap between computational outcomes and practical applications. Another advantage rooted in its interpretability is its simplicity; it does not necessi- tate extensive preprocessing, such as pre-filtering or windowing, nor does it require the level of tuning typically associated with deep learning approaches.
Beyond the development of ConceFT-S, the primary advantage of the proposed analysis framework, ConceFT, lies in its capacity to investigate spindle IFs. The IF of an oscillatory time series, along with its associated phase, encapsulates intricate physiological dynamics. The quantification of these features through TF analy- sis tools has paved the way for various clinical applications, including automatic sleep apnea detection [57], early prediction of acute hemorrhage [58], and numer- ous others. In the context of spindle analysis, it’s crucial to remember that the alternation between the acceleration and deceleration of sleep spindle frequencies is linked to sleep-related disorders, such as sleep apnea [12]. Consequently, it becomes an intriguing subject to further investigate whether the more detailed dynamics of spindle IF contain valuable physiological or clinical insights.
The present study has certain limitations, and topens avenues for future explo- ration. First, it is important to note that our analysis was conducted using a single EEG channel. While we demonstrated its performance in this context, its applica- bility to different channels remains unvalidated. Furthermore, it is well-established that sleep spindles originating from various brain regions are associated with dis- tinct generators [2]. Prior research, such as that presented in [59], has shown that spatio-spectral-temporal analysis can illuminate the potential involvement of spindles in coordinating cortical activity during consolidation. The high-resolution TFR generated by ConceFT might be valuable in distinguishing the characteristics of spindles recorded from different channels. Second, our focus has been on epochs labeled as N2 stages by experts, which may limit practical application, given that expert sleep stage annotations may not always be readily available. In such cases, it would be beneficial to employ automatic sleep stage classification algorithms [60]. Third, although the algorithm has been validated using two publicly available datasets, further validation on larger datasets and application to clinical scenarios is warranted. Lastly, our model 1 is purely phenomenological and serves the objec- tives outlined in this paper. However, its potential for further exploration is not | |
210210252/27 | Table1: Clustering goodness in Simulation 1; here, $k$ in $k$ -mer and $n$ in nTreeClus is equal to the square root of the length of sequences.
patterns is dynamic. Here, patterns $x$ and $y$ with the length of $\mathscr{N}_{(P_x,P_y)}$ = {(2,3), (4,6), (6,9)}, respectively, have been inserted in each half of the data while the gap in between changes randomly. For $x$ = {A,C} and $y$ = {A,B,E}, it is expected to have them in sequences with different gap in between; for instance,.. AC AABDBA ABE..,.. ACABE.., or.. AC BE ABE... The other hyperparameters, including $\mathscr{N}$, $a$, $\mathscr{A}$, $\mathscr{L}$, and $\mathscr{B}_S$, are the same as the first simulation. The second simulation with a dynamic gap and shorter lengths of patterns aims to identify intricate patterns in a given categorical sequence.
Table 2 indicates that the general performance of the six methods has reduced as the structure of the data becomes more complicated. Once again, it shows that nTreeClus results in considerably higher indices than those of other algorithms. Here, 3-mer and MHMM are the closest opponents of the current method. On account of the greater complexity of this new dataset, the ASW index for all the methods decreases, whereas the external indices | |
171004438/36 | trivial monodromy around the joints. For instance, going around the top branch-point gives the constraint (see Figure 16)
$$
\begin{aligned}
\widetilde{D}_3 \, \ensuremath{\mathcal S}_{w_1} D_1 \, \ensuremath{\mathcal S}_{w_2} D_2 \, \ensuremath{\mathcal S}_{w_2}^{-1} = \mathbf{1}.
\end{aligned}
$$
Furthermore, we need to enforce the boundary conditions at the punctures. For instance, going around the minimal puncture gives the constraint (see Figure 17)
$$
\begin{aligned}
\ensuremath{\mathcal S}_{w_{3,a}} D_4 \, D_5 \, \ensuremath{\mathcal S}_{w_{4}}^{-1} \, D_6 \, D_7 \, \ensuremath{\mathcal S}_{w_{3,b}} = \mathrm{Hol}_1 \nabla^\mathrm{ab},
\end{aligned}
$$
where $\ensuremath{\mathcal S}_{w_{4}} =\ensuremath{\mathcal S}_{w_{4,b}} \ensuremath{\mathcal S}_{w_{4,a}}$.
Solving all these constraints shows that the matrices $\ensuremath{\mathcal S}_w$ have a canonical solution, which (just like for $K=2$ abelianizations) have an interpretation as parallel transport along auxiliary paths.
The resulting expressions for the matrices $\ensuremath{\mathcal S}_w$ depend on the choice of framing at the (maximal) holes through the choice of ordening the eigenvalues in the abelian holon- omy matrices $\mathrm{Hol}_{0}{\nabla^\mathrm{ab}}$ and $\mathrm{Hol}_{\infty}{\nabla^\mathrm{ab}}$. In contrast, the particular choice of eigenlines $(l_{1,\alpha},l_{1,\beta};l_{1,\gamma})$ at the minimal puncture doesn’t play any role in computing the $\ensuremath{\mathcal S}_w$.
Yet, the canonical solution for the transformations $\ensuremath{\mathcal S}_w$ implies that the basis in any re- gion $C\backslash \ensuremath{\mathcal W}$, and in particular near the minimal puncture, is uniquely determined in terms of the choice of eigenlines at the boundary components. That is, the abelianization of $\nabla$ canonically determines the choice of eigenlines $(l_{1,\alpha},l_{1,\beta};l_{1,\gamma})$ at the minimal puncture. In particular, there is no framing ambiguity at the minimal puncture after all.
Note that this is consistent with the interpretation of the framing data in terms of the S-wall matrices. Indeed, whereas a change in framing at an annulus $A_l$ corresponds to a permutation of the mass parameters $M_{l,i}$, the permutation group of mass parameters at the minimal puncture is trivial. Generalizing this argument to any regular puncture, we expect that the framing ambiguity at a regular puncture with Young diagram $Y$ is given by the group $S_{n_1} \times \ldots \times S_{n_k}$, where $n_1, \ldots, n_k$ counts columns of $Y$ with the same height and $S_n$ is the permutation group with $n$ elements.
## Gluing
Fix a length-twist type network $\ensuremath{\mathcal W}$ built out of two molecules. (The same argument can be extended to more molecules.) Say that $\nabla$ is a $\ensuremath{\mathcal W}$ -framed flat connection on $C$ and suppose that $\nabla$ admits a $\ensuremath{\mathcal W}$ -abelianization.
Choose a pants cycle $\alpha$ relative to $\ensuremath{\mathcal W}$ and fix a marked point $z_\alpha$ on $\alpha$. The monodromy $\nabla$ along $\alpha$ is diagonal in the “ abelian ” gauge. Cut the surface $C$ along $\alpha$ into two pair of pants $C_1$ and $C_2$. Say $\nabla_1$ is the restriction of $\nabla$ to $C_1$, and $\nabla_2$ the restriction to $C_2$. $\nabla_1$ and $\nabla_2$ are both flat $SL(K)$ -connections with trivialization at the marked point $z_\alpha$.
The $\ensuremath{\mathcal W}$ -abelianization of $\nabla_1$ (as well as $\nabla_2$) is almost the same as described in the previous subsection. In particular, we still find the same unique solution to the S-wall | |
220602623/11 | condition (20) should be regarded as the equation of state for the black hole system. In this sense, the generalized free energy (16) or (19) should be regarded as the off-shell free energy function at the arbitrary temperature $T$. By substituting Eq.(20) into Eq.(16) or Eq.(19), one can obtain the on-shell value of the free energy for the SAdS black hole [42]. Here, the on-shell or off-shell refers to whether the ensemble temperature is equal to the Hawking temperature or not.
The free energy landscape provides a pictorial and quantitative description of the gener- alized free energy topography. One can refer to Ref. [1] for the detailed discussion on the free energy landscape. The minimum principle of the generalized free energy indicates that the local stable black holes are the extreme points on the free energy landscape. The thermody- namic stability of the local stable black hole and the kinetics of black hole phase transition were also discussed based on the free energy topography in [1].
## GENERALIZED FREE ENERGY LANDSCAPE OF RNADS BLACK HOLES
In this section, we derive the generalized free energy landscape of the the small/large RNAdS black hole phase transition in extended phase space [50–52]. As shown in the last section, to evaluate the partition function in the semiclassical approximation, one just needs to compute the gravitational action of the Euclidean gravitational instanton with the conical singularity.
We start with the Euclidean metric of RNAdS black hole in four dimensions
$$
\begin{aligned}
ds^2=\left(1-\frac{2M}{r}+\frac{Q^2}{r^2}+\frac{r^2}{L^2}\right)d\tau^2+
\left(1-\frac{2M}{r}+\frac{Q^2}{r^2}+\frac{r^2}{L^2}\right)^{-1}dr^2+r^2d\Omega_2^2\;,
\end{aligned}
$$
where $M$ is the mass, $Q$ is the electric charge, and $L$ is the AdS curvature radius. The mass of the fluctuating RNAdS black hole as the function of the black hole radius can be expressed as
$$
\begin{aligned}
M=\frac{r_+}{2}\left(1+\frac{r_+^2}{L^2}+\frac{Q^2}{r_+^2}\right)\;.
\end{aligned}
$$
Note that the Euclidean time has the period $\beta$ determined by the ensemble temperature $T$. Therefore, the metric (21) with the general time period also describes the Euclidean gravitational instanton with the conical singularity, the geometry of which is depicted in Figure 2. Analogous to the case of Hawking-Page phase transition, one can derive the | |
230802803/6 | two components $\Omega_1$ and $\Omega_2$ such that $\partial \Omega_1=\partial \Omega_2=\Sigma$. Let $u$ be the first eigenfunction of $\Sigma$, that is,
$$
{\Delta} u+\lambda_1 u=0.
$$
Let $\nu$ be the unit outward normal of $\Sigma$ in $\Omega=\Omega_1$. We denote by $h(X, Y)=$
$g\left(\nabla_X \nu_, Y\right)$ and
$H=\operatorname{tr}_{g_{\Sigma}} h$ the second fundamental form and the mean curvature of $\Sigma$. Without loss of generality, assuming that
$$
\int_{\Sigma}h\left(\nabla_{\Sigma} f, \nabla_{\Sigma} f\right)\geq0.
$$
Let $f$ be the solution of the Dirichlet problem such that:
$$
\begin{cases}
\Delta f=0, & \text {in} \Omega;
\\ f=u, & \text {in} \partial \Omega=\Sigma.
\end{cases}
$$
Then $u$ is a function defined on $\Omega$ smooth up to $\partial \Omega$. By Reilly’s formula, we have
$$
\int_{\Omega} (\Delta f)^2-\left|{\rm Hess} f\right|^2-\mathrm{Ric}\left(\nabla f,\nabla f \right)
\geq \int_{\Sigma} 2 f_\nu \Delta_{\Sigma} f+H\left(f_\nu\right)^2.
$$
Due to (2.4), one has
$$
- \int_{\Sigma} f_\nu \Delta_{\Sigma} f=
\lambda_1
\int_{\Sigma} f_\nu f=
\lambda_1
\int_{\Omega} f \Delta f+|\nabla f|^2=
\lambda_1 \int_{\Omega} |\nabla f|^2.
$$
Next, we choose the normal coordinate $\left\lbrace e_1,e_2,\cdots,e_{n+1}\right\rbrace$ at the point $p\in \Omega$ such that $\nabla f(p)=f_1(p)=\left\langle \nabla f,e_1\right\rangle e_1(p)$. At $p\in \Omega$, one has
$$
\nabla_{e_i}|\nabla f|^2=\sum_{j}\nabla_{e_i}f_j^2=2\sum_{j}f_jf_{ij}=2f_1f_{1i},
$$
and
$$
|\nabla|\nabla f|^2|^2=\sum_{i=1}^{n+1}|\nabla_{e_i}|\nabla f|^2|^2
=4\sum_{i=1}^{n+1}f_1^2f_{1i}^2=4|\nabla f|^2\sum_{i=1}^{n+1}f_{1i}^2.
$$
Since $\Delta f=0$ in $\Omega$, we have
\begin{align*} 2\sum_{i=1}^{n+1}f_{1i}^2 &=\frac{2n}{n+1}f_{11}^2+ \frac{2}{n+1} f_{11}^2+ \sum_{i=2}^{n+1}\left(f_{1i}^2+f_{i1}^2\right) \\
&=\frac{2n}{n+1}f_{11}^2+ \frac{2}{n+1} \left(\sum_{i=2}^{n+1}f_{ii}\right) ^2+ \sum_{i=2}^{n+1}\left(f_{1i}^2+f_{i1}^2\right)\\
&\leq \frac{2n}{n+1}\sum_{i=1}^{n+1}f_{ii}^2+\sum_{i=2}^{n+1}\left(f_{1i}^2+f_{i1}^2\right)\\
&\leq\frac{2n}{n+1}\sum_{i,j=1}^{n+1}f_{ij}^2=\frac{2n}{n+1}|{\rm Hess}f|^2. \end{align*}
| |
230406470/9 | Figure 6: Some samples of generated eyeglasses with poor quality.
Figure 7: Examples of poorly generated teeth.
Figure 8: Clues that can reveal fake ears, here through earrings.
when viewed from different perspectives, resulting in a flat or two-dimensional appearance. Additionally, incorrect or inconsistent motion blur may suggest that certain parts of the image have been manipulated. The absence of occlusion, i.e., the overlapping of objects in the scene, is another telltale sign of generated images, as it can make the image look flat or unrealistic. Lastly, generated images may display improper image alignment, with objects seeming misaligned or out of place.
## Physics
Generated images that violate physics rules exhibit various cues that can give them away as unrealistic or physically impossible. These cues include objects appearing to float in mid-air without support, shadows that are inconsistent with the light source, reflections or refractions that break the laws of optics, objects passing | |
200401021/73 | donde $\epsilon^\alpha$, el objeto que parametriza la transformación, es un Weyl de dos componentes infi- nitesimal que anticonmuta. Como trabajamos en el marco de supersimetría global, $\epsilon^\alpha$ es una constante.
Para que la acción que obtuvimos sea invariante ante transformaciones de supersimetría $\delta\textsl{L}_{\text{escalar}}+\delta\textsl{L}_{\text{fermiónico}}$ debe cancelarse, a menos de una derivada total. Entonces encontramos que la transformación del campo fermiónico debe ser
$$
\delta\psi_{\alpha}=i(\sigma^{\mu}\epsilon^{\dag})_{\alpha}\partial_{\mu}\varphi, \hspace{1cm} \delta\psi^{\dag}_{\alpha}=-i(\epsilon\sigma^{\mu})_{\alpha}\partial_{\mu}\varphi^*.
$$
Por lo que arribamos a:
$$
\delta S=\int(\delta\textsl{L}_{\text{escalar}}+\delta\textsl{L}_{\text{fermiónico}})d^4x=0.
$$
Como anteriormente mencionamos, el álgebra supersimétrica debe ser cerrada, en otras pa- labras, el conmutador de dos transformaciones supersimétricas distintas debe ser otra simetría de la teoría. Usando Ec. (4.6) y (4.7) encontramos
$$
\begin{aligned}
(\delta_{\epsilon_{2}}\delta_{\epsilon_{1}}-\delta_{\epsilon_{1}}\delta_{\epsilon_{2}})\varphi\equiv\delta_{\epsilon_{2}}(\delta_{\epsilon_{1}}\varphi)-\delta_{\epsilon_{1}}(\delta_{\epsilon_{2}}\varphi)=i(\epsilon_{1}\sigma^{\mu}\epsilon_{2}^{\dag}-\epsilon_{2}\sigma^{\mu}\epsilon_{1}^{\dag})\partial_{\mu}\varphi.
\end{aligned}
$$
Obtenemos que el conmutador de dos transformaciones supersimétricas nos da la derivada del campo original. $\partial_{\mu}$ corresponde al generador de traslaciones espacio temporales $P_{\mu}$.
Debemos ahora encontrar el mismo resultado para el campo fermiónico. Utilizando Ec. (4.6) y (4.7) junto a la identidad de Fierz encontramos:
$$
\begin{aligned}
(\delta_{\epsilon_{2}}\delta_{\epsilon_{1}}-\delta_{\epsilon_{1}}\delta_{\epsilon_{2}})\psi_{\alpha}=i(\epsilon_{1}\sigma^{\mu}\epsilon_{2}^{\dag}-\epsilon_{2}\sigma^{\mu}\epsilon_{1}^{\dag})\partial_{\mu}\psi_{\alpha}-i\epsilon_{1\alpha}\epsilon^{\dag}_{2}\overline{\sigma}^{\mu}\partial_{\mu}\psi+i\epsilon_{2\alpha}\epsilon^{\dag}_{1}\overline{\sigma}^{\mu}\partial_{\mu}\psi.
\end{aligned}
$$
Los últimos dos términos en (
4.10
) se anulan on-shell, es decir, si la ecuación de movimiento se cumple. Para que el álgebra sea cerrada off-shell, incluimos un campo escalar complejo F sin | |
230617773/7 | Given $d\in D,$ we will denote by $\varphi _{d}(P)$ the (only) contract in $%
\varphi (P)$involving$d$.
## Stability
An essential property to be considered in matching models is the stability of allocations. In order to introduce the notion of stability, it is necessary to consider the preferences of the hospitals over the subsets of contracts. As we mentioned before, each hospital $h\in H$ has an antisymmetric, transitive and complete preference relation over $%
A (\mathbf{X}_{h}) $that we will denote by$\succ _{h}.$A preference profile for hospitals is denoted by$\succ :=(\succ _{h})_{h\in
H} $. Along the paper$\succ$ is assumed to be arbitrary but fixed and known by doctors.
The choice set of $h\in H$ given a preference $\succ _{h}$ and $Y\subseteq \mathbf{X,}$ is the subset of $Y_{h}$ that $h$ likes best according to $\succ _{h}$
\begin{equation*} C_{h} (\succ _{h},Y) =\mathrm{max}_{\succ _{h}}A(Y_{h}) \end{equation*}
Similarly, the choice set of $d\in D$ given a preference $%
P_{d}\in \mathcal{P}_d$and$Y\subseteq \mathbf{X,}$is the subset of$Y_{d}$that$d$likes best according to$P_{d}$
\begin{equation*} C_{d} (P_{d},Y) =\mathrm{max}_{P_{d}}A(Y_{d}) \end{equation*}
Remark1 To keep the notation simpler, we will omit describing the preferences used to obtain the choice sets unless this information were relevant and not obvious from the context. We will write $C_{h} (Y)$ and $%
C_{d} (Y) $instead of$C_{h} (Y,\succ _{h})$and$%
C_{d} (Y,P_{d}) ,$ respectively.
In addition, given $Y\subseteq \mathbf{X}$ we define $C_{H} (Y)
=\cup_{h\in H}C_{h} (Y) $and$C_{D} (Y)
=\cup_{d\in D}C_{d} (Y) .$
Now, we are ready to introduce the concept of individually rational and stable allo- cation. Given $P\in \mathcal{P}$, an allocation $Y\in A(\mathbf{X})$ is individually rational at $P$ if it does not contain unwanted contracts, i.e.,
\begin{equation*} C_{D} (Y) =C_{H} (Y) =Y. \end{equation*}
7 Observe that for each $d\in D,$ the choice set $C_{d} (Y,P_{d})$ contains only the best element in $Y_{d}$ according to $P_{d}.$ However, we introduce this definition here to have a symmetric notation for hospitals and doctors which makes the exposition simpler. | |
190411476/3 | Fig. 2: (a) Visualization of scanning radar operation from a bird’s-eye view. The radar (green dot) sequentially observes $M$ azimuths (dotted lines), collecting readings for $N$ range bins along $r$. For each azimuth $a$, it outputs a power-range spectrum, an example of which is given on the bottom right. (b) The polar radar scan formed by combining the spectra of a full rotation of azimuths, as observed in Oxford city center, UK. (c) The corresponding Cartesian scan with equivalent regions in (b) and (c) highlighted. Note the high levels of speckle noise and false positives due to saturation.
be categorized as indirect or direct. Indirect methods first extract salient keypoints, then associate those that correspond to the same location. Direct methods [9]–[11], which forego keypoint extraction and operate on minimally pre-processed sensor outputs, are discussed in [2],[3] and not in this paper. All methods assume the majority of observed objects are static.
The first step of indirect methods is keypoint extraction, for which the most popular approach is constant false- alarm rate (CFAR) detection [12], which distinguishes peaks from noise using sliding-window thresholding. CFAR and its variants generally require at least three tunable parameters, which are based on assumed noise characteristics and do not behave consistently across datasets (see [2] for comparison). Some works leverage the knowledge that coherent structures make good keypoints by clustering or detecting the edges of bright regions in scans [13]. Others elect to represent the surroundings using predetermined geometric primitives [14] or models, like the normal distribution transform (NDT) [15]. Vision-inspired works treat radar scans as images and extract features, like SIFT and FAST [16].
These keypoints must then undergo data association, also known as scan matching [1], [14] in robotics. The most common technique is iterative closest point (ICP) [17], which iteratively matches points using naive methods until the alignment between keypoint sets is sufficiently close [14], [18]. ICP relies on a good estimate of the relative displacement (i.e., motion prior) between scans. Other data association techniques search for motion parameters that optimize some objective function, like maximizing similarity (e.g., overlap of Gaussian distributions for NDT [19]) or min- imizing distance (e.g., cluster edge difference [9]) between keypoint sets. Two further examples of objective functions characterize map quality [20] and radar scan distortion [21] in terms of motion. Feature-based approaches associate keypoints using descriptors, like BASD [22] and SURF [16].
Many of the methods discussed do not generalize well due to the high levels of noise in radar scans. Adequate performance often requires fine tuning, a priori knowledge, restrictive assumptions, or outlier detection. Several of the works rely heavily on other sensors for robustness, which compromises performance under conditions that cause these sensors to fail, or use simultaneous localization and mapping (SLAM), which is accompanied by overhead costs and model-reliant motion filters [16], [23].
We utilize an indirect method to explicitly select salient information from noisy artifacts. We present two algorithms that improve on our previous work [2]. The new pipeline is more efficient and easily adaptable across diverse settings. Accordingly, our keypoint extraction algorithm returns inter- pretable features with minimal redundancy, only one input parameter, and no assumptions about the scene structure or noise. Our data association algorithm does not need a motion prior or parameter tuning, and it is robust to large amounts of noise and false detections. An in-depth literature review and description of our contributions can be found in [3].
## III. R ADAR - ONLY EGO - MOTION ESTIMATION
## A. FMCW scanning radar
FMCW radar, which is becoming more compact, safe, and affordable than alternative radars, collects long-range measurements with high accuracy and remains resilient under variable lighting and weather [3]. Scanning radar, as shown in Fig. 2, sequentially observes narrow angular regions as it rotates, allowing it to locate an object that falls inside the transmitted beam by its range and azimuth. Received power depends on object reflectivity, surface area, orientation, and material. A wide spreadbeam in elevation and long wave- length allow multiple objects to be detected per transmission.
For each azimuth, radar outputs a one-dimensional signal, termed the power-range spectrum, which encodes the power reflected by the scatterers within the beam at each range. After a full rotation, radar returns a two-dimensional scan, as shown in Fig. 2. Radar scans are both information-rich and data-efficient, but they contain several unwanted artifacts visible in Fig. 2, including noise and false detections [3]. Another consideration of radar is its lower resolution and slower measurement update speeds compared to lidar. | |
220402385/3 | and use the RGB channels as axes of the encoded quaternion, following a computer vision-oriented approach. Parcollet et al. [0] presented two learning-based approaches to map real- valued vectors into the quaternion domain, by producing through a network four-channel representations of the input data that present meaningful intra-channel correlations. On the one hand, the Real to H-space encoder [0], applied to speech recognition tasks, consists of a simple real-valued dense layer applied at the beginning of a quaternion classifier network, which is trained jointly with the classifier. On the other hand, the Real to H-space Autoencoder, tested in the natural language processing field (conversation theme identification) [0] operates in an unsupervised way. Such a method contains a real-valued encoder and a quaternion-valued decoder, where the latter is expected to enable both the network’s embeddings and output to present meaningful intra-channel correlations that can be exploited by a quaternion-valued classifier network.
In this paper, we introduce RH-emo, a hybrid real- quaternion autoencoder-classifier architecture that is trained in a semi-supervised fashion in order to optimize each axis of the embedding dimension to different emotional characteristics: the first channel is optimized for discrete emotion recognition and the 3 other channels are individually optimized for the classification of valence, arousal, and dominance (as shown in Figure 1). RH-emo is intended to be used as a feature extractor that permits using QNNs for SER tasks with real-valued sig- nals without additional preprocessing. This approach has two advantages: it improves the performance of SER models even in situations where data is scarce and it drastically reduces the number of network parameters, consequently reducing the resource demand. We extend the approach of the quaternion autoencoder in [0] by specializing the learned quaternion representation for our specific task (SER), where the different axes are optimized for the detection of different emotional characteristics that are coherent with the most used criteria of emotion classification. Moreover, we implement it with a more complex architecture (deep convolutional autoencoder) and we apply it to a different domain: emotion recognition from speech audio.
## III. Q UATERNION CONVOLUTIONAL NEURAL NETWORKS
Operations between quaternion numbers are defined in the quaternions algebra $\mathbb{H}$. A quaternion Q is a four-dimensional extension of a complex number, defined as $\mathbf{q} = q_0 + q_1 \,\mathrm{i}\, + q_2 \ij + q_3 {k} = q_0 + q$, where, $q_0$, $q_1$, $q_2$ are real numbers, and $\,\mathrm{i}\,$, $\ij$ and ${k}$ are the quaternion unit basis. In this representation $q_0$ is the real part and $q_1 \,\mathrm{i}\, + q_2 \ij + q_3 {k}$ is the imaginary part, where $\,\mathrm{i}\,^{2} = \ij^{2} = {k}^{2} = -1$ and $\,\mathrm{i}\, \ij = - \ij \,\mathrm{i}\,$. From the latter assumption follows that the quaternion vector multiplication is not commutative. A quaternion can also be represented as a matrix of real numbers:
$$
\mathbf{q} = \mathopen{}\mathclose\bgroup\originalleft[{\begin{array}{*{20}c}
\hfill {q_0} & \hfill {- q_1} & \hfill {- q_2} & \hfill {- q_3} \\
\hfill {q_1} & \hfill {q_0} & \hfill {- q_3} & \hfill {q_2} \\
\hfill {q_2} & \hfill {q_3} & \hfill {q_0} & \hfill {- q_1} \\
\hfill {q_3} & \hfill {- q_2} & \hfill {q_1} & \hfill {q_0} \\
\end{array}} \aftergroup\egroup\originalright].
$$
Analogously to real and complex numbers, a set of opera- tions can be defined in the quaternion space:
- Addition: $\mathbf{q}+\mathbf{p} = (q_0 + p_0) + (q_1 + p_1)\,\mathrm{i}\, + \\ (q_2 + p_2)\ij + (q_3 + p_3){k}$
- Conjugation: $\mathbf{q}^{*} = q_0 - q_1 \,\mathrm{i}\, − q_2 \ij - q_3 {k}$
- Scalar multiplication: $\lambda \mathbf{q} = \lambda q_0 + \lambda q_1 \,\mathrm{i}\,+ \lambda q_2 \ij + \lambda q_3 {k}$
- Element multiplication (or Hamilton product):
$$
\begin{split}
\mathbf{q} \otimes \mathbf{p} &= \mathopen{}\mathclose\bgroup\originalleft(q_0 + q_1\,\mathrm{i}\, + q_2\ij + q_3{k}\aftergroup\egroup\originalright)\mathopen{}\mathclose\bgroup\originalleft(p_0 + p_1\,\mathrm{i}\, + p_2\ij + p_3{k}\aftergroup\egroup\originalright) \\
&= \mathopen{}\mathclose\bgroup\originalleft(q_0 p_0 - q_1 p_1 - q_2 p_2 - q_3 p_3\aftergroup\egroup\originalright) \\
&+ \mathopen{}\mathclose\bgroup\originalleft(q_0 p_1 + q_1 p_0 + q_2 p_3 - q_3 p_2\aftergroup\egroup\originalright)\,\mathrm{i}\, \\
&+ \mathopen{}\mathclose\bgroup\originalleft(q_0 p_2 - q_1 p_3 + q_2 p_0 + q_3 p_1\aftergroup\egroup\originalright)\ij \\
&+ \mathopen{}\mathclose\bgroup\originalleft(q_0 p_3 + q_1 p_2 - q_2 p_1 + q_3 p_0\aftergroup\egroup\originalright){k}.
\end{split}
$$
The quaternion convolutional neural network (QCNN) is an extension of the real-valued convolutional neural network to the quaternion domain. For each input vector of a quaternion layer, the dimensions are split into four parts to compose a quaternion representation. In a quaternion-valued fully- connected layer the parameters matrices are treated as a single quaternion entity with four components, even though they are manipulated as matrices of real numbers [0]. In a quaternion layer, the dot product operations used in real layers are replaced with the Hamilton product (eq. (2)) between the input vector and a quaternion-represented weight matrix. This allows the processing of all input channels together as a single entity maintaining original intra-channels dependencies because the weights submatrices are shared among the input channels. Consequently, quaternion layers permit to spare the 75% of free parameters compared to their real-valued equivalents because, as shown in eq. (2), the same components are re-used to build the output matrix.
In a QCNN, the convolution of a quaternion filter matrix with a quaternion vector is performed as the Hamilton product between the real-valued matrices representation of the input vector and filters. A quaternion convolution between a quater- nion input vector $\mathbf{x} = x_0 + x_1 \,\mathrm{i}\, + x_2 \ij + x_3 {k}$ and a quaternion filter $W = W_0 + W_1 \,\mathrm{i}\, + W_2 \ij + W_3 {k}$ can be defined as:
$$
W * x = \mathopen{}\mathclose\bgroup\originalleft[{\begin{array}{*{20}c}
\hfill {W_0} & \hfill {- W_1} & \hfill {- W_2} & \hfill {- W_3} \\
\hfill {W_1} & \hfill {W_0} & \hfill {- W_3} & \hfill {W_2} \\
\hfill {W_2} & \hfill {W_3} & \hfill {W_0} & \hfill {- W_1} \\
\hfill {W_3} & \hfill {- W_2} & \hfill {W_1} & \hfill {W_0} \\
\end{array}} \aftergroup\egroup\originalright] * \mathopen{}\mathclose\bgroup\originalleft[{\begin{array}{*{20}c}
{x_0} \hfill \\
{x_1} \hfill \\
{x_2} \hfill \\
{x_3} \hfill \\
\end{array}} \aftergroup\egroup\originalright].
$$
The optimization of quaternion-valued networks is identical to the one of a real network and can be achieved through regular backpropagation. This is possible because of the use of split activation and loss functions, as introduced in [0], [0]. These functions map a quaternion-like entity back to the real domain, consequently enabling the use of standard loss functions for the network training.
## IV. T HE P ROPOSED RH- EMO M ODEL
## A. Approach | |
210801937/4 | KAMIL NIEDZIA ℓ OMSKI
1.3.
## Complex approach
. We say that a complex $2$ – dimensional subspace $V$ in $\Delta$ is if for any $\varphi\in V^{\bot}$ we have $V\subset W_{\varphi}$.
Lemma 1.2. Fix (unit) $\varphi\in \Delta$. Then $V_{\varphi}$ is admissible.
Proof. By Lemma $V_{\varphi}=V_{\tilde{\varphi}}$, where $\tilde{\varphi}$ is as in the proof of Lemma 1.1. Take any linear combination $\psi$ of $\varphi$ and $\tilde{\varphi}$. Then, $y_{\psi}=y_{\varphi}$, hence $D_{\psi}=D_{\varphi}$. Moreover, for $x$ orthogonal to $y_{\psi}$ we have
\begin{equation*} x\cdot\psi=a x\cdot\varphi+b x\cdot\tilde{\varphi}\in V_{\varphi} \end{equation*}
for some $a,b\in\mathbb{C}$. Thus $V_{\psi}=V_{\varphi}$.
Lemma 1.3. Let $V$ be an admissible subspace. Then
- (1) there is unique unit vector $y\in\mathbb{R}^5$ such that $y\cdot\varphi=i\varphi$ for any $\varphi\in V^{\bot}$, $D_{\varphi}$ coincide for all $\varphi\in V^{\bot}$, $V=V_{\varphi}$ for any $\varphi\in V^{\bot}$.
Proof. Fix $\varphi\in V^{\bot}$. By Lemma 1.1 there is unique $y$ such that $y\cdot \varphi=i\varphi$ and $W_{\varphi}=V_{\varphi}\oplus\langle i\varphi\rangle$, where $V_{\varphi}=\{x\cdot\varphi\mid \langle x,y\rangle=0\}$. Thus $D_{\varphi}=\langle y\rangle^{\bot}$. Moreover, there is $\tilde{\varphi}$ which is $\mathbb{C}$ – linearly independent with $\varphi$ and such that $y\cdot\tilde{\varphi}=i\tilde{\varphi}$ and $\Delta=V_{\varphi}\oplus\langle \varphi,\tilde{\varphi}\rangle_{\mathbb{C}}$. Since $V$ is maximal complex in $W_{\varphi}$, by admissibility we have $V_{\varphi}=V$. This proves the third condition.
Now, take any $\psi\in V^{\bot}$. Then, $\psi=a\varphi+b\tilde{\varphi}$ for some $a,b\in\mathbb{C}$. Thus $y\cdot\psi=i\psi$, which implies $D_{\psi}=D_{\varphi}$, what proves the first and the second condition.
Theorem 1.4. Assume $V$ is admissible. Then $\mathfrak{su}(2)_{\varphi}$ coincide for all $\varphi\in V^{\bot}$. Conversely, for the maximal space $V^{\bot}$ such that all $\mathfrak{su}(2)_{\varphi}$ coincide for $\varphi\in V^{\bot}$ the orthogonal complement $V$ is admissible.
Proof. Assume $V$ is admissible. Take an orthonormal $\mathbb{C}$ – basis $\{\varphi, \tilde{\varphi}\}$ of $V^{\bot}$. By Lemma $D_{\varphi}=D_{\tilde{\varphi}}$. We will write just $D$. Choose an orthonormal basis $(e_j)$ of $D$ and let $(u_j)$ be a basis of $D$ such that $e_j\cdot\varphi=u_j\cdot\tilde{\varphi}$. Then
\begin{equation*} \langle u_j,u_k\rangle=\langle u_j\cdot\tilde{\varphi},u_k\cdot\tilde{\varphi}\rangle=\langle e_j\cdot\varphi,e_k\cdot\varphi\rangle=\langle e_j,e_k\rangle. \end{equation*}
Hence $(u_j)$ is also orthonormal. Moreover,
\begin{equation*} 0=\langle \varphi,\tilde{\varphi}\rangle=\langle e_j\cdot\varphi,e_j\cdot\tilde{\varphi}\rangle=\langle u_j\cdot\tilde{\varphi},e_j\cdot\tilde{\varphi}\rangle=\langle e_j,u_j\rangle. \end{equation*}
Fixing $e_1$ and corresponding $u_1$, we may take $e_2=u_1$. Hence, $u_1\cdot\varphi=u_2\cdot\tilde{\varphi}$. We have
\begin{equation*} \langle e_1,u_2\rangle=\langle e_1\cdot\varphi,u_2\cdot\varphi\rangle=\langle u_1\cdot\tilde{\varphi},u_2\cdot\varphi\rangle=-\langle u_2\cdot\tilde{\varphi},u_1\cdot\varphi\rangle=-\langle u_1\cdot\varphi,u_1\cdot\varphi\rangle=-1. \end{equation*}
Since both $e_1$ and $u_2$ are unit, it follows that $u_2=-e_1$. In particular, ${\rm span}\{e_1,e_2\}={\rm span}\{u_1,u_2\}$. Analogously, we set $e_4=u_3$ and we obtain $u_4=-e_3$. Consider the following $2$ – forms
$$
\omega_0=e_1\wedge u_1-e_3\wedge u_3,\quad
\omega_1=e_1\wedge e_3+u_1\wedge u_3,\quad
\omega_2=e_1\wedge u_3-u_1\wedge e_3.
$$
It is not hard to check that $\omega_j\cdot\varphi=\omega_j\cdot\tilde{\varphi}=0$ and $\omega_j$ are linearly independent. In particular $\omega_j\cdot\psi=0$ for any $\mathbb{C}$ linear combination of $\varphi$ and $\tilde{\varphi}$. Hence, all $\mathfrak{su}(2)_{\varphi}$ for $\varphi\in V^{\bot}$ coincide.
Conversely, let $\omega_0,\omega_1,\omega_2$ be $2$ – forms in $\mathbb{R}^5$ defining a Lie algebra $\mathfrak{g}$ isomorphic to $\mathfrak{su}(2)$ and let $V^{\bot}$ be a subspace of these $\varphi\in\Delta$ such that $\omega_j\cdot\varphi=0$. Clearly, $V^{\bot}$ is complex.
Fix $\varphi\in V^{\bot}$. By Lemma 1.1, $V_{\varphi}$ is complex $2$ – dimensional and orthogonal to $\varphi$. By Lemma 1.3, $V_{\varphi}$ is admissible. Since $\tilde{\varphi}\in V_{\varphi}^{\bot}$, where $\tilde{\varphi}$ is as above, by the first part $\mathfrak{su}(2)_{\tilde{\varphi}}=\mathfrak{su}(2)_{\varphi}=\mathfrak{g}$, i.e., $\omega_j\cdot\tilde{\varphi}=0$. Thus $\tilde{\varphi}\in V^{\bot}$. We have shown that $V_{\varphi}^{\bot}\subset V^{\bot}$. In other words, $V\subset V_{\varphi}\subset W_{\varphi}$. It suffices to show that $\dim_{\mathbb{C}}V=\dim_{\mathbb{C}}V^{\bot}=2$. | |
231017785/4 | The end-effector displacement $(d, z, \theta_y)$ corresponds to action space of the low-level agent, where $P_s$ is the starting pose and $p_{a^l}$ is the current end-effector task-space pose.
with the high-level agent from the previous section. In this paper we train only one such low-level agent — for the flipping primitive — though extensions to multiple learnable low-level agents are in principle possible.
Observation and action spaces. The action space for the low-level agent is discrete. Each action corresponds to a specific end-effector displacement $(d, z, \theta_y)$, shown in Fig. 3. In this context: $d \in \{0, a_d\}$ encodes the forward distance, $z \in \{0, a_z\}$ encodes the vertical movement, and $\theta_y \in \{-r_y, 0, r_y\}$ encodes the angular deviation along the $Y$ -axis relative to the primitive’s initial pose. The observation space of the low-level agent is a combination of the end- effector pose and the contact force. We formulate the state as $s_l = (p_{a^l}, f_d, f_{max})$, where $p_{a^l} \in \mathbb{R}^3$ is the current end- effector task-space pose; $f_d$ is the contact force along the $d$ - axis; and $f_{max}$ is the maximum of the current contact force. To facilitate efficient training, we develop an observation space designed to remain invariant to the starting pose of the low-level agent, ensuring state consistency across different initial poses.
Reward. To learn the low-level action policy, we could in a straightforward manner define the reward sparsely on successful flips. However, as sparse rewards result in less efficient learning [28], we choose to instead design a reward function that considers the contact force and end-effector position:
$$
r_\tau = r^{H_f}_t + r^L_\tau
$$
$$
r^L_\tau =
\begin{cases}
\min(\sigma, \frac{z_\tau \sigma}{w}) & \text{, if $f_c>0$} \\
-1 & \text{, if $f_c>f_{limit}$} \\
0 & \text{, otherwise}
\end{cases}
$$
where $f_c$ is the current contact force, $f_\mathit{limit}$ is the maximum safety contact force, $z_\tau$ is the current end-effector height, and $\sigma$ and $w$ are hyper-parameters that normalize the $z_\tau$ and limit the upper bound of the reward. Based on this reward function, we encourage the agent not only to flip up the object but also to raise it up with contact and avoid applying too much contact force. We find that without the penalty for applying too much contact force, the robot may trigger emergency stops in the real world, making the transfer of policies learned in simulation more challenging.
Initial pose invariant state. To further improve the train- ing efficiency, we ensure the low-level model doesn’t need to adapt to various initial poses. We establish the starting pose provided by the high-level agent as the reference base frame for $p_{a^l}$. Consequently, $p_{a^l}$ represents a projection of the end- effector pose with respect to the forward direction, vertical displacement, and rotation along the $Y$ -axis, all based on the initial pose frame.
Model architecture. For more details of our low-level agent, we employ a multilayer perceptron (MLP) as the underlying model. As indicated in Fig. 2, we combine $p_{a^l}$, $f_d$, and $f_{max}$ as inputs to the MLP, which in turn outputs Q values corresponding to a specific end-effector displacement to control the robot. Once selected for execution, the low- level agent controls the robot iteratively, taking a fixed number of actions within a horizon $T$.
## C. Curriculum Learning and Domain Randomization
We follow a curriculum learning [13] approach to train our high-level and low-level agents separately. Learning the high- level agent is hard when the low-level agent is not competent in its sub-task because the high-level reward is conditioned on whether the low-level task is completed successfully. We train the low-level agent first, devising progressively more complex interaction scenarios and only include the high-level model once the low-level policy can successfully flip objects.
To adapt to the domain shift between simulation and the real world and perform zero-shot sim2real transfer, we employ Automatic Domain Randomization [14] during the simulation training. In each episode, we randomly sample the object size, friction, and mass to make our method able to address varied box-shaped objects. To overcome the noise of the height map in the real world, we add Gaussian noise and randomly block a few regions as fake reflections in the simulated depth image.
## A. Setup and Evaluation Metrics
To demonstrate our model’s ability to learn extrinsic dexterity, we use a Franka Emika Panda arm with a 2- finger gripper that does not open wide enough to grasp the target objects (shown in Table II) from above directly. For the environmental setup, we use the inside of a $44.8 \times 44.8$ (cm) box as the robot’s workspace, with the four boundaries serving as potential locations for performing an object pivot. An overhead depth image is captured by a Kinect v2 camera and transformed into a height map as input to the high-level model. To avoid the robot occluding the workspace, we move the robot to the bottom-left corner before acquiring a depth image.
We use the completion rate as the evaluation metric and test 10 episodes for each object. An episode is considered successful if the robot picks up the object within a certain number of primitive actions. Similarly to above, the success rate for each primitive is the number of successful actions divided by the total number of attempted actions with that primitive, averaged over the 10 episodes. This metric does | |
170402102/6 | the emergence of the Fe-Fe exchange parameter only at $x=0.5$. The strength of this interaction lies in between that of the dominant Cr $_{B}$ -Fe $_{B}$ and weaker Cr $_{B}$ -Cr $_{B}$ in- teractions, and would add to the strength of the total J $_{BB}$, thus competing closely with total J $_{AB}$ interactions. Thus, with the ”normal” configuration, one would not achieve a collinear magnetic structure in this system. In Figs. 5, 6 we observe the effects of the cation disorder on the relative strengths of the exchange interactions. In the ”half-inverse” configuration ($y=0.5$), the A-B interactions overwhelmingly dominate over the B-B in- teractions for $0.5 \leqslant x \leqslant 1$. It may be noted that modelling of a ”half-inverse” structure for $x < 0.5$ would have required larger unit cells and hence we refrained from modelling the ”inverse” structures for smaller $x$ values. This would not pose any serious problem as we are interested only to find the trends in different exchange interactions when the state of cation disorder changes, in order to under- stand the origin of the emergence of the collinear struc- ture at one end. Analysing Fig. 5, we further see that the dominant A-B interaction for $y=0.5$ comes from the Fe atoms at crystallographic sites of different sym- metry. The three other dominant A-B exchange interac- tions Co $_{A}$ -Fe $_{B}$,Fe $_{A}$ -Co $_{B}$ and Fe $_{A}$ -Cr $_{B}$ compete with each other. On the other hand, the dominant B-B interaction is between Fe and Cr atoms at B sites as was the case for $y=0$. The picture remains same for $y=1$, that is, for ”complete inverse” configuration. Thus, it can be con- cluded that in Co $\left(Cr_{1-x}Fe_{x} \right)_{2}$ O $_{4}$, a collinear magnetic structure would emerge due to the cation disorder among A and B sites, and would certainly be achieved when Fe composition would reach $50 \%$.
The question now re- mains as to why a particular magnetic exchange parame- ter dominates over the others and how that changes with changes in the degree of cationic disorder. In the next section, we address this question by analysing the elec- tronic structures as a function of composition and degree of cation disorder.
## Electronic structures and their variations with $x$ and $y$
We present results on the electronic structures with variations in $x$ for three degrees of cation disorder repre- sented by $y=0,0.5$ and $1$ in Figures 7, 8, 9 respectively. Quite interestingly, the qualitative features in the atom- projected densities of states remain absolutely same irre- spective of $x$ and $y$. A careful look at the figures suggest the following: for any $x$ and $y$, Co $_{A}$ t $_{2g}$ states are half- filled and the $e_{g}$ states are completely filled, Cr t $_{2g}$ states are half-filled while e $_{g}$ states are completely empty, Fe states at both A and B sites have both t $_{2g}$ and e $_{g}$ states half-filled while Co $_{B}$ t $_{2g}$ states are more than half-filled and the e $_{g}$ states are half-filled. Thus, the band filling as- sociated with the atoms resemble that of pure CoCr $_{2}$ O $_{4}$ and CoFe $_{2}$ O $_{4}$ 17. Such robustness of the electronic struc- tures originates from strong and comparable crystal fields
The absolute values of different magnetic exchange parameters calculated in the present work as a function of Fe concentration $x$ in Co $\left(Cr_{1-x}Fe_{x} \right)_{2}$ O $_{4}$. The B-B exchange interactions(in meV) are indicated by $|J_{BB}|$, while the A-B exchange interactions(in meV) are indicated by $|J_{AB}|$. The results are for the ”normal” spinel structure, that is, for $y=0$.
associated with the magnetic cations as was seen in cases of pure CoCr $_{2}$ O $_{4}$ and CoFe $_{2}$ O $_{4}$ 17. The positions of the major peaks associated with each individual atoms also do not change their positions considerably with changes in $x$ and $y$. The only change that one can see is the re- distributions of the weights associated with a particular atom type when $x$ changes, keeping $y$ fixed.
The features in the densities of states can now be con- veniently used to explain the trends in the magnetic ex- change interactions. For $y=0$, the results presented in Fig.?? show that Co $_{A}$ -Fe $_{B}$ exchange interaction closely compete with Cr $_{B}$ -Fe $_{B}$ one. Since Co $_{A}$ and Fe $_{B}$ have half-filled t $_{2g}$ and e $_{g}$ states respectively, this interaction is strongly antiferromagnetic 26. On the other hand, both Cr $_{B}$ and Fe $_{B}$ t $_{2g}$ orbitals are half-filled. Direct inter- action between them is possible which strengthens the overall B-B interaction 26. As a result, there is close com- petition between $J_{AB}$ and $J_{BB}$ exchange interactions for $y=0$. The strength of the dominant $J_{BB}$, however, is weaker in comparison to the dominant $J_{AB}$ due to the fact that the inter-cation distance at the B site is larger than that at the A site as can be understood from the bond distances presented in Table I. For $y=0.5,1$, the overwhelmingly dominant A-B interaction is due to the Fe atoms at the A and the B sites. In this case, both A site t $_{2g}$ and B site e $_{g}$ orbitals are half-filled mak- ing the interaction strongly antiferromagnetic. The rela- tively much weaker B-B interaction, mostly coming from Cr-Fe pairs are due to the fact that due to inversion, the Fe content at B site decreases in comparison to the | |
180808228/6 | Initial constraint violations on the $z=0$ plane near the larger hole of a BBH system, for horizon boundary conditions (left) and negative expansion boundary conditions (right). Colors show the magnitude of the Hamiltonian-momentum constraint energy (cf. Eq. 42), the yellow circle is the apparent horizon, and the large black area inside the horizon is the excision region. Here superposed Kerr-Schild free data are used to construct a BBH with mass ratio $q=1.1$ and spins $\chi_{1z}=-0.3$ and $\chi_{2z}=-0.4$ along the direction of orbital angular momentum. Unlike the horizon boundary conditions, the negative expansion boundary conditions require no extrapolation inside the horizon, and thus yield constraints near and inside the apparent horizon that are about 3 orders of magnitude smaller.
The idea behind the new boundary conditions is to set the expansion not to zero, but to some nonzero value that ensures that the excision boundary is inside an apparent horizon rather than on one. We use Eq. (14) to modify the conformal factor boundary condition at $\mathcal{B}_{E}$ to:
$$
\begin{aligned}
\bar{s}^k~\partial_k \psi &=& - \frac{\psi^{3}}{8 N} \bar{s}^i \bar{s}^j
\left((\bar{\mathbb{L}} N)_{ij} - \bar{u}_{ij} \right) \\
&&-\frac{\psi}{4} \bar{h}^{ij} \bar{\nabla}_i \bar{s}_j
+\frac{1}{6} K \psi^3 + \frac{\psi^3}{4} \Theta_{\alpha},
\end{aligned}
$$
where $\alpha$ denotes the particular BH and $\Theta_{\alpha}$ is computed from the single BH metrics used in Eq. 8. As we choose the excision surface to be slightly inside the single BH horizons, $\Theta_{\alpha}$ is negative on the surface. Henceforth we refer to this boundary condition as a negative expansion boundary condition.
When imposing the negative expansion condition, we also need to modify the shift boundary condition, as Eq. (19) holds only on a horizon. Noting that for a single BH, $\epsilon = N_{\bot} - N$ is positive inside the horizon and neg- ative outside, we modify the boundary condition at $\mathcal{B}_{E}$ for the normal component of shift to:
\begin{gather} N_{\bot} = N + \epsilon_{\alpha}, \end{gather}
where $\epsilon_{\alpha} = N_{\bot\alpha} - N_{\alpha}$ are again obtained from the single BH solutions of the individual holes.
For negative expansion boundary conditions, we con- tinue to use Eq. (20) for the tangential part of the shift. We also continue to use Eq. (24) for the boundary condi- tion on the lapse, with $N_{\alpha}$ evaluated at the new location of the inner boundary. We find that the procedure for setting the spin via iteration over $\Omega_r^{(k)}$, as described in Sec. IIB1, works just as well in the case of a negative expansion BC as it does for a horizon BC.
Figure 1 demonstrates the efficacy of these new bound- ary conditions; shown are the constraints near the larger black hole when using horizon boundary conditions and the new negative expansion boundary conditions. When using negative expansion boundary conditions, the con- straints improve by about 3 orders of magnitude inside and near the apparent horizon. Note, however, that once the evolution begins, most of this constraint violation propagates inwards into the excision surfaces and out of the computational domain. This is because in the gen- eralized harmonic formalism the evolution of constraint violations is governed by a wave equation [20], which ensures that constraint violations propagate causally. Hence, we do not expect the new boundary conditions to reduce constraint violations during the evolution nearly as much as they improve initial constraint violations.
## Gauge choices
SpEC uses the generalized harmonic evolution sys- tem [20 – 23] to evolve the initial data. In this formalism, the gauge choice is set by requiring the coordinates to satisfy an inhomogeneous wave equation,
$$
-{^{(4)}}\Gamma^a = \nabla^c \nabla_c x^a = H^a,
$$
where $^{(4)}\Gamma^a=\psi^{bc} \, {^{(4)}}\Gamma^a_{bc}$, $\psi_{ab}$ is the spacetime metric, | |
220409569/11 | as the 90% C.L. limit on $T$).
As a final step, one could include a joint probability distribution or a $\Delta\chi^2$ penalty defined over the NME landscape $(M_\mathrm{Xe},\,M_\mathrm{Ge},\,M_\mathrm{Te})$, and numerically minimize the total $\Delta\chi^2$ function. This exercise was performed in [3] by assuming a conservative characterization of the NME and their correlated uncertainties, derived within QRPA calculations [27]. A limit $m_{\beta\beta}<110$ meV was obtained at $2\sigma$ [3]. By repeating the same exercise with the updated (Xe+Ge+Te) combination considered herein, we get the following marginalized bounds (in meV): $m_{\beta\beta}<79.5$ at 90% C.L., $m_{\beta\beta}<99.8$ at $2\sigma$, and $m_{\beta\beta}<169$ at $3\sigma$. Roughly speaking, from these results and from the summary in Eq. (20) one can state that the combination of current $0\nu\beta\beta$ experiments sets $2\sigma$ upper bounds on $m_{\beta\beta}$ at the level of $\sim 90$ – 100 meV for “ average ” NME values, possibly lowered to $\sim 40$ – 50 meV for favorable NME values.
A final remark is in order. In principle, one should replace the QRPA input from [27] with more general and up-to-date estimates of the NME’s and their uncertainties, characterizing also the spread among different models and calculations. However, no consensus estimates exist yet for NME fiducial values and covariances, although relevant work is in progress toward this goal [1, 31]. Part of the planned strategy involves benchmarking nuclear models for $0\nu\beta\beta$ decay against a variety of data, coming from related electroweak and strong interaction processes or from
## CONCLUSIONS AND PERSPECTIVES
We have discussed an approach to the analysis of neutrinoless double beta decay experiments, in terms of $\Delta\chi^2$ profiles for the signal strength $S_i$ (inverse of the half-life $T_i$) in the isotopes $i=$ Xe, Ge and Te, building upon previous work [3]. The approach becomes exceedingly simple for quadratic approximations to such profiles, implying quadratic (in)equalities in the landscape of nuclear matrix elements $M_i$ that connect the $S_i$ to the Majorana mass $m_{\beta\beta}$. For convenience, some results have been discussed in terms of pseudo data for nan signal in Te (dubbed Te $^*$). Simple relations among the $M_i$ have been derived to gauge the relative contributions of different isotopic data in setting upper limits to $m_{\beta\beta}$ (for nan best fits in Xe, Ge and Te $^*$), and to identify the conditions leading to a preference for nonzero $m_{\beta\beta}$ (for generic Xe, Ge and Te data). Using the latest available $0\nu\beta\beta$ data, as well as representative values of the NME from different models, we have discussed current constraints on $m_{\beta\beta}$ at several confidence levels and in various combinations, both numerically and graphically. Global $2\sigma$ upper limits on $m_{\beta\beta}$ are found in the range from 40.5 to 135 meV, depending on the NME.
The approach can be easily extended to nonstandard processes for $0\nu\beta\beta$ decay [1, 36, 37] by replacing the relation $S_i=G_i M^2_i m^2_{\beta\beta}$ with the appropriate phase space, NME and particle physics parameter characterizing the process. Also, the approach can be extended to generic $\Delta\chi^2(S_i)$ functions, with a modest price to pay in terms of numerical
FIG.7: Significance of upper limits on $m_{\beta\beta}$ in terms of $N_\sigma=(\Delta\chi^2)^{1/2}$, from the combination of current Xe+Ge+Te data, for the representative NME calculations considered in this work. | |
201200671/1 | # ML-based Flood Forecasting: Advances in Scale, Accuracy and Reach
Sella Nevo Google Research [email protected] Gal Elidan Google Research [email protected] Avinatan Hassidim Google Research [email protected] Guy Shalev Google Research [email protected] Oren Gilon Google Research [email protected] Grey Nearing Google Research [email protected] Yossi Matias Google Research [email protected]
Floods are among the most common and deadly natural disasters in the world, and flood warning systems have been shown to be effective in reducing harm. Yet the majority of the world’s vulnerable population does not have access to reliable and actionable warning systems, due to core challenges in scalability, computational costs, and data availability. In this paper we present two components of flood forecasting systems which were developed over the past year, providing access to these critical systems to 75 million people who didn’t have this access before.
## Introduction
## Flooding harms and early warning impact
Floods are the most common, and among the most deadly, natural disaster on the planet, affecting hundreds of millions of people every year, and causing between thousands and tens of thousands of fatalities annually [1]. Recent research shows that those who receive flood warnings prior to the arrival of the flood are twice as likely to evacuate or take other protective measures than those who don’t receive such warning, with 65% of those warned taking action [2].
As a result, warning systems have the potential to save lives and reduce harm to livestock and other assets. For this reason, the World Bank has identified early warning systems for floods as the most cost-effective tool for climate change adaptation amongst all options evaluated. It estimated that for each dollar spent on such systems, nine dollars of damages are prevented [3]. Machine learning is well suited to help provide reliable operational flood forecasts at large scales [4].
## Challenges
The vast majority of populations vulnerable to flooding do not have access to reliable, actionable flood warnings [5]. There are several reasons why this is a difficult challenge, including the following:
- Local calibration: Existing high-accuracy flood models often require large amounts of per-site data and effort to set up and calibrate [6]. This limits the scale at which such systems | |
180804872/17 | From (28), applying triangle inequality,
$$
\begin{aligned}
&&\|\rho^{\star}\rho (\psi_{n,j})-|\rho_{j}|^{2}\psi_{n,j}\|_{H}\leq \|\rho^{\star}\rho (\psi_{n,j})-
[\mathcal{D}_{n}\mathcal{C}_{n}^{-1}]^{\star}[\mathcal{D}_{n}\mathcal{C}_{n}^{-1}](\psi_{n,j})\|_{H}
\\
&&+\|[\mathcal{D}_{n}\mathcal{C}_{n}^{-1}]^{\star}[\mathcal{D}_{n}\mathcal{C}_{n}^{-1}](\psi_{n,j})-|\rho_{j}|^{2}\psi_{n,j}\|_{H}\\
&&\leq 2\|\rho^{\star}\rho-[\mathcal{D}_{n}\mathcal{C}_{n}^{-1}]^{\star}[\mathcal{D}_{n}\mathcal{C}_{n}^{-1}]\|_{\mathcal{L}(H)}.
\end{aligned}
$$
On the other hand,
$$
\begin{aligned}
&&\|\psi_{n,j}-\psi_{n,j}^{\prime}\|_{H}^{2}=
\sum_{l=1}^{\infty}\left[\left\langle \psi_{n,j},\psi_{l} \right\rangle_{H}-\mbox{sgn}\left\langle \psi_{n,j} , \psi_{l} \right\rangle_{H}\left\langle \psi_{j}, \psi_{l}\right\rangle_{H}\right]^{2}
\\
&&=\sum_{l\neq j}\left[\left\langle \psi_{n,j},\psi_{l} \right\rangle_{H}\right]^{2}+\left[\left\langle \psi_{n,j},\psi_{j} \right\rangle_{H}-\mbox{sgn}\left\langle \psi_{n,j} , \psi_{j} \right\rangle_{H}\right]^{2}\\
&&=\sum_{l\neq j}\left[\left\langle \psi_{n,j},\psi_{l} \right\rangle_{H}\right]^{2}+\left[1-\left|\left\langle \psi_{n,j},\psi_{j} \right\rangle_{H}\right|\right]^{2}\\
&&=\sum_{l\neq j}\left[\left\langle \psi_{n,j},\psi_{l} \right\rangle_{H}\right]^{2}+\sum_{l=1}^{\infty}\left[\left\langle \psi_{n,j},\psi_{l} \right\rangle_{H}\right]^{2}-2\left|\left\langle \psi_{n,j},\psi_{j} \right\rangle_{H}\right|+\left|\left\langle \psi_{n,j},\psi_{j} \right\rangle_{H}\right|^{2}\\
&&\leq 2\sum_{l\neq j}\left[\left\langle \psi_{n,j},\psi_{l} \right\rangle_{H}\right]^{2}.
\end{aligned}
$$
Furthermore,
$$
\begin{aligned}
&&\|\rho^{\star}\rho (\psi_{n,j})-|\rho_{j}|^{2}\psi_{n,j}\|_{H}^{2}=
\sum_{l=1}^{\infty}\left[\left\langle \psi_{n,j},|\rho_{l}|^{2}\psi_{l}\right\rangle_{H}-\left\langle \psi_{n,j},|\rho_{j}|^{2}\psi_{l}\right\rangle_{H}\right]^{2}\\
&&\geq \min_{l\neq j}\left||\rho_{l}|^{2}-|\rho_{j}|^{2}\right|^{2}\sum_{l\neq j}\left[\left\langle \psi_{n,j},\psi_{l} \right\rangle_{H}\right]^{2}\\
&&\geq \min_{l\neq j}\left||\rho_{l}|^{2}-|\rho_{j}|^{2}\right|^{2}\frac{1}{2}\|\psi_{n,j}-\psi_{n,j}^{\prime}\|_{H}^{2}\\
&&\geq \alpha_{j}^{2}\frac{1}{2}\|\psi_{n,j}-\psi_{n,j}^{\prime}\|_{H}^{2},
\end{aligned}
$$
| |
220803121/21 | In this paper we introduced MAP-independence as a formal notion, rel- evant for decision support and justification of decisions. In a sense, MAP- independence is a relaxation of conditional independence, suggested by Pearl [30] to be a scaffold for human context-dependent reasoning. We suggest that MAP-independence may be a useful notion to further explicate the variables that are relevant for the establishment of a particular MAP explanation. Es- tablishing whether the MAP explanation is MAP-independent from a set of variables given the evidence (and so, whether these variables are relevant for justifying the MAP explanation) is a computationally intractable problem; we expect that the problem of finding the maximal set for which the MAP explanation is MAP-independent is even harder, modulo commonly assumed separation properties of the counting hierarchy. However, for a specific vari- able of interest $I$ (or a small set of these variables together) the problem is tractable whenever MAP can be computed tractably; in practice, this may suffice for usability in typical decision support systems.
There are many related problems of interest that one can identify, but which will be delegated to future work. For example, if the set of relevant variables is large, one might be interested in deciding whether observing one variable can bring down this set (by more than one, obviously). Initial work in this area has been proposed by [14] who looked at several heuristic approaches (e.g., Gini-index [33] and linear-value utility [10]) to decide which variable may best be observed.
Another related problem would be to decide upon the observations that are relevant for the MAP explanation (i.e., had we not observed $E \in \mathbf{E}$ or had we observed a different value, would that change the MAP explanation?). This would extend previous work [26] where the relevance of $E$ for computing a posterior probability (conditioned on $\mathbf{E}$) was established.
The notion of MAP-independence can, for practical purposes, be to strict; furthermore, when an explanation is MAP-dependent on a set of variables, the ‘amount of impact’ can be relevant. Various authors have proposed a form of quantification of MAP-independence, for example, based on the summed probability of joint value assignments to $\mathbf{R}$ for which the MAP explanation does not change [37], the proportion of such joint value assignments (ignoring their probability) [7], or the average structural impact (in terms of Hamming distance) on the best explanation [13]. A thorough overview of these and possibly other ways of quantifying MAP-independence would be welcome.
Finally, in order to test its practical usage, the formal concept introduced in this paper should be put to the empirical test in an actual decision support | |
230410708/3 | Moreover, when $G$ is the finite unitary group, we prove same re- sult (Theorem 1.4) by “ Ennola duality ”. V. Ennola [?enn] conjectured certain class functions on the finite unitary group $\mathrm{U}_n$ obtained from the irreducible characters of $\mathrm{GL}_n(q)$ is the irreducible characters of $\mathrm{U}_n$. Ennola’s conjecture is proved by N. Kawanaka [?ka] and called Ennola duality.
Theorem Let $q$ be a power of a prime. Then $h(\mathrm{U}_n(q)) \in \mathbb{Z}[q^{-1}].$ Moreover, for any fixed $n$, there exists an integer $q_n$ such that if $q>q_n$, then $h(\mathrm{U}_n) \in \mathbb{Z}$.
I would like to thank my advisor, Scott Carnahan for his continuous guidance. This work was supported by JST SPRING, Grant Number JPMJSP2124.
## 2. Representations of $\mathrm{GL}_n(k)$
First, we classify conjugacy classes of $\mathrm{GL}_n(k)$ by characteristic poly- nomials and partitions.
The order of $\mathrm{GL}_n(k)$ is
$$
|\mathrm{GL}_n(k)|=q^{\binom{n}{2}}\psi_n(q)
$$
where $\psi_n(q)=(q-1)(q^2-1)\cdots(q^n-1)$.
For a partition $\lambda= (\lambda_1, \cdots \lambda_l)$ with $\lambda_1 \geq \cdots \geq \lambda_l >0$, we write $l(\lambda)=l$ for the length of $\lambda$ and $\lambda'= (\lambda'_1, \cdots \lambda'_{l'})$ for the conjugate partition of $\lambda$.
Let $f$ be a monic polynomial $f(t) = t^d - a_{d-1}t^{d-1} -\cdots - a_0 \in k[t]$. We define the $d\times d$ matrix
$$
U_1(f)=
\begin{pmatrix}
& 1 & & \\
& & \ddots & \\
& & & 1 \\
a_0 & a_1 & \cdots & a_{d-1} \\
\end{pmatrix}
$$
whose characteristic polynomial is $f(t)$, the $rd \times rd$ matrix
$$
U_r(f)=
\begin{pmatrix}
U_1(f) & I_d & & \\
& \ddots & \ddots & \\
& & & I_d \\
& & & U_1(f) \\
\end{pmatrix}
$$
| |
170803480/9 | Figure2: Unstrained premixed flamelet solution at $\phi = 1$ for the number density of electrons obtained using mechanism “ A ” (symbols); number density of e $^-$ (), of O $_2^-$ (), of OH $^-$ () and of O $^-$ () obtained using mechanism “ B ”. A close-up view of the lower left corner of the graph is shown by the inset.
Figure 3: Comparison of the unstrained premixed flamelet solution at $\phi = 1$ obtained using the FlameMaster code and the proposed model considering mechanism “ A ”. Symbols represent the profiles of P () and temperature (), computed using the detailed chemistry. The dashed and solid lines correspond to the solutions obtained with our CFD code.
These species have multiple effects on the mixture. Firstly, they absorb part of the electrons preventing the recombination process of the H $_3$ O $^+$ and slightly increasing the number of charged particles in the mixture. Secondly, they strongly reduce the mobility of the negative charges being much heavier than the electrons. This kinetic mechanism has been employed to test the capabilities of the present model to handle these phenomena without increasing the numerical cost of the computation. For both mechanisms, the transport properties of the heavy ions have been assumed equal to those of the corresponding neutral species given by the GriMech 3.0 [32] database. The thermal properties of the ions are instead taken from the database by Burcat and Ruscic [34]. Although the validation of these two mechanisms has been already provided in Belhi et al. [20, 21], we analyze the results for the stoichiometric ($\phi = 1$) unstrained flamelets computed for the conditions of the test case described in Section 6. The calculations have been performed using the freely distributed C++ code FlameMaster V3.3.10 [35].
Figure 1 shows temperature and P profiles for both the mechanisms. The two temperature profiles coincide, confirming that the change of the ionization mechanism has a negligible influence on the combustion process. Indeed, the charged species constitute only a minor part of the mixture and there are two or, in some cases, three orders of magnitude between the molar fractions of the combustion radicals and those of the anions and cations. On the other hand, the peak value of P, which, as expected, is located close to the flame-front for both mechanisms, is slightly lower for mechanism “ A ”. The difference between the two peak values is about $1\%$ and is due to the employment of the electrons in the production of the heavy charged species. Figure 2 shows the breakdown of the negative species produced by the two mechanisms, in order to observe their influence on the transport properties of the charged particles. As expected the main negative species for the mechanism “ B ” is the electron, whose number density has a profile almost identical to that predicted by mechanism “ A ”. The main difference between the two profiles is in the upstream part the flame front, which is shown in the inset. In this region, the dominant negative species in the mixture is OH $^-$, whose density is in some points even higher than that of the electrons. For this reason, the mobility of the negative species computed with Eq. (19) is much lower for the mechanism “ B ”, entailing a different response of the mixture to an applied electric field.
## 4.2. 1D verification of the method
In order to verify the correct implementation of the model in our CFD solver and to provide a first validation of the present approach, we have computed the two flamelet solutions provided in the previous subsection with the proposed FPV approach. Even though this may seem a trivial test case, it requires that the production rates, the diffusive fluxes as well as the drift induced by the electric field have to be well resolved in order to obtain a good agreement between the results of the FlameMaster code and of the CFD solver.
The CFD solution has been obtained using a one-dimensional grid composed of 3001 nodes evenly distributed over a total length of 40mm. Dirichlet conditions have been imposed at the first left point of the computational domain, enforcing the laminar planar flame speed, the mixture fraction relative to the case at $\phi = 1$ and the progress-variable equal to zero. The concentration of positive and negative ions has also been set to zero at this point, considering that this configuration should not produce any ion-flux. Moreover, the large distance considered between the inlet plane and the flame-front (about 1. 6mm) ensures | |
171003913/14 | Since $R_G$ preserves the fibers of $\xi_{O_0},$ then $(R_G)_*$ maps $V_P \xi_{O_0} (\mathbb{H})$ into $V_{R_G(P)} \xi_{O_0} (\mathbb{H}).$
Now, we are ready to define the concept of quantum connection over the observable space.
Definition6.5. Fix $O_0 \in \mathcal{W} (\mathbb{H}).$ A connection on the principal fiber bundle $\xi_{O_0}= (\mathcal{U} (\mathbb{H}), \mathcal{W} (\mathbb{H}), \Pi_{O_0}, \mathcal{G}_{O_0})$ is a family of linear functionals $\Omega = \{\Omega_P:\; P \in \mathcal{U} (\mathbb{H}) \},$ where for each $P \in \mathcal{U} (\mathbb{H}),$
$\Omega_P$ is
a linear functional in $T_P \xi_{O_0} (\mathbb{H})$ with values in $\mathrm{g}_{O_0},$ satisfying the following conditions:
- (1) For any $P \in \mathcal{U} (\mathbb{H})$ and for all vertically tangent vectors $Q \in V_P \xi_{O_0} (\mathbb{H}),$ one has
$$
\Omega_P (Q) = P^{-1} Q.
$$
$\Omega_P$ depends
- continuously on $P,$ in the sense that if $P_n$ converges to $P$ as well as $Q_n \in T_{P_n} \xi_{O_0} (\mathbb{H})$ converges $Q_0 \in T_P \xi_{O_0} (\mathbb{H})$ in the operator topology of $\mathcal{B} (\mathbb{H}),$ then $\lim_{n \rightarrow \infty} \Omega_{P_n} (Q_n) = \Omega_P (Q_0)$ in $\mathrm{g}_{O_0}.$
- (3) Under the right action of $\mathcal{G}_{O_0}$ on $\xi_{O_0} (\mathbb{H}),$
$\Omega$ transforms
- according to
$$
\Omega_{R_G(P)} [(R_G)_* (Q)] = G^{-1} \Omega_P (Q) G,
$$
- for $G \in \mathcal{G}_{O_0},$
$P \in \mathcal{U} (\mathbb{H}),$ and
- $Q \in T_P \xi_{O_0} (\mathbb{H}).$
Such a connection is simply called an $O_0$ -connection.
Next, we present a canonical example of such quantum connections, which plays a crucial role in the expression of observable-geometric phases.
Example Given a fixed $O_0 = \{|e_n \rangle\langle e_n|: 1 \le n \le d \} \in \mathcal{W} (\mathbb{H}),$ we define $\check{\Omega} = \{\check{\Omega}_P: P \in \mathcal{U} (\mathbb{H}) \}$ as follows: For each $P \in \mathcal{U} (\mathbb{H}),$
$\check{\Omega}_P : T_P \xi_{O_0} (\mathbb{H}) \mapsto \mathrm{g}_{O_0}$ is
defined by
$$
\check{\Omega}_P (Q) = P^{-1} \star Q
$$
for any $Q \in T_P \xi_{O_0} (\mathbb{H}),$ where
$$
P^{-1} \star Q = \sum^d_{n = 1} \langle e_n | P^{-1} Q | e_n \rangle |e_n \rangle \langle e_n|.
$$
By, one has $P^{-1} \star Q = P^{-1} Q \in \mathrm{g}_{O_0}$ for any $Q \in V_P \xi_{O_0} (\mathbb{H}),$ namely $\check{\Omega}_P$ satisfies. The conditions (2) and (3) of Definition 6.5 are clearly satisfied by $\check{\Omega}.$ Hence, $\check{\Omega}$ is an $O_0$ -connection on $\xi_{O_0}.$ In this case, we write $\check{\Omega}_P = P^{-1} \star d P$ for any $P \in \mathcal{U} (\mathbb{H}).$
6.4.
## Quantum parallel transportation
. The quantum parallel transportation in the state space was introduced in [14,1] and studied in [3] in details. This section is devoted to the study of quantum parallel transport over the observable space.
Definition Fix a point $O_0 \in \mathcal{W} (\mathbb{H}).$ For a continuous curve $C_W: [a, b] \ni t \longmapsto O (t) \in \mathcal{W} (\mathbb{H}),$ a lift of $C_W$ with respect to $O_0$ is defined to be a continuous curve
$$
C_P: [a, b] \ni t \longmapsto U(t) \in \mathcal{U} (\mathbb{H})
$$
such that $U(t) \in \mathcal{F}^{O(t)}_{O_0}$ for any $t\in [a, b].$
Remark 6.3. Note that, a lift of $C_W$ depends on the choice of the point $O_0;$ for the same curve $C_W,$ lifts are distinct for different points $O_0.$ For this reason, such a lift $C_P$ is called a $O_0$ -lift of $C_W.$
Definition Fix a point $O_0 \in \mathcal{W} (\mathbb{H}).$ A continuous curve $C_W: [a, b] \ni t \longmapsto O (t) \in \mathcal{W} (\mathbb{H})$ is said to be smooth, if it has a $O_0$ -lift $C_P: [a, b] \ni t \longmapsto U(t) \in \mathcal{U} (\mathbb{H})$ which is a smooth curve. In this case, $C_P$ is called a smooth $O_0$ -lift of $C_W.$
Note that, if a continuous curve $C_W: [a, b] \ni t \longmapsto O (t) \in \mathcal{W} (\mathbb{H})$ is smooth, then it is differential at every point $t \in [a, b].$ Indeed, suppose that $C_P: [a, b] \ni t \longmapsto U(t) \in \mathcal{U} (\mathbb{H})$ is a smooth $O_0$ -lift of $C_W.$ For each $t \in [a, b],$ we have $\frac{d C_P (t)}{d t} \in \frac{d O(t)}{d t},$ namely $\frac{d O(t)}{d t}$ is a nonempty subset of $\mathcal{B} (\mathbb{H}),$ and hence $C_W$ is differential.
Definition Fix $O_0 \in \mathcal{W} (\mathbb{H})$ and let $\Omega$ be an $O_0$ -connection on $\xi_{O_0} (\mathbb{H}).$ Let $C_W: [0, T] \ni t \longmapsto O (t) \in \mathcal{W} (\mathbb{H})$ be a smooth curve. If $C_P: [0, T] \ni t \longmapsto \tilde{U} (t)$ is a smooth $O_0$ -lift of $C_W$ such that
$$
\Omega_{\tilde{U} (t)} \Big [\frac{d \tilde{U} (t)}{d t}\Big] =0
$$
| |
230617381/5 | Fig.1. Density structure as a function of the enclosed mass and radius for initial models of super-Chandrasekhar WDs. Each color corresponds to WD models Model1.0 (blue), Model2.0 (orange), Model4.0 (green) and Model10 (red). The color convention for each model is kept the same throughout our paper.
Fig.2. Temperature profiles of the initial WD models. The vertical axis is the radius.
to simulate collapse triggered by electron capture in sec- tion 4.1. The procedure allows us to rule out initial models which collapse triggered by the other process, for example, photodissociation in section 4.2. A super-Chandrasekhar WD with high density does not take time to start its col- lapse. Note that there is an exception, which employed a lower density model of $4\times10^{9}\,$ g cm $^{-3}$. It took 200 ms for the low density model to collapse. We check stability of WD for longer time Abdikamalovetal.2010. Our numer- ical simulation should reproduce the onset of WD collapse more correctly than previous simulations.
## 3 Simulation Setup
## 3.1 Simulation Overview
Our simulation uses the open source code for 1D core- collapse, GR1D (O’Connor & Ott 2010; O’Connor 2015). GR1D is implemented with both the Newtonian and the GR gravity and includes the M1 scheme for the neutrino- radiation transport, which have been used in Mori et al. 2021; Mori et al. 2023 to conduct simulations for core- collapse supernovae from iron cores.
Our simulations are performed in three steps: First, we check the validity of the hydrostatic equilibrium for initial conditions using Newtonian hydrodynamic calcula- tions without the weak interactions (Section 4.1). Then, by taking into account weak interactions, we calculate the process from triggering gravitational instability to the col- lapse of the WD (Section 4.2, ‘collapse phase’), and finally, we calculate the explosion process using GR hydrodynam- ics simulations with neutrino transport (Section 4.3, ‘ex- plosion phase’).
The computational domain is taken from 0 km to the radius at which the density structure of the initial WD decreases to $10^5$ g cm $^{-3}$. The maximum radius is between $1.5-2.0\times 10^8$ cm. In our study, we perform Newtonian simulations on a grid comprising six uniformly spaced zones, extending up to 60 km, and an additional 594 radial zones that are logarithmically spaced, reaching up to the radius where density reaches $10^5{\,\mathrm g\,cm^{-3}}$. The maximum radius is about 2,000 km. Conversely, the GR simulations are carried out on a grid with 40 uniformly spaced zones spanning up to 20 km and logarithmically spaced radial zones in the outer region. The maximum radius in the GR simulations is 10,000 km, and we have to extend the WD profiles to it. Hence, we put the low-density atmosphere of $2.0{\,\mathrm g cm^{-3}}$. See Section 3.5 for a detailed discussion of the influences of the atmosphere. | |
220809372/2 | be impeded by such transfer inability. Resolving transfer inability is crucial to the advancement of practical dynamic pricing policy.
Online hypothesis testing is the key statistical instrument to assess and monitor the credibility of learned demand information. These test statistics measure the discrepancy between the learned and current market environment model, suggesting possible pricing optimality disruption. The funda- mental rationality of such alarm is to utilize the unlikely occurrence of large deviations given the nan hypothesis, the non-existence of environmental changes. Yet, these tools are far from both dynamic pricing and bandit literature, encouraging our investigation of the non-stationary dynamic pricing problems.
Solving this fundamental problem based on existing knowledge of MAB pricing [Misraetal.,2019] requires overcoming three major challenges:
(a) Market environment shift. Depart from optimality in the stationary environment of existing bandit algorithm design, our approach should allow environment shift alarm and quick reaction to change sub-optimal price to the new optimal price to reduce regrettable opportunity cost.
(b) Indirect customer feedback. Depart from independent assumption between available arms in multi-armed bandit problems, our approach should utilize microeconomic choice theory to infer indirect customer feedback through strategical pricing experiments and incorporate such indirect information into the pricing pipeline.
(c) Pricing performance monitoring. Being aware of potential market environment shifts, our ap- proach should build a statistical method-based auditing pipeline to monitor the pricing performance to ensure the transferability of the pricing decision mechanism in the current market environment.
## Main contributions
We summarize the knowledge advancement of this work contributing to existing bandit-based dy- namic pricing algorithm design (that learns a stationary-environment-optimal pricing policy without awareness of potential market environment change):
(I) ACIDP non-stationary dynamic pricing algorithm. (Figure 1, left) We deliver the Actor-Critic Information Directed Pricing (ACIDP) algorithm that learns profit-maximizing price via incomplete demand information and is able to transfer pricing optimality in the presence of market environ- ment shift. The key of pricing mechanism transfer ability inside ACIDP is to construct the Multi- Universe, a space of synthetic bandit environments shaped by environment exploration, exploitation, and experimentation. By evaluating the expected profit and demand information in each universe of Multi-Universe, we adopt the Bayesian rule to maintain a posterior belief distribution over universes to implement the well-known information directed sampling (IDS) [RussoandVanRoy,2018] for pricing decision. Then, we develop an auditing pipeline to ensure pricing optimality transferability.
(II) Transferability auditing pipeline (Figure 1, right) We design a transferability auditing pipeline to monitor market environment changes and inform pricing agents if the credibility of the learned consumer demand model is no longer accurate enough to inform profit-maximizing price. The auditing pipeline consists a 3 stage sequence of events: yellow card (credibility inspection; Section 3.3.A), audit sampling (optimality inspection; Section 3.3.B) and red card (transfer-ability inspection; Section 3.3.C). This auditing pipeline contributes a novel framework in dynamic pricing literature.
(III) Empirical results. Via a combination of synthetic market environments (stationary, structure shift, upside-down and real-world scenario), we show the success of ACIDP is rooted in (a) benefits of synthetic bandit environments and (b) benefits of auditing pipeline. The empirical results support the superiority of ACIDP over the existing bandit-based pricing method. We show the established en- vironment synthesis pipeline and environment auditing pipeline leads to substantial regret reduction across different market environment shifts.
## Related works
Bandit-based dynamic pricing and demand learning. Dynamic pricing is acknowledged to be challenging without knowing demand function [Besbes and Zeevi, 2009]. While MAB offers a cutting-edge partial monitoring setup [CohenandTreetanthiploet,2021,Chen,2021], bandits liter- ature on dealing with correlated and non-stationary rewards are still far from practical use. [Gupta | |
211009398/33 | SIX OPERATIONS FOR $\wideparen{\mathcal{D}}$ -MODULES ON RIGID ANALYTIC SPACES
$A$ -module,
then we can also endow $M\otimes_A N$ with a locally convex topology by considering it as the cokernel of
\begin{align*} M\otimes_K A\otimes_K N&\rightarrow M\otimes_K N\
\
m\otimes a\otimes n&\mapsto ma\otimes n-m\otimes an, \end{align*}
where the tensor products over $K$ are equipped with the projective tensor product topology.
Let $A$ be a Fr
é
chet $K$ -algebra. Let $M$ and $N$ be a right resp. left Fr
é
chet $A$ -module. Then the natural morphism
\begin{equation*} (M^b{\otimes}_{A^b} N^b)^t\rightarrow M{\otimes}_AN \end{equation*}
is an isomorphism. In particular, the natural morphism
\begin{equation*} M^b\otimes_{A^b}N^b\rightarrow (M\otimes_A N)^b \end{equation*}
is a bounded bijection.
Proof.
The bornological vector space $M^b\otimes_{A^b}N^b$ is the cokernel of the natural morphism
\begin{equation*} (M\otimes_K A\otimes_K N)^b\rightarrow (M\otimes_K N)^b, \end{equation*}
thanks to Lemma
3.1
. Note that all these tensor products are metrisable spaces, so applying first $(-)^b$ and then $(-)^t$ yields the identity functor. As $(-)^t$ preserves cokernels, we can deduce that
\begin{equation*} (M^b\otimes_{A^b}N^b)^t\cong \mathrm{coker}(M\otimes_K A\otimes_K N\rightarrow M\otimes_K N), \end{equation*}
as required.
□
It is harder to make any general statements comparing the completed tensor products. We will return to this issue in subsection 3.4. We note the following special case:
Let $A$ be a Banach $K$ -algebra. Let $M$ and $N$ be a right resp. left Banach module. Then there is a natural isomorphism
\begin{equation*} M^b\widehat{\otimes}_{A^b}N^b\cong\left(M\widehat{\otimes}_AN\right)^b. \end{equation*}
Proof.
Let $f: M\widehat{\otimes}_K A\widehat{\otimes}_K N\rightarrow M\widehat{\otimes}_K N$ be the usual morphism. Then $(M\widehat{\otimes}_A N)^b\cong (\mathrm{coker} f)^b$, and by Lemma
3.1
, $M^b\widehat{\otimes}_{A^b}N^b\cong \mathrm{coker} (f^b)$. The result thus follows from the exactness of $(-)^b$ on Banach spaces (Lemma
3.1
).
□
3.3.
## Nuclearity
. Coadmissible modules over any Fr
é
chet
–
Stein algebra carry a natural Fr
é
chet topology, and many of their elementary properties given e.g. in
[32, section 3]
are expressed within this Fr
é
chet (topological) context. To have the same results at our disposal when working in the complete bornological framework, we need the functor $(-)^b$ to preserve various constructions. We first collect a couple of problems which we have already encountered in the previous two subsections.
(a)
- We might ask whether $(-)^b$ is exact on a suitable class of Fr
é
- chet spaces. Let $f: V\rightarrow W$ be a continuous morphism of Fr
é
- chet spaces. When is $\mathrm{coker}(f^b)\cong (\mathrm{coker} f)^b$? We note that this question becomes particularly relevant with regard to completed tensor products of Fr
é
- chet modules, see Lemma
3.2
- .
(b)
- Let $V\in \mathscr{M}$. Is there some topological condition which might ensure that the natural morphism $\widehat{V^b}\rightarrow \widehat{V}^b$ is an isomorphism? Note that as we do not know a priori whether $\widehat{V^b}\in \mathscr{M}^b$, this is far from clear. Lemma
3.1
- only gives some partial information. | |
181004947/4 | incident photon angular frequency or energy and we used the expression $k_{ij}=\omega^2 n^2(\delta_{ij}-\hat k_i\hat k_j)$ with $k_{ij}$ being the photon momentum tensor, $n=|\boldsymbol k|/\omega$ is the index of refraction and $\hat k_i$ are the components of a unit vector along the electromagnetic direction of propagation wave-vector $\boldsymbol k$. We may see that the role of $\Pi_{ij}$ in (1) is to give to photons an ” effective mass ” in the medium. In the case when the medium is isotropic, we have that $k_{ij}$ is a diagonal tensor with diagonal entries corresponding to the photon indexes of refraction in medium where $k_{ii}\neq 1$. In the case when photons propagate in vacuum, we have that $k_{ij}=\omega^2 \delta_{ij}$ and we get the on-shell photon relation $\omega=\boldsymbol k^2$ where $\Pi_{ij}=0$.
The explicit expression of the photon polarization tensor $\Pi_{ij}$ depends on the induced currents that enter a given problem. In this work we are interested in a cold magnetized plasma which is quite common situation in astrophysics and cosmology. We assume that the magnetized plasma is with almost no collisions, globally neutral and homogeneous. In addition, there is not an external electric field, namely $\boldsymbol E_e=0$ and the presence of the external magnetic field $\boldsymbol B_e$ locally breaks the isotropy of the plasma since it singles out a preferred direction in a given region of space where the plasma is located.
In the cold magnetized plasma approximation, consider now an incident electromagnetic wave propagating along the observer’s $z$ axis which points to the East, in a magnetized plasma with external magnetic field vector $\boldsymbol B_e = B_e \hat{\boldsymbol n}$. Here $\hat{\boldsymbol n}=[\cos(\Theta), \sin(\Theta)\cos(\Phi), \sin(\Theta)\sin(\Phi)]$ is a unit vector in the direction of the external magnetic field $\boldsymbol B_e$ and $\Theta, \Phi$ are, respectively, the polar and azimutal angles between the magnetic field $\boldsymbol B_e$ and $x$ and $y$ axes. As shown in Ref. [19], the medium polarization vector $\boldsymbol P$ satisfies the equation of motion
$$
\ddot{\boldsymbol P}= \omega_\text{pl}^2 \boldsymbol E-\omega_c\,\dot{\boldsymbol P}\times \hat{\boldsymbol n},
$$
where $\boldsymbol E$ is the electric field of the incident electromagnetic wave, $\omega_\text{pl}^2=4 \pi \alpha n_e/m_e$ is the plasma frequency, $n_e$ is the free electron number density, $m_e$ is the electron mass and $\omega_c=e B_e/m_e$ is the cyclotron frequency. In Eq. (2) the dot symbol ($\cdot$) above $\boldsymbol P$ denotes the derivative with respect to the time $t$.
Assume that the fields evolve in time harmonically at a given point $\boldsymbol x$
$$
\boldsymbol P(\boldsymbol x, t)=\boldsymbol P(\boldsymbol x, \omega) e^{-i\omega t}, \qquad \boldsymbol E(\boldsymbol x, t)=\boldsymbol E(\boldsymbol x, \omega) e^{-i\omega t},
$$
By using the expressions in (3) in Eq. (2) and then solving for the components of $\boldsymbol P$, after we get the following solution in terms of the incident electric field components $E_j$, in the case when $\omega\neq 0$ and $\omega\neq \pm\omega_c$
$$
P_i(\boldsymbol x, \omega)=\chi_{ij}(\omega) E_j(\boldsymbol x, \omega), \qquad (i, j=x, y, z),
$$
where $\chi_{ij}(\omega)$ are the components of the electric susceptibility tensor
$$
\begin{gathered}
\chi_{xx}=-\frac{\omega_\text{pl}^2}{\omega^2-\omega_c^2}+\frac{\omega_\text{pl}^2\omega_c^2 \cos^2(\Theta)}{\omega^2(\omega^2-\omega_c^2)}, \quad \chi_{xy}=\frac{\omega_\text{pl}^2\,\omega_c^2 \sin(2 \Theta)\cos(\Phi)}{2\, \omega^2(\omega^2-\omega_c^2)}+i \frac{\omega_\text{pl}^2\omega_c \sin(\Theta)\sin(\Phi)}{\omega(\omega^2-\omega_c^2)}, \\
\chi_{xz}=\frac{\omega_\text{pl}^2\,\omega_c^2 \sin(2 \Theta)\sin(\Phi)}{2\,\omega^2(\omega^2-\omega_c^2)}-i \frac{\omega_\text{pl}^2\omega_c \sin(\Theta)\cos(\Phi)}{\omega(\omega^2-\omega_c^2)}, \quad \chi_{yx} = \chi_{xy}^*,\\
\chi_{yy}=-\frac{\omega_\text{pl}^2}{\omega^2-\omega_c^2}+\frac{\omega_\text{pl}^2\omega_c^2 \sin^2(\Theta)\cos^2(\Phi)}{\omega^2(\omega^2-\omega_c^2)}, \quad \chi_{yz} = \frac{\omega_\text{pl}^2\,\omega_c^2 \sin(2\Phi)\sin^2(\Theta)}{2\,\omega^2(\omega^2-\omega_c^2)} + i \frac{\omega_\text{pl}^2\omega_c \cos(\Theta)}{\omega(\omega^2-\omega_c^2)},\\
\chi_{zx}=\chi_{xz}^*, \quad \chi_{zy}=\chi_{yz}^*, \quad \chi_{zz}=-\frac{\omega_\text{pl}^2}{\omega^2-\omega_c^2}+\frac{\omega_\text{pl}^2\omega_c^2 \sin^2(\Theta)\sin^2(\Phi)}{\omega^2(\omega^2-\omega_c^2)}.
\end{gathered}
$$
The expressions for the components of $\chi_{ij}$ in (5) are valid for an incident electromagnetic wave with an arbitrary direction of propagation with respect to $\boldsymbol B_e$. In addition, the components $\chi_{ij}$ do not explicitly depend on $\boldsymbol x$ but only implicitly through $B_e(\boldsymbol x, t)$ which enters in $\omega_c$. After these general comments about (5), let us find the | |
200615062/11 | ## RESULTS AND DISCUSSION
Typical successive RHSLs measured by LS01 after nearly 1 year at LEO are shown in Fig. 6. The DUT still shows bipolar resistive switching and the counter-clockwise circulation is maintained as well as the shape “ Table with Legs”, indicating that the Au/YBCO interface is still the dominant with a lower resistive switching contribution of the Pt/YBCO interface. A small change in the remanent resistance can be observed after each cycle. This can be a consequence of the temperature variation of the DUT or can be related to a relaxation of the final resistance, considering the 24 h delay between successive measurements. This effect is characteristic of the YBCO-based interfaces, indicating the high mobility of oxygens along specific crystallographic orientations or in grain boundaries [30, 31]. Additionally, it can be noticed that the remanent resistance values are 15-30% higher than those measured before lift-off.
FIG. 6: (Color online) Typical succesive RHSLs of DUT measured after almost 1 year at LEO, showing similar features as those observed before lift-off, such as the counter-clockwise circulation and the table with legs structure.
In order to gain insight on the origin of the variations observed in the RHSLs, we plotted in Fig. 7 the evolution of the maximum and minimum R $_{rem}$ as a function of the days at LEO. A noisy behavior with an overall tendency to increase with increasing the number of | |
201000768/2 | corresponding query, DeepCT [2] and Doc2Query [11] learn PLM- based models to enhance the performance of traditional BoW meth- ods. The difference is that DeepCT learns a regression model to re- weight terms with contextualized representations, while Doc2query learns an encoder-decoder generative model to expand query terms for passage. Both of these two methods train an auxiliary interme- diate model and then help refine the final sparse representations to achieve better text ranking performance.
In this paper, we propose a novel framework SparTerm to learn Term -based Spa rse r epresentations directly in the full vocabu- lary space. Equipped with the pre-trained language model, the proposed SparTerm learns a function to map the frequency-based BoW representation to a sparse term importance distribution in the whole vocabulary, which offers the flexibility to involve both term-weighting and expansion in the same framework. As shown in Figure 1, compared with BoW representation, SparTerm assigns more weights to the term of high distinguishability given the con- text, and expand extra terms hopefully bridging the lexical gap with future queries. We empirically show that SparTerm signifi- cantly increase the upper limit of sparse retrieval methods, and gives new insights of transferring deep knowledge from PLM-based representation to simple BoW representations.
More specifically, SparTerm comprises an importance predictor and a gating controller. The importance predictor maps the raw in- put text to a dense importance distribution in the vocabulary space, which is different from traditional term weighting methods that only consider literal terms of the input text. To ensure the sparsity and flexibility of the final representation, the gating controller is introduced to generate a binary and sparse gating signal across the dimension of vocabulary size, indicating which tokens should be activated. These two modules cooperatively yield a term-based sparse representation based on the semantic relationship of the input text with each term in the vocabulary.
Our contributions. In summary, we propose to directly learn term-based sparse representation in the full vocabulary space. The proposed SparTerm indicates that there is much space for improving the ranking performance of termed-based representations, while still keeping the interpretability and efficiency of BoW methods. Evaluated on MSMARCO [10] dataset, SparTerm significantly out- performs previous sparse models based on the comparable size of PLMs. The top-ranking performance of SparTerm even outper- forms Doc2Query-T5, which is based on the pre-trained model of 2x model size and 70x pre-training corpus size. Moreover, we conduct further empirical analysis about how the deep knowledge of PLMs can be transferred to the sparse method, which gives new insights for sparse representation learning.
## RELATED WORK
Our work relates to two research fields: bag-of-words representa- tions and pre-trained language model for text retrieval.
## Bag-of-words Methods
Bag-of-words(BoW) methods have played a central role in the first- stage retrieval. These methods convert a document or query into a set of single terms, and each term associates a weight to characterize its weight. Most of the early common practice adopted TF-IDF style models to calculate weights. Robertson [14] proposed the well- known method BM25, which further improve the performance of the original TF-IDF. Later proposed methods, such as [7], [18], [16], did not show much advantage over BM25. More recently, Hamed Zamani [17] proposed SRNM to learn a sparse coding in hidden space using weak supervision, which shows good potential for solving the “ lexical mismatch ” problem. However, the latent unexplainable tokens can not ensure that documents with exact matched terms can be retrieved.
## PLMs for dense text retrieval
The pre-trained language models like BERT [4] show new possi- bilities for text retrieval. Based on dense representations, Lee [8] proposed ORQA with bi-encoder architecture to retrieve candi- date passages for question answering using FAISS [5]. However, analysis from [9] concludes that bi-encoders based on dense rep- resentation suffer from its capacity limitation in scenarios that emphasize long document retrieval and exact matching. Follow- ing the late-interaction paradigm, Khattab [6] proposed Col-BERT to conduct efficient interaction between the query and document, which can run 150x faster than fully-interactive BERT but achieve comparable precision. Though much faster than BERT, Col-BERT is still not computationally feasible for large scale first-stage retrieval, for the existence of the late interaction layer.
## PLMs for sparse text retrieval
Several PLM-based models have emerged to improve the traditional sparse BoW representations. Dai [2] proposed DeepCT to estimate a term’s weight considering its contextualized information, and this work was later extended to generate document-level term weights [3]. Another work Doc2query [11] tries to “ translate ” potential queries to expand document content, which also shows a large im- provement compared to the traditional BM25 method. The biggest difference between our work and these two methods is that DeepCT and Doc2Query train an auxiliary intermediate model to help refine the sparse representations, while SparTerm is desinged to directly learn sparse representations within the whole vocabulary.
## SPARSE REPRESENTATION LEARNING
This section presents the model architecture of SparTerm and the corresponding training strategy.
## Overview
Figure 2 (a) depicts the general architecture of SparTerm which comprises an importance predictor and a gating controller. Given the original textual passage $p$, we aim to map it into a deep and contextualized sparse representation $p{'}$ in the vocabulary space. The mapping process can be formulated as:
$$
p{'} = \mathcal{F}(p) \odot \mathcal{G}(p)
$$
where $\mathcal{F}$ is the item importance predictor and $\mathcal{G}$ the gating con- troller. The importance predictor $\mathcal{F}$ generates a dense vector rep- resenting the semantic importance of each item in the vocabulary. The gating controller $\mathcal{G}$ generates a binary gating vector to control which terms to appear in the final sparse representation. To achieve | |
231202309/4 | items, to a continuous parameter space such as the environment parameters in UED. From here, we drop the subscript for the notations $a$, $d$, and $r$ to indicate our move away from discretized items and students, as originally formulated in IRT.
## Generating New Levels for Curricula
The objective of VIBO is to learn the latent representation of student ability and difficulty of items. In order for us to generate the next set of environment parameters $\lambda_{t+1}$ for the student to train on, PERM include an additional decoder to generate $\lambda$ given a desired difficulty estimate $d$.
PERM makes a core assumption that optimal learning takes place when the difficulty of the environ- ment matches the ability of the student. In the continuous response model given in Eq. 2.1, we see that when ability and difficulty is matched (i.e. $a = d$), the probability which the student achieves a normalized average score $r = 0$ is 0.5. This is a useful property to operationalize ZPD, as we can see that the model estimates an equal probability of the student overperforming or underperforming.
Training is initialized by uniformly sampling across the range of environment parameters. After each interaction between the student and the environment, PERM estimates the ability $a_t$ of the student given the recent episodic rewards and parameters of the environment. PERM then generates the parameters of the next environment $\lambda_{t+1} \sim p_\theta(\lambda|d_{t+1})$ where $d_{t+1} = a_t$.
We refer motivated readers to the Appendix and TioandVarakantham[2023] for more details.
## Deploying PERM as a Training System: A Two-Stage Process
Equipped with a understanding of PERM’s role and capabilities, we now describe a two-stage process which we propose as a generalized framework for training. The process can be visualized in Figure 1.
## Stage 1: Collecting Data to Train PERM
One of the primary challenges facing IRT-based solutions lies in the availability of item-student interaction data. While established standardized tests possess abundant data due to its rich history and wide student base [Harlen,2001], applying IRT to new domains remains problematic. Commonly referred to as the cold-start problem [Bassenetal.,2020], such models often provide a sub-optimal experiences during the infant stages, potentially leading to poor user experience.
In contrast, training RL agents in simulated environments provides a substantial advantage by offering access to a virtually limitless pool of item-student interaction data. This advantage enables us to address these challenges effectively.
In the first stage of our approach, we initialize new RL agents to gather the necessary item-student interaction data in a given environment. The RL Agents would be exposed to a variety of levels, all of which were generated using Domain Randomization [Tobinetal.,2017]. We then collect their performance results and the associated level parameters used to generate the levels.
After the RL agent has completed its training, we train PERM with the entire history of performance- environment parameters to yield a model that is capable of inferring student ability and level difficulty.
## Stage 2: Deploying PERM to Train Others
In this stage, we implement PERM as the teacher algorithm, in accordance with the UED framework.
During this phase, the pre-trained PERM actively observes the student’s interactions with the learning environment, allowing it to make accurate assessments of the student’s current ability level. Based on this assessment, PERM then suggests the next set of environment parameters that are deemed appropriate for the student’s current stage of learning.
Importantly, since PERM was pre-trained prior to the initiation of the student’s training, we effectively circumvent the cold-start problem. It’s noteworthy that even when human students commence their training with varying levels of competency, PERM can quickly infer their abilities using the initial training interactions and subsequently offer appropriately challenging content tailored to each student. | |
181009575/17 | Moreover, since $a, b, k_\mu\in G$ we have $s,s'\in \mathcal Z_G(S) \propto G\cap S$ and thus
$$
\mathcal D_S(a \varrho b^\dagger)=
\mathcal D_S(k_\mu s \varrho {s'}^\dagger k_\mu^\dagger)\propto
\mathcal D_S(k_\mu \varrho k_\mu^\dagger) = k_\mu \varrho k_\mu^\dagger.
$$
b) Armed with the above expresion for $\varrho '$ we have for any $s\in S$
$$
\begin{aligned}
\langle K^\dagger s K \rangle_{\varrho}
= \langle s \rangle_{\varrho '}
&= \sum_\mu c_\mu \langle k_\mu^\dagger s k_\mu\rangle_\varrho
\\
&= \sum_\mu c_\mu \mu(s) \langle s\rangle_\varrho
= \sum_\mu c_\mu \mu(s),
\end{aligned}
$$
where the sum is over the characters $\mu$ of $S_G$. By a well known property of characters:
$$
c_\mu = \frac 1{|S_G|}\sum_{s\in S_G} \langle K^\dagger s K\rangle_\varrho \,\mu(s)
=\frac 1 {|S_G|}=\frac 1 {|G_S|}.
$$
Finally, using again that $Z_G(S)\propto S\cap G$,
$$
\varrho '
= \frac 1 {|G_S|}\sum_\mu k_\mu \varrho k_\mu^\dagger
= \frac 1 {|G|}\sum_{g\in G} g \varrho g^\dagger.\qedhere
$$
Proof of theorem 3. We set $X_\alpha = x$ and consider two separate cases.
(X α tolerable) Since $A_\alpha P_0$ cannot be zero, by lemma 14 there exist $z\in E_\alpha$ and $K \in\mathcal L(G_\alpha)$ such that
$$
A_\alpha^\dagger P_0
= z K P_0.
$$
Combining this with the fact that $U\varrho U^\dagger$ is an encoded state of $S_Z$ and using lemma 15 we get for any $s\in S$
$$
\begin{aligned}
\langle K^\dagger s K \rangle_{U\varrho U^\dagger}
&= (z, s) \langle A_\alpha s A_\alpha^\dagger \rangle_{U\varrho U^\dagger}
\\
&=\begin{cases}
1, & \text{if $s\in Z_S(G_\alpha)\subset S_Z \mathcal Z_X(G_{\alpha})$},\\
0, & \text{otherwise}.
\end{cases}
\end{aligned}
$$
A few manipulations give
$$
\begin{aligned}
(\mathcal D_{S_X}\circ \hat U\circ \hat X_\alpha) (\varrho)
&= (\mathcal D_S\circ \hat U\circ \hat X_\alpha) (\varrho)
\\
&=
(\mathcal D_S \circ \hat X_\alpha \circ \hat U \circ \hat A^\dagger_\alpha) (\varrho)
\\
&= (\hat X_\alpha \circ \hat z \circ\mathcal D_S \circ \hat K\circ \hat U) (\varrho).
\end{aligned}
$$
Since $X_\alpha$ is tolerable
$$
\mathcal Z_{G_\alpha}(S) = G_\alpha\cap S
$$
and thus we can apply lemma 16 to the encoded state $U\varrho U^\dagger$, recovering the result.
(X α not tolerable) By lemma 2 we can choose a logical operator $X_\lambda\in\mathcal Z(H(\varphi))$. By lemma 2 (iv), $X_{\alpha+\lambda}$ is tolerable. For $w$ we can choose any logical operator that does the job, and in particular we set
$$
w = U^\dagger X_\lambda U X_\lambda.
$$
The result follows from the first case: since $X_\lambda \varrho X_\lambda$ is an encoded state
$$
\begin{aligned}
(\mathcal D_{S_X}\circ \hat U\circ \hat X_\alpha) (\varrho)
&= (\mathcal D_{S_X}\circ \hat U\circ \hat X_{\alpha+\lambda}\circ\hat X_\lambda) (\varrho)
\\
&= (\hat X_{\alpha+\lambda} \circ \mathcal D_{E_{\alpha+\lambda}}\circ \hat U \circ \hat X_\lambda)(\varrho)
\\
&= (\hat X_{\alpha} \circ \mathcal D_{E_{\alpha+\lambda}}\circ \hat U \circ \hat w)(\varrho),
\end{aligned}
$$
where, in the last equality, $\hat X_\lambda$ commutes with $\mathcal D_{E_{\alpha+\lambda}}$ because $G_{\alpha+\lambda}=H_{\alpha}$.
Sketch of proof of lemma 4. (i) Trivially
$$
G_\alpha\subset \prod_i G_{\alpha_i},
$$
and, via lemma 2,
$$
G_{\alpha_i}=H_{\alpha_i}\subset H_\alpha\subset G_\alpha.
$$
The result follows, again via lemma 2, because
$$
G_\alpha = \prod_i G_{\alpha_i} = \prod_i H_{\alpha_i}=H_\alpha.
$$
(ii) This follows combining
$$
\mathcal Z(G_\alpha)= \bigcap_i \mathcal Z(G_{\alpha_i}),
$$
with
$$
\frac {g(\alpha\cap \gamma)} 2
= \sum_i \frac {g(\alpha_i\cap \gamma)} 2 - \sum_{i\neq j} g(\alpha_i\cap\alpha_j\cap \gamma)
$$
and, for any $\beta$,
$$
g(\beta) \equiv |\beta|
\mod 2.\qedhere
$$
Proof Let $i=1,\dots,n$ and $j=1,\dots,m$. The condition on the generators $X_{\beta_j}$ amounts to the ex- istence of a function
$$
I:\{1,\dots,m\}\rightarrow \{1,\dots,n\}
$$
such that
$$
i\neq I(j) \implies Z_{\alpha_i\cap \beta_j}\in S.
$$
Denote by $I^{-1}[i]$ the preimage of $i$. When $j\in I^{-1}[i]$ there exists some $s\in S_Z$ such that
$$
Z_{\alpha_i\cap \beta_j}
= s \prod_{k=1}^n Z_{\alpha_k\cap \beta}
= s Z_{\alpha\cap\beta_j}.
$$
Then:
$$
\begin{aligned}
H_{\alpha_i}
&= S_Z\cdot\langle Z_{\alpha_i\cap\beta_j} \rangle_{j=1}^m
= S_Z\cdot\langle Z_{\alpha_i\cap\beta_j} \rangle_{j\in I^{-1}[i]}
\\
&= S_Z\cdot\langle Z_{\alpha\cap\beta_j} \rangle_{j\in I^{-1}[i]}
\subset S_Z\cdot\langle Z_{\alpha\cap\beta_j} \rangle_{j=1}^k = H_\alpha.
\end{aligned}
$$
The converse, i.e.
$$
H_\alpha \subset \prod_i H_{\alpha_i},
$$
is trivial. | |
190410332/18 | Proof. The first point follows after a straightforward calculation. For the second, use that, by defini- tion
\begin{eqnarray*} \mathrm{interior}_0^\infty \bigg| \frac{d}{d \, r} \Big\{e^{-\beta \frac{r^2}{u}} r^{\alpha} \Big\} \bigg| \, \Phi(r) \, d r & \leq & \mathrm{interior}_0^\infty \bigg(\frac{2 r}{u} e^{-\beta \frac{r^2}{u}} + \alpha r^{\alpha - 1} e^{-\beta \frac{r^2}{u}} \bigg) \Phi(r) \, d r\\
& = & u^\frac{\alpha}{2} \, \mathrm{interior}_0^\infty \bigg(2 r e^{-\beta r^2} + \alpha r^{\alpha - 1} e^{-\beta r^2} \bigg) \Phi(r\sqrt{u}) \, d r\\
& \leq & u^\frac{\alpha}{2} \Phi(\sqrt{u}) \bigg(\sum_{k = 0}^\infty D_0^k \, \mathrm{interior}_{2^k-1}^{2^{k+1}-1} \big(2 r e^{-\beta r^2} + \alpha r^{\alpha - 1} e^{-\beta r^2} \big) \, d r \bigg). \end{eqnarray*}
But the series on the last expression are summable.
Let us denote by $\| \cdot \|_\alpha$ the classical H ¨o lder seminorm over functions in $M$ given by
$$
\| f \|_\alpha = \sup_{x \neq y} \left\{\frac{|f(x) - f(y)|}{d(x,y)^\alpha} \right\}.
$$
- (i) If $M$ is doubling and satisfies the Poincare inequality (Po), then it holds that $\displaystyle{\| f \|_{\Lambda_\alpha^\circ({\mathcal T})} \lesssim \| f \|_\alpha}$. (ii) If $\Ric \geq 0$ and $M$ is connected, then we have that $\displaystyle{\| f \|_\alpha \lesssim \| f \|_{\Lambda_\alpha^\circ({\mathcal T})}}$.
Proof. We will prove first (i). Fix a function $f$ with $\| f \|_\alpha < \infty$. We have, by the definition of the integral kernel of $P_s$, that
$$
s \frac{d P_s f}{d \, s}(x) = \mathrm{interior}_M s \frac{\partial}{\partial \, s} p_s(x,y) \, f(y) \, d {\mathrm{vol}}(y).
$$
Now, using that $\partial_s p_s(x,y)$ has $0$ -integral over $y$ for every $x \in M$, we obtain that
$$
s \frac{d P_s f}{d \, s}(x)
=
\mathrm{interior}_M \frac{\partial}{\partial \, s} p_s(x,y) \big(f(x) - f(y) \big) \, d {\mathrm{vol}}(y)
$$
and so
$$
\begin{aligned}
\left| s \, \frac{d P_s f}{d \, s}(x) \right|
& \leq & \mathrm{interior}_M \left| \frac{\partial}{\partial \, s} p_s(x,y) \right| \, \big| f(x) - f(y) \big| \, d {\mathrm{vol}}(y) \\
& \leq & \| f \|_\alpha \mathrm{interior}_M \left| \frac{\partial}{\partial \, s} p_s(x,y) \right| \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \\
& = & \| f \|_\alpha \mathrm{interior}_M \left| \mathrm{interior}_0^\infty \frac{\partial}{\partial \, s} \varphi_s(u) \, h_u(x,y) \, d u \right| \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \\
& \leq & \| f \|_\alpha \mathrm{interior}_M \mathrm{interior}_0^\infty \left| \frac{\partial}{\partial \, s} \varphi_s(u) \right| \, h_u(x,y) \, d u \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \\
& = & \| f \|_\alpha \mathrm{interior}_0^\infty \left| \frac{\partial}{\partial \, s} \varphi_s(u) \right| \underbrace{\bigg(\mathrm{interior}_M h_u(x,y) \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \bigg)}_{\mathrm{(I)}} d u.
\end{aligned}
$$
$$
\begin{aligned}
\left| s \, \frac{d P_s f}{d \, s}(x) \right|
& \leq & \mathrm{interior}_M \left| \frac{\partial}{\partial \, s} p_s(x,y) \right| \, \big| f(x) - f(y) \big| \, d {\mathrm{vol}}(y) \\
& \leq & \| f \|_\alpha \mathrm{interior}_M \left| \frac{\partial}{\partial \, s} p_s(x,y) \right| \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \\
& = & \| f \|_\alpha \mathrm{interior}_M \left| \mathrm{interior}_0^\infty \frac{\partial}{\partial \, s} \varphi_s(u) \, h_u(x,y) \, d u \right| \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \\
& \leq & \| f \|_\alpha \mathrm{interior}_M \mathrm{interior}_0^\infty \left| \frac{\partial}{\partial \, s} \varphi_s(u) \right| \, h_u(x,y) \, d u \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \\
& = & \| f \|_\alpha \mathrm{interior}_0^\infty \left| \frac{\partial}{\partial \, s} \varphi_s(u) \right| \underbrace{\bigg(\mathrm{interior}_M h_u(x,y) \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \bigg)}_{\mathrm{(I)}} d u.
\end{aligned}
$$
$$
\begin{aligned}
\left| s \, \frac{d P_s f}{d \, s}(x) \right|
& \leq & \mathrm{interior}_M \left| \frac{\partial}{\partial \, s} p_s(x,y) \right| \, \big| f(x) - f(y) \big| \, d {\mathrm{vol}}(y) \\
& \leq & \| f \|_\alpha \mathrm{interior}_M \left| \frac{\partial}{\partial \, s} p_s(x,y) \right| \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \\
& = & \| f \|_\alpha \mathrm{interior}_M \left| \mathrm{interior}_0^\infty \frac{\partial}{\partial \, s} \varphi_s(u) \, h_u(x,y) \, d u \right| \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \\
& \leq & \| f \|_\alpha \mathrm{interior}_M \mathrm{interior}_0^\infty \left| \frac{\partial}{\partial \, s} \varphi_s(u) \right| \, h_u(x,y) \, d u \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \\
& = & \| f \|_\alpha \mathrm{interior}_0^\infty \left| \frac{\partial}{\partial \, s} \varphi_s(u) \right| \underbrace{\bigg(\mathrm{interior}_M h_u(x,y) \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \bigg)}_{\mathrm{(I)}} d u.
\end{aligned}
$$
We have used the definition of $\| f \|_\alpha$ in (4.1), the subordination formula in (4.2) and Fubini’s theorem in (4.3). We can estimate the integral $\mathrm{(I)}$ as follows. Denote by $\Phi_x(r)$ the doubling function given by $\Phi_x(r)= {\mathrm{vol}}(B_x(r))$. Taking a change of variable by calling $r$ to $d(x,y)$, we get that
\begin{eqnarray*} \mathrm{(I)} & = & \mathrm{interior}_M h_u(x,y) \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \\
& \leq & \frac1{\Phi_x(\sqrt{u})} \mathrm{interior}_M e^{- \beta_1 \, \frac{d(x,y)^2}{u}} \, d(x,y)^\alpha \, d {\mathrm{vol}}(y) \, = \, \frac1{\Phi_x(\sqrt{u})} \mathrm{interior}_0^\infty e^{- \beta_1 \, \frac{r^2}{u}} \, r^\alpha \, d m_{\Phi_x}(r) \end{eqnarray*}
| |
230201531/4 | Figure2. Stacked radial density profiles of $N$ -body particles with even number of apocenter passages, ranging from $p=4$ to $40$. The four mass bins are displayed in the upper three and the lower left panel for S, M, L and XL, respectively. Additionally, the lower middle and right panels show the results obtained from $460$ halos in the mass range $[4.10\time10^{11},\,2.30\times10^{12}]\,h^{-1}M_\odot$, which are further divided into the two subsamples based on the concentration parameter $c_{\rm vir}$ and accretion rate $\Gamma_{\rm dyn}$, respectively (see text in detail). In each panel, the fitted results with Equation (1) are depicted as solid lines.
Dependence of the characteristic density $A$ (upper) and scale $S$ (lower) on the number of apocenter passages, $p$, as deter- mined by fitting to Equation (1) in different symbols four mass bins (see legend). The thin solid curves represent the fitting formulae, Eqs. (2) and (3). For comparison, predictions of the Fillmore- Goldreich self-similar solutions are also shown, for specific values of the parameter $\epsilon$ ($1/15$, $1/6$ and $1$). In plotting these predictions, we identify the position of radial caustics in the self-similar solutions with the characteristic scale $S(p)$, and derive $A(p)$ by equating the masses contained in each stream. The shaded regions for the pre- dictions indicate uncertainty in identifying $S(p)$ with the position of the $p$ -th or $(p+1)$ -th radial caustics of the self-similar solutions.
large amplitude $A(p)$ and a large characteristic scale $S(p)$. These trends are particularly evident for larger $p$, suggest- ing that the universal double power-law feature is established in a self-regulated manner during the orbital motion in the multi-stream region, where the diversity of mass accretion and merger histories tend to be erased and only be imprinted in $A(p)$ and $S(p)$.
## 4.2. Comparison with self-similar solutions
The results in section 3 and 4.1 strongly indicate that the in- ner structure of halos is built up dynamically in a self-similar manner. Here, we compare our results with self-similar so- lutions. While self-similar solutions are only valid in the Einstein-de Sitter universe, the secondary infall model of Bertschinger (1985) has been shown to reproduce the pseudo phase-space density of $Q(r)\propto r^{-1.875}$ found in simulations in the $\Lambda$ CDM model. Along the line of this, we consider the spherically symmetric solutions put forth by Fillmore&Gol- dreich (1984), which include the Bertschinger’s secondary infall model as a special case. Recent work by Sugiuraetal. (2020) has made a direct comparison of these predictions with radial multi-stream structures obtained from simulations up to $p=5$. Identifying the position of radial caustics in self- similar solutions with the characteristic scale of the double power-law profile in Equation (1), it is possible to make pre- dictions for both $A(p)$ and $S(p)$.
In Figure 3, we compare the predictions of self-similar solutions with our $N$ -body results for three values of the model parameter $\epsilon$, which describes the power-law slope of initial density contrast. Note that the parameter $\epsilon$ is restricted to the range $[0,\,1]$, and the solution with $\epsilon=1$ corresponds to Bertschinger’s secondary infall model. Figure 3 shows that none of the solutions consistently explain the trends in | |
220800715/10 | ourselves to $\|\boldsymbol{m}\beta\|\leqslant M$ for $t$ sufficiently large. Consider the set $A_M$ defined in (4.8). Due to condition (A) and the fact that $\mathbf{X}$ has full rank, with probability one, for any $\eta>0$, we can find an $M>0$, such that $\varrho_1(\sqrt{M}/(2\sqrt{\lambda_1}))P(A_M)>R_P(\textbf{0})+2\eta$. On the other hand, for any $\eta>0$, we have
$$
|R_{P_t}(\mathbf{0})-R_P(\mathbf{0})|\leqslant \eta,
$$
for $t$ sufficiently large. If $\boldsymbol{m}\beta$ minimizes $R_{P_t}(\boldsymbol{m}\beta)$, for $t$ sufficiently large, we must have $\|\boldsymbol{m}\beta\|\leqslant M$, since otherwise, according to (4.7),
$$
R_{P_t}(\boldsymbol{m}\beta)
\geqslant
\varrho_1(\sqrt{M}/(2\sqrt{\lambda_1}))P(A_M)
>
R_P(\textbf{0})+2\eta
\geqslant
R_{P_t}(\textbf{0})+\eta
>
R_{P_t}(\textbf{0}).
$$
Hence, for minimizing $R_{P_t}(\boldsymbol{m}\beta)$, we can restrict to the compact set $K=\{\boldsymbol{m}\beta\in\mathbb{R}^q:\|\boldsymbol{m}\beta\|\leqslant M\}$. Furthermore, as in the proof of Theorem 1, the function $R_{P_t}(\boldsymbol{m}\beta)$ is continuous on the compact set $K$, and must therefore attain a minimum $\boldsymbol{m}\beta_{1}(P_t)$.
According to Theorem there exists at least one $\boldsymbol{m}\beta_1(P)$ that minimizes $R_P(\boldsymbol{m}\beta)$. Now, suppose that $\boldsymbol{m}\beta_1(P)$ is unique. Because $\boldsymbol{m}\beta_1(P)$ is regression equivariant, we may assume that $\boldsymbol{m}\beta_1(P)=\mathbf{0}$. For the sake of brevity, let us write $\boldsymbol{m}\beta_{1,t}=\boldsymbol{m}\beta_1(P_t)$, $\mathbf{V}_{0,t}=\mathbf{V}_0(P_t)$, and $R_t=R_{P_t}$. From (5.2) it follows that for $t$ sufficiently large,
$$
0<\lambda_k(\mathbf{V}_0(P))/4\leqslant \lambda_k(\mathbf{V}_{0,t})\leqslant\lambda_1(\mathbf{V}_{0,t})\leqslant 4\lambda_1(\mathbf{V}_0(P))<\infty.
$$
Now, consider a sequence $\{(\boldsymbol{m}\beta_{1,t},\mathbf{V}_{0,t})\}$, such that $\|\boldsymbol{m}\beta_{1,t}\|\leqslant M$ and $\mathbf{V}_{0,t}$ satisfies. Then the sequence $\{(\boldsymbol{m}\beta_{1,t},\mathbf{V}_{0,t})\}$ lies in a compact set, so it has a convergent subsequence $(\boldsymbol{m}\beta_{1,t_j},\mathbf{V}_{0,t_j})\rightarrow(\boldsymbol{m}\beta_{1,L},\mathbf{V}_0(P))$. According to, it follows that
$$
\begin{split}
\lim_{j\rightarrow\infty}
R_{t_j}(\boldsymbol{m}\beta_{1,t_j})
&=
\lim_{j\rightarrow\infty}
\int
\varrho_1\mathopen{}\mathclose\bgroup\originalleft(
d(\mathbf{s},\boldsymbol{m}\beta_{1,t_j},\mathbf{V}_{0,t_j})
\aftergroup\egroup\originalright)
\,\mathrm{d} P_{t_j}(\mathbf{s})\\
&=
\int
\varrho_1\mathopen{}\mathclose\bgroup\originalleft(
d(\mathbf{s},\boldsymbol{m}\beta_{1,L},\mathbf{V}_0(P))
\aftergroup\egroup\originalright)
\,\mathrm{d} P(\mathbf{s})
=
R_P(\boldsymbol{m}\beta_{1,L}).
\end{split}
$$
Now, suppose that $\boldsymbol{m}\beta_{1,L}\neq \mathbf{0}$. Then, since $R_P(\boldsymbol{m}\beta)$ is uniquely minimized at $\boldsymbol{m}\beta=\mathbf{0}$, this would mean that there exists $\eta>0$, such that together with (5.5),
$$
R_{t_j}(\boldsymbol{m}\beta_{1,t_j})
>
R_P(\boldsymbol{m}\beta_{1,l})+2\eta
\geqslant
R_P(\mathbf{0})+2\eta
\geqslant
R_{t_j}(\mathbf{0})+\eta
>
R_{t_j}(\mathbf{0}),
$$
for $t_j$ sufficiently large, This would mean that $\boldsymbol{m}\beta_{1,t_j}$ is not the minimizer of $R_{t_j}(\boldsymbol{m}\beta)$. We conclude that $\boldsymbol{m}\beta_{1,L}=\mathbf{0}$, which proves the theorem.
There are several examples of covariance functionals that satisfy (5.2), such as the Mini- mum Covariance Determinant functional (see Cator and Lopuha¨a (2012)) and the covariance S-functional (see Lopuha¨a (1989)), including the Minimum Volume Ellipsoid functional. For a structured covariance functional $\mathbf{V}(\boldsymbol{m}\vartheta_0(P))$ to satisfy (5.2), it is required that the mapping $\boldsymbol{m}\vartheta\mapsto\mathbf{V}(\boldsymbol{m}\vartheta)$ is continuous. This is true for all the examples mentioned in Section 2. In ad- dition, the functional $\boldsymbol{m}\vartheta_0(P)$ needs to be continuous. An example is the S-functional $\boldsymbol{m}\vartheta_0(P)$ defined in Lopuha¨a et al. (2022).
A direct corollary of $\boldsymbol{m}\beta_1(P)$ being continuous, is the consistency of the estimator $\boldsymbol{m}\beta_{1,n}$.
Corollary 2. Suppose that $\varrho_1:\mathbb{R}\rightarrow[0,\infty)$ satisfies (R-BND) and suppose that $P$ is such that (A) holds and that $\mathbf{X}$ has full rank with probability one. Suppose that $\mathbf{V}_{0,n}\rightarrow\mathbf{V}_0(P)$, with probability one. Then for $n$ sufficiently large, there is at least one $\boldsymbol{m}\beta_{1,n}$ that minimizes $R_{n}(\boldsymbol{m}\beta)$, with probability one. If $\boldsymbol{m}\beta_1(P)$ is the unique minimizer of $R_P(\boldsymbol{m}\beta)$, then for any sequence $\boldsymbol{m}\beta_{1,n}$, $n=1,2...$, it holds that
$$
\lim_{n\rightarrow\infty}\boldsymbol{m}\beta_{1,n}=\boldsymbol{m}\beta_1(P),
$$
with probability one. | |
200803841/16 | MARCELO M. DISCONZI, VU HOANG, MARIA RADOSZ
The first major part of the proof consists in deriving an a-priori estimate for the viscous pressure $\Pi$. This will be done in the upcoming Proposition 10. A key idea will be to introduce a transport equation in the variables $(\rho, n, \Pi)$ which encodes the evolution of $\Pi$ relative to $\rho, n$ and for which an estimate will be derived in Proposition 9.
Suppose $(\rho, n, \Pi, u)$ is a smooth solution of (2.9)–(2.12). Then along a flowline defined by
$$
\begin{aligned}
\frac{d}{d\tau} X^\mu(\tau; x_0) = u^\mu(X(\tau; x_0)), ~~X(0; x_0)=(0, x_0)
\end{aligned}
$$
with given $x_0\in \mathbb{R}^3$, the following equations hold for $\rho = \rho(X(\tau;x_0)), n = n(X(\tau;x_0)), \Pi = \Pi(X(\tau;x_0))$
$$
\begin{aligned}
\begin{split}
&\dot{\rho} = \frac{(\rho+p(\rho, n)+\Pi)\dot{n}}{n}\\
&\dot{\Pi} = \frac{\zeta(\rho, n) \dot{n}}{\tau_0(\rho, n) n} - \frac{(1+\lambda(\rho, n)\Pi)\Pi}{\tau_0(\rho, n)}
\end{split}
\end{aligned}
$$
where a dot denotes differentiation with respect to $\tau$. To derive these equations, eliminate $\nabla_\alpha u^\alpha$ from (2.9) and (2.12) using (2.11). Note also
$$
\begin{aligned}
\frac{d}{d\tau}(p+\Pi) = (\partial_\rho p) \dot{\rho} + (\partial_n p) \dot{n} + \dot{\Pi} = c_s^2 \frac{(\rho+p+\Pi)\dot{n}}{n} - \frac{(1+\lambda \Pi)\Pi}{\tau_0},
\end{aligned}
$$
where we used the equation for $\dot \Pi$ in (3.3) and (2.13).
Before we can address the estimates for $\Pi$, we need the fact that the weak energy condition propagates for admissible solutions, provided it is satisfied at $t= 0$.
Proposition 8. Assume that the inital data satisfy $\mathring{\rho}+p(\mathring{\rho}, \mathring{n})+\mathring{\Pi} \geq 0$ for all $x\in \mathbb{R}^3$. Then $\rho+p+\Pi \geq 0$ for all $(t, x)$.
Proof. Along a trajectory, we have for $e = \rho+p+\Pi$ the following equation:
$$
\begin{aligned}
\dot{e} = e \left[(1+c_s^2) \frac{\dot{n}}{n} - \frac{\lambda e + 1}{\tau_0} \right] + \frac{(1-\lambda (\rho+p)) (\rho+p)}{\tau_0}
\end{aligned}
$$
which is seen by using (3.3). We therefore have
$$
\begin{aligned}
e(\tau) = E(\tau) \left(e(0) + \int_0^\tau \frac{(1-\lambda (\rho+p))) (\rho+p)}{E(s)\tau_0}~ds \right)
\end{aligned}
$$
with
$$
\begin{aligned}
E(\tau) = \exp\left(\int_0^\tau (1+c_s^2) \frac{\dot{n}}{n} - \frac{\lambda e + 1}{\tau_0}~ds\right).
\end{aligned}
$$
By (A1) and (A5), $\rho+p \geq 0$ and $1-\lambda (\rho+p) \geq 0$. It is now immediate that $e(\tau) \geq 0$, since $e(0) \geq 0$.
We will bound solutions of (3.3) by solving a transport equation in the variables $(\rho, n, \Pi)\in \mathbb{R}\times\mathbb{R}^+\times \mathbb{R}$. To shorten notation, define the folllowing differential operator $\mathcal{T}$, acting on smooth functions $F: \mathbb{R}\times \mathbb{R}^+\times \mathbb{R}\rightarrow \mathbb{R}$:
$$
\begin{aligned}
(\mathcal{T} F)(\rho, n, \Pi) = n \partial_n F + (\rho+p(\rho, n)+\Pi)\partial_\rho F + \frac{\zeta}{\tau_0}\partial_{\Pi} F.
\end{aligned}
$$
Using (3.3), we have the identity
$$
\begin{aligned}
\frac{\dot n}{n} \mathcal{T}F = \frac{d F}{d\tau} + \frac{(1+\lambda \Pi)\Pi }{\tau_0}\partial_\Pi F
\end{aligned}
$$
along a fixed flow-line of the smooth solution. | |
230214698/3 | ## Introduction
Community detection (CD), the process of inductively identifying communi- ties within a network, is a core problem in computational sciences, particularly, in physics, computer science, biology, and computational social science [50,19]. Among common approaches for CD are the algorithms which are designed to maximize a utility function, modularity [37], across all possible ways that the nodes of the input network can be partitioned into communities. Modularity measures the fraction of edges within communities minus the expected fraction if the edges were distributed randomly; with the random distribution of the edges being a nan model that preserves the node degrees. Despite their name and de- sign philosophy, current modularity maximization algorithms, which are used by no less than tens of thousands of peer-reviewed studies [28], are not guaran- teed to maximize modularity [38,24,35]. This has led to uncertainty [20,24] in the extent to which they succeed in returning a maximum-modularity (optimal) partition or something similar.
Modularity is among the first objective functions proposed for optimization- based detection of communities [37,18]. Several limitations [21,16,18,41] of mod- ularity including the resolution limit [17] have led researchers to develop al- ternative CD methods using stochastic block modeling [23,40,32,45], informa- tion theoretic approaches [43,44], and alternative objective functions [2,47,36,34]. Modularity-based algorithms are the most commonly used method for CD [46,19]. Despite the widespread adoption of modularity-based heuristics, there is uncer- tainty [20,24] in their success in maximizing modularity. This study aims to address this uncertainty by quantifying the extent to which eight commonly used heuristics [13,7,30,46,50,8,48,31] succeed in returning an optimal partition or a partition resembling an optimal partition. After describing the methods and materials, we present the main results followed by a discussion of the method- ological ramifications and future directions.
## Methods and Materials
This study aims to investigate the extent to which eight commonly used heuristic modularity maximization algorithms [13,7,30,46,50,8,48,31] succeed in returning an optimal partition or a partition similar to an optimal partition. To achieve this objective, we quantify the proximity of their results to the globally optimal partition(s), which we obtain using an exact Integer Programming (IP) model for maximizing modularity [9,1,15]. We do not claim that maximum-modularity partitions represent best partitions. Throughout the paper, we use the terms network and graph interchangeably.
## Modularity
Consider the simple graph $G=(V,E)$ with $|V|=n$ nodes, $|E|=m$ edges, adjacency matrix entries $a_{ij}$, and a partition $X=\{V_1,V_2, \dots, V_k \}$ of the node | |
200405806/13 | (a) The EM-scalar perturbation with $l=1$.
(b) The Gravitational-scalar perturbation with $l=2$.
FIG. 3: The effective potentials for EM-scalar and Gravitational-scalar perturbations in 5-dimensional asymptotic AdS spherically symmetric black holes.
However, it is important to note that even though the asymptotic behaviors of our po- tentials fit exactly with the classification of Ishibashi and Wald’s work, there exist some intrinsic differences with our work here, including the physical phenomena and the back- ground spacetimes. The self-adjoint extended dynamical solutions in purely AdS spacetimes were their focus, and a new set of QNMs in $SAdS_{D}$ is ours. As a consequence, the constraint equations for the boundary conditions in their work may not fit exactly with those in our studies, and this was also commented on in their conclusion section. However, a common argument is that in such cases a more general Robin boundary condition, where the outgoing wave boundary condition we use falls under this category, is allowed to be used beyond the box-like boundary conditions.
## QUASI-NORMAL MODES
In the previous section we have demonstrated cases where there exist barrier-like effective potentials in asymptotic AdS black hole spacetimes. From this, the behavior of the related particle interactions with the black hole spacetimes seems to dominate the shape of these effective potentials. Note that the previously studied shape of the effective potentials for AdS black hole spacetimes had strict boundary conditions (b.c.) imposed upon them, which meant a box-like shape was used for the calculation of QNMs. These conditions were the Dirichlet and the vanishing energy flux (VEF) b.c.’s, where the QNMs obtained from these b.c. had linear relations to the black hole Hawking temperature with variations of the black hole mass. However, due to the barrier-like behavior of the effective potentials listed in | |
200801185/6 | SARA BROFFERIO,
DARIUSZ BURACZEWSKI,AND
TOMASZ SZAREK
2.5.
## Ergodic measures and Ratio Ergodic Theorem
. Among $\mu$ -invariant measures, ergodic measures play a special role. We present here the main facts and we refer to Section 6 for a more detailed discussion.
For any measurable $A\subseteq \mathbb R$ denote by $\nu_A$ the restriction of $\nu$ to $A$. The restriction is called trivial if either $\nu(A)=0$ or $\nu(\mathbb R\setminus A)=0$. We say that a measure $\nu$ is if for any $A\in {\mathcal B}(\mathbb R)$ such that the restriction $\nu_A$ is invariant, it must be also trivial. In our setting we can say that if an invariant measure is ergodic, then any closed $\Gamma$ -invariant set $M$ is either nan or has full measure: $\nu(M)=0$ or $\nu(\mathbb R\setminus M)=0.$ In Section we give a more detailed discussion of other equivalent characterisations of the ergodic measures.
Ergodic measures can be seen as atomic bricks that are used to construct any invariant measure. In fact, any invariant measure $\nu$ can be decomposed into ergodic components, in the sense that there exists a measurable set $\mathcal E_\nu$ of ergodic measures and a finite measure $\eta_\nu$ on $\mathcal E_\nu$ such that
$$
\nu(f)=\int_{\mathcal E_\nu}\nu_e(f) d\eta_\nu(e) \quad\mbox{for all $f\in C_C(\mathbb R)$}.
$$
In Theorem 6.6 we provide a proof of this decomposition for conservative Markov – Feller processes. Note that the above decomposition entails that if there are two dif- ferent invariant measures, there must exist at least two different ergodic measures. An- other consequence is that if $\nu$ is invariant, there exists an ergodic measure $\nu_e$ such that $\supp\nu_e\subseteq \supp\nu$. In fact, for $\eta_\nu$ -almost all $e\in \mathcal E_\nu$ we have $\nu_e(\mathbb R\setminus\supp\nu)=0$. Hence we have $\supp\nu_e\subseteq \supp\nu$.
A fundamental property of ergodic $\mu$ -invariant Radon measures, that we will often use in the sequel, is the Ratio Ergodic Theorem (or the Chacon-Ornstein Theorem), that gives the asymptotic behaviour of the partial sum defined by:
$$
S_n\varphi(x):=\varphi (\mathbf{g}_{n}\cdots \mathbf{g}_1(x))+\ldots+ \varphi(\mathbf{g}_{1}(x))+\varphi(x)=\sum_{k=0}^n\varphi(X_n^x)
$$
for any measurable function $\varphi\in L^1(\mathbb R,\nu)$ and $x\in \mathbb R$. Observe that if $\varphi$ is the indicator function of some set $A$, then $S_n\varphi(x)=S_n{\bf 1}_A(x)$ is the number of visits in $A$ up to time $n$ for the Markov chain $(X_n^x)_{n\in\mathbb N}$ starting at $x$.
Whenever recurrence condition ($\mathfrak{R}$) is satisfied, it follows that for any arbitrary function $\Phi$ whose support contains a recurrent interval $\mathcal I$ we have $S_n\Phi(x)\rightarrow +\infty$ for any $x\in\mathbb R$, as $n\rightarrow\infty$.
If $\nu$ is ergodic for any nonnegative function $\Phi\in L^1(\mathbb R,\nu)$ we have $\nu(\Phi)>0$ if and only if $S_n\Phi(x)\rightarrow +\infty$ for $\nu$ -almost all $x$ and in this case the Chacon-Ornstein Theorem [11] guarantees that for any $\varphi\in L^1(\mathbb R, \nu)$ the following limit exists
$$
\lim_{n\rightarrow\infty}\frac{S_n\varphi(x)}{S_n \Phi(x)} =\frac{\nu(\varphi)}{\nu(\Phi)}
$$
for $\mu^\mathbb N$ -almost all sequences $(g_1, g_2,\ldots)\in\Gamma^{\mathbb N}$ and $\nu$ -almost all $x\in\mathbb R$. This is a consequence of the fact that the shift $\tau$ is a contraction on the space $L^1(\mathbb R^\mathbb N,\mathbb P_\nu)$ and that $\mathbb P_\nu$ is ergodic, if $\nu$ is ergodic (see Section 6 and in particular Corollary 6.5 for a more complete discussion of these results).
2.6.
## Measures with atoms
. The following lemma is useful when we have to deal with some invariant measures $\nu$ that have atoms, that is, for which there exists $x\in\mathbb R$ such that $\nu(\{x\})>0$. It essentially says that one can have invariant measures with atoms only if the orbits of action of $\Gamma^{-1}$ are somehow discrete.
Lemma 2.2. Assume that condition ($\mathfrak{R}$) is satisfied. Let $\nu$ be a $\mu$ -invariant Radon mea- sure with atoms and let $K$ be a compact interval that contains the recurrence interval $\mathcal I$ and some atoms. Then there exists $x_0\in K$ such that the orbit $(\Gamma^{-1})^*x_0\cap K$ is finite. | |
220804752/15 | INCOMPLETENESS FOR STABLY COMPUTABLE FORMAL SYSTEMS
Define a Turing decision map ${D}$ by
$$
{D} (T,n):= (Dec _{\mathcal{A}} ({C (M
_{e})))'} (s (T),n)
$$
for $Dec _{\mathcal{A}}$ as in (3.9) defined with respect to $B=\mathcal{A}$, and where $C$ is as in Section 3. Then by construction, and by Axiom 4 in particular, ${D}$ is computable by some Turing machine $({D}, {D} _{e})$, we make this more explicit in the following Section 5.
Now ${D}$ is stably sound by Lemma 3.11 and the assumption that $F$ is sound. So by Lemma 4.3:
\begin{equation*} \neg ({D} _{e} \text{is ${D}$-decided}). \end{equation*}
In particular, $s ({D} _{e})$ is not $Dec _{\mathcal{A}} (C
({M} _{e}))
$-decided, and so$s ({D} _{e})$is not$C (M
_{e}) $-stable (Lemma), i.e.$M \nvdash s ({D} _{e})$.
On the other hand,
$$
\neg ({D} _{e} \text{is
${D}$-decided}) \models \neg ({D} _{e}
\text{is ${D} _{e}$-decided}),
$$
by definition. And so since ${D} _{e} \in \mathcal{D} _{\mathcal{T}}$ by construction, $s ({D} _{e})$ is satisfied. Set $\alpha (M):=s ({D} _{e})$ and we are done.
## 5. Syntactic incompleteness for stably computable formal systems
Let $s: \mathcal{T} \rightarrow \mathcal{A}$, $T \mapsto
s (T) $ be as in the previous section. Define
\begin{equation*} H: \mathcal{T} \times \mathcal{T} \times \mathbb{N} \rightarrow \{\mathbb{\pm}\}, \end{equation*}
by $H (M,T,n):= (Dec
_{\mathcal{A}} ({C(Spec(M))))'} (s (T),n)$. We can express$H$ as the composition of the sequence of maps:
$$
\mathcal{T} \times \mathcal{T} \times
\mathbb{N} \xrightarrow{Dec _{\mathcal{A}} \circ C
\circ Spec \times
s \times id} \mathcal{T} \times \mathcal{A} \times
\mathbb{N} \xrightarrow{id \times e _{\mathcal{A}
\times \mathbb{N}}} \mathcal{T} \times
\mathcal{U} \xrightarrow{U} \mathcal{U}
\xrightarrow{e _{\{\pm \}} ^{-1}} \{\pm
\},
$$
where the last map is:
\begin{equation*} \Sigma \mapsto \begin{cases} undefined, &\text{if} \Sigma \notin \{\pm \} _{e} \\
e _{\{\pm \}} ^{-1} (\Sigma), & \text {otherwise}. \end{cases} \end{equation*}
So $H$ is a composition of maps that are computable by the axioms of $\mathcal{S}$ and so $H$ is computable.
Hence by Axiom 7 there is an associated computable map:
$$
Tur: \mathcal{T} \rightarrow \mathcal{T},
$$
s.t. for each $M \in \mathcal{T}$, $Tur (M)$ encodes the map $D ^{M}: \mathcal{T} \times
\mathbb{N} \rightarrow \{\pm \} $,$D ^{M} (T,n) = H
(M,T,n) $.
In what follows, $(M,M _{e}): \mathbb{N} \rightarrow
\mathcal{A} \times \{\pm \} $will be a fixed total Turing machine. We abbreviate$D ^{M _{e}}$by$D$and$Tur (M _{e})$by$D _{e}$. As usual notation of the form$M \vdash \alpha$means$M ^{s} \vdash \alpha$.
Proposition For $(M,M _{e}), ({D}, {D} _{e})$ as above:
\begin{align*} \text{$M ^{s}$ is 1-consistent} \implies M \nvdash s ({D} _{e}). \end{align*}
\begin{align*} \text{$M ^{s}$ is 2-consistent} \implies M \nvdash \neg s ({D} _{e}). \end{align*}
Moreover, the sentence:
$$
\text{$M ^{s}$ is 1-consistent} \implies s
({D} _{e})
$$
is a theorem of $PA$ under standard interpretation of all terms, (this will be further formalized in the course of the proof).
Proof. This proposition is meant to just be a theorem of set theory $ZF$, however we avoid complete set theoretic formalization, as is common. Arithmetic is interpreted in set theory the standard way, using the standard set $\mathbb{N}$ of natural numbers. So for example, for $M:
\mathbb{N} \rightarrow \mathcal{A} \times \{\pm \} $a sentence of the form$M \vdash \alpha$is a priori interpreted as a sentence of$ZF$, however if$M$is a Turing machine this also can be interpreted as a sentence of$PA$, once G ¨o del encodings are invoked.
Set $N:= (Spec (M _{e}))'$, in particular this is a speculative total Turing machine $\mathbb{N} \rightarrow
\mathcal{A} \times \{\pm \} $. Set$s:=s ({D} _{e})$. Suppose that$M
\vdash s$. Hence,$N \vdash s$and so$s$is$C(Spec (M _{e}))$ -stable, and so by Lemma 3.11 | |
230212504/2 | is typically estimated by comparing event rates between the treated and control arms using the Proportional Hazards (Cox, 1972) model and its extensions.
Identification of subgroups in such scenarios has been the subject of a large amount of tradi- tional statistical literature. Large number of such approaches involve estimation of the factual and counterfactual outcomes using separate regression models (T-learner) followed by regressing the difference between these estimated potential outcomes. Within this category of approaches, Lip- kovichetal. (2011) propose the subgroup identification based on differential effect search (SIDES) algorithm, Suetal. (2009) propose a recursive partitioning method for subgroup discovery, Dussel- dorpandMechelen (2014) propose the qualitative interaction trees (QUINT) algorithm, and Foster etal. (2011) propose the virtual twins (VT) method for subgroup discovery involving decision tree ensembles. We include a parametric version of such an approach as a competing baseline.
Identification of heterogeneous treatment effects (HTE) is also of growing interest to the ma- chine learning community with multiple approaches involving deep neural networks with balanced representations (Shalitetal., 2017; Johanssonetal., 2020), generative models Louizosetal. (2017) as well as Non-Parametric methods involving random-forests (Wager and Athey, 2018) and Gaus- sian Processes (Alaa and Van Der Schaar, 2017). There is a growing interest in estimating HTEs from an interpretable and trustworthy standpoint (Lee et al., 2020; Nagpal et al., 2020; Morucci ´ e et al., 2022). Wang and Rudin (2022) propose a sampling based approach to discovering interpretable rule sets demonstrating HTEs.
However large part of this work has focused extensively on outcomes that are binary or continu- ous. The estimation of HTEs in the presence of censored time-to-events has been limited. Xuetal. (2022) explore the problem and describe standard approaches to estimate treatment effect hetero- geneity with survival outcomes. They also describe challenges associated with existing risk models when assessing treatment effect heterogeneity in the case of cardiovascular health.
There has been some initial attempt to use neural network for causal inference with censored time-to-event outcomes. Curth et al. (2021) propose a discrete time method along with regulariza- tion to match the treated and control representations. Chapfuwa et al. (2021) ’s approach is related and involves the use of normalizing flows to estimate the potential time-to-event distributions under treatment and control. While our contributions are similar to Nagpal et al. (2022a), in that we as- sume treatment effect heterogeneity through a latent variable model, our contribution differs in that 1) Our approach is free of the expensive Monte-Carlo sampling procedure and 2) Our generalized EM inference procedure allows us to naturally incorporate structured sparsity regularization, which helps recovers phenogroups that are parsimonious in the features they recover that define subgroups.
Survival and time-to-event outcomes occur pre-eminently in areas of cardiovascular health. One such area is reducing combined risk of adverse outcomes from atherosclerotic disease 1 (Herrington, 2016; Furberg et al., 2002; Group, 2009; Buse et al., 2007) The ability of recovering groups with differential benefits to interventions can thus lead to improved patient care through framing of optimal clinical guidelines.
## 3. Proposed Model: Sparse Cox Subgrouping
- Notation As is standard in survival analysis, we assume that we either observe the true time-to- event or the time of censoring $U = \mathrm{min}\{T, C\}$ indicated by the censoring indicator defined as
1. A class of related clinical conditions from increasing deposits of plaque in the arteries, leading to Stroke, Myorcardial Infarction and other Coronary Heart Diseases. | |
190407322/16 | Remark 2.18. Clearly we do not need to verify (2) for all objects in $\mbT$: It suffices to do so for a subcollection $\mbT' \subseteq \mbT$ such that $\add \mbT' = \mbT$.
Similarly we do not need to check (3) for all projectives, it suffices to check it for a subcategory $\mathbb{P} \subseteq \Proj \mcA$ such that $\add \mathbb{P} = \Proj \mcA$. Indeed, the collection of objects having two-term $\mbT$ coresolutions is clearly closed under direct sums. Thus it remains to show that it is also closed under summands. So assume $P \oplus Q$ has a two-term $\mbT$ coresolutions as in the top row of the following diagram.
$$
\begin{tikzcd}[ampersand replacement=\&]
P \oplus Q \ar[r,right hook->,"{(f \; g)}"] \ar[d,equal] \& T^0 \ar[r,->>] \& T^1 \\
P \oplus Q \ar[r,right hook->,"{\left(\begin{smallmatrix} f & 0 \\ 0 & g \end{smallmatrix} \right)}"] \& (T^0)^2 \ar[r,->>] \ar[u,->>,"{(1\;1)}" swap] \& \coker f \oplus \coker g \ar[u,->>,densely dashed] \\
\& T_0 \ar[r,equal] \ar[u,right hook->,"{\left(\begin{smallmatrix} \phantom-1 \\ -1 \end{smallmatrix} \right)}" swap] \& T_0 \ar[u,right hook->,densely dashed]
\end{tikzcd}
$$
We can clearly draw the remaining solid part of this diagram, and it follows that there is the dashed short exact sequence in the right column. Since $\Ext_{\mcA}^1(T^1, T_0) \subseteq \Ext_{\mcA}^1(\mbT, \mbT) = 0$ this sequence splits, and thus both $\coker f$ and $\coker g$ are in $\mbT$.
Remark 2.19. The existence of enough projectives is not essential for tilting theory. However we keep this assumption here since it simplifies the formulation of the definition, and it will be satisfied in the examples we are interested in.
Conditions (1) to (3) are the standard conditions for tilting, while assuming contravariant finiteness is less commonly made explicit. In fact, we proceed to show that this is automatically satisfied if the abelian category $\mcA$ is noetherian, that is, any ascending chain of subobjects of a given object eventually becomes stationary. We will see below, however, that it is not automatically satisfied in general, even for fairly reasonable categories.
Proposition 2.20. Let $\mcA$ be a noetherian abelian category with enough projectives and assume that $\mbT$ is a weak tilting subcategory of $\mcA$. Then $\mbT$ is automatically contravariantly finite, and therefore already tilting.
A related result can be found in the proof of [20 20, Theorem 2.5(ii)].
Proof. We need to show that $\mbT \subseteq \mcA$ is contravariantly finite, i.e., for any given object $M \in \mcA$, the functor
$$
\hom_\mcA(-, M)|_{\mbT}\colon \mbT^\mathsf{op} \longrightarrow \Ab
$$
is finitely generated. To this end, first observe that by noetherianity, $M$ has a (unique) maximal subobject $\mathsf{t} M$ that is an epimorphic image of an object $T \in \mbT$. This means that for $X\in \mbT$ and $f\colon X\longrightarrow M$ the factorization through $\mathsf{t} M$ in
$$
\begin{tikzpicture}[scale=.75]
\node (1) at (0,2) {$X$};
\node (2) at (2,1) {$\im f$};
\node (3) at (4,2) {$M$};
\node (4) at (4,0) {$\mathsf{t} M$};
\path[->,font=\scriptsize]
(1) edge[->>] node[auto] {} (2)
(2) edge[right hook->] node[auto] {} (3)
(1) edge[bend left=15] node[auto] {$f$} (3)
(2) edge[right hook->, densely dashed] node[auto] {} (4)
(4) edge[right hook->] node[auto] {} (3);
\end{tikzpicture}
$$
exists and is unique. From this follows that $\hom_\mcA(-, M)|_{\mbT} \cong \hom_\mcA(-, \mathsf{t} M)|_{\mbT}$: Any morphism $f\colon X \longrightarrow M$ is sent to the composition $X\longrightarrow\mathrel{\mspace{-25mu}}\longrightarrow \im f\lhook\joinrel\longrightarrow \mathsf{t} M$. The other direction is then clearly given by sending $X\longrightarrow \mathsf{t} M$ to $X\longrightarrow \mathsf{t} M\lhook\joinrel\longrightarrow M$. | |
170801973/1 | # ACTIONS OF CATEGORIES BY LIPSCHITZ MORPHISMS ON LIMITS FOR THE GROMOV-HAUSDORFF PROPINQUITY
FR É D É RIC LATR É MOLI È RE
We prove a compactness result for classes of actions of many small categories on quantum compact metric spaces by Lipschitz linear maps, for the topology of the covariant Gromov-Hausdorff propinquity. In particular, our result applies to actions of proper groups by Lipschitz isomorphisms on quantum compact spaces. Our result provides a first example of a structure which passes to the limit of quantum metric spaces for the propinquity, as well as a new method to construct group actions, including from non-locally com- pact groups seen as inductive limits of compact groups, on unital C*-algebras. We apply our techniques to obtain some properties of closure of certain classes of quasi-Leibniz quantum compact metric space s for the propinquity.
The noncommutative Gromov-Hausdorff propinquity [0, 0, 0, 0, 0] is a distance which induces a topology on the hyperspace of quantum compact metric spaces, with sights on applications to mathematical physics.
A natural yet delicate question is to determine which properties do pass to the limit of a sequence of quantum metric spaces. Partial answers to this question can for instance help us determine the closure of a class of quantum metric spaces. Most importantly, the propinquity is intended to be used to construct theories for physics from finite dimensional models [0, 0, 0] or by using other form of approximations [0, 0], possibly using the fact that the propinquity is complete. Thus, we would like to know which properties of the finite approximations would indeed carry to the eventual limit.
The existence of group actions on limits of sequences of quantum compact metric spaces is thus of prime interest. The question can be informally asked: if a sequence of metrized compact groups act on a Cauchy sequence of quasi-Leibniz quantum compact metric space s, and if these groups converge in some appropriate sense, then does the limit carry an action of the limit group? Asked in such a manner, the answer is trivially yes — one may involve trivial actions — but it is easy and rather natural to impose a form of nontriviality using the underlying quantum metrics. Indeed, we can control the dilation of the automorphisms [0] arising from the actions at each level to ensure that any sort of limit would also be non-trivial. We also note that much efforts have been dedicated to constructing full matrix approximations of certain homogeneous spaces [0, 0, 0, 0] with the idea that finite dimensional quantum metric spaces are the noncommutative analogues of finite
Key words and phrases. Noncommutative metric geometry, Gromov-Hausdorff convergence, Monge-Kantorovich distance, Quantum Metric Spaces, Lip-norms, semigroup, groupoid and group actions.
This work is part of the project supported by the grant H2020-MSCA-RISE-2015-691246- QUANTUM DYNAMICS. | |
211209331/5 | Fig.2. Illustration of sampling strategies.
Fig.3. Illustration of image and text embeddings.
random sampling. Then, we delve into the gradient of InfoNCE loss and provide evidences that data bias influences the model optimization.
Dataset bias leads to biased representation distributions. In the upper part of Figure 3, we visualize image and text embeddings learned with random sampling. Intra-dataset representations are closely gathered, while inter-dataset representations are separated. Representations are separated to three parts, i.e., VG, SBU and “ CC3M+CC12M ”. Since CC3M and CC12M are composed of similar image-text pairs, representations of CC3M and CC12M are slightly overlapped. It demonstrates that the model is driven to separate representations from different datasets, and, within a training batch, the model will easily distinguish negative samples.
Dataset bias influences the optimization of InfoNCE. Since the dual- encoder model is optimized by InfoNCE loss, we first formulate the loss function and its gradient for further explorations:
$$
\mathcal{L} = \sum_{j} \sum_{k} y_{jk} \text{log} \left(\frac{\text{exp}(s_{jk})}{\sum_{l}\text{exp}(s_{jl})} \right),
\nabla_{\theta} \mathcal{L} = - \sum_{j} \sum_{k} y_{jk} \nabla_{\theta} \text{log} \left(\frac{\text{exp}(s_{jk})}{\sum_{l}\text{exp}(s_{jl})} \right),
$$
where the similarity between the query $j$ and the key $k$ as $s_{jk}$. The ground- truth label corresponding to $s_{jk}$ is represented by $y_{jk} \in \{0,1\}$. We omit the temperature parameter for simplification. Then, we derive the gradient item as 1:
$$
\begin{aligned}
\nabla_{\theta} \mathcal{L} & = \sum_{j} \sum_{k} \left(\frac{\text{exp}(s_{jk})}{\sum_{l} \text{exp}(s_{jl})} - y_{jk} \right) \nabla_{\theta} s_{jk} \\
& = \sum_{j} \sum_{k} \left({\bar{p}_{jk}} - y_{jk} \right) \nabla_{\theta} s_{jk} \\
\end{aligned}
$$
where we could observe that the gradient term is related to the stop-gradient term $\bar{p}_{jk}$, which reflects the similarities among training samples. Negative pairs are essential for self-supervised learning methods which are based on the InfoNCE loss [35]. However, as suggested in Figure 3, dataset bias makes the model easily separate negative samples from different data sources, resulting in the small
1 Detailed deriviations are attached in Appendix A.1. | |
170407903/5 |
\begin{align*} & \textstyle\bigwedge^0(\mathfrak{p}) \cong \textstyle\bigwedge^4(\mathfrak{p}) \cong V_{(0,0)} \\
& \textstyle\bigwedge^1(\mathfrak{p})\cong\textstyle\bigwedge^3(\mathfrak{p}) \cong V_{(1,0)} \\
& \operatorname{span} \{E_3 \wedge E_4, E_3 \wedge F_3 + E_4 \wedge F_4, F_3 \wedge F_4 \} \cong V_{(1,1)} \\
& \operatorname{span} \{E_3 \wedge F_4, E_3 \wedge F_3 - E_4 \wedge F_4, E_4 \wedge F_3 \} \cong V_{(1, -1)}. \end{align*}
It follows that $\textstyle\bigwedge(\mathfrak{p})$ decomposes under $\mathfrak{k}$ as
\begin{align*} \textstyle\bigwedge (\mathfrak{p}) & = \overbrace{V_{(0, 0)}}^{0} \\
& \oplus \overbrace{V_{(1,0)}}^{1} \\
& \oplus \overbrace{V_{(1,1)}}^{2} \oplus \overbrace{V_{(1,-1)}}^{2} \\
& \oplus \overbrace{V_{(1,0)}}^{3} \\
& \oplus \overbrace{V_{(0,0)}}^{4}, \end{align*}
where each of the numbers over braces denotes the degree in which the corresponding $\mathfrak{k}$ -module is appearing.
If $V$ is a $\mathfrak{k}_1$ – module and $W$ is a $\mathfrak{k}_2$ – module, then we denote by $V \boxtimes W$ the external tensor product of $V$ and $W$. As a vector space, $V \boxtimes W$ is equal to the direct sum of $V$ and $W$, and the action of $\mathfrak{k}=\mathfrak{k}_1\oplus\mathfrak{k}_2$ – module is defined so that $\mathfrak{k}_1$ acts on $V$ and $\mathfrak{k}_2$ acts on $W$. Then we have
$$
V_{(n,m)} \simeq V_{n+m} \boxtimes V_{n-m} \quad\text{as $\mathfrak{k}$--modules,}
$$
where $V_k$ denotes the $\mathfrak{sl}(2,\mathbb{C})$ – module with highest weight $k$. It follows that
\begin{align*} \mathcal{H}_{\mathfrak{k}_1} \otimes \mathcal{H}_{\mathfrak{k}_2} & = \mathopen{}\mathclose\bgroup\originalleft (\bigoplus_{n \in \mathbb{Z}_{+}} V_{(n,n)} \aftergroup\egroup\originalright) \otimes \mathopen{}\mathclose\bgroup\originalleft (\bigoplus_{m \in \mathbb{Z}_{+}} V_{(m,-m)} \aftergroup\egroup\originalright) \\
& \simeq \mathopen{}\mathclose\bgroup\originalleft (\bigoplus_{n \in \mathbb{Z}_{+}} V_{2n} \boxtimes V_{0} \aftergroup\egroup\originalright) \otimes \mathopen{}\mathclose\bgroup\originalleft (\bigoplus_{m \in \mathbb{Z}_{+}} V_{0} \boxtimes V_{2m} \aftergroup\egroup\originalright) \\
& \simeq \bigoplus_{n,m \in \mathbb{Z}_{+}}\mathopen{}\mathclose\bgroup\originalleft (V_{2n} \otimes V_0 \aftergroup\egroup\originalright) \boxtimes \mathopen{}\mathclose\bgroup\originalleft (V_{0} \otimes V_{2m} \aftergroup\egroup\originalright) \\
& \simeq \bigoplus_{n,m \in \mathbb{Z}_{+}} V_{2n} \boxtimes V_{2m} \\
& \simeq \bigoplus_{n,m \in \mathbb{Z}_{+}} V_{(n+m, n-m)}. \end{align*}
| |
200613113/7 | ## Photoelectron spectroscopy
In-situ synchrotron measurements were recorded using 700 eV photon energy for the survey spectra and 104 eV for the simultaneous detection of tungsten 4f core levels and the valence band. During hydrogenation experiments, the photoemission spectra were recorded continuously in single sweep mode, with each spectrum taking approximately 120 s. The applied hydrogen pressure was measured by an external pressure gauge. As soon as the partial hydrogen pressure in the chamber was constant, the external hydrogen pressure was increased. This lead to approximately ten spectra being recorded for each external hydrogen pressure. These spectra were averaged for the subsequent data evaluation to improve signal to noise ratios. Data analysis was performed using the CasaXPS software, employing the GL(30) line shape, Shirley background subtraction and a 2.1 eV spin orbit splitting for both W $^{5+}$ and W $^{6+}$ doublets [23]. The W5p $_{3/2}$ core level is not fitted in the spectra recorded at 104 eV, due to its small photoionization cross section at this photon energy [24]. The binding energy calibration was performed on the gold coated sample holder, setting the Au 4f $_{7/2}$ peak to 84.0 eV. Due to small photon energy instabilities and work function changes during hydrogenation, all spectra were subsequently aligned by shifting the W $^{6+}$ 4f $_{7/2}$ peak to 35.8 eV. These shifts are on the order of 0.2-0.3 eV.
Ex-situ measurements were performed using a PHI Quantes spectrometer (ULVAC-PHI) equipped with a monochromatic Al K $\alpha$ (1486.6 eV) X-ray source. Charge neutralization was accomplished by a dual beam charge neutralization system, employing low energy electron and argon ion beams. Detailed acquisition parameters are given in table I.
## Electronic structure calculation
Calculation of electronic structure by Density Functional Theory (DFT) and hybrid- DFT was performed using the Vienna Ab initio Simulation Package (VASP) [25–28]. The calculation used Projector Augmented Wave (PAW) method [29,30] to describe the effects of core electrons, and Perdew-Burke-Ernzerhof (PBE) [31] implementation of the Generalized Gradient Approximation (GGA) for the exchange-correlation functional. Energy cutoff was 600 eV for the plane-wave basis of the valence electrons. The electronic structure was calculated on a 15x15x15 $\Gamma$ -centered mesh for WO $_3$ (unit cell), and a 4x4x4 $\Gamma$ -centered mesh | |
181003623/7 | at a distance $a$. Its radius is estimated as
$$
r_{\rm H} = \left(\frac{q}{3} \right)^{1/3} a.
$$
As gas flows across the gap, it can continue to replenish the net angular momentum of material in the Hill sphere of the secondary. This can cause an asymmetry within the Hill sphere, and a slight increase in the gas density upstream from the BH, leading to a consistently positive (outward) torque. This torque has been seen in other works (e.g. D’Angeloetal.2005). Accurately capturing the gas morphology within the Hill sphere requires high spatial resolution. Further work investigating the gas streamlines, particularly with high-resolution 3-dimensional simulations, will likely be necessary to accurately measure the torque arising from this asymmetry, and to fully understand the reason for the small front-back asymmetry that yields the net torque.
We note that in disc-satellite calculations, the torque within the Hill sphere is often ignored or “ damped ”, based on the assumption that material within the Hill sphere is bound to the perturber (e.g. de Val-Borroetal.2006; Dürmann&Kley2017). However, gas flows through the Hill sphere $-$ while some of it may become bound or accreted, certainly the majority flows across the gap to supply the inner disc. Asymmetrically distributed gas in this region can exert a net torque on the BH, which must be included if the BH and the gas is separately resolved and followed. This asymmetry, unlike a bound ‘orbiter’, is constantly supplied as gas flows across the gap. In our case the accumulation of gas leading the BH’s orbit occurs because gas preferentially flows ahead of the BH and/or because it slows down (and hence spends more time) in this region.
Further evidence for torque inside the Hill sphere is present in simulations by Cridaetal. (2009), in which a live BH experiences changes in migration rate when torques within the Hill sphere are truncated. In their case the planet migrates inward more quickly when the Hill region is included, suggesting an accumulation of gas in the trailing side of the planet’s orbit. This is likely due to a difference in disc parameters since the dynamics are sensitive to viscosity and mach number. Ultimately truncating the torque in the Hill sphere gives a poor approximation to a self-consistent, highly resolved torque calculation, which suggests that this region is important. Our simulations sufficiently resolve gas flow within the Hill sphere, and thus we choose to not omit this region when computing the total torque.
Fig. 3 shows the torque broken into two components – inside and outside the secondary’s Hill sphere – along with smooth fits to the data which we use for calculating detectability in § 5. The Hill sphere torque is qualitatively different in that it settles to a positive value (it pushes outward). The gas from elsewhere in the disc, including the inner and outer discs, as well as the streams connecting the two regions, exerts a negative (inward) torque that is 1-5% of the viscous torque. Fig. 4 shows 2D torque density contours of each of these components. The left panel illustrates the torque density outside of the Hill sphere, which is dominated by the streams flowing across the secondary BH’s orbit. The right panel zooms in on the torque density inside the Hill sphere, showing that there is a slight density increase in the gas in front of the BH (for a counterclockwise orbit). This asymmetry is responsible for the positive torque.
## Accretion torque
For the fiducial accretion rate, we find that the accretion torque is negligible compared to the gravitational torque from the gas – their
Figure3. Gravitational torque $T_{\rm g}$ exerted by different regions of the gas disc onto the secondary BH, as a function of binary separation. The red (upper) curve shows torques from within the Hill sphere, and the green (lower) curve shows torques from outside this region. Both torques are scaled by the viscous torque. The dashed curve shows fitting formula we adopt for our LISA SNR computations (§ 5).
magnitudes differ by $\sim6$ orders of magnitude, as seen in Fig. 2. This is because the relative velocity of the gas near the black hole is effectively negligible. We show this in Fig. 5, with a snapshot of the velocity field in a frame co-rotating with the binary. The gas close to the BH has low angular momentum and resembles a quasi-stationary atmosphere, rather than a near-Keplerian mini-disc.
Accretion plays a minimal role in our simulations. The most conspicuous effect of accretion is to reduce the density of the gas near the BH, which leads to a decrease in the positive component of the torque. Without including a sink, our results show the same asymmetry within the Hill sphere, albeit with more gas (and a larger positive component of the torque). The torque from elsewhere in the disc is unaffected by our sink prescription. We hypothesize that more efficient accretion (as well as feedback) would lead to more depleted gas density and less positive torque. A drastic increase in accretion rate may lead to other differences, since it would steal gas that would otherwise flow to the inner disc. We leave an investigation of the dependence of the torques on the accretion rate in the $q\neq 1$, GW-driven case to future work.
## DETECTABILITY OF GAS IMPRINT BY LISA
## Drift in the accumulated GW phase
In this section, we estimate the deviation from the vacuum GW signal caused by the gas disc torques, and assess its detectability by LISA.
The total accumulated phase of a gravitational wave event can be obtained by integrating over the total frequency evolution
$$
\varphi_{\rm tot} = \int_{t_0}^{t_0 + t_{\rm obs}}\! \dot{\varphi}_{\rm GW} \, {\rm dt} = 2 \pi \int_{t(f_{\rm min})}^{t(f_{\rm max})} f_{\rm GW} \, {\rm dt}
$$
where $t_0$ is an arbitrary reference time when the LISA observation begins, $t_{\rm obs}$ is the total observation time, $f_{\rm min}$ and $f_{\rm max}$ bracket the corresponding observed frequency range, $f_{\rm GW}(t) = \Omega (t)/\pi$ is the GW frequency, which is twice the binary’s orbital frequency |
End of preview. Expand
in Dataset Viewer.
README.md exists but content is empty.
Use the Edit dataset card button to edit it.
- Downloads last month
- 0